Upload folder using huggingface_hub

.gitignore

@@ -32,11 +32,10 @@
 ### PreCI ###
 .codecc
 
-app_hg.py
 outputs
 weights
 .vscode/
-baking
 inference.py
-third_party/weights
-third_party/dust3r
+# third_party/weights
+# third_party/dust3r
+# app_hg.py

README.md

@@ -1,3 +1,15 @@
+---
+title: Hunyuan3D-1.0
+emoji: 😻
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.5.0
+app_file: app_hg.py
+pinned: false
+short_description: Text-to-3D and Image-to-3D Generation
+---
+
 [English](README.md) | [简体中文](README_zh_cn.md)
 
 <!-- ## **Hunyuan3D-1.0** -->

app_hg.py

@@ -23,7 +23,6 @@
 # by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
 import spaces
 import os
-os.environ['CUDA_HOME'] = '/usr/local/cuda-11*'
 import warnings
 import argparse
 import gradio as gr
@@ -33,10 +32,22 @@ import torch
 import numpy as np
 from PIL import Image
 from einops import rearrange
+import pandas as pd
 from huggingface_hub import snapshot_download
 
 from infer import seed_everything, save_gif
 from infer import Text2Image, Removebg, Image2Views, Views2Mesh, GifRenderer
+from third_party.check import check_bake_available
+
+try:
+    from third_party.mesh_baker import MeshBaker
+    BAKE_AVAILEBLE = True
+except Exception as err:
+    print(err)
+    print("import baking related fail, run without baking")
+    check_bake_available()
+    BAKE_AVAILEBLE = False
+
 
 warnings.simplefilter('ignore', category=UserWarning)
 warnings.simplefilter('ignore', category=FutureWarning)
@@ -47,43 +58,14 @@ parser.add_argument("--use_lite", default=False, action="store_true")
 parser.add_argument("--mv23d_cfg_path", default="./svrm/configs/svrm.yaml", type=str)
 parser.add_argument("--mv23d_ckt_path", default="weights/svrm/svrm.safetensors", type=str)
 parser.add_argument("--text2image_path", default="weights/hunyuanDiT", type=str)
-parser.add_argument("--save_memory", default=False) # , action="store_true")
+parser.add_argument("--save_memory", default=False, action="store_true")
 parser.add_argument("--device", default="cuda:0", type=str)
 args = parser.parse_args()
 
-@spaces.GPU
-def find_cuda():
-    # Check if CUDA_HOME or CUDA_PATH environment variables are set
-    cuda_home = os.environ.get('CUDA_HOME') or os.environ.get('CUDA_PATH')
-
-    if cuda_home and os.path.exists(cuda_home):
-        return cuda_home
-
-    # Search for the nvcc executable in the system's PATH
-    nvcc_path = shutil.which('nvcc')
-
-    if nvcc_path:
-        # Remove the 'bin/nvcc' part to get the CUDA installation path
-        cuda_path = os.path.dirname(os.path.dirname(nvcc_path))
-        return cuda_path
-
-    return None
-
-cuda_path = find_cuda()
-
-if cuda_path:
-    print(f"CUDA installation found at: {cuda_path}")
-else:
-    print("CUDA installation not found")
-
-
-
 def download_models():
-    # Create weights directory if it doesn't exist
     os.makedirs("weights", exist_ok=True)
     os.makedirs("weights/hunyuanDiT", exist_ok=True)
-
-    # Download Hunyuan3D-1 model
+    os.makedirs("third_party/weights/DUSt3R_ViTLarge_BaseDecoder_512_dpt", exist_ok=True)
     try:
         snapshot_download(
             repo_id="tencent/Hunyuan3D-1",
@@ -93,8 +75,6 @@ def download_models():
         print("Successfully downloaded Hunyuan3D-1 model")
     except Exception as e:
         print(f"Error downloading Hunyuan3D-1: {e}")
-
-    # Download HunyuanDiT model
     try:
         snapshot_download(
             repo_id="Tencent-Hunyuan/HunyuanDiT-v1.1-Diffusers-Distilled",
@@ -109,25 +89,27 @@ def download_models():
 download_models()
 
 ################################################################
+# initial setting
+################################################################
 
-CONST_PORT = 8080
-CONST_MAX_QUEUE = 1
-CONST_SERVER = '0.0.0.0'
 
 CONST_HEADER = '''
 <h2><a href='https://github.com/tencent/Hunyuan3D-1' target='_blank'><b>Tencent Hunyuan3D-1.0: A Unified Framework for Text-to-3D and Image-to-3D Generation</b></a></h2>
 ⭐️Technical report: <a href='https://arxiv.org/pdf/2411.02293' target='_blank'>ArXiv</a>. ⭐️Code: <a href='https://github.com/tencent/Hunyuan3D-1' target='_blank'>GitHub</a>.
+'''
 
-❗️❗️❗️**Important Notes**
+CONST_NOTE = '''
+❗️❗️❗️Usage❗️❗️❗️<br>
 
-Our demo allows you to export models in various formats:
-- By default, export as a *.obj mesh with vertex colors or a *.glb mesh.
-- Select "texture mapping" to export a *.obj mesh with a texture map or a *.glb mesh.
-- Select "render GIF" to export a GIF rendering of the *.glb file.
+Limited by format, the model can only export *.obj mesh with vertex colors. The "texture" mod can only work on *.glb.<br>
+Please click "Do Rendering" to export a GIF.<br>
+You can click "Do Baking" to bake multi-view imgaes onto the shape.<br>
 
-If the results aren't satisfactory, try using a different seed value (default is 0).
+If the results aren't satisfactory, please try a different radnom seed (default is 0).
 '''
 
+################################################################
+# prepare text examples and image examples
 ################################################################
 
 def get_example_img_list():
@@ -143,6 +125,9 @@ def get_example_txt_list():
 
 example_is = get_example_img_list()
 example_ts = get_example_txt_list()
+
+################################################################
+# initial models
 ################################################################
 
 worker_xbg = Removebg()
@@ -166,6 +151,12 @@ worker_v23 = Views2Mesh(
 )
 worker_gif = GifRenderer(args.device)
 
+
+if BAKE_AVAILEBLE:
+    worker_baker = MeshBaker()
+
+
+### functional modules    
 @spaces.GPU
 def stage_0_t2i(text, image, seed, step):
     os.makedirs('./outputs/app_output', exist_ok=True)
@@ -177,7 +168,7 @@ def stage_0_t2i(text, image, seed, step):
     save_folder = f'./outputs/app_output/{cur_id}'
     os.makedirs(save_folder, exist_ok=True)
 
-    dst = os.path.join(save_folder, 'img.png')
+    dst = save_folder + '/img.png'
     
     if not text:
         if image is None: 
@@ -190,16 +181,16 @@ def stage_0_t2i(text, image, seed, step):
     image.save(dst)
     dst = worker_xbg(image, save_folder)
     return dst, save_folder
-    
+
 @spaces.GPU
-def stage_1_xbg(image, save_folder): 
+def stage_1_xbg(image, save_folder, force_remove): 
     if isinstance(image, str):
         image = Image.open(image)
     dst =  save_folder + '/img_nobg.png'
-    rgba = worker_xbg(image)
+    rgba = worker_xbg(image, force=force_remove)
     rgba.save(dst)
     return dst
-    
+
 @spaces.GPU
 def stage_2_i2v(image, seed, step, save_folder):
     if isinstance(image, str):
@@ -222,12 +213,9 @@ def stage_3_v23(
     seed, 
     save_folder,
     target_face_count = 30000,
-    do_texture_mapping = True,
-    do_render =True
+    texture_color = 'texture'
 ): 
-    do_texture_mapping = do_texture_mapping or do_render
-    obj_dst = save_folder + '/mesh_with_colors.obj'
-    glb_dst = save_folder + '/mesh.glb'
+    do_texture_mapping = texture_color == 'texture'
     worker_v23(
         views_pil, 
         cond_pil, 
@@ -236,95 +224,197 @@ def stage_3_v23(
         target_face_count = target_face_count,
         do_texture_mapping = do_texture_mapping
     )
+    glb_dst = save_folder + '/mesh.glb' if do_texture_mapping else None
+    obj_dst =  save_folder + '/mesh.obj'
+    obj_dst = save_folder + '/mesh_vertex_colors.obj' # gradio just only can show vertex shading
     return obj_dst, glb_dst
 
 @spaces.GPU
-def stage_4_gif(obj_dst, save_folder, do_render_gif=True):
-    if not do_render_gif: return None
-    gif_dst = save_folder + '/output.gif'
-    worker_gif(
-        save_folder + '/mesh.obj',
-        gif_dst_path = gif_dst
-    )
+def stage_3p_baking(save_folder, color, bake):
+    if color == "texture" and bake:
+        obj_dst = worker_baker(save_folder)
+        glb_dst = obj_dst.replace(".obj", ".glb")
+        return glb_dst
+    else:
+        return None
+    
+@spaces.GPU
+def stage_4_gif(save_folder, color, bake, render):
+    if not render: return None
+    if os.path.exists(save_folder + '/view_1/bake/mesh.obj'):
+        obj_dst = save_folder + '/view_1/bake/mesh.obj'
+    elif os.path.exists(save_folder + '/view_0/bake/mesh.obj'):
+        obj_dst = save_folder + '/view_0/bake/mesh.obj'
+    elif os.path.exists(save_folder + '/mesh.obj'):
+        obj_dst = save_folder + '/mesh.obj'
+    else:
+        print(save_folder)
+        raise FileNotFoundError("mesh obj file not found")
+    gif_dst = obj_dst.replace(".obj", ".gif")
+    worker_gif(obj_dst, gif_dst_path=gif_dst)
     return gif_dst
 
-#===============================================================
+
+def check_image_available(image):
+    if image.mode == "RGBA":
+        data = np.array(image)
+        alpha_channel = data[:, :, 3]
+        unique_alpha_values = np.unique(alpha_channel)
+        if len(unique_alpha_values) == 1:
+            msg = "The alpha channel is missing or invalid. The background removal option is selected for you."
+            return msg, gr.update(value=True, interactive=False)
+        else:
+            msg = "The image has four channels, and you can choose to remove the background or not."
+            return msg, gr.update(value=False, interactive=True)
+    elif image.mode == "RGB":
+        msg = "The alpha channel is missing or invalid. The background removal option is selected for you."
+        return msg, gr.update(value=True, interactive=False)
+    else:
+        raise Exception("Image Error")
+    
+def update_bake_render(color):
+    if color == "vertex":
+        return gr.update(value=False, interactive=False), gr.update(value=False, interactive=False)
+    else:
+        return gr.update(interactive=True), gr.update(interactive=True)
+    
+# ===============================================================
+# gradio display
+# ===============================================================
+
 with gr.Blocks() as demo:
     gr.Markdown(CONST_HEADER)
     with gr.Row(variant="panel"):
+        
+        ###### Input region
+        
         with gr.Column(scale=2):
+            
+            ### Text iutput region
+            
             with gr.Tab("Text to 3D"):
                 with gr.Column():
-                    text = gr.TextArea('一只黑白相间的熊猫在白色背景上居中坐着，呈现出卡通风格和可爱氛围。', lines=1, max_lines=10, label='Input text')
+                    text = gr.TextArea('一只黑白相间的熊猫在白色背景上居中坐着，呈现出卡通风格和可爱氛围。', 
+                                       lines=3, max_lines=20, label='Input text')
                     with gr.Row():
-                        textgen_seed = gr.Number(value=0, label="T2I seed", precision=0)
-                        textgen_step = gr.Number(value=25, label="T2I steps", precision=0, minimum=10, maximum=50)
-                        textgen_SEED = gr.Number(value=0, label="Gen seed", precision=0)
-                        textgen_STEP = gr.Number(value=50, label="Gen steps", precision=0, minimum=40, maximum=100)
-                        textgen_max_faces = gr.Number(value=90000, label="Face number", precision=0, minimum=5000, maximum=1000000)
-                        
+                        textgen_color = gr.Radio(choices=["vertex", "texture"], label="Color", value="texture")
+                    with gr.Row():
+                        textgen_render = gr.Checkbox(label="Do Rendering", value=True, interactive=True)
+                        if BAKE_AVAILEBLE:
+                            textgen_bake = gr.Checkbox(label="Do Baking", value=True, interactive=True)
+                        else:
+                            textgen_bake = gr.Checkbox(label="Do Baking", value=False, interactive=False)
+                    
+                    textgen_color.change(fn=update_bake_render, inputs=textgen_color, outputs=[textgen_bake, textgen_render])
+
+                    with gr.Row():
+                        textgen_seed = gr.Number(value=0, label="T2I seed", precision=0, interactive=True)
+                        textgen_step = gr.Number(value=25, label="T2I steps", precision=0, 
+                                                 minimum=10, maximum=50, interactive=True)
+                        textgen_SEED = gr.Number(value=0, label="Gen seed", precision=0, interactive=True)
+                        textgen_STEP = gr.Number(value=50, label="Gen steps", precision=0, 
+                                                 minimum=40, maximum=100, interactive=True)
+                        textgen_max_faces = gr.Number(value=90000, label="Face number", precision=0, 
+                                                      minimum=5000, maximum=1000000, interactive=True)
                     with gr.Row():
-                        # textgen_do_texture_mapping = gr.Checkbox(label="Texture mapping", value=False, interactive=True)
-                        # textgen_do_render_gif = gr.Checkbox(label="Render GIF", value=False, interactive=True)
                         textgen_submit = gr.Button("Generate", variant="primary")
 
                     with gr.Row():
                         gr.Examples(examples=example_ts, inputs=[text], label="Text examples", examples_per_page=10)
                     
+            ### Image iutput region
+            
             with gr.Tab("Image to 3D"):
-                with gr.Column():
-                    input_image = gr.Image(label="Input image",
-                                           width=256, height=256, type="pil",
-                                           image_mode="RGBA", sources="upload",
-                                           interactive=True)
-                    with gr.Row(): 
-                        imggen_SEED = gr.Number(value=0, label="Gen seed", precision=0)
-                        imggen_STEP = gr.Number(value=50, label="Gen steps", precision=0, minimum=40, maximum=100)
-                        imggen_max_faces = gr.Number(value=90000, label="Face number", precision=0, minimum=5000, maximum=1000000)
+                with gr.Row():
+                    input_image = gr.Image(label="Input image", width=256, height=256, type="pil",
+                                           image_mode="RGBA", sources="upload", interactive=True)
+                with gr.Row():
+                    alert_message = gr.Markdown("")  # for warning 
+                with gr.Row():
+                    imggen_color = gr.Radio(choices=["vertex", "texture"], label="Color", value="texture")
+                with gr.Row():
+                    imggen_removebg = gr.Checkbox(label="Remove Background", value=True, interactive=True)
+                    imggen_render = gr.Checkbox(label="Do Rendering", value=True, interactive=True)
+                    if BAKE_AVAILEBLE:
+                        imggen_bake = gr.Checkbox(label="Do Baking", value=True, interactive=True)
+                    else:
+                        imggen_bake = gr.Checkbox(label="Do Baking", value=False, interactive=False)
+                    
+                input_image.change(fn=check_image_available, inputs=input_image, outputs=[alert_message, imggen_removebg])
+                imggen_color.change(fn=update_bake_render, inputs=imggen_color, outputs=[imggen_bake, imggen_render])
+                        
+                with gr.Row():
+                    imggen_SEED = gr.Number(value=0, label="Gen seed", precision=0, interactive=True)
+                    imggen_STEP = gr.Number(value=50, label="Gen steps", precision=0, 
+                                            minimum=40, maximum=100, interactive=True)
+                    imggen_max_faces = gr.Number(value=90000, label="Face number", precision=0, 
+                                                     minimum=5000, maximum=1000000, interactive=True)
+                with gr.Row():
+                    imggen_submit = gr.Button("Generate", variant="primary")      
+
+                with gr.Row():
+                    gr.Examples(examples=example_is,  inputs=[input_image], 
+                        label="Img examples", examples_per_page=10)
+                    
+            gr.Markdown(CONST_NOTE)
+                    
+        ###### Output region
 
-                    with gr.Row():
-                        # imggen_do_texture_mapping = gr.Checkbox(label="Texture mapping", value=False, interactive=True)
-                        # imggen_do_render_gif = gr.Checkbox(label="Render GIF", value=False, interactive=True)
-                        imggen_submit = gr.Button("Generate", variant="primary")       
-                    with gr.Row():
-                        gr.Examples(examples=example_is, inputs=[input_image], label="Img examples", examples_per_page=10)
-           
         with gr.Column(scale=3):
             with gr.Row():
                 with gr.Column(scale=2):
-                    rem_bg_image = gr.Image(label="Image without background", type="pil",
-                                           image_mode="RGBA", interactive=False)
+                    rem_bg_image = gr.Image(
+                        label="Image without background", 
+                        type="pil",
+                        image_mode="RGBA", 
+                        interactive=False
+                    )
                 with gr.Column(scale=3):
-                    result_image = gr.Image(label="Multi-view images", type="pil", interactive=False)
-                
-            with gr.Row():                
+                    result_image = gr.Image(
+                        label="Multi-view images", 
+                        type="pil", 
+                        interactive=False
+                    )
+            
+            with gr.Row():          
                 result_3dobj = gr.Model3D(
                     clear_color=[0.0, 0.0, 0.0, 0.0],
-                    label="OBJ",
+                    label="OBJ vertex color",
                     show_label=True,
                     visible=True,
                     camera_position=[90, 90, None],
                     interactive=False
                 )
+                result_gif = gr.Image(label="GIF", interactive=False)
+                
+            with gr.Row():
+                result_3dglb_texture = gr.Model3D(
+                    clear_color=[0.0, 0.0, 0.0, 0.0],
+                    label="GLB texture color",
+                    show_label=True,
+                    visible=True,
+                    camera_position=[90, 90, None],
+                    interactive=False)
 
-                result_3dglb = gr.Model3D(
+                result_3dglb_baked = gr.Model3D(
                     clear_color=[0.0, 0.0, 0.0, 0.0],
-                    label="GLB",
+                    label="GLB baked color",
                     show_label=True,
                     visible=True,
                     camera_position=[90, 90, None],
-                    interactive=False
-                )
-                # result_gif = gr.Image(label="Rendered GIF", interactive=False)
+                    interactive=False)
                 
-            with gr.Row():    
-                gr.Markdown("""Due to Gradio limitations, OBJ files are displayed with vertex shading only, while GLB files can be viewed with texture shading. For the best experience, we recommend downloading the GLB files and opening them with 3D software like Blender or MeshLab.""")
-
-#===============================================================
-    textgen_do_texture_mapping = gr.State(False)
-    textgen_do_render_gif = gr.State(False)
-    imggen_do_texture_mapping = gr.State(False)
-    imggen_do_render_gif = gr.State(False)
+            with gr.Row():
+                gr.Markdown(
+                    "Due to Gradio limitations, OBJ files are displayed with vertex shading only, "
+                    "while GLB files can be viewed with texture shading. <br>For the best experience, "
+                    "we recommend downloading the GLB files and opening them with 3D software "
+                    "like Blender or MeshLab."
+                )
+
+    #===============================================================
+    # gradio running code
+    #===============================================================
     
     none = gr.State(None)
     save_folder = gr.State()
@@ -332,41 +422,60 @@ with gr.Blocks() as demo:
     views_image = gr.State()
     text_image = gr.State()
     
+    
     textgen_submit.click(
-        fn=stage_0_t2i, inputs=[text, none, textgen_seed, textgen_step], 
+        fn=stage_0_t2i, 
+        inputs=[text, none, textgen_seed, textgen_step], 
         outputs=[rem_bg_image, save_folder],
     ).success(
-        fn=stage_2_i2v, inputs=[rem_bg_image, textgen_SEED, textgen_STEP, save_folder], 
+        fn=stage_2_i2v, 
+        inputs=[rem_bg_image, textgen_SEED, textgen_STEP, save_folder], 
         outputs=[views_image, cond_image, result_image],
     ).success(
-        fn=stage_3_v23, inputs=[views_image, cond_image, textgen_SEED, save_folder, textgen_max_faces, textgen_do_texture_mapping, textgen_do_render_gif], 
-        outputs=[result_3dobj, result_3dglb],
+        fn=stage_3_v23, 
+        inputs=[views_image, cond_image, textgen_SEED, save_folder, textgen_max_faces, textgen_color], 
+        outputs=[result_3dobj, result_3dglb_texture],
+    ).success(
+        fn=stage_3p_baking, 
+        inputs=[save_folder, textgen_color, textgen_bake], 
+        outputs=[result_3dglb_baked],
+    ).success(
+        fn=stage_4_gif, 
+        inputs=[save_folder, textgen_color, textgen_bake, textgen_render], 
+        outputs=[result_gif],
     ).success(lambda: print('Text_to_3D Done ...'))
-    # .success(
-    #     fn=stage_4_gif, inputs=[result_3dglb, save_folder, textgen_do_render_gif], 
-    #     outputs=[result_gif],
-    # ).success(lambda: print('Text_to_3D Done ...'))
 
+    
     imggen_submit.click(
-        fn=stage_0_t2i, inputs=[none, input_image, textgen_seed, textgen_step], 
+        fn=stage_0_t2i, 
+        inputs=[none, input_image, textgen_seed, textgen_step], 
         outputs=[text_image, save_folder],
     ).success(
-        fn=stage_1_xbg, inputs=[text_image, save_folder], 
+        fn=stage_1_xbg, 
+        inputs=[text_image, save_folder, imggen_removebg], 
         outputs=[rem_bg_image],
     ).success(
-        fn=stage_2_i2v, inputs=[rem_bg_image, imggen_SEED, imggen_STEP, save_folder], 
+        fn=stage_2_i2v, 
+        inputs=[rem_bg_image, imggen_SEED, imggen_STEP, save_folder], 
         outputs=[views_image, cond_image, result_image],
     ).success(
-        fn=stage_3_v23, inputs=[views_image, cond_image, imggen_SEED, save_folder, imggen_max_faces, imggen_do_texture_mapping, imggen_do_render_gif], 
-        outputs=[result_3dobj, result_3dglb],
+        fn=stage_3_v23, 
+        inputs=[views_image, cond_image, imggen_SEED, save_folder, imggen_max_faces, imggen_color],
+        outputs=[result_3dobj, result_3dglb_texture],
+    ).success(
+        fn=stage_3p_baking, 
+        inputs=[save_folder, imggen_color, imggen_bake], 
+        outputs=[result_3dglb_baked],
+    ).success(
+        fn=stage_4_gif, 
+        inputs=[save_folder, imggen_color, imggen_bake, imggen_render], 
+        outputs=[result_gif],
     ).success(lambda: print('Image_to_3D Done ...'))
-    # success(
-    #     fn=stage_4_gif, inputs=[result_3dglb, save_folder, imggen_do_render_gif], 
-    #     outputs=[result_gif],
-    # ).success(lambda: print('Image_to_3D Done ...'))
     
-#===============================================================
+    #===============================================================
+    # start gradio server
+    #===============================================================
 
-    demo.queue()
-    demo.launch()
+    demo.queue(max_size=CONST_MAX_QUEUE)
+    demo.launch(server_name=CONST_SERVER, server_port=CONST_PORT)
 

third_party/dust3r/.gitignore

@@ -0,0 +1,132 @@
+data/
+checkpoints/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/

third_party/dust3r/.gitmodules

@@ -0,0 +1,3 @@
+[submodule "croco"]
+	path = croco
+	url = https://github.com/naver/croco

third_party/dust3r/LICENSE

@@ -0,0 +1,7 @@
+DUSt3R, Copyright (c) 2024-present Naver Corporation, is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 license.
+
+A summary of the CC BY-NC-SA 4.0 license is located here:
+	https://creativecommons.org/licenses/by-nc-sa/4.0/
+
+The CC BY-NC-SA 4.0 license is located here:
+	https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode

third_party/dust3r/NOTICE

@@ -0,0 +1,12 @@
+DUSt3R
+Copyright 2024-present NAVER Corp.
+
+This project contains subcomponents with separate copyright notices and license terms. 
+Your use of the source code for these subcomponents is subject to the terms and conditions of the following licenses.
+
+====
+
+naver/croco
+https://github.com/naver/croco/
+
+Creative Commons Attribution-NonCommercial-ShareAlike 4.0

third_party/dust3r/README.md

@@ -0,0 +1,390 @@
+![demo](assets/dust3r.jpg)
+
+Official implementation of `DUSt3R: Geometric 3D Vision Made Easy`  
+[[Project page](https://dust3r.europe.naverlabs.com/)], [[DUSt3R arxiv](https://arxiv.org/abs/2312.14132)]  
+
+> **Make sure to also check [MASt3R](https://github.com/naver/mast3r): Our new model with a local feature head, metric pointmaps, and a more scalable global alignment!**
+
+![Example of reconstruction from two images](assets/pipeline1.jpg)
+
+![High level overview of DUSt3R capabilities](assets/dust3r_archi.jpg)
+
+```bibtex
+@inproceedings{dust3r_cvpr24,
+      title={DUSt3R: Geometric 3D Vision Made Easy}, 
+      author={Shuzhe Wang and Vincent Leroy and Yohann Cabon and Boris Chidlovskii and Jerome Revaud},
+      booktitle = {CVPR},
+      year = {2024}
+}
+
+@misc{dust3r_arxiv23,
+      title={DUSt3R: Geometric 3D Vision Made Easy}, 
+      author={Shuzhe Wang and Vincent Leroy and Yohann Cabon and Boris Chidlovskii and Jerome Revaud},
+      year={2023},
+      eprint={2312.14132},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
+}
+```
+
+## Table of Contents
+
+- [Table of Contents](#table-of-contents)
+- [License](#license)
+- [Get Started](#get-started)
+  - [Installation](#installation)
+  - [Checkpoints](#checkpoints)
+  - [Interactive demo](#interactive-demo)
+  - [Interactive demo with docker](#interactive-demo-with-docker)
+- [Usage](#usage)
+- [Training](#training)
+  - [Datasets](#datasets)
+  - [Demo](#demo)
+  - [Our Hyperparameters](#our-hyperparameters)
+
+## License
+
+The code is distributed under the CC BY-NC-SA 4.0 License.
+See [LICENSE](LICENSE) for more information.
+
+```python
+# Copyright (C) 2024-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+```
+
+## Get Started
+
+### Installation
+
+1. Clone DUSt3R.
+```bash
+git clone --recursive https://github.com/naver/dust3r
+cd dust3r
+# if you have already cloned dust3r:
+# git submodule update --init --recursive
+```
+
+2. Create the environment, here we show an example using conda.
+```bash
+conda create -n dust3r python=3.11 cmake=3.14.0
+conda activate dust3r 
+conda install pytorch torchvision pytorch-cuda=12.1 -c pytorch -c nvidia  # use the correct version of cuda for your system
+pip install -r requirements.txt
+# Optional: you can also install additional packages to:
+# - add support for HEIC images
+# - add pyrender, used to render depthmap in some datasets preprocessing
+# - add required packages for visloc.py
+pip install -r requirements_optional.txt
+```
+
+3. Optional, compile the cuda kernels for RoPE (as in CroCo v2).
+```bash
+# DUST3R relies on RoPE positional embeddings for which you can compile some cuda kernels for faster runtime.
+cd croco/models/curope/
+python setup.py build_ext --inplace
+cd ../../../
+```
+
+### Checkpoints
+
+You can obtain the checkpoints by two ways:
+
+1) You can use our huggingface_hub integration: the models will be downloaded automatically.
+
+2) Otherwise, We provide several pre-trained models:
+
+| Modelname   | Training resolutions | Head | Encoder | Decoder |
+|-------------|----------------------|------|---------|---------|
+| [`DUSt3R_ViTLarge_BaseDecoder_224_linear.pth`](https://download.europe.naverlabs.com/ComputerVision/DUSt3R/DUSt3R_ViTLarge_BaseDecoder_224_linear.pth) | 224x224 | Linear | ViT-L | ViT-B |
+| [`DUSt3R_ViTLarge_BaseDecoder_512_linear.pth`](https://download.europe.naverlabs.com/ComputerVision/DUSt3R/DUSt3R_ViTLarge_BaseDecoder_512_linear.pth)   | 512x384, 512x336, 512x288, 512x256, 512x160 | Linear | ViT-L | ViT-B |
+| [`DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth`](https://download.europe.naverlabs.com/ComputerVision/DUSt3R/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth) | 512x384, 512x336, 512x288, 512x256, 512x160 | DPT | ViT-L | ViT-B |
+
+You can check the hyperparameters we used to train these models in the [section: Our Hyperparameters](#our-hyperparameters)
+
+To download a specific model, for example `DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth`:
+```bash
+mkdir -p checkpoints/
+wget https://download.europe.naverlabs.com/ComputerVision/DUSt3R/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth -P checkpoints/
+```
+
+For the checkpoints, make sure to agree to the license of all the public training datasets and base checkpoints we used, in addition to CC-BY-NC-SA 4.0. Again, see [section: Our Hyperparameters](#our-hyperparameters) for details.
+
+### Interactive demo
+
+In this demo, you should be able run DUSt3R on your machine to reconstruct a scene.
+First select images that depicts the same scene.
+
+You can adjust the global alignment schedule and its number of iterations.
+
+> [!NOTE]
+> If you selected one or two images, the global alignment procedure will be skipped (mode=GlobalAlignerMode.PairViewer)
+
+Hit "Run" and wait.
+When the global alignment ends, the reconstruction appears.
+Use the slider "min_conf_thr" to show or remove low confidence areas.
+
+```bash
+python3 demo.py --model_name DUSt3R_ViTLarge_BaseDecoder_512_dpt
+
+# Use --weights to load a checkpoint from a local file, eg --weights checkpoints/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth
+# Use --image_size to select the correct resolution for the selected checkpoint. 512 (default) or 224
+# Use --local_network to make it accessible on the local network, or --server_name to specify the url manually
+# Use --server_port to change the port, by default it will search for an available port starting at 7860
+# Use --device to use a different device, by default it's "cuda"
+```
+
+### Interactive demo with docker
+
+To run DUSt3R using Docker, including with NVIDIA CUDA support, follow these instructions:
+
+1. **Install Docker**: If not already installed, download and install `docker` and `docker compose` from the [Docker website](https://www.docker.com/get-started).
+
+2. **Install NVIDIA Docker Toolkit**: For GPU support, install the NVIDIA Docker toolkit from the [Nvidia website](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html).
+
+3. **Build the Docker image and run it**: `cd` into the `./docker` directory and run the following commands: 
+
+```bash
+cd docker
+bash run.sh --with-cuda --model_name="DUSt3R_ViTLarge_BaseDecoder_512_dpt"
+```
+
+Or if you want to run the demo without CUDA support, run the following command:
+
+```bash 
+cd docker
+bash run.sh --model_name="DUSt3R_ViTLarge_BaseDecoder_512_dpt"
+```
+
+By default, `demo.py` is lanched with the option `--local_network`.  
+Visit `http://localhost:7860/` to access the web UI (or replace `localhost` with the machine's name to access it from the network).  
+
+`run.sh` will launch docker-compose using either the [docker-compose-cuda.yml](docker/docker-compose-cuda.yml) or [docker-compose-cpu.ym](docker/docker-compose-cpu.yml) config file, then it starts the demo using [entrypoint.sh](docker/files/entrypoint.sh).
+
+
+![demo](assets/demo.jpg)
+
+## Usage
+
+```python
+from dust3r.inference import inference
+from dust3r.model import AsymmetricCroCo3DStereo
+from dust3r.utils.image import load_images
+from dust3r.image_pairs import make_pairs
+from dust3r.cloud_opt import global_aligner, GlobalAlignerMode
+
+if __name__ == '__main__':
+    device = 'cuda'
+    batch_size = 1
+    schedule = 'cosine'
+    lr = 0.01
+    niter = 300
+
+    model_name = "naver/DUSt3R_ViTLarge_BaseDecoder_512_dpt"
+    # you can put the path to a local checkpoint in model_name if needed
+    model = AsymmetricCroCo3DStereo.from_pretrained(model_name).to(device)
+    # load_images can take a list of images or a directory
+    images = load_images(['croco/assets/Chateau1.png', 'croco/assets/Chateau2.png'], size=512)
+    pairs = make_pairs(images, scene_graph='complete', prefilter=None, symmetrize=True)
+    output = inference(pairs, model, device, batch_size=batch_size)
+
+    # at this stage, you have the raw dust3r predictions
+    view1, pred1 = output['view1'], output['pred1']
+    view2, pred2 = output['view2'], output['pred2']
+    # here, view1, pred1, view2, pred2 are dicts of lists of len(2)
+    #  -> because we symmetrize we have (im1, im2) and (im2, im1) pairs
+    # in each view you have:
+    # an integer image identifier: view1['idx'] and view2['idx']
+    # the img: view1['img'] and view2['img']
+    # the image shape: view1['true_shape'] and view2['true_shape']
+    # an instance string output by the dataloader: view1['instance'] and view2['instance']
+    # pred1 and pred2 contains the confidence values: pred1['conf'] and pred2['conf']
+    # pred1 contains 3D points for view1['img'] in view1['img'] space: pred1['pts3d']
+    # pred2 contains 3D points for view2['img'] in view1['img'] space: pred2['pts3d_in_other_view']
+
+    # next we'll use the global_aligner to align the predictions
+    # depending on your task, you may be fine with the raw output and not need it
+    # with only two input images, you could use GlobalAlignerMode.PairViewer: it would just convert the output
+    # if using GlobalAlignerMode.PairViewer, no need to run compute_global_alignment
+    scene = global_aligner(output, device=device, mode=GlobalAlignerMode.PointCloudOptimizer)
+    loss = scene.compute_global_alignment(init="mst", niter=niter, schedule=schedule, lr=lr)
+
+    # retrieve useful values from scene:
+    imgs = scene.imgs
+    focals = scene.get_focals()
+    poses = scene.get_im_poses()
+    pts3d = scene.get_pts3d()
+    confidence_masks = scene.get_masks()
+
+    # visualize reconstruction
+    scene.show()
+
+    # find 2D-2D matches between the two images
+    from dust3r.utils.geometry import find_reciprocal_matches, xy_grid
+    pts2d_list, pts3d_list = [], []
+    for i in range(2):
+        conf_i = confidence_masks[i].cpu().numpy()
+        pts2d_list.append(xy_grid(*imgs[i].shape[:2][::-1])[conf_i])  # imgs[i].shape[:2] = (H, W)
+        pts3d_list.append(pts3d[i].detach().cpu().numpy()[conf_i])
+    reciprocal_in_P2, nn2_in_P1, num_matches = find_reciprocal_matches(*pts3d_list)
+    print(f'found {num_matches} matches')
+    matches_im1 = pts2d_list[1][reciprocal_in_P2]
+    matches_im0 = pts2d_list[0][nn2_in_P1][reciprocal_in_P2]
+
+    # visualize a few matches
+    import numpy as np
+    from matplotlib import pyplot as pl
+    n_viz = 10
+    match_idx_to_viz = np.round(np.linspace(0, num_matches-1, n_viz)).astype(int)
+    viz_matches_im0, viz_matches_im1 = matches_im0[match_idx_to_viz], matches_im1[match_idx_to_viz]
+
+    H0, W0, H1, W1 = *imgs[0].shape[:2], *imgs[1].shape[:2]
+    img0 = np.pad(imgs[0], ((0, max(H1 - H0, 0)), (0, 0), (0, 0)), 'constant', constant_values=0)
+    img1 = np.pad(imgs[1], ((0, max(H0 - H1, 0)), (0, 0), (0, 0)), 'constant', constant_values=0)
+    img = np.concatenate((img0, img1), axis=1)
+    pl.figure()
+    pl.imshow(img)
+    cmap = pl.get_cmap('jet')
+    for i in range(n_viz):
+        (x0, y0), (x1, y1) = viz_matches_im0[i].T, viz_matches_im1[i].T
+        pl.plot([x0, x1 + W0], [y0, y1], '-+', color=cmap(i / (n_viz - 1)), scalex=False, scaley=False)
+    pl.show(block=True)
+
+```
+![matching example on croco pair](assets/matching.jpg)
+
+## Training
+
+In this section, we present a short demonstration to get started with training DUSt3R.
+
+### Datasets
+At this moment, we have added the following training datasets:
+  - [CO3Dv2](https://github.com/facebookresearch/co3d) - [Creative Commons Attribution-NonCommercial 4.0 International](https://github.com/facebookresearch/co3d/blob/main/LICENSE)
+  - [ARKitScenes](https://github.com/apple/ARKitScenes) - [Creative Commons Attribution-NonCommercial-ShareAlike 4.0](https://github.com/apple/ARKitScenes/tree/main?tab=readme-ov-file#license)
+  - [ScanNet++](https://kaldir.vc.in.tum.de/scannetpp/) - [non-commercial research and educational purposes](https://kaldir.vc.in.tum.de/scannetpp/static/scannetpp-terms-of-use.pdf)
+  - [BlendedMVS](https://github.com/YoYo000/BlendedMVS) - [Creative Commons Attribution 4.0 International License](https://creativecommons.org/licenses/by/4.0/)
+  - [WayMo Open dataset](https://github.com/waymo-research/waymo-open-dataset) - [Non-Commercial Use](https://waymo.com/open/terms/)
+  - [Habitat-Sim](https://github.com/facebookresearch/habitat-sim/blob/main/DATASETS.md)
+  - [MegaDepth](https://www.cs.cornell.edu/projects/megadepth/)
+  - [StaticThings3D](https://github.com/lmb-freiburg/robustmvd/blob/master/rmvd/data/README.md#staticthings3d)
+  - [WildRGB-D](https://github.com/wildrgbd/wildrgbd/)
+
+For each dataset, we provide a preprocessing script in the `datasets_preprocess` directory and an archive containing the list of pairs when needed.
+You have to download the datasets yourself from their official sources, agree to their license, download our list of pairs, and run the preprocessing script.
+
+Links:  
+  
+[ARKitScenes pairs](https://download.europe.naverlabs.com/ComputerVision/DUSt3R/arkitscenes_pairs.zip)  
+[ScanNet++ pairs](https://download.europe.naverlabs.com/ComputerVision/DUSt3R/scannetpp_pairs.zip)  
+[BlendedMVS pairs](https://download.europe.naverlabs.com/ComputerVision/DUSt3R/blendedmvs_pairs.npy)  
+[WayMo Open dataset pairs](https://download.europe.naverlabs.com/ComputerVision/DUSt3R/waymo_pairs.npz)  
+[Habitat metadata](https://download.europe.naverlabs.com/ComputerVision/DUSt3R/habitat_5views_v1_512x512_metadata.tar.gz)  
+[MegaDepth pairs](https://download.europe.naverlabs.com/ComputerVision/DUSt3R/megadepth_pairs.npz)  
+[StaticThings3D pairs](https://download.europe.naverlabs.com/ComputerVision/DUSt3R/staticthings_pairs.npy)  
+
+> [!NOTE]
+> They are not strictly equivalent to what was used to train DUSt3R, but they should be close enough.
+
+### Demo
+For this training demo, we're going to download and prepare a subset of [CO3Dv2](https://github.com/facebookresearch/co3d) - [Creative Commons Attribution-NonCommercial 4.0 International](https://github.com/facebookresearch/co3d/blob/main/LICENSE) and launch the training code on it.
+The demo model will be trained for a few epochs on a very small dataset.
+It will not be very good.
+
+```bash
+# download and prepare the co3d subset
+mkdir -p data/co3d_subset
+cd data/co3d_subset
+git clone https://github.com/facebookresearch/co3d
+cd co3d
+python3 ./co3d/download_dataset.py --download_folder ../ --single_sequence_subset
+rm ../*.zip
+cd ../../..
+
+python3 datasets_preprocess/preprocess_co3d.py --co3d_dir data/co3d_subset --output_dir data/co3d_subset_processed  --single_sequence_subset
+
+# download the pretrained croco v2 checkpoint
+mkdir -p checkpoints/
+wget https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo_V2_ViTLarge_BaseDecoder.pth -P checkpoints/
+
+# the training of dust3r is done in 3 steps.
+# for this example we'll do fewer epochs, for the actual hyperparameters we used in the paper, see the next section: "Our Hyperparameters"
+# step 1 - train dust3r for 224 resolution
+torchrun --nproc_per_node=4 train.py \
+    --train_dataset "1000 @ Co3d(split='train', ROOT='data/co3d_subset_processed', aug_crop=16, mask_bg='rand', resolution=224, transform=ColorJitter)" \
+    --test_dataset "100 @ Co3d(split='test', ROOT='data/co3d_subset_processed', resolution=224, seed=777)" \
+    --model "AsymmetricCroCo3DStereo(pos_embed='RoPE100', img_size=(224, 224), head_type='linear', output_mode='pts3d', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12)" \
+    --train_criterion "ConfLoss(Regr3D(L21, norm_mode='avg_dis'), alpha=0.2)" \
+    --test_criterion "Regr3D_ScaleShiftInv(L21, gt_scale=True)" \
+    --pretrained "checkpoints/CroCo_V2_ViTLarge_BaseDecoder.pth" \
+    --lr 0.0001 --min_lr 1e-06 --warmup_epochs 1 --epochs 10 --batch_size 16 --accum_iter 1 \
+    --save_freq 1 --keep_freq 5 --eval_freq 1 \
+    --output_dir "checkpoints/dust3r_demo_224"	  
+
+# step 2 - train dust3r for 512 resolution
+torchrun --nproc_per_node=4 train.py \
+    --train_dataset "1000 @ Co3d(split='train', ROOT='data/co3d_subset_processed', aug_crop=16, mask_bg='rand', resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter)" \
+    --test_dataset "100 @ Co3d(split='test', ROOT='data/co3d_subset_processed', resolution=(512,384), seed=777)" \
+    --model "AsymmetricCroCo3DStereo(pos_embed='RoPE100', patch_embed_cls='ManyAR_PatchEmbed', img_size=(512, 512), head_type='linear', output_mode='pts3d', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12)" \
+    --train_criterion "ConfLoss(Regr3D(L21, norm_mode='avg_dis'), alpha=0.2)" \
+    --test_criterion "Regr3D_ScaleShiftInv(L21, gt_scale=True)" \
+    --pretrained "checkpoints/dust3r_demo_224/checkpoint-best.pth" \
+    --lr 0.0001 --min_lr 1e-06 --warmup_epochs 1 --epochs 10 --batch_size 4 --accum_iter 4 \
+    --save_freq 1 --keep_freq 5 --eval_freq 1 \
+    --output_dir "checkpoints/dust3r_demo_512"
+
+# step 3 - train dust3r for 512 resolution with dpt
+torchrun --nproc_per_node=4 train.py \
+    --train_dataset "1000 @ Co3d(split='train', ROOT='data/co3d_subset_processed', aug_crop=16, mask_bg='rand', resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter)" \
+    --test_dataset "100 @ Co3d(split='test', ROOT='data/co3d_subset_processed', resolution=(512,384), seed=777)" \
+    --model "AsymmetricCroCo3DStereo(pos_embed='RoPE100', patch_embed_cls='ManyAR_PatchEmbed', img_size=(512, 512), head_type='dpt', output_mode='pts3d', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12)" \
+    --train_criterion "ConfLoss(Regr3D(L21, norm_mode='avg_dis'), alpha=0.2)" \
+    --test_criterion "Regr3D_ScaleShiftInv(L21, gt_scale=True)" \
+    --pretrained "checkpoints/dust3r_demo_512/checkpoint-best.pth" \
+    --lr 0.0001 --min_lr 1e-06 --warmup_epochs 1 --epochs 10 --batch_size 2 --accum_iter 8 \
+    --save_freq 1 --keep_freq 5 --eval_freq 1 --disable_cudnn_benchmark \
+    --output_dir "checkpoints/dust3r_demo_512dpt"
+
+```
+
+### Our Hyperparameters
+
+Here are the commands we used for training the models:
+
+```bash
+# NOTE: ROOT path omitted for datasets
+# 224 linear
+torchrun --nproc_per_node 8 train.py \
+    --train_dataset=" + 100_000 @ Habitat(1_000_000, split='train', aug_crop=16, resolution=224, transform=ColorJitter) + 100_000 @ BlendedMVS(split='train', aug_crop=16, resolution=224, transform=ColorJitter) + 100_000 @ MegaDepth(split='train', aug_crop=16, resolution=224, transform=ColorJitter) + 100_000 @ ARKitScenes(aug_crop=256, resolution=224, transform=ColorJitter) + 100_000 @ Co3d(split='train', aug_crop=16, mask_bg='rand', resolution=224, transform=ColorJitter) + 100_000 @ StaticThings3D(aug_crop=256, mask_bg='rand', resolution=224, transform=ColorJitter) + 100_000 @ ScanNetpp(split='train', aug_crop=256, resolution=224, transform=ColorJitter) + 100_000 @ InternalUnreleasedDataset(aug_crop=128, resolution=224, transform=ColorJitter) " \
+    --test_dataset=" Habitat(1_000, split='val', resolution=224, seed=777) + 1_000 @ BlendedMVS(split='val', resolution=224, seed=777) + 1_000 @ MegaDepth(split='val', resolution=224, seed=777) + 1_000 @ Co3d(split='test', mask_bg='rand', resolution=224, seed=777) " \
+    --train_criterion="ConfLoss(Regr3D(L21, norm_mode='avg_dis'), alpha=0.2)" \
+    --test_criterion="Regr3D_ScaleShiftInv(L21, gt_scale=True)" \
+    --model="AsymmetricCroCo3DStereo(pos_embed='RoPE100', img_size=(224, 224), head_type='linear', output_mode='pts3d', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12)" \
+    --pretrained="checkpoints/CroCo_V2_ViTLarge_BaseDecoder.pth" \
+    --lr=0.0001 --min_lr=1e-06 --warmup_epochs=10 --epochs=100 --batch_size=16 --accum_iter=1 \
+    --save_freq=5 --keep_freq=10 --eval_freq=1 \
+    --output_dir="checkpoints/dust3r_224"
+
+# 512 linear
+torchrun --nproc_per_node 8 train.py \
+    --train_dataset=" + 10_000 @ Habitat(1_000_000, split='train', aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ BlendedMVS(split='train', aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ MegaDepth(split='train', aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ ARKitScenes(aug_crop=256, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ Co3d(split='train', aug_crop=16, mask_bg='rand', resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ StaticThings3D(aug_crop=256, mask_bg='rand', resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ ScanNetpp(split='train', aug_crop=256, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ InternalUnreleasedDataset(aug_crop=128, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) " \
+    --test_dataset=" Habitat(1_000, split='val', resolution=(512,384), seed=777) + 1_000 @ BlendedMVS(split='val', resolution=(512,384), seed=777) + 1_000 @ MegaDepth(split='val', resolution=(512,336), seed=777) + 1_000 @ Co3d(split='test', resolution=(512,384), seed=777) " \
+    --train_criterion="ConfLoss(Regr3D(L21, norm_mode='avg_dis'), alpha=0.2)" \
+    --test_criterion="Regr3D_ScaleShiftInv(L21, gt_scale=True)" \
+    --model="AsymmetricCroCo3DStereo(pos_embed='RoPE100', patch_embed_cls='ManyAR_PatchEmbed', img_size=(512, 512), head_type='linear', output_mode='pts3d', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12)" \
+    --pretrained="checkpoints/dust3r_224/checkpoint-best.pth" \
+    --lr=0.0001 --min_lr=1e-06 --warmup_epochs=20 --epochs=100 --batch_size=4 --accum_iter=2 \
+    --save_freq=10 --keep_freq=10 --eval_freq=1 --print_freq=10 \
+    --output_dir="checkpoints/dust3r_512"
+
+# 512 dpt
+torchrun --nproc_per_node 8 train.py \
+    --train_dataset=" + 10_000 @ Habitat(1_000_000, split='train', aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ BlendedMVS(split='train', aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ MegaDepth(split='train', aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ ARKitScenes(aug_crop=256, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ Co3d(split='train', aug_crop=16, mask_bg='rand', resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ StaticThings3D(aug_crop=256, mask_bg='rand', resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ ScanNetpp(split='train', aug_crop=256, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) + 10_000 @ InternalUnreleasedDataset(aug_crop=128, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)], transform=ColorJitter) " \
+    --test_dataset=" Habitat(1_000, split='val', resolution=(512,384), seed=777) + 1_000 @ BlendedMVS(split='val', resolution=(512,384), seed=777) + 1_000 @ MegaDepth(split='val', resolution=(512,336), seed=777) + 1_000 @ Co3d(split='test', resolution=(512,384), seed=777) " \
+    --train_criterion="ConfLoss(Regr3D(L21, norm_mode='avg_dis'), alpha=0.2)" \
+    --test_criterion="Regr3D_ScaleShiftInv(L21, gt_scale=True)" \
+    --model="AsymmetricCroCo3DStereo(pos_embed='RoPE100', patch_embed_cls='ManyAR_PatchEmbed', img_size=(512, 512), head_type='dpt', output_mode='pts3d', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12)" \
+    --pretrained="checkpoints/dust3r_512/checkpoint-best.pth" \
+    --lr=0.0001 --min_lr=1e-06 --warmup_epochs=15 --epochs=90 --batch_size=4 --accum_iter=2 \
+    --save_freq=5 --keep_freq=10 --eval_freq=1 --print_freq=10 --disable_cudnn_benchmark \
+    --output_dir="checkpoints/dust3r_512dpt"
+
+```

third_party/dust3r/croco/LICENSE

@@ -0,0 +1,52 @@
+CroCo, Copyright (c) 2022-present Naver Corporation, is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 license.
+
+A summary of the CC BY-NC-SA 4.0 license is located here:
+	https://creativecommons.org/licenses/by-nc-sa/4.0/
+
+The CC BY-NC-SA 4.0 license is located here:
+	https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode
+	
+	
+SEE NOTICE BELOW WITH RESPECT TO THE FILE: models/pos_embed.py, models/blocks.py
+
+***************************
+
+NOTICE WITH RESPECT TO THE FILE: models/pos_embed.py
+
+This software is being redistributed in a modifiled form. The original form is available here:
+
+https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py
+
+This software in this file incorporates parts of the following software available here:
+
+Transformer: https://github.com/tensorflow/models/blob/master/official/legacy/transformer/model_utils.py
+available under the following license: https://github.com/tensorflow/models/blob/master/LICENSE
+
+MoCo v3: https://github.com/facebookresearch/moco-v3
+available under the following license: https://github.com/facebookresearch/moco-v3/blob/main/LICENSE
+
+DeiT: https://github.com/facebookresearch/deit
+available under the following license: https://github.com/facebookresearch/deit/blob/main/LICENSE
+
+
+ORIGINAL COPYRIGHT NOTICE AND PERMISSION NOTICE AVAILABLE HERE IS REPRODUCE BELOW:
+
+https://github.com/facebookresearch/mae/blob/main/LICENSE
+
+Attribution-NonCommercial 4.0 International
+
+***************************
+
+NOTICE WITH RESPECT TO THE FILE: models/blocks.py
+
+This software is being redistributed in a modifiled form. The original form is available here:
+
+https://github.com/rwightman/pytorch-image-models
+
+ORIGINAL COPYRIGHT NOTICE AND PERMISSION NOTICE AVAILABLE HERE IS REPRODUCE BELOW:
+
+https://github.com/rwightman/pytorch-image-models/blob/master/LICENSE
+
+Apache License
+Version 2.0, January 2004
+http://www.apache.org/licenses/

third_party/dust3r/croco/NOTICE

@@ -0,0 +1,21 @@
+CroCo
+Copyright 2022-present NAVER Corp.
+
+This project contains subcomponents with separate copyright notices and license terms. 
+Your use of the source code for these subcomponents is subject to the terms and conditions of the following licenses.
+
+====
+
+facebookresearch/mae
+https://github.com/facebookresearch/mae
+
+Attribution-NonCommercial 4.0 International
+
+====
+
+rwightman/pytorch-image-models
+https://github.com/rwightman/pytorch-image-models
+
+Apache License
+Version 2.0, January 2004
+http://www.apache.org/licenses/

third_party/dust3r/croco/README.MD

@@ -0,0 +1,124 @@
+# CroCo + CroCo v2 / CroCo-Stereo / CroCo-Flow
+
+[[`CroCo arXiv`](https://arxiv.org/abs/2210.10716)] [[`CroCo v2 arXiv`](https://arxiv.org/abs/2211.10408)] [[`project page and demo`](https://croco.europe.naverlabs.com/)]
+
+This repository contains the code for our CroCo model presented in our NeurIPS'22 paper [CroCo: Self-Supervised Pre-training for 3D Vision Tasks by Cross-View Completion](https://openreview.net/pdf?id=wZEfHUM5ri) and its follow-up extension published at ICCV'23 [Improved Cross-view Completion Pre-training for Stereo Matching and Optical Flow](https://openaccess.thecvf.com/content/ICCV2023/html/Weinzaepfel_CroCo_v2_Improved_Cross-view_Completion_Pre-training_for_Stereo_Matching_and_ICCV_2023_paper.html), refered to as CroCo v2:
+
+![image](assets/arch.jpg)
+
+```bibtex
+@inproceedings{croco,
+  title={{CroCo: Self-Supervised Pre-training for 3D Vision Tasks by Cross-View Completion}},
+  author={{Weinzaepfel, Philippe and Leroy, Vincent and Lucas, Thomas and Br\'egier, Romain and Cabon, Yohann and Arora, Vaibhav and Antsfeld, Leonid and Chidlovskii, Boris and Csurka, Gabriela and Revaud J\'er\^ome}},
+  booktitle={{NeurIPS}},
+  year={2022}
+}
+
+@inproceedings{croco_v2,
+  title={{CroCo v2: Improved Cross-view Completion Pre-training for Stereo Matching and Optical Flow}},
+  author={Weinzaepfel, Philippe and Lucas, Thomas and Leroy, Vincent and Cabon, Yohann and Arora, Vaibhav and Br{\'e}gier, Romain and Csurka, Gabriela and Antsfeld, Leonid and Chidlovskii, Boris and Revaud, J{\'e}r{\^o}me}, 
+  booktitle={ICCV},
+  year={2023}
+}
+```
+
+## License
+
+The code is distributed under the CC BY-NC-SA 4.0 License. See [LICENSE](LICENSE) for more information.
+Some components are based on code from [MAE](https://github.com/facebookresearch/mae) released under the CC BY-NC-SA 4.0 License and [timm](https://github.com/rwightman/pytorch-image-models) released under the Apache 2.0 License.
+Some components for stereo matching and optical flow are based on code from [unimatch](https://github.com/autonomousvision/unimatch) released under the MIT license.
+
+## Preparation
+
+1. Install dependencies on a machine with a NVidia GPU using e.g. conda. Note that `habitat-sim` is required only for the interactive demo and the synthetic pre-training data generation. If you don't plan to use it, you can ignore the line installing it and use a more recent python version.
+
+```bash
+conda create -n croco python=3.7 cmake=3.14.0
+conda activate croco
+conda install habitat-sim headless -c conda-forge -c aihabitat
+conda install pytorch torchvision -c pytorch
+conda install notebook ipykernel matplotlib
+conda install ipywidgets widgetsnbextension
+conda install scikit-learn tqdm quaternion opencv # only for pretraining / habitat data generation
+
+```
+
+2. Compile cuda kernels for RoPE
+
+CroCo v2 relies on RoPE positional embeddings for which you need to compile some cuda kernels.
+```bash
+cd models/curope/
+python setup.py build_ext --inplace
+cd ../../
+```
+
+This can be a bit long as we compile for all cuda architectures, feel free to update L9 of `models/curope/setup.py` to compile for specific architectures only.
+You might also need to set the environment `CUDA_HOME` in case you use a custom cuda installation.
+
+In case you cannot provide, we also provide a slow pytorch version, which will be automatically loaded.
+
+3. Download pre-trained model
+
+We provide several pre-trained models:
+
+| modelname                                                                                                                          | pre-training data | pos. embed. | Encoder | Decoder |
+|------------------------------------------------------------------------------------------------------------------------------------|-------------------|-------------|---------|---------|
+| [`CroCo.pth`](https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo.pth)                                                 | Habitat           | cosine      | ViT-B   | Small   |
+| [`CroCo_V2_ViTBase_SmallDecoder.pth`](https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo_V2_ViTBase_SmallDecoder.pth) | Habitat + real    | RoPE        | ViT-B   | Small   |
+| [`CroCo_V2_ViTBase_BaseDecoder.pth`](https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo_V2_ViTBase_BaseDecoder.pth)   | Habitat + real    | RoPE        | ViT-B   | Base    |
+| [`CroCo_V2_ViTLarge_BaseDecoder.pth`](https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo_V2_ViTLarge_BaseDecoder.pth) | Habitat + real    | RoPE        | ViT-L   | Base    |
+
+To download a specific model, i.e., the first one (`CroCo.pth`)
+```bash
+mkdir -p pretrained_models/
+wget https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo.pth -P pretrained_models/
+```
+
+## Reconstruction example
+
+Simply run after downloading the `CroCo_V2_ViTLarge_BaseDecoder` pretrained model (or update the corresponding line in `demo.py`)
+```bash
+python demo.py
+```
+
+## Interactive demonstration of cross-view completion reconstruction on the Habitat simulator
+
+First download the test scene from Habitat:
+```bash
+python -m habitat_sim.utils.datasets_download --uids habitat_test_scenes --data-path habitat-sim-data/
+```
+
+Then, run the Notebook demo `interactive_demo.ipynb`.
+
+In this demo, you should be able to sample a random reference viewpoint from an [Habitat](https://github.com/facebookresearch/habitat-sim) test scene. Use the sliders to change viewpoint and select a masked target view to reconstruct using CroCo.
+![croco_interactive_demo](https://user-images.githubusercontent.com/1822210/200516576-7937bc6a-55f8-49ed-8618-3ddf89433ea4.jpg)
+
+## Pre-training 
+
+### CroCo 
+
+To pre-train CroCo, please first generate the pre-training data from the Habitat simulator, following the instructions in [datasets/habitat_sim/README.MD](datasets/habitat_sim/README.MD) and then run the following command:
+```
+torchrun --nproc_per_node=4 pretrain.py --output_dir ./output/pretraining/
+```
+
+Our CroCo pre-training was launched on a single server with 4 GPUs.
+It should take around 10 days with A100 or 15 days with V100 to do the 400 pre-training epochs, but decent performances are obtained earlier in training.
+Note that, while the code contains the same scaling rule of the learning rate as MAE when changing the effective batch size, we did not experimented if it is valid in our case.
+The first run can take a few minutes to start, to parse all available pre-training pairs.
+
+### CroCo v2 
+
+For CroCo v2 pre-training, in addition to the generation of the pre-training data from the Habitat simulator above, please pre-extract the crops from the real datasets following the instructions in [datasets/crops/README.MD](datasets/crops/README.MD).
+Then, run the following command for the largest model (ViT-L encoder, Base decoder):
+```
+torchrun --nproc_per_node=8 pretrain.py --model "CroCoNet(enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_num_heads=12, dec_depth=12, pos_embed='RoPE100')" --dataset "habitat_release+ARKitScenes+MegaDepth+3DStreetView+IndoorVL" --warmup_epochs 12 --max_epoch 125 --epochs 250 --amp 0 --keep_freq 5 --output_dir ./output/pretraining_crocov2/
+```
+
+Our CroCo v2 pre-training was launched on a single server with 8 GPUs for the largest model, and on a single server with 4 GPUs for the smaller ones, keeping a batch size of 64 per gpu in all cases.
+The largest model should take around 12 days on A100.
+Note that, while the code contains the same scaling rule of the learning rate as MAE when changing the effective batch size, we did not experimented if it is valid in our case.
+
+## Stereo matching and Optical flow downstream tasks
+
+For CroCo-Stereo and CroCo-Flow, please refer to [stereoflow/README.MD](stereoflow/README.MD).

third_party/dust3r/croco/datasets/crops/README.MD

@@ -0,0 +1,104 @@
+## Generation of crops from the real datasets
+
+The instructions below allow to generate the crops used for pre-training CroCo v2 from the following real-world datasets: ARKitScenes, MegaDepth, 3DStreetView and IndoorVL.
+
+### Download the metadata of the crops to generate 
+
+First, download the metadata and put them in `./data/`:
+```
+mkdir -p data
+cd data/
+wget https://download.europe.naverlabs.com/ComputerVision/CroCo/data/crop_metadata.zip
+unzip crop_metadata.zip
+rm crop_metadata.zip
+cd ..
+```
+
+### Prepare the original datasets 
+
+Second, download the original datasets in `./data/original_datasets/`.
+```
+mkdir -p data/original_datasets
+```
+
+##### ARKitScenes
+
+Download the `raw` dataset from https://github.com/apple/ARKitScenes/blob/main/DATA.md and put it in `./data/original_datasets/ARKitScenes/`.
+The resulting file structure should be like:
+```
+./data/original_datasets/ARKitScenes/
+└───Training
+    └───40753679
+     │  │   ultrawide
+     │  │   ...
+     └───40753686
+     │   
+      ...
+```
+
+##### MegaDepth
+
+Download `MegaDepth v1 Dataset` from https://www.cs.cornell.edu/projects/megadepth/ and put it in `./data/original_datasets/MegaDepth/`.
+The resulting file structure should be like:
+
+```
+./data/original_datasets/MegaDepth/
+└───0000
+│   └───images
+│    │      │   1000557903_87fa96b8a4_o.jpg
+│    │      └ ...
+│    └─── ...
+└───0001
+│   │   
+│   └ ...
+└─── ...
+```
+
+##### 3DStreetView
+
+Download `3D_Street_View` dataset from https://github.com/amir32002/3D_Street_View and put it in `./data/original_datasets/3DStreetView/`.
+The resulting file structure should be like:
+
+``` 
+./data/original_datasets/3DStreetView/
+└───dataset_aligned
+│   └───0002
+│    │      │   0000002_0000001_0000002_0000001.jpg
+│    │      └ ...
+│    └─── ...
+└───dataset_unaligned
+│   └───0003
+│    │      │   0000003_0000001_0000002_0000001.jpg
+│    │      └ ...
+│    └─── ...
+```
+
+##### IndoorVL
+
+Download the `IndoorVL` datasets using [Kapture](https://github.com/naver/kapture).
+
+```
+pip install kapture
+mkdir -p ./data/original_datasets/IndoorVL
+cd ./data/original_datasets/IndoorVL
+kapture_download_dataset.py update
+kapture_download_dataset.py install  "HyundaiDepartmentStore_*"
+kapture_download_dataset.py install  "GangnamStation_*"
+cd -
+```
+
+### Extract the crops
+
+Now, extract the crops for each of the dataset:
+```
+for dataset in ARKitScenes MegaDepth 3DStreetView IndoorVL; 
+do 
+  python3 datasets/crops/extract_crops_from_images.py --crops ./data/crop_metadata/${dataset}/crops_release.txt --root-dir ./data/original_datasets/${dataset}/ --output-dir ./data/${dataset}_crops/ --imsize 256 --nthread 8 --max-subdir-levels 5 --ideal-number-pairs-in-dir 500;
+done
+```
+
+##### Note for IndoorVL
+
+Due to some legal issues, we can only release 144,228 pairs out of the 1,593,689 pairs used in the paper.
+To account for it in terms of number of pre-training iterations, the pre-training command in this repository uses 125 training epochs including 12 warm-up epochs and learning rate cosine schedule of 250, instead of 100, 10 and 200 respectively.
+The impact on the performance is negligible.

third_party/dust3r/croco/datasets/crops/extract_crops_from_images.py

@@ -0,0 +1,159 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+# 
+# --------------------------------------------------------
+# Extracting crops for pre-training
+# --------------------------------------------------------
+
+import os
+import argparse
+from tqdm import tqdm
+from PIL import Image
+import functools
+from multiprocessing import Pool
+import math
+
+
+def arg_parser():
+    parser = argparse.ArgumentParser('Generate cropped image pairs from image crop list')
+
+    parser.add_argument('--crops', type=str, required=True, help='crop file')
+    parser.add_argument('--root-dir', type=str, required=True, help='root directory')
+    parser.add_argument('--output-dir', type=str, required=True, help='output directory')
+    parser.add_argument('--imsize', type=int, default=256, help='size of the crops')
+    parser.add_argument('--nthread', type=int, required=True, help='number of simultaneous threads')
+    parser.add_argument('--max-subdir-levels', type=int, default=5, help='maximum number of subdirectories')
+    parser.add_argument('--ideal-number-pairs-in-dir', type=int, default=500, help='number of pairs stored in a dir')
+    return parser
+
+
+def main(args):
+    listing_path = os.path.join(args.output_dir, 'listing.txt')
+
+    print(f'Loading list of crops ... ({args.nthread} threads)')
+    crops, num_crops_to_generate = load_crop_file(args.crops)
+
+    print(f'Preparing jobs ({len(crops)} candidate image pairs)...')
+    num_levels = min(math.ceil(math.log(num_crops_to_generate, args.ideal_number_pairs_in_dir)), args.max_subdir_levels)
+    num_pairs_in_dir = math.ceil(num_crops_to_generate ** (1/num_levels))
+
+    jobs = prepare_jobs(crops, num_levels, num_pairs_in_dir)
+    del crops
+
+    os.makedirs(args.output_dir, exist_ok=True)
+    mmap = Pool(args.nthread).imap_unordered if args.nthread > 1 else map
+    call = functools.partial(save_image_crops, args)
+
+    print(f"Generating cropped images to {args.output_dir} ...")
+    with open(listing_path, 'w') as listing:
+        listing.write('# pair_path\n')
+        for results in tqdm(mmap(call, jobs), total=len(jobs)):
+            for path in results:
+                listing.write(f'{path}\n')
+    print('Finished writing listing to', listing_path)
+
+
+def load_crop_file(path):
+    data = open(path).read().splitlines()
+    pairs = []
+    num_crops_to_generate = 0
+    for line in tqdm(data):
+        if line.startswith('#'):
+            continue
+        line = line.split(', ')
+        if len(line) < 8:
+            img1, img2, rotation = line
+            pairs.append((img1, img2, int(rotation), []))
+        else:
+            l1, r1, t1, b1, l2, r2, t2, b2 = map(int, line)
+            rect1, rect2 = (l1, t1, r1, b1), (l2, t2, r2, b2)
+            pairs[-1][-1].append((rect1, rect2))
+            num_crops_to_generate += 1
+    return pairs, num_crops_to_generate
+
+
+def prepare_jobs(pairs, num_levels, num_pairs_in_dir):
+    jobs = []
+    powers = [num_pairs_in_dir**level for level in reversed(range(num_levels))]
+
+    def get_path(idx):
+        idx_array = []
+        d = idx
+        for level in range(num_levels - 1):
+            idx_array.append(idx // powers[level])
+            idx = idx % powers[level]
+        idx_array.append(d)
+        return '/'.join(map(lambda x: hex(x)[2:], idx_array))
+
+    idx = 0
+    for pair_data in tqdm(pairs):
+        img1, img2, rotation, crops = pair_data
+        if -60 <= rotation and rotation <= 60:
+            rotation = 0  # most likely not a true rotation
+        paths = [get_path(idx + k) for k in range(len(crops))]
+        idx += len(crops)
+        jobs.append(((img1, img2), rotation, crops, paths))
+    return jobs
+
+
+def load_image(path):
+    try:
+        return Image.open(path).convert('RGB')
+    except Exception as e:
+        print('skipping', path, e)
+        raise OSError()
+
+
+def save_image_crops(args, data):
+    # load images
+    img_pair, rot, crops, paths = data
+    try:
+        img1, img2 = [load_image(os.path.join(args.root_dir, impath)) for impath in img_pair]
+    except OSError as e:
+        return []
+
+    def area(sz):
+        return sz[0] * sz[1]
+
+    tgt_size = (args.imsize, args.imsize)
+
+    def prepare_crop(img, rect, rot=0):
+        # actual crop
+        img = img.crop(rect)
+
+        # resize to desired size
+        interp = Image.Resampling.LANCZOS if area(img.size) > 4*area(tgt_size) else Image.Resampling.BICUBIC
+        img = img.resize(tgt_size, resample=interp)
+
+        # rotate the image
+        rot90 = (round(rot/90) % 4) * 90
+        if rot90 == 90:
+            img = img.transpose(Image.Transpose.ROTATE_90)
+        elif rot90 == 180:
+            img = img.transpose(Image.Transpose.ROTATE_180)
+        elif rot90 == 270:
+            img = img.transpose(Image.Transpose.ROTATE_270)
+        return img
+
+    results = []
+    for (rect1, rect2), path in zip(crops, paths):
+        crop1 = prepare_crop(img1, rect1)
+        crop2 = prepare_crop(img2, rect2, rot)
+
+        fullpath1 = os.path.join(args.output_dir,  path+'_1.jpg')
+        fullpath2 = os.path.join(args.output_dir,  path+'_2.jpg')
+        os.makedirs(os.path.dirname(fullpath1), exist_ok=True)
+
+        assert not os.path.isfile(fullpath1), fullpath1
+        assert not os.path.isfile(fullpath2), fullpath2
+        crop1.save(fullpath1)
+        crop2.save(fullpath2)
+        results.append(path)
+
+    return results
+
+
+if __name__ == '__main__':
+    args = arg_parser().parse_args()
+    main(args)
+

third_party/dust3r/croco/datasets/habitat_sim/README.MD

@@ -0,0 +1,76 @@
+## Generation of synthetic image pairs using Habitat-Sim
+
+These instructions allow to generate pre-training pairs from the Habitat simulator.
+As we did not save metadata of the pairs used in the original paper, they are not strictly the same, but these data use the same setting and are equivalent.
+
+### Download Habitat-Sim scenes
+Download Habitat-Sim scenes:
+- Download links can be found here: https://github.com/facebookresearch/habitat-sim/blob/main/DATASETS.md
+- We used scenes from the HM3D, habitat-test-scenes, Replica, ReplicaCad and ScanNet datasets.
+- Please put the scenes under `./data/habitat-sim-data/scene_datasets/` following the structure below, or update manually paths in `paths.py`.
+```
+./data/
+└──habitat-sim-data/
+   └──scene_datasets/
+      ├──hm3d/
+      ├──gibson/
+      ├──habitat-test-scenes/
+      ├──replica_cad_baked_lighting/
+      ├──replica_cad/
+      ├──ReplicaDataset/
+      └──scannet/
+```
+
+### Image pairs generation
+We provide metadata to generate reproducible images pairs for pretraining and validation.
+Experiments described in the paper used similar data, but whose generation was not reproducible at the time.
+
+Specifications:
+- 256x256 resolution images, with 60 degrees field of view .
+- Up to 1000 image pairs per scene.
+- Number of scenes considered/number of images pairs per dataset:
+  - Scannet: 1097 scenes / 985 209 pairs
+  - HM3D:
+    - hm3d/train: 800 / 800k pairs
+    - hm3d/val: 100 scenes / 100k pairs
+    - hm3d/minival: 10 scenes / 10k pairs
+  - habitat-test-scenes: 3 scenes / 3k pairs
+  - replica_cad_baked_lighting: 13 scenes / 13k pairs
+
+- Scenes from hm3d/val and hm3d/minival pairs were not used for the pre-training but kept for validation purposes.
+
+Download metadata and extract it:
+```bash
+mkdir -p data/habitat_release_metadata/
+cd data/habitat_release_metadata/
+wget https://download.europe.naverlabs.com/ComputerVision/CroCo/data/habitat_release_metadata/multiview_habitat_metadata.tar.gz
+tar -xvf multiview_habitat_metadata.tar.gz
+cd ../..
+# Location of the metadata
+METADATA_DIR="./data/habitat_release_metadata/multiview_habitat_metadata"
+```
+
+Generate image pairs from metadata:
+- The following command will print a list of commandlines to generate image pairs for each scene:
+```bash
+# Target output directory
+PAIRS_DATASET_DIR="./data/habitat_release/"
+python datasets/habitat_sim/generate_from_metadata_files.py --input_dir=$METADATA_DIR --output_dir=$PAIRS_DATASET_DIR
+```
+- One can launch multiple of such commands in parallel e.g. using GNU Parallel:
+```bash
+python datasets/habitat_sim/generate_from_metadata_files.py --input_dir=$METADATA_DIR --output_dir=$PAIRS_DATASET_DIR | parallel -j 16
+```
+
+## Metadata generation
+
+Image pairs were randomly sampled using the following commands, whose outputs contain randomness and are thus not exactly reproducible:
+```bash
+# Print commandlines to generate image pairs from the different scenes available.
+PAIRS_DATASET_DIR=MY_CUSTOM_PATH
+python datasets/habitat_sim/generate_multiview_images.py --list_commands --output_dir=$PAIRS_DATASET_DIR
+
+# Once a dataset is generated, pack metadata files for reproducibility.
+METADATA_DIR=MY_CUSTON_PATH
+python datasets/habitat_sim/pack_metadata_files.py $PAIRS_DATASET_DIR  $METADATA_DIR
+```

third_party/dust3r/croco/datasets/habitat_sim/generate_from_metadata.py

@@ -0,0 +1,92 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+"""
+Script to generate image pairs for a given scene reproducing poses provided in a metadata file.
+"""
+import os
+from datasets.habitat_sim.multiview_habitat_sim_generator import MultiviewHabitatSimGenerator
+from datasets.habitat_sim.paths import SCENES_DATASET
+import argparse
+import quaternion
+import PIL.Image
+import cv2
+import json
+from tqdm import tqdm
+
+def generate_multiview_images_from_metadata(metadata_filename,
+                                            output_dir,
+                                            overload_params = dict(),
+                                            scene_datasets_paths=None,
+                                            exist_ok=False):   
+    """
+    Generate images from a metadata file for reproducibility purposes.
+    """
+    # Reorder paths by decreasing label length, to avoid collisions when testing if a string by such label
+    if scene_datasets_paths is not None:
+        scene_datasets_paths = dict(sorted(scene_datasets_paths.items(), key= lambda x: len(x[0]), reverse=True))
+
+    with open(metadata_filename, 'r') as f:
+        input_metadata = json.load(f)
+    metadata = dict()
+    for key, value in input_metadata.items():
+        # Optionally replace some paths
+        if key in ("scene_dataset_config_file", "scene", "navmesh") and value != "":
+            if scene_datasets_paths is not None:
+                for dataset_label, dataset_path in scene_datasets_paths.items():
+                    if value.startswith(dataset_label):
+                        value = os.path.normpath(os.path.join(dataset_path, os.path.relpath(value, dataset_label)))
+                        break
+        metadata[key] = value
+
+    # Overload some parameters
+    for key, value in overload_params.items():
+        metadata[key] = value
+
+    generation_entries = dict([(key, value) for key, value in metadata.items() if not (key in ('multiviews', 'output_dir', 'generate_depth'))])
+    generate_depth = metadata["generate_depth"]
+
+    os.makedirs(output_dir, exist_ok=exist_ok)
+ 
+    generator = MultiviewHabitatSimGenerator(**generation_entries)
+
+    # Generate views
+    for idx_label, data in tqdm(metadata['multiviews'].items()):
+        positions = data["positions"]
+        orientations = data["orientations"]
+        n = len(positions)
+        for oidx in range(n):
+            observation = generator.render_viewpoint(positions[oidx], quaternion.from_float_array(orientations[oidx]))
+            observation_label = f"{oidx + 1}" # Leonid is indexing starting from 1
+            # Color image saved using PIL
+            img = PIL.Image.fromarray(observation['color'][:,:,:3])
+            filename = os.path.join(output_dir, f"{idx_label}_{observation_label}.jpeg")
+            img.save(filename)
+            if generate_depth:
+                # Depth image as EXR file
+                filename = os.path.join(output_dir, f"{idx_label}_{observation_label}_depth.exr")
+                cv2.imwrite(filename, observation['depth'], [cv2.IMWRITE_EXR_TYPE, cv2.IMWRITE_EXR_TYPE_HALF])
+                # Camera parameters
+                camera_params = dict([(key, observation[key].tolist()) for key in ("camera_intrinsics", "R_cam2world", "t_cam2world")])
+                filename = os.path.join(output_dir, f"{idx_label}_{observation_label}_camera_params.json")
+                with open(filename, "w") as f:
+                    json.dump(camera_params, f)
+                # Save metadata
+    with open(os.path.join(output_dir, "metadata.json"), "w") as f:
+        json.dump(metadata, f)
+
+    generator.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--metadata_filename", required=True)
+    parser.add_argument("--output_dir", required=True)
+    args = parser.parse_args()
+
+    generate_multiview_images_from_metadata(metadata_filename=args.metadata_filename,
+                             output_dir=args.output_dir,
+                             scene_datasets_paths=SCENES_DATASET,
+                             overload_params=dict(),
+                             exist_ok=True)
+
+ 

third_party/dust3r/croco/datasets/habitat_sim/generate_from_metadata_files.py

@@ -0,0 +1,27 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+"""
+Script generating commandlines to generate image pairs from metadata files.
+"""
+import os
+import glob
+from tqdm import tqdm
+import argparse
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--input_dir", required=True)
+    parser.add_argument("--output_dir", required=True)
+    parser.add_argument("--prefix", default="", help="Commanline prefix, useful e.g. to setup environment.")
+    args = parser.parse_args()
+
+    input_metadata_filenames = glob.iglob(f"{args.input_dir}/**/metadata.json", recursive=True)
+
+    for metadata_filename in tqdm(input_metadata_filenames):
+        output_dir = os.path.join(args.output_dir, os.path.relpath(os.path.dirname(metadata_filename), args.input_dir))
+        # Do not process the scene if the metadata file already exists
+        if os.path.exists(os.path.join(output_dir, "metadata.json")):
+            continue
+        commandline = f"{args.prefix}python datasets/habitat_sim/generate_from_metadata.py --metadata_filename={metadata_filename} --output_dir={output_dir}"
+        print(commandline)

third_party/dust3r/croco/datasets/habitat_sim/generate_multiview_images.py

@@ -0,0 +1,177 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+import os
+from tqdm import tqdm
+import argparse
+import PIL.Image
+import numpy as np
+import json
+from datasets.habitat_sim.multiview_habitat_sim_generator import MultiviewHabitatSimGenerator, NoNaviguableSpaceError
+from datasets.habitat_sim.paths import list_scenes_available
+import cv2
+import quaternion
+import shutil
+
+def generate_multiview_images_for_scene(scene_dataset_config_file,
+                                        scene,
+                                        navmesh,
+                                        output_dir, 
+                                        views_count,
+                                        size, 
+                                        exist_ok=False, 
+                                        generate_depth=False,
+                                        **kwargs):
+    """
+    Generate tuples of overlapping views for a given scene.
+    generate_depth: generate depth images and camera parameters.
+    """
+    if os.path.exists(output_dir) and not exist_ok:
+        print(f"Scene {scene}: data already generated. Ignoring generation.")
+        return
+    try:
+        print(f"Scene {scene}: {size} multiview acquisitions to generate...")
+        os.makedirs(output_dir, exist_ok=exist_ok)
+
+        metadata_filename = os.path.join(output_dir, "metadata.json")
+
+        metadata_template = dict(scene_dataset_config_file=scene_dataset_config_file,
+            scene=scene, 
+            navmesh=navmesh,
+            views_count=views_count,
+            size=size,
+            generate_depth=generate_depth,
+            **kwargs)
+        metadata_template["multiviews"] = dict()
+
+        if os.path.exists(metadata_filename):
+            print("Metadata file already exists:", metadata_filename)
+            print("Loading already generated metadata file...")
+            with open(metadata_filename, "r") as f:
+                metadata = json.load(f)
+
+            for key in metadata_template.keys():
+                if key != "multiviews":
+                    assert metadata_template[key] == metadata[key], f"existing file is inconsistent with the input parameters:\nKey: {key}\nmetadata: {metadata[key]}\ntemplate: {metadata_template[key]}."
+        else:
+            print("No temporary file found. Starting generation from scratch...")
+            metadata = metadata_template
+
+        starting_id = len(metadata["multiviews"])
+        print(f"Starting generation from index {starting_id}/{size}...")
+        if starting_id >= size:
+            print("Generation already done.")
+            return
+
+        generator = MultiviewHabitatSimGenerator(scene_dataset_config_file=scene_dataset_config_file,
+                                                scene=scene,
+                                                navmesh=navmesh,
+                                                views_count = views_count,
+                                                size = size,
+                                                **kwargs)
+
+        for idx in tqdm(range(starting_id, size)):
+            # Generate / re-generate the observations
+            try:
+                data = generator[idx]
+                observations = data["observations"]
+                positions = data["positions"]
+                orientations = data["orientations"]
+
+                idx_label = f"{idx:08}"
+                for oidx, observation in enumerate(observations):
+                    observation_label = f"{oidx + 1}" # Leonid is indexing starting from 1
+                    # Color image saved using PIL
+                    img = PIL.Image.fromarray(observation['color'][:,:,:3])
+                    filename = os.path.join(output_dir, f"{idx_label}_{observation_label}.jpeg")
+                    img.save(filename)
+                    if generate_depth:
+                        # Depth image as EXR file
+                        filename = os.path.join(output_dir, f"{idx_label}_{observation_label}_depth.exr")
+                        cv2.imwrite(filename, observation['depth'], [cv2.IMWRITE_EXR_TYPE, cv2.IMWRITE_EXR_TYPE_HALF])
+                        # Camera parameters
+                        camera_params = dict([(key, observation[key].tolist()) for key in ("camera_intrinsics", "R_cam2world", "t_cam2world")])
+                        filename = os.path.join(output_dir, f"{idx_label}_{observation_label}_camera_params.json")
+                        with open(filename, "w") as f:
+                            json.dump(camera_params, f)
+                metadata["multiviews"][idx_label] = {"positions": positions.tolist(),
+                                                    "orientations": orientations.tolist(),
+                                                    "covisibility_ratios": data["covisibility_ratios"].tolist(),
+                                                    "valid_fractions": data["valid_fractions"].tolist(),
+                                                    "pairwise_visibility_ratios": data["pairwise_visibility_ratios"].tolist()}
+            except RecursionError:
+                print("Recursion error: unable to sample observations for this scene. We will stop there.")
+                break
+
+            # Regularly save a temporary metadata file, in case we need to restart the generation
+            if idx % 10 == 0:
+                with open(metadata_filename, "w") as f:
+                    json.dump(metadata, f)
+
+        # Save metadata
+        with open(metadata_filename, "w") as f:
+            json.dump(metadata, f)
+
+        generator.close()
+    except NoNaviguableSpaceError:
+        pass
+
+def create_commandline(scene_data, generate_depth, exist_ok=False):
+    """
+    Create a commandline string to generate a scene.
+    """
+    def my_formatting(val):
+        if val is None or val == "":
+            return '""'
+        else:
+            return val
+    commandline = f"""python {__file__} --scene {my_formatting(scene_data.scene)} 
+    --scene_dataset_config_file {my_formatting(scene_data.scene_dataset_config_file)} 
+    --navmesh {my_formatting(scene_data.navmesh)} 
+    --output_dir {my_formatting(scene_data.output_dir)} 
+    --generate_depth {int(generate_depth)} 
+    --exist_ok {int(exist_ok)}
+    """
+    commandline = " ".join(commandline.split())
+    return commandline
+
+if __name__ == "__main__":
+    os.umask(2)
+
+    parser = argparse.ArgumentParser(description="""Example of use -- listing commands to generate data for scenes available:
+    > python datasets/habitat_sim/generate_multiview_habitat_images.py --list_commands
+    """)
+    
+    parser.add_argument("--output_dir", type=str, required=True)
+    parser.add_argument("--list_commands", action='store_true', help="list commandlines to run if true")
+    parser.add_argument("--scene", type=str, default="")
+    parser.add_argument("--scene_dataset_config_file", type=str, default="")
+    parser.add_argument("--navmesh", type=str, default="")
+    
+    parser.add_argument("--generate_depth", type=int, default=1)
+    parser.add_argument("--exist_ok", type=int, default=0)
+
+    kwargs = dict(resolution=(256,256), hfov=60, views_count = 2, size=1000)
+
+    args = parser.parse_args()
+    generate_depth=bool(args.generate_depth)
+    exist_ok = bool(args.exist_ok)
+
+    if args.list_commands:
+        # Listing scenes available...
+        scenes_data = list_scenes_available(base_output_dir=args.output_dir)
+        
+        for scene_data in scenes_data:
+            print(create_commandline(scene_data, generate_depth=generate_depth, exist_ok=exist_ok))
+    else:
+        if args.scene == "" or args.output_dir == "":
+            print("Missing scene or output dir argument!")
+            print(parser.format_help())
+        else:
+            generate_multiview_images_for_scene(scene=args.scene,
+                                                scene_dataset_config_file = args.scene_dataset_config_file,
+                                                navmesh = args.navmesh,
+                                                output_dir = args.output_dir,
+                                                exist_ok=exist_ok,
+                                                generate_depth=generate_depth,
+                                                **kwargs)

third_party/dust3r/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py

@@ -0,0 +1,390 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+import os
+import numpy as np
+import quaternion
+import habitat_sim
+import json
+from sklearn.neighbors import NearestNeighbors
+import cv2
+
+# OpenCV to habitat camera convention transformation
+R_OPENCV2HABITAT = np.stack((habitat_sim.geo.RIGHT, -habitat_sim.geo.UP, habitat_sim.geo.FRONT), axis=0)
+R_HABITAT2OPENCV = R_OPENCV2HABITAT.T
+DEG2RAD = np.pi / 180
+
+def compute_camera_intrinsics(height, width, hfov):
+    f = width/2 / np.tan(hfov/2 * np.pi/180)
+    cu, cv = width/2, height/2
+    return f, cu, cv
+
+def compute_camera_pose_opencv_convention(camera_position, camera_orientation):
+    R_cam2world = quaternion.as_rotation_matrix(camera_orientation) @ R_OPENCV2HABITAT
+    t_cam2world = np.asarray(camera_position)
+    return R_cam2world, t_cam2world
+
+def compute_pointmap(depthmap, hfov):
+    """ Compute a HxWx3 pointmap in camera frame from a HxW depth map."""
+    height, width = depthmap.shape
+    f, cu, cv = compute_camera_intrinsics(height, width, hfov)
+    # Cast depth map to point
+    z_cam = depthmap
+    u, v = np.meshgrid(range(width), range(height))
+    x_cam = (u - cu) / f * z_cam
+    y_cam = (v - cv) / f * z_cam
+    X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1)
+    return X_cam
+
+def compute_pointcloud(depthmap, hfov, camera_position, camera_rotation):
+    """Return a 3D point cloud corresponding to valid pixels of the depth map"""
+    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(camera_position, camera_rotation)
+
+    X_cam = compute_pointmap(depthmap=depthmap, hfov=hfov)
+    valid_mask = (X_cam[:,:,2] != 0.0)
+
+    X_cam = X_cam.reshape(-1, 3)[valid_mask.flatten()]
+    X_world = X_cam @ R_cam2world.T + t_cam2world.reshape(1, 3)
+    return X_world
+
+def compute_pointcloud_overlaps_scikit(pointcloud1, pointcloud2, distance_threshold, compute_symmetric=False):
+    """
+    Compute 'overlapping' metrics based on a distance threshold between two point clouds.
+    """
+    nbrs = NearestNeighbors(n_neighbors=1, algorithm = 'kd_tree').fit(pointcloud2)
+    distances, indices = nbrs.kneighbors(pointcloud1)
+    intersection1 = np.count_nonzero(distances.flatten() < distance_threshold)
+
+    data = {"intersection1": intersection1,
+            "size1": len(pointcloud1)}
+    if compute_symmetric:
+        nbrs = NearestNeighbors(n_neighbors=1, algorithm = 'kd_tree').fit(pointcloud1)
+        distances, indices = nbrs.kneighbors(pointcloud2)
+        intersection2 = np.count_nonzero(distances.flatten() < distance_threshold)
+        data["intersection2"] = intersection2
+        data["size2"] = len(pointcloud2)
+
+    return data
+
+def _append_camera_parameters(observation, hfov, camera_location, camera_rotation):
+    """
+    Add camera parameters to the observation dictionnary produced by Habitat-Sim
+    In-place modifications.
+    """
+    R_cam2world, t_cam2world = compute_camera_pose_opencv_convention(camera_location, camera_rotation)
+    height, width = observation['depth'].shape
+    f, cu, cv = compute_camera_intrinsics(height, width, hfov)
+    K = np.asarray([[f, 0, cu],
+                    [0, f, cv],
+                    [0, 0, 1.0]])
+    observation["camera_intrinsics"] = K
+    observation["t_cam2world"] = t_cam2world
+    observation["R_cam2world"] = R_cam2world
+
+def look_at(eye, center, up, return_cam2world=True):
+    """
+    Return camera pose looking at a given center point.
+    Analogous of gluLookAt function, using OpenCV camera convention.
+    """
+    z = center - eye
+    z /= np.linalg.norm(z, axis=-1, keepdims=True)
+    y = -up
+    y = y - np.sum(y * z, axis=-1, keepdims=True) * z
+    y /= np.linalg.norm(y, axis=-1, keepdims=True)
+    x = np.cross(y, z, axis=-1)
+
+    if return_cam2world:
+        R = np.stack((x, y, z), axis=-1)
+        t = eye
+    else:
+        # World to camera transformation
+        # Transposed matrix
+        R = np.stack((x, y, z), axis=-2)
+        t = - np.einsum('...ij, ...j', R, eye)
+    return R, t
+
+def look_at_for_habitat(eye, center, up, return_cam2world=True):
+    R, t = look_at(eye, center, up)
+    orientation = quaternion.from_rotation_matrix(R @ R_OPENCV2HABITAT.T)
+    return orientation, t
+
+def generate_orientation_noise(pan_range, tilt_range, roll_range):
+    return (quaternion.from_rotation_vector(np.random.uniform(*pan_range) * DEG2RAD * habitat_sim.geo.UP)
+            * quaternion.from_rotation_vector(np.random.uniform(*tilt_range) * DEG2RAD * habitat_sim.geo.RIGHT)
+            * quaternion.from_rotation_vector(np.random.uniform(*roll_range) * DEG2RAD * habitat_sim.geo.FRONT))
+
+
+class NoNaviguableSpaceError(RuntimeError):
+    def __init__(self, *args):
+            super().__init__(*args)
+
+class MultiviewHabitatSimGenerator:
+    def __init__(self,
+                scene,
+                navmesh,
+                scene_dataset_config_file,
+                resolution = (240, 320),
+                views_count=2,
+                hfov = 60,
+                gpu_id = 0,
+                size = 10000,
+                minimum_covisibility = 0.5,
+                transform = None):
+        self.scene = scene
+        self.navmesh = navmesh
+        self.scene_dataset_config_file = scene_dataset_config_file
+        self.resolution = resolution
+        self.views_count = views_count
+        assert(self.views_count >= 1)
+        self.hfov = hfov
+        self.gpu_id = gpu_id
+        self.size = size
+        self.transform = transform
+
+        # Noise added to camera orientation
+        self.pan_range = (-3, 3)
+        self.tilt_range = (-10, 10)
+        self.roll_range = (-5, 5)
+
+        # Height range to sample cameras
+        self.height_range = (1.2, 1.8)
+
+        # Random steps between the camera views
+        self.random_steps_count = 5
+        self.random_step_variance = 2.0
+
+        # Minimum fraction of the scene which should be valid (well defined depth)
+        self.minimum_valid_fraction = 0.7
+
+        # Distance threshold to see  to select pairs
+        self.distance_threshold = 0.05
+        # Minimum IoU of a view point cloud with respect to the reference view to be kept.
+        self.minimum_covisibility = minimum_covisibility
+
+        # Maximum number of retries.
+        self.max_attempts_count = 100
+
+        self.seed = None
+        self._lazy_initialization()
+
+    def _lazy_initialization(self):
+        # Lazy random seeding and instantiation of the simulator to deal with multiprocessing properly
+        if self.seed == None:
+            # Re-seed numpy generator
+            np.random.seed()
+            self.seed = np.random.randint(2**32-1)
+            sim_cfg = habitat_sim.SimulatorConfiguration()
+            sim_cfg.scene_id = self.scene
+            if self.scene_dataset_config_file is not None and self.scene_dataset_config_file != "":
+                    sim_cfg.scene_dataset_config_file = self.scene_dataset_config_file
+            sim_cfg.random_seed = self.seed
+            sim_cfg.load_semantic_mesh = False
+            sim_cfg.gpu_device_id = self.gpu_id
+
+            depth_sensor_spec = habitat_sim.CameraSensorSpec()
+            depth_sensor_spec.uuid = "depth"
+            depth_sensor_spec.sensor_type = habitat_sim.SensorType.DEPTH
+            depth_sensor_spec.resolution = self.resolution
+            depth_sensor_spec.hfov = self.hfov
+            depth_sensor_spec.position = [0.0, 0.0, 0]
+            depth_sensor_spec.orientation
+
+            rgb_sensor_spec = habitat_sim.CameraSensorSpec()
+            rgb_sensor_spec.uuid = "color"
+            rgb_sensor_spec.sensor_type = habitat_sim.SensorType.COLOR
+            rgb_sensor_spec.resolution = self.resolution
+            rgb_sensor_spec.hfov = self.hfov
+            rgb_sensor_spec.position = [0.0, 0.0, 0]
+            agent_cfg = habitat_sim.agent.AgentConfiguration(sensor_specifications=[rgb_sensor_spec, depth_sensor_spec])
+
+            cfg = habitat_sim.Configuration(sim_cfg, [agent_cfg])
+            self.sim = habitat_sim.Simulator(cfg)
+            if self.navmesh is not None and self.navmesh != "":
+                # Use pre-computed navmesh when available (usually better than those generated automatically)
+                self.sim.pathfinder.load_nav_mesh(self.navmesh)
+
+            if not self.sim.pathfinder.is_loaded:
+                # Try to compute a navmesh
+                navmesh_settings = habitat_sim.NavMeshSettings()
+                navmesh_settings.set_defaults()
+                self.sim.recompute_navmesh(self.sim.pathfinder, navmesh_settings, True)
+
+            # Ensure that the navmesh is not empty
+            if not self.sim.pathfinder.is_loaded:
+                raise NoNaviguableSpaceError(f"No naviguable location (scene: {self.scene} -- navmesh: {self.navmesh})")
+
+            self.agent = self.sim.initialize_agent(agent_id=0)
+
+    def close(self):
+        self.sim.close()
+
+    def __del__(self):
+        self.sim.close()
+
+    def __len__(self):
+        return self.size
+
+    def sample_random_viewpoint(self):
+        """ Sample a random viewpoint using the navmesh """
+        nav_point = self.sim.pathfinder.get_random_navigable_point()
+
+        # Sample a random viewpoint height
+        viewpoint_height = np.random.uniform(*self.height_range)
+        viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP
+        viewpoint_orientation = quaternion.from_rotation_vector(np.random.uniform(0, 2 * np.pi) * habitat_sim.geo.UP) * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range)
+        return viewpoint_position, viewpoint_orientation, nav_point
+
+    def sample_other_random_viewpoint(self, observed_point, nav_point):
+        """ Sample a random viewpoint close to an existing one, using the navmesh and a reference observed point."""
+        other_nav_point = nav_point
+
+        walk_directions = self.random_step_variance * np.asarray([1,0,1])
+        for i in range(self.random_steps_count):
+            temp = self.sim.pathfinder.snap_point(other_nav_point + walk_directions * np.random.normal(size=3))
+            # Snapping may return nan when it fails
+            if not np.isnan(temp[0]):
+                    other_nav_point = temp
+
+        other_viewpoint_height = np.random.uniform(*self.height_range)
+        other_viewpoint_position = other_nav_point + other_viewpoint_height * habitat_sim.geo.UP
+
+        # Set viewing direction towards the central point
+        rotation, position = look_at_for_habitat(eye=other_viewpoint_position, center=observed_point, up=habitat_sim.geo.UP, return_cam2world=True)
+        rotation = rotation * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range)
+        return position, rotation, other_nav_point
+
+    def is_other_pointcloud_overlapping(self, ref_pointcloud, other_pointcloud):
+        """ Check if a viewpoint is valid and overlaps significantly with a reference one. """
+        # Observation
+        pixels_count = self.resolution[0] * self.resolution[1]
+        valid_fraction = len(other_pointcloud) / pixels_count
+        assert valid_fraction <= 1.0 and valid_fraction >= 0.0
+        overlap = compute_pointcloud_overlaps_scikit(ref_pointcloud, other_pointcloud, self.distance_threshold, compute_symmetric=True)
+        covisibility = min(overlap["intersection1"] / pixels_count, overlap["intersection2"] / pixels_count)
+        is_valid = (valid_fraction >= self.minimum_valid_fraction) and (covisibility >= self.minimum_covisibility)
+        return is_valid, valid_fraction, covisibility
+
+    def is_other_viewpoint_overlapping(self, ref_pointcloud, observation, position, rotation):
+        """ Check if a viewpoint is valid and overlaps significantly with a reference one. """
+        # Observation
+        other_pointcloud = compute_pointcloud(observation['depth'], self.hfov, position, rotation)
+        return self.is_other_pointcloud_overlapping(ref_pointcloud, other_pointcloud)
+
+    def render_viewpoint(self, viewpoint_position, viewpoint_orientation):
+        agent_state = habitat_sim.AgentState()
+        agent_state.position = viewpoint_position
+        agent_state.rotation = viewpoint_orientation
+        self.agent.set_state(agent_state)
+        viewpoint_observations = self.sim.get_sensor_observations(agent_ids=0)
+        _append_camera_parameters(viewpoint_observations, self.hfov, viewpoint_position, viewpoint_orientation)
+        return viewpoint_observations
+
+    def __getitem__(self, useless_idx):
+        ref_position, ref_orientation, nav_point = self.sample_random_viewpoint()
+        ref_observations = self.render_viewpoint(ref_position, ref_orientation)
+        # Extract point cloud
+        ref_pointcloud = compute_pointcloud(depthmap=ref_observations['depth'], hfov=self.hfov,
+                                        camera_position=ref_position, camera_rotation=ref_orientation)
+
+        pixels_count = self.resolution[0] * self.resolution[1]
+        ref_valid_fraction = len(ref_pointcloud) / pixels_count
+        assert ref_valid_fraction <= 1.0 and ref_valid_fraction >= 0.0
+        if ref_valid_fraction < self.minimum_valid_fraction:
+                # This should produce a recursion error at some point when something is very wrong.
+                return self[0]
+        # Pick an reference observed point in the point cloud
+        observed_point = np.mean(ref_pointcloud, axis=0)
+
+        # Add the first image as reference
+        viewpoints_observations = [ref_observations]
+        viewpoints_covisibility = [ref_valid_fraction]
+        viewpoints_positions = [ref_position]
+        viewpoints_orientations = [quaternion.as_float_array(ref_orientation)]
+        viewpoints_clouds = [ref_pointcloud]
+        viewpoints_valid_fractions = [ref_valid_fraction]
+
+        for _ in range(self.views_count - 1):
+            # Generate an other viewpoint using some dummy random walk
+            successful_sampling = False
+            for sampling_attempt in range(self.max_attempts_count):
+                position, rotation, _ = self.sample_other_random_viewpoint(observed_point, nav_point)
+                # Observation
+                other_viewpoint_observations = self.render_viewpoint(position, rotation)
+                other_pointcloud = compute_pointcloud(other_viewpoint_observations['depth'], self.hfov, position, rotation)
+
+                is_valid, valid_fraction, covisibility = self.is_other_pointcloud_overlapping(ref_pointcloud, other_pointcloud)
+                if is_valid:
+                        successful_sampling = True
+                        break
+            if not successful_sampling:
+                print("WARNING: Maximum number of attempts reached.")
+                # Dirty hack, try using a novel original viewpoint
+                return self[0]
+            viewpoints_observations.append(other_viewpoint_observations)
+            viewpoints_covisibility.append(covisibility)
+            viewpoints_positions.append(position)
+            viewpoints_orientations.append(quaternion.as_float_array(rotation)) # WXYZ convention for the quaternion encoding.
+            viewpoints_clouds.append(other_pointcloud)
+            viewpoints_valid_fractions.append(valid_fraction)
+
+        # Estimate relations between all pairs of images
+        pairwise_visibility_ratios = np.ones((len(viewpoints_observations), len(viewpoints_observations)))
+        for i in range(len(viewpoints_observations)):
+            pairwise_visibility_ratios[i,i] = viewpoints_valid_fractions[i]
+            for j in range(i+1, len(viewpoints_observations)):
+                overlap = compute_pointcloud_overlaps_scikit(viewpoints_clouds[i], viewpoints_clouds[j], self.distance_threshold, compute_symmetric=True)
+                pairwise_visibility_ratios[i,j] = overlap['intersection1'] / pixels_count
+                pairwise_visibility_ratios[j,i] = overlap['intersection2'] / pixels_count
+
+        # IoU is relative to the image 0
+        data = {"observations": viewpoints_observations,
+                "positions": np.asarray(viewpoints_positions),
+                "orientations": np.asarray(viewpoints_orientations),
+                "covisibility_ratios": np.asarray(viewpoints_covisibility),
+                "valid_fractions": np.asarray(viewpoints_valid_fractions, dtype=float),
+                "pairwise_visibility_ratios": np.asarray(pairwise_visibility_ratios, dtype=float),
+                }
+
+        if self.transform is not None:
+            data = self.transform(data)
+        return  data
+
+    def generate_random_spiral_trajectory(self, images_count = 100, max_radius=0.5, half_turns=5, use_constant_orientation=False):
+        """
+        Return a list of images corresponding to a spiral trajectory from a random starting point.
+        Useful to generate nice visualisations.
+        Use an even number of half turns to get a nice "C1-continuous" loop effect 
+        """
+        ref_position, ref_orientation, navpoint = self.sample_random_viewpoint()
+        ref_observations = self.render_viewpoint(ref_position, ref_orientation)
+        ref_pointcloud = compute_pointcloud(depthmap=ref_observations['depth'], hfov=self.hfov,
+                                                        camera_position=ref_position, camera_rotation=ref_orientation)
+        pixels_count = self.resolution[0] * self.resolution[1]
+        if len(ref_pointcloud) / pixels_count < self.minimum_valid_fraction:
+            # Dirty hack: ensure that the valid part of the image is significant
+            return self.generate_random_spiral_trajectory(images_count, max_radius, half_turns, use_constant_orientation)
+
+        # Pick an observed point in the point cloud
+        observed_point = np.mean(ref_pointcloud, axis=0)
+        ref_R, ref_t = compute_camera_pose_opencv_convention(ref_position, ref_orientation)
+
+        images = []
+        is_valid = []
+        # Spiral trajectory, use_constant orientation
+        for i, alpha in enumerate(np.linspace(0, 1, images_count)):
+            r = max_radius * np.abs(np.sin(alpha * np.pi)) # Increase then decrease the radius
+            theta = alpha * half_turns * np.pi 
+            x = r * np.cos(theta)
+            y = r * np.sin(theta)
+            z = 0.0
+            position = ref_position + (ref_R @ np.asarray([x, y, z]).reshape(3,1)).flatten()
+            if use_constant_orientation:
+                orientation = ref_orientation
+            else:
+                # trajectory looking at a mean point in front of the ref observation
+                orientation, position = look_at_for_habitat(eye=position, center=observed_point, up=habitat_sim.geo.UP)
+            observations = self.render_viewpoint(position, orientation)
+            images.append(observations['color'][...,:3])
+            _is_valid, valid_fraction, iou = self.is_other_viewpoint_overlapping(ref_pointcloud, observations, position, orientation)
+            is_valid.append(_is_valid)
+        return images, np.all(is_valid)

third_party/dust3r/croco/datasets/habitat_sim/pack_metadata_files.py

@@ -0,0 +1,69 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+"""
+Utility script to pack metadata files of the dataset in order to be able to re-generate it elsewhere.
+"""
+import os
+import glob
+from tqdm import tqdm
+import shutil
+import json
+from datasets.habitat_sim.paths import *
+import argparse
+import collections
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("input_dir")
+    parser.add_argument("output_dir")
+    args = parser.parse_args()
+
+    input_dirname = args.input_dir
+    output_dirname = args.output_dir
+
+    input_metadata_filenames = glob.iglob(f"{input_dirname}/**/metadata.json", recursive=True)
+
+    images_count = collections.defaultdict(lambda : 0)
+    
+    os.makedirs(output_dirname)
+    for input_filename in tqdm(input_metadata_filenames):
+        # Ignore empty files
+        with open(input_filename, "r") as f:
+            original_metadata = json.load(f)
+            if "multiviews" not in original_metadata or len(original_metadata["multiviews"]) == 0:
+                print("No views in", input_filename)
+                continue
+
+        relpath = os.path.relpath(input_filename, input_dirname)
+        print(relpath)
+
+        # Copy metadata, while replacing scene paths by generic keys depending on the dataset, for portability.
+        # Data paths are sorted by decreasing length to avoid potential bugs due to paths starting by the same string pattern.
+        scenes_dataset_paths = dict(sorted(SCENES_DATASET.items(), key=lambda x: len(x[1]), reverse=True))
+        metadata = dict()
+        for key, value in original_metadata.items():
+            if key in ("scene_dataset_config_file", "scene", "navmesh") and value != "":
+                known_path = False
+                for dataset, dataset_path in scenes_dataset_paths.items():
+                    if value.startswith(dataset_path):
+                        value = os.path.join(dataset, os.path.relpath(value, dataset_path))
+                        known_path = True
+                        break
+                if not known_path:
+                    raise KeyError("Unknown path:" + value)
+            metadata[key] = value
+
+        # Compile some general statistics while packing data
+        scene_split = metadata["scene"].split("/")
+        upper_level = "/".join(scene_split[:2]) if scene_split[0] == "hm3d" else scene_split[0]
+        images_count[upper_level] += len(metadata["multiviews"])
+        
+        output_filename = os.path.join(output_dirname, relpath)
+        os.makedirs(os.path.dirname(output_filename), exist_ok=True)
+        with open(output_filename, "w") as f:
+            json.dump(metadata, f)
+
+    # Print statistics
+    print("Images count:")
+    for upper_level, count in images_count.items():
+        print(f"- {upper_level}: {count}")

third_party/dust3r/croco/datasets/habitat_sim/paths.py

@@ -0,0 +1,129 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+"""
+Paths to Habitat-Sim scenes
+"""
+
+import os
+import json
+import collections
+from tqdm import tqdm
+
+
+# Hardcoded path to the different scene datasets
+SCENES_DATASET = {
+    "hm3d": "./data/habitat-sim-data/scene_datasets/hm3d/",
+    "gibson": "./data/habitat-sim-data/scene_datasets/gibson/",
+    "habitat-test-scenes": "./data/habitat-sim/scene_datasets/habitat-test-scenes/",
+    "replica_cad_baked_lighting": "./data/habitat-sim/scene_datasets/replica_cad_baked_lighting/",
+    "replica_cad": "./data/habitat-sim/scene_datasets/replica_cad/",
+    "replica": "./data/habitat-sim/scene_datasets/ReplicaDataset/",
+    "scannet": "./data/habitat-sim/scene_datasets/scannet/"
+}
+
+SceneData = collections.namedtuple("SceneData", ["scene_dataset_config_file", "scene", "navmesh", "output_dir"])
+
+def list_replicacad_scenes(base_output_dir, base_path=SCENES_DATASET["replica_cad"]):
+    scene_dataset_config_file = os.path.join(base_path, "replicaCAD.scene_dataset_config.json")
+    scenes = [f"apt_{i}" for i in range(6)] + ["empty_stage"]
+    navmeshes = [f"navmeshes/apt_{i}_static_furniture.navmesh" for i in range(6)] + ["empty_stage.navmesh"]
+    scenes_data = []
+    for idx in range(len(scenes)):
+        output_dir = os.path.join(base_output_dir, "ReplicaCAD", scenes[idx])
+        # Add scene
+        data = SceneData(scene_dataset_config_file=scene_dataset_config_file,
+                    scene = scenes[idx] + ".scene_instance.json",
+                    navmesh = os.path.join(base_path, navmeshes[idx]),
+                    output_dir = output_dir)
+        scenes_data.append(data)
+    return scenes_data
+
+def list_replica_cad_baked_lighting_scenes(base_output_dir, base_path=SCENES_DATASET["replica_cad_baked_lighting"]):
+    scene_dataset_config_file = os.path.join(base_path, "replicaCAD_baked.scene_dataset_config.json")
+    scenes = sum([[f"Baked_sc{i}_staging_{j:02}" for i in range(5)] for j in range(21)], [])
+    navmeshes = ""#[f"navmeshes/apt_{i}_static_furniture.navmesh" for i in range(6)] + ["empty_stage.navmesh"]
+    scenes_data = []
+    for idx in range(len(scenes)):
+        output_dir = os.path.join(base_output_dir, "replica_cad_baked_lighting", scenes[idx])
+        data = SceneData(scene_dataset_config_file=scene_dataset_config_file,
+                    scene = scenes[idx],
+                    navmesh = "",
+                    output_dir = output_dir)
+        scenes_data.append(data)
+    return scenes_data    
+
+def list_replica_scenes(base_output_dir, base_path):
+    scenes_data = []
+    for scene_id in os.listdir(base_path):
+        scene = os.path.join(base_path, scene_id, "mesh.ply")
+        navmesh = os.path.join(base_path, scene_id, "habitat/mesh_preseg_semantic.navmesh") # Not sure if I should use it
+        scene_dataset_config_file = ""
+        output_dir = os.path.join(base_output_dir, scene_id)
+        # Add scene only if it does not exist already, or if exist_ok
+        data = SceneData(scene_dataset_config_file = scene_dataset_config_file,
+                    scene = scene,
+                    navmesh = navmesh,
+                    output_dir = output_dir)
+        scenes_data.append(data)
+    return scenes_data
+
+
+def list_scenes(base_output_dir, base_path):
+    """
+    Generic method iterating through a base_path folder to find scenes.
+    """
+    scenes_data = []
+    for root, dirs, files in os.walk(base_path, followlinks=True):
+        folder_scenes_data = []
+        for file in files:
+            name, ext = os.path.splitext(file)
+            if ext == ".glb":
+                scene = os.path.join(root, name + ".glb")
+                navmesh = os.path.join(root, name + ".navmesh")
+                if not os.path.exists(navmesh):
+                    navmesh = ""
+                relpath = os.path.relpath(root, base_path)
+                output_dir = os.path.abspath(os.path.join(base_output_dir, relpath, name))
+                data = SceneData(scene_dataset_config_file="",
+                    scene = scene,
+                    navmesh = navmesh,
+                    output_dir = output_dir)
+                folder_scenes_data.append(data)
+
+        # Specific check for HM3D:
+        # When two meshesxxxx.basis.glb and xxxx.glb are present, use the 'basis' version.
+        basis_scenes = [data.scene[:-len(".basis.glb")] for data in folder_scenes_data if data.scene.endswith(".basis.glb")]
+        if len(basis_scenes) != 0:
+            folder_scenes_data = [data for data in folder_scenes_data if not (data.scene[:-len(".glb")] in basis_scenes)]
+
+        scenes_data.extend(folder_scenes_data)
+    return scenes_data
+
+def list_scenes_available(base_output_dir, scenes_dataset_paths=SCENES_DATASET):
+    scenes_data = []
+
+    # HM3D
+    for split in ("minival", "train", "val", "examples"):
+        scenes_data += list_scenes(base_output_dir=os.path.join(base_output_dir, f"hm3d/{split}/"),
+                                    base_path=f"{scenes_dataset_paths['hm3d']}/{split}")
+
+    # Gibson
+    scenes_data += list_scenes(base_output_dir=os.path.join(base_output_dir, "gibson"),
+                                base_path=scenes_dataset_paths["gibson"])
+
+    # Habitat test scenes (just a few)
+    scenes_data += list_scenes(base_output_dir=os.path.join(base_output_dir, "habitat-test-scenes"),
+                                base_path=scenes_dataset_paths["habitat-test-scenes"])
+
+    # ReplicaCAD (baked lightning)
+    scenes_data += list_replica_cad_baked_lighting_scenes(base_output_dir=base_output_dir)
+
+    # ScanNet
+    scenes_data += list_scenes(base_output_dir=os.path.join(base_output_dir, "scannet"), 
+                            base_path=scenes_dataset_paths["scannet"])
+    
+    # Replica
+    list_replica_scenes(base_output_dir=os.path.join(base_output_dir, "replica"),
+                        base_path=scenes_dataset_paths["replica"])
+    return scenes_data    

third_party/dust3r/croco/datasets/pairs_dataset.py

@@ -0,0 +1,109 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+
+from datasets.transforms import get_pair_transforms
+
+def load_image(impath):
+    return Image.open(impath)
+
+def load_pairs_from_cache_file(fname, root=''):
+    assert os.path.isfile(fname), "cannot parse pairs from {:s}, file does not exist".format(fname)
+    with open(fname, 'r') as fid:
+        lines = fid.read().strip().splitlines()
+    pairs = [ (os.path.join(root,l.split()[0]), os.path.join(root,l.split()[1])) for l in lines]
+    return pairs
+    
+def load_pairs_from_list_file(fname, root=''):
+    assert os.path.isfile(fname), "cannot parse pairs from {:s}, file does not exist".format(fname)
+    with open(fname, 'r') as fid:
+        lines = fid.read().strip().splitlines()
+    pairs = [ (os.path.join(root,l+'_1.jpg'), os.path.join(root,l+'_2.jpg')) for l in lines if not l.startswith('#')]
+    return pairs
+    
+    
+def write_cache_file(fname, pairs, root=''):
+    if len(root)>0:
+        if not root.endswith('/'): root+='/'
+        assert os.path.isdir(root)
+    s = ''
+    for im1, im2 in pairs:
+        if len(root)>0:
+            assert im1.startswith(root), im1
+            assert im2.startswith(root), im2
+        s += '{:s} {:s}\n'.format(im1[len(root):], im2[len(root):])
+    with open(fname, 'w') as fid:
+        fid.write(s[:-1])
+    
+def parse_and_cache_all_pairs(dname, data_dir='./data/'):
+    if dname=='habitat_release':
+        dirname = os.path.join(data_dir, 'habitat_release')
+        assert os.path.isdir(dirname), "cannot find folder for habitat_release pairs: "+dirname
+        cache_file = os.path.join(dirname, 'pairs.txt')
+        assert not os.path.isfile(cache_file), "cache file already exists: "+cache_file
+        
+        print('Parsing pairs for dataset: '+dname)
+        pairs = []
+        for root, dirs, files in os.walk(dirname):
+            if 'val' in root: continue
+            dirs.sort()
+            pairs += [ (os.path.join(root,f), os.path.join(root,f[:-len('_1.jpeg')]+'_2.jpeg')) for f in sorted(files) if f.endswith('_1.jpeg')]
+        print('Found {:,} pairs'.format(len(pairs)))
+        print('Writing cache to: '+cache_file)
+        write_cache_file(cache_file, pairs, root=dirname)
+
+    else:
+        raise NotImplementedError('Unknown dataset: '+dname)
+    
+def dnames_to_image_pairs(dnames, data_dir='./data/'):
+    """
+    dnames: list of datasets with image pairs, separated by +
+    """
+    all_pairs = []
+    for dname in dnames.split('+'):
+        if dname=='habitat_release':
+            dirname = os.path.join(data_dir, 'habitat_release')
+            assert os.path.isdir(dirname), "cannot find folder for habitat_release pairs: "+dirname
+            cache_file = os.path.join(dirname, 'pairs.txt')
+            assert os.path.isfile(cache_file), "cannot find cache file for habitat_release pairs, please first create the cache file, see instructions. "+cache_file
+            pairs = load_pairs_from_cache_file(cache_file, root=dirname)
+        elif dname in ['ARKitScenes', 'MegaDepth', '3DStreetView', 'IndoorVL']:
+            dirname = os.path.join(data_dir, dname+'_crops')
+            assert os.path.isdir(dirname), "cannot find folder for {:s} pairs: {:s}".format(dname, dirname)
+            list_file = os.path.join(dirname, 'listing.txt')
+            assert os.path.isfile(list_file), "cannot find list file for {:s} pairs, see instructions. {:s}".format(dname, list_file)
+            pairs = load_pairs_from_list_file(list_file, root=dirname)            
+        print('  {:s}: {:,} pairs'.format(dname, len(pairs)))
+        all_pairs += pairs 
+    if '+' in dnames: print(' Total: {:,} pairs'.format(len(all_pairs)))
+    return all_pairs 
+
+
+class PairsDataset(Dataset):
+
+    def __init__(self, dnames, trfs='', totensor=True, normalize=True, data_dir='./data/'):
+        super().__init__()
+        self.image_pairs = dnames_to_image_pairs(dnames, data_dir=data_dir)
+        self.transforms = get_pair_transforms(transform_str=trfs, totensor=totensor, normalize=normalize)
+              
+    def __len__(self):
+        return len(self.image_pairs)
+            
+    def __getitem__(self, index):
+        im1path, im2path = self.image_pairs[index]
+        im1 = load_image(im1path)
+        im2 = load_image(im2path)
+        if self.transforms is not None: im1, im2 = self.transforms(im1, im2)
+        return im1, im2
+
+        
+if __name__=="__main__":
+    import argparse
+    parser = argparse.ArgumentParser(prog="Computing and caching list of pairs for a given dataset")
+    parser.add_argument('--data_dir', default='./data/', type=str, help="path where data are stored")
+    parser.add_argument('--dataset', default='habitat_release', type=str, help="name of the dataset")
+    args = parser.parse_args()
+    parse_and_cache_all_pairs(dname=args.dataset, data_dir=args.data_dir)

third_party/dust3r/croco/datasets/transforms.py

@@ -0,0 +1,95 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+import torch
+import torchvision.transforms
+import torchvision.transforms.functional as F
+
+# "Pair": apply a transform on a pair
+# "Both": apply the exact same transform to both images
+
+class ComposePair(torchvision.transforms.Compose):
+    def __call__(self, img1, img2):
+        for t in self.transforms:
+            img1, img2 = t(img1, img2)
+        return img1, img2
+
+class NormalizeBoth(torchvision.transforms.Normalize):
+    def forward(self, img1, img2):
+        img1 = super().forward(img1)
+        img2 = super().forward(img2)
+        return img1, img2
+
+class ToTensorBoth(torchvision.transforms.ToTensor):
+    def __call__(self, img1, img2):
+        img1 = super().__call__(img1)
+        img2 = super().__call__(img2)
+        return img1, img2
+        
+class RandomCropPair(torchvision.transforms.RandomCrop): 
+    # the crop will be intentionally different for the two images with this class
+    def forward(self, img1, img2):
+        img1 = super().forward(img1)
+        img2 = super().forward(img2)
+        return img1, img2
+
+class ColorJitterPair(torchvision.transforms.ColorJitter): 
+    # can be symmetric (same for both images) or assymetric (different jitter params for each image) depending on assymetric_prob  
+    def __init__(self, assymetric_prob, **kwargs):
+        super().__init__(**kwargs)
+        self.assymetric_prob = assymetric_prob
+    def jitter_one(self, img, fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor):
+        for fn_id in fn_idx:
+            if fn_id == 0 and brightness_factor is not None:
+                img = F.adjust_brightness(img, brightness_factor)
+            elif fn_id == 1 and contrast_factor is not None:
+                img = F.adjust_contrast(img, contrast_factor)
+            elif fn_id == 2 and saturation_factor is not None:
+                img = F.adjust_saturation(img, saturation_factor)
+            elif fn_id == 3 and hue_factor is not None:
+                img = F.adjust_hue(img, hue_factor)
+        return img
+        
+    def forward(self, img1, img2):
+
+        fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor = self.get_params(
+            self.brightness, self.contrast, self.saturation, self.hue
+        )
+        img1 = self.jitter_one(img1, fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor)
+        if torch.rand(1) < self.assymetric_prob: # assymetric:
+            fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor = self.get_params(
+                self.brightness, self.contrast, self.saturation, self.hue
+            )
+        img2 = self.jitter_one(img2, fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor)
+        return img1, img2
+
+def get_pair_transforms(transform_str, totensor=True, normalize=True):
+    # transform_str is eg    crop224+color
+    trfs = []
+    for s in transform_str.split('+'):
+        if s.startswith('crop'):
+            size = int(s[len('crop'):])
+            trfs.append(RandomCropPair(size))
+        elif s=='acolor':
+            trfs.append(ColorJitterPair(assymetric_prob=1.0, brightness=(0.6, 1.4), contrast=(0.6, 1.4), saturation=(0.6, 1.4), hue=0.0))
+        elif s=='': # if transform_str was ""
+            pass
+        else:
+            raise NotImplementedError('Unknown augmentation: '+s)
+            
+    if totensor:
+        trfs.append( ToTensorBoth() )
+    if normalize:
+        trfs.append( NormalizeBoth(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) )
+
+    if len(trfs)==0:
+        return None
+    elif len(trfs)==1:
+        return trfs
+    else:
+        return ComposePair(trfs)
+        
+        
+        
+        
+        

third_party/dust3r/croco/demo.py

@@ -0,0 +1,55 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+import torch
+from models.croco import CroCoNet
+from PIL import Image
+import torchvision.transforms
+from torchvision.transforms import ToTensor, Normalize, Compose
+
+def main():
+    device = torch.device('cuda:0' if torch.cuda.is_available() and torch.cuda.device_count()>0 else 'cpu')
+    
+    # load 224x224 images and transform them to tensor 
+    imagenet_mean = [0.485, 0.456, 0.406]
+    imagenet_mean_tensor = torch.tensor(imagenet_mean).view(1,3,1,1).to(device, non_blocking=True)
+    imagenet_std = [0.229, 0.224, 0.225]
+    imagenet_std_tensor = torch.tensor(imagenet_std).view(1,3,1,1).to(device, non_blocking=True)
+    trfs = Compose([ToTensor(), Normalize(mean=imagenet_mean, std=imagenet_std)])
+    image1 = trfs(Image.open('assets/Chateau1.png').convert('RGB')).to(device, non_blocking=True).unsqueeze(0)
+    image2 = trfs(Image.open('assets/Chateau2.png').convert('RGB')).to(device, non_blocking=True).unsqueeze(0)
+    
+    # load model 
+    ckpt = torch.load('pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth', 'cpu')
+    model = CroCoNet( **ckpt.get('croco_kwargs',{})).to(device)
+    model.eval()
+    msg = model.load_state_dict(ckpt['model'], strict=True)
+    
+    # forward 
+    with torch.inference_mode():
+        out, mask, target = model(image1, image2)
+        
+    # the output is normalized, thus use the mean/std of the actual image to go back to RGB space 
+    patchified = model.patchify(image1)
+    mean = patchified.mean(dim=-1, keepdim=True)
+    var = patchified.var(dim=-1, keepdim=True)
+    decoded_image = model.unpatchify(out * (var + 1.e-6)**.5 + mean)
+    # undo imagenet normalization, prepare masked image
+    decoded_image = decoded_image * imagenet_std_tensor + imagenet_mean_tensor
+    input_image = image1 * imagenet_std_tensor + imagenet_mean_tensor
+    ref_image = image2 * imagenet_std_tensor + imagenet_mean_tensor
+    image_masks = model.unpatchify(model.patchify(torch.ones_like(ref_image)) * mask[:,:,None])
+    masked_input_image = ((1 - image_masks) * input_image)
+
+    # make visualization
+    visualization = torch.cat((ref_image, masked_input_image, decoded_image, input_image), dim=3) # 4*(B, 3, H, W) -> B, 3, H, W*4
+    B, C, H, W = visualization.shape
+    visualization = visualization.permute(1, 0, 2, 3).reshape(C, B*H, W)
+    visualization = torchvision.transforms.functional.to_pil_image(torch.clamp(visualization, 0, 1))
+    fname = "demo_output.png"
+    visualization.save(fname)
+    print('Visualization save in '+fname)
+    
+
+if __name__=="__main__":
+    main()

third_party/dust3r/croco/models/blocks.py

@@ -0,0 +1,241 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+
+# --------------------------------------------------------
+# Main encoder/decoder blocks
+# --------------------------------------------------------
+# References: 
+# timm
+# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
+# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/helpers.py
+# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py
+# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/mlp.py
+# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/patch_embed.py
+
+
+import torch
+import torch.nn as nn 
+
+from itertools import repeat
+import collections.abc
+
+
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
+            return x
+        return tuple(repeat(x, n))
+    return parse
+to_2tuple = _ntuple(2)
+
+def drop_path(x, drop_prob: float = 0., training: bool = False, scale_by_keep: bool = True):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+    """
+    if drop_prob == 0. or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+    if keep_prob > 0.0 and scale_by_keep:
+        random_tensor.div_(keep_prob)
+    return x * random_tensor
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob: float = 0., scale_by_keep: bool = True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)
+
+    def extra_repr(self):
+        return f'drop_prob={round(self.drop_prob,3):0.3f}'
+
+class Mlp(nn.Module):
+    """ MLP as used in Vision Transformer, MLP-Mixer and related networks"""
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, bias=True, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        bias = to_2tuple(bias)
+        drop_probs = to_2tuple(drop)
+
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop1(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+class Attention(nn.Module):
+
+    def __init__(self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0., proj_drop=0.):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim ** -0.5
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.rope = rope 
+
+    def forward(self, x, xpos):
+        B, N, C = x.shape
+
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).transpose(1,3)
+        q, k, v = [qkv[:,:,i] for i in range(3)]
+        # q,k,v = qkv.unbind(2)  # make torchscript happy (cannot use tensor as tuple)
+               
+        if self.rope is not None:
+            q = self.rope(q, xpos)
+            k = self.rope(k, xpos)
+               
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, rope=None):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(dim, rope=rope, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x, xpos):
+        x = x + self.drop_path(self.attn(self.norm1(x), xpos))
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+class CrossAttention(nn.Module):
+    
+    def __init__(self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0., proj_drop=0.):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim ** -0.5
+
+        self.projq = nn.Linear(dim, dim, bias=qkv_bias)
+        self.projk = nn.Linear(dim, dim, bias=qkv_bias)
+        self.projv = nn.Linear(dim, dim, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        
+        self.rope = rope
+        
+    def forward(self, query, key, value, qpos, kpos):
+        B, Nq, C = query.shape
+        Nk = key.shape[1]
+        Nv = value.shape[1]
+        
+        q = self.projq(query).reshape(B,Nq,self.num_heads, C// self.num_heads).permute(0, 2, 1, 3)
+        k = self.projk(key).reshape(B,Nk,self.num_heads, C// self.num_heads).permute(0, 2, 1, 3)
+        v = self.projv(value).reshape(B,Nv,self.num_heads, C// self.num_heads).permute(0, 2, 1, 3)
+        
+        if self.rope is not None:
+            q = self.rope(q, qpos)
+            k = self.rope(k, kpos)
+            
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, Nq, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+class DecoderBlock(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, norm_mem=True, rope=None):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(dim, rope=rope, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)
+        self.cross_attn = CrossAttention(dim, rope=rope, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        self.norm3 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+        self.norm_y = norm_layer(dim) if norm_mem else nn.Identity()
+
+    def forward(self, x, y, xpos, ypos):
+        x = x + self.drop_path(self.attn(self.norm1(x), xpos))
+        y_ = self.norm_y(y)
+        x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, y_, xpos, ypos))
+        x = x + self.drop_path(self.mlp(self.norm3(x)))
+        return x, y
+        
+        
+# patch embedding
+class PositionGetter(object):
+    """ return positions of patches """
+
+    def __init__(self):
+        self.cache_positions = {}
+        
+    def __call__(self, b, h, w, device):
+        if not (h,w) in self.cache_positions:
+            x = torch.arange(w, device=device)
+            y = torch.arange(h, device=device)
+            self.cache_positions[h,w] = torch.cartesian_prod(y, x) # (h, w, 2)
+        pos = self.cache_positions[h,w].view(1, h*w, 2).expand(b, -1, 2).clone()
+        return pos
+
+class PatchEmbed(nn.Module):
+    """ just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed"""
+
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None, flatten=True):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
+        self.num_patches = self.grid_size[0] * self.grid_size[1]
+        self.flatten = flatten
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
+        
+        self.position_getter = PositionGetter()
+        
+    def forward(self, x):
+        B, C, H, W = x.shape
+        torch._assert(H == self.img_size[0], f"Input image height ({H}) doesn't match model ({self.img_size[0]}).")
+        torch._assert(W == self.img_size[1], f"Input image width ({W}) doesn't match model ({self.img_size[1]}).")
+        x = self.proj(x)
+        pos = self.position_getter(B, x.size(2), x.size(3), x.device)
+        if self.flatten:
+            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
+        x = self.norm(x)
+        return x, pos
+        
+    def _init_weights(self):
+        w = self.proj.weight.data
+        torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1])) 
+

third_party/dust3r/croco/models/criterion.py

@@ -0,0 +1,37 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+# 
+# --------------------------------------------------------
+# Criterion to train CroCo
+# --------------------------------------------------------
+# References:
+# MAE: https://github.com/facebookresearch/mae
+# --------------------------------------------------------
+
+import torch
+
+class MaskedMSE(torch.nn.Module):
+
+    def __init__(self, norm_pix_loss=False, masked=True):
+        """
+            norm_pix_loss: normalize each patch by their pixel mean and variance
+            masked: compute loss over the masked patches only 
+        """
+        super().__init__()
+        self.norm_pix_loss = norm_pix_loss
+        self.masked = masked 
+        
+    def forward(self, pred, mask, target):
+        
+        if self.norm_pix_loss:
+            mean = target.mean(dim=-1, keepdim=True)
+            var = target.var(dim=-1, keepdim=True)
+            target = (target - mean) / (var + 1.e-6)**.5
+            
+        loss = (pred - target) ** 2
+        loss = loss.mean(dim=-1)  # [N, L], mean loss per patch
+        if self.masked:
+            loss = (loss * mask).sum() / mask.sum()  # mean loss on masked patches
+        else:
+            loss = loss.mean()  # mean loss
+        return loss

third_party/dust3r/croco/models/croco.py

@@ -0,0 +1,249 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+
+# --------------------------------------------------------
+# CroCo model during pretraining
+# --------------------------------------------------------
+
+
+
+import torch
+import torch.nn as nn
+torch.backends.cuda.matmul.allow_tf32 = True # for gpu >= Ampere and pytorch >= 1.12
+from functools import partial
+
+from models.blocks import Block, DecoderBlock, PatchEmbed
+from models.pos_embed import get_2d_sincos_pos_embed, RoPE2D 
+from models.masking import RandomMask
+
+
+class CroCoNet(nn.Module):
+
+    def __init__(self,
+                 img_size=224,           # input image size
+                 patch_size=16,          # patch_size 
+                 mask_ratio=0.9,         # ratios of masked tokens 
+                 enc_embed_dim=768,      # encoder feature dimension
+                 enc_depth=12,           # encoder depth 
+                 enc_num_heads=12,       # encoder number of heads in the transformer block 
+                 dec_embed_dim=512,      # decoder feature dimension 
+                 dec_depth=8,            # decoder depth 
+                 dec_num_heads=16,       # decoder number of heads in the transformer block 
+                 mlp_ratio=4,
+                 norm_layer=partial(nn.LayerNorm, eps=1e-6),
+                 norm_im2_in_dec=True,   # whether to apply normalization of the 'memory' = (second image) in the decoder 
+                 pos_embed='cosine',     # positional embedding (either cosine or RoPE100)
+                ):
+                
+        super(CroCoNet, self).__init__()
+                
+        # patch embeddings  (with initialization done as in MAE)
+        self._set_patch_embed(img_size, patch_size, enc_embed_dim)
+
+        # mask generations
+        self._set_mask_generator(self.patch_embed.num_patches, mask_ratio)
+
+        self.pos_embed = pos_embed
+        if pos_embed=='cosine':
+            # positional embedding of the encoder 
+            enc_pos_embed = get_2d_sincos_pos_embed(enc_embed_dim, int(self.patch_embed.num_patches**.5), n_cls_token=0)
+            self.register_buffer('enc_pos_embed', torch.from_numpy(enc_pos_embed).float())
+            # positional embedding of the decoder  
+            dec_pos_embed = get_2d_sincos_pos_embed(dec_embed_dim, int(self.patch_embed.num_patches**.5), n_cls_token=0)
+            self.register_buffer('dec_pos_embed', torch.from_numpy(dec_pos_embed).float())
+            # pos embedding in each block
+            self.rope = None # nothing for cosine 
+        elif pos_embed.startswith('RoPE'): # eg RoPE100 
+            self.enc_pos_embed = None # nothing to add in the encoder with RoPE
+            self.dec_pos_embed = None # nothing to add in the decoder with RoPE
+            if RoPE2D is None: raise ImportError("Cannot find cuRoPE2D, please install it following the README instructions")
+            freq = float(pos_embed[len('RoPE'):])
+            self.rope = RoPE2D(freq=freq)
+        else:
+            raise NotImplementedError('Unknown pos_embed '+pos_embed)
+
+        # transformer for the encoder 
+        self.enc_depth = enc_depth
+        self.enc_embed_dim = enc_embed_dim
+        self.enc_blocks = nn.ModuleList([
+            Block(enc_embed_dim, enc_num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer, rope=self.rope)
+            for i in range(enc_depth)])
+        self.enc_norm = norm_layer(enc_embed_dim)
+        
+        # masked tokens 
+        self._set_mask_token(dec_embed_dim)
+
+        # decoder 
+        self._set_decoder(enc_embed_dim, dec_embed_dim, dec_num_heads, dec_depth, mlp_ratio, norm_layer, norm_im2_in_dec)
+        
+        # prediction head 
+        self._set_prediction_head(dec_embed_dim, patch_size)
+        
+        # initializer weights
+        self.initialize_weights()           
+
+    def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768):
+        self.patch_embed = PatchEmbed(img_size, patch_size, 3, enc_embed_dim)
+
+    def _set_mask_generator(self, num_patches, mask_ratio):
+        self.mask_generator = RandomMask(num_patches, mask_ratio)
+        
+    def _set_mask_token(self, dec_embed_dim):
+        self.mask_token = nn.Parameter(torch.zeros(1, 1, dec_embed_dim))
+        
+    def _set_decoder(self, enc_embed_dim, dec_embed_dim, dec_num_heads, dec_depth, mlp_ratio, norm_layer, norm_im2_in_dec):
+        self.dec_depth = dec_depth
+        self.dec_embed_dim = dec_embed_dim
+        # transfer from encoder to decoder 
+        self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True)
+        # transformer for the decoder 
+        self.dec_blocks = nn.ModuleList([
+            DecoderBlock(dec_embed_dim, dec_num_heads, mlp_ratio=mlp_ratio, qkv_bias=True, norm_layer=norm_layer, norm_mem=norm_im2_in_dec, rope=self.rope)
+            for i in range(dec_depth)])
+        # final norm layer 
+        self.dec_norm = norm_layer(dec_embed_dim)
+        
+    def _set_prediction_head(self, dec_embed_dim, patch_size):
+         self.prediction_head = nn.Linear(dec_embed_dim, patch_size**2 * 3, bias=True)
+        
+        
+    def initialize_weights(self):
+        # patch embed 
+        self.patch_embed._init_weights()
+        # mask tokens
+        if self.mask_token is not None: torch.nn.init.normal_(self.mask_token, std=.02)
+        # linears and layer norms
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            # we use xavier_uniform following official JAX ViT:
+            torch.nn.init.xavier_uniform_(m.weight)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+            
+    def _encode_image(self, image, do_mask=False, return_all_blocks=False):
+        """
+        image has B x 3 x img_size x img_size 
+        do_mask: whether to perform masking or not
+        return_all_blocks: if True, return the features at the end of every block 
+                           instead of just the features from the last block (eg for some prediction heads)
+        """
+        # embed the image into patches  (x has size B x Npatches x C) 
+        # and get position if each return patch (pos has size B x Npatches x 2)
+        x, pos = self.patch_embed(image)              
+        # add positional embedding without cls token  
+        if self.enc_pos_embed is not None: 
+            x = x + self.enc_pos_embed[None,...]
+        # apply masking 
+        B,N,C = x.size()
+        if do_mask:
+            masks = self.mask_generator(x)
+            x = x[~masks].view(B, -1, C)
+            posvis = pos[~masks].view(B, -1, 2)
+        else:
+            B,N,C = x.size()
+            masks = torch.zeros((B,N), dtype=bool)
+            posvis = pos
+        # now apply the transformer encoder and normalization        
+        if return_all_blocks:
+            out = []
+            for blk in self.enc_blocks:
+                x = blk(x, posvis)
+                out.append(x)
+            out[-1] = self.enc_norm(out[-1])
+            return out, pos, masks
+        else:
+            for blk in self.enc_blocks:
+                x = blk(x, posvis)
+            x = self.enc_norm(x)
+            return x, pos, masks
+ 
+    def _decoder(self, feat1, pos1, masks1, feat2, pos2, return_all_blocks=False):
+        """
+        return_all_blocks: if True, return the features at the end of every block 
+                           instead of just the features from the last block (eg for some prediction heads)
+                           
+        masks1 can be None => assume image1 fully visible 
+        """
+        # encoder to decoder layer 
+        visf1 = self.decoder_embed(feat1)
+        f2 = self.decoder_embed(feat2)
+        # append masked tokens to the sequence
+        B,Nenc,C = visf1.size()
+        if masks1 is None: # downstreams
+            f1_ = visf1
+        else: # pretraining 
+            Ntotal = masks1.size(1)
+            f1_ = self.mask_token.repeat(B, Ntotal, 1).to(dtype=visf1.dtype)
+            f1_[~masks1] = visf1.view(B * Nenc, C)
+        # add positional embedding
+        if self.dec_pos_embed is not None:
+            f1_ = f1_ + self.dec_pos_embed
+            f2 = f2 + self.dec_pos_embed
+        # apply Transformer blocks
+        out = f1_
+        out2 = f2 
+        if return_all_blocks:
+            _out, out = out, []
+            for blk in self.dec_blocks:
+                _out, out2 = blk(_out, out2, pos1, pos2)
+                out.append(_out)
+            out[-1] = self.dec_norm(out[-1])
+        else:
+            for blk in self.dec_blocks:
+                out, out2 = blk(out, out2, pos1, pos2)
+            out = self.dec_norm(out)
+        return out
+
+    def patchify(self, imgs):
+        """
+        imgs: (B, 3, H, W)
+        x: (B, L, patch_size**2 *3)
+        """
+        p = self.patch_embed.patch_size[0]
+        assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0
+
+        h = w = imgs.shape[2] // p
+        x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))
+        x = torch.einsum('nchpwq->nhwpqc', x)
+        x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))
+        
+        return x
+
+    def unpatchify(self, x, channels=3):
+        """
+        x: (N, L, patch_size**2 *channels)
+        imgs: (N, 3, H, W)
+        """
+        patch_size = self.patch_embed.patch_size[0]
+        h = w = int(x.shape[1]**.5)
+        assert h * w == x.shape[1]
+        x = x.reshape(shape=(x.shape[0], h, w, patch_size, patch_size, channels))
+        x = torch.einsum('nhwpqc->nchpwq', x)
+        imgs = x.reshape(shape=(x.shape[0], channels, h * patch_size, h * patch_size))
+        return imgs
+
+    def forward(self, img1, img2):
+        """
+        img1: tensor of size B x 3 x img_size x img_size
+        img2: tensor of size B x 3 x img_size x img_size
+        
+        out will be    B x N x (3*patch_size*patch_size)
+        masks are also returned as B x N just in case 
+        """
+        # encoder of the masked first image 
+        feat1, pos1, mask1 = self._encode_image(img1, do_mask=True)
+        # encoder of the second image 
+        feat2, pos2, _ = self._encode_image(img2, do_mask=False)
+        # decoder 
+        decfeat = self._decoder(feat1, pos1, mask1, feat2, pos2)
+        # prediction head 
+        out = self.prediction_head(decfeat)
+        # get target
+        target = self.patchify(img1)
+        return out, mask1, target

third_party/dust3r/croco/models/croco_downstream.py

@@ -0,0 +1,122 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+# --------------------------------------------------------
+# CroCo model for downstream tasks
+# --------------------------------------------------------
+
+import torch
+
+from .croco import CroCoNet
+
+
+def croco_args_from_ckpt(ckpt):
+    if 'croco_kwargs' in ckpt: # CroCo v2 released models
+        return ckpt['croco_kwargs']
+    elif 'args' in ckpt and hasattr(ckpt['args'], 'model'): # pretrained using the official code release
+        s = ckpt['args'].model # eg "CroCoNet(enc_embed_dim=1024, enc_num_heads=16, enc_depth=24)"
+        assert s.startswith('CroCoNet(')
+        return eval('dict'+s[len('CroCoNet'):]) # transform it into the string of a dictionary and evaluate it
+    else: # CroCo v1 released models
+        return dict()
+
+class CroCoDownstreamMonocularEncoder(CroCoNet):
+
+    def __init__(self,
+                 head,
+                 **kwargs):
+        """ Build network for monocular downstream task, only using the encoder.
+        It takes an extra argument head, that is called with the features 
+          and a dictionary img_info containing 'width' and 'height' keys
+        The head is setup with the croconet arguments in this init function
+        NOTE: It works by *calling super().__init__() but with redefined setters
+        
+        """
+        super(CroCoDownstreamMonocularEncoder, self).__init__(**kwargs)
+        head.setup(self)
+        self.head = head
+
+    def _set_mask_generator(self, *args, **kwargs):
+        """ No mask generator """
+        return
+
+    def _set_mask_token(self, *args, **kwargs):
+        """ No mask token """
+        self.mask_token = None
+        return
+
+    def _set_decoder(self, *args, **kwargs):
+        """ No decoder """
+        return
+
+    def _set_prediction_head(self, *args, **kwargs):
+        """ No 'prediction head' for downstream tasks."""
+        return
+
+    def forward(self, img):
+        """
+        img if of size batch_size x 3 x h x w
+        """
+        B, C, H, W = img.size()
+        img_info = {'height': H, 'width': W}
+        need_all_layers = hasattr(self.head, 'return_all_blocks') and self.head.return_all_blocks
+        out, _, _ = self._encode_image(img, do_mask=False, return_all_blocks=need_all_layers)
+        return self.head(out, img_info)
+        
+        
+class CroCoDownstreamBinocular(CroCoNet):
+
+    def __init__(self,
+                 head,
+                 **kwargs):
+        """ Build network for binocular downstream task
+        It takes an extra argument head, that is called with the features 
+          and a dictionary img_info containing 'width' and 'height' keys
+        The head is setup with the croconet arguments in this init function
+        """
+        super(CroCoDownstreamBinocular, self).__init__(**kwargs)
+        head.setup(self)
+        self.head = head
+
+    def _set_mask_generator(self, *args, **kwargs):
+        """ No mask generator """
+        return
+
+    def _set_mask_token(self, *args, **kwargs):
+        """ No mask token """
+        self.mask_token = None
+        return
+
+    def _set_prediction_head(self, *args, **kwargs):
+        """ No prediction head for downstream tasks, define your own head """
+        return
+        
+    def encode_image_pairs(self, img1, img2, return_all_blocks=False):
+        """ run encoder for a pair of images
+            it is actually ~5% faster to concatenate the images along the batch dimension 
+             than to encode them separately
+        """
+        ## the two commented lines below is the naive version with separate encoding
+        #out, pos, _ = self._encode_image(img1, do_mask=False, return_all_blocks=return_all_blocks)
+        #out2, pos2, _ = self._encode_image(img2, do_mask=False, return_all_blocks=False)
+        ## and now the faster version
+        out, pos, _ = self._encode_image( torch.cat( (img1,img2), dim=0), do_mask=False, return_all_blocks=return_all_blocks )
+        if return_all_blocks:
+            out,out2 = list(map(list, zip(*[o.chunk(2, dim=0) for o in out])))
+            out2 = out2[-1]
+        else:
+            out,out2 = out.chunk(2, dim=0)
+        pos,pos2 = pos.chunk(2, dim=0)            
+        return out, out2, pos, pos2
+
+    def forward(self, img1, img2):
+        B, C, H, W = img1.size()
+        img_info = {'height': H, 'width': W}
+        return_all_blocks = hasattr(self.head, 'return_all_blocks') and self.head.return_all_blocks
+        out, out2, pos, pos2 = self.encode_image_pairs(img1, img2, return_all_blocks=return_all_blocks)
+        if return_all_blocks:
+            decout = self._decoder(out[-1], pos, None, out2, pos2, return_all_blocks=return_all_blocks)
+            decout = out+decout
+        else:
+            decout = self._decoder(out, pos, None, out2, pos2, return_all_blocks=return_all_blocks)
+        return self.head(decout, img_info)

third_party/dust3r/croco/models/curope/__init__.py

@@ -0,0 +1,4 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+from .curope2d import cuRoPE2D

third_party/dust3r/croco/models/curope/curope.cpp

@@ -0,0 +1,69 @@
+/* 
+  Copyright (C) 2022-present Naver Corporation. All rights reserved.
+  Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+*/
+
+#include <torch/extension.h>
+
+// forward declaration
+void rope_2d_cuda( torch::Tensor tokens, const torch::Tensor pos, const float base, const float fwd );
+
+void rope_2d_cpu( torch::Tensor tokens, const torch::Tensor positions, const float base, const float fwd )
+{
+    const int B = tokens.size(0);
+    const int N = tokens.size(1);
+    const int H = tokens.size(2);
+    const int D = tokens.size(3) / 4;
+
+    auto tok = tokens.accessor<float, 4>();
+    auto pos = positions.accessor<int64_t, 3>();
+
+    for (int b = 0; b < B; b++) {
+      for (int x = 0; x < 2; x++) { // y and then x (2d)
+        for (int n = 0; n < N; n++) {
+        
+            // grab the token position
+            const int p = pos[b][n][x];
+
+            for (int h = 0; h < H; h++) {
+                for (int d = 0; d < D; d++) {
+                    // grab the two values
+                    float u = tok[b][n][h][d+0+x*2*D];
+                    float v = tok[b][n][h][d+D+x*2*D];
+
+                    // grab the cos,sin
+                    const float inv_freq = fwd * p / powf(base, d/float(D));
+                    float c = cosf(inv_freq);
+                    float s = sinf(inv_freq);
+
+                    // write the result
+                    tok[b][n][h][d+0+x*2*D] = u*c - v*s;
+                    tok[b][n][h][d+D+x*2*D] = v*c + u*s;
+                }
+            }
+        }
+      }
+    }
+}
+
+void rope_2d( torch::Tensor tokens,     // B,N,H,D
+        const torch::Tensor positions,  // B,N,2
+        const float base, 
+        const float fwd )
+{
+    TORCH_CHECK(tokens.dim() == 4, "tokens must have 4 dimensions");
+    TORCH_CHECK(positions.dim() == 3, "positions must have 3 dimensions");
+    TORCH_CHECK(tokens.size(0) == positions.size(0), "batch size differs between tokens & positions");
+    TORCH_CHECK(tokens.size(1) == positions.size(1), "seq_length differs between tokens & positions");
+    TORCH_CHECK(positions.size(2) == 2, "positions.shape[2] must be equal to 2");
+    TORCH_CHECK(tokens.is_cuda() == positions.is_cuda(), "tokens and positions are not on the same device" );
+
+    if (tokens.is_cuda())
+        rope_2d_cuda( tokens, positions, base, fwd );
+    else
+        rope_2d_cpu( tokens, positions, base, fwd );
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("rope_2d", &rope_2d, "RoPE 2d forward/backward");
+}

third_party/dust3r/croco/models/curope/curope2d.py

@@ -0,0 +1,40 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+import torch
+
+try:
+    import curope as _kernels # run `python setup.py install`
+except ModuleNotFoundError:
+    from . import curope as _kernels # run `python setup.py build_ext --inplace`
+
+
+class cuRoPE2D_func (torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, tokens, positions, base, F0=1):
+        ctx.save_for_backward(positions)
+        ctx.saved_base = base
+        ctx.saved_F0 = F0
+        # tokens = tokens.clone() # uncomment this if inplace doesn't work
+        _kernels.rope_2d( tokens, positions, base, F0 )
+        ctx.mark_dirty(tokens)
+        return tokens
+
+    @staticmethod
+    def backward(ctx, grad_res):
+        positions, base, F0 = ctx.saved_tensors[0], ctx.saved_base, ctx.saved_F0
+        _kernels.rope_2d( grad_res, positions, base, -F0 )
+        ctx.mark_dirty(grad_res)
+        return grad_res, None, None, None
+
+
+class cuRoPE2D(torch.nn.Module):
+    def __init__(self, freq=100.0, F0=1.0):
+        super().__init__()
+        self.base = freq 
+        self.F0 = F0
+
+    def forward(self, tokens, positions): 
+        cuRoPE2D_func.apply( tokens.transpose(1,2), positions, self.base, self.F0 )
+        return tokens

third_party/dust3r/croco/models/curope/kernels.cu

@@ -0,0 +1,108 @@
+/* 
+  Copyright (C) 2022-present Naver Corporation. All rights reserved.
+  Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+*/
+
+#include <torch/extension.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <vector>
+
+#define CHECK_CUDA(tensor) {\
+    TORCH_CHECK((tensor).is_cuda(), #tensor " is not in cuda memory"); \
+    TORCH_CHECK((tensor).is_contiguous(), #tensor " is not contiguous"); }
+void CHECK_KERNEL() {auto error = cudaGetLastError(); TORCH_CHECK( error == cudaSuccess, cudaGetErrorString(error));}
+
+
+template < typename scalar_t  >
+__global__ void rope_2d_cuda_kernel( 
+        //scalar_t* __restrict__ tokens, 
+        torch::PackedTensorAccessor32<scalar_t,4,torch::RestrictPtrTraits> tokens,
+        const int64_t* __restrict__ pos, 
+        const float base, 
+        const float fwd )
+        // const int N, const int H, const int D )
+{
+    // tokens shape = (B, N, H, D)
+    const int N = tokens.size(1);
+    const int H = tokens.size(2);
+    const int D = tokens.size(3);
+    
+    // each block update a single token, for all heads
+    // each thread takes care of a single output
+    extern __shared__ float shared[];
+    float* shared_inv_freq = shared + D;
+
+    const int b = blockIdx.x / N;
+    const int n = blockIdx.x % N;
+
+    const int Q = D / 4; 
+    // one token = [0..Q : Q..2Q : 2Q..3Q : 3Q..D]
+    //              u_Y     v_Y     u_X      v_X
+
+    // shared memory: first, compute inv_freq
+    if (threadIdx.x < Q)
+        shared_inv_freq[threadIdx.x] = fwd / powf(base, threadIdx.x/float(Q));
+    __syncthreads();
+
+    // start of X or Y part
+    const int X = threadIdx.x < D/2 ? 0 : 1; 
+    const int m = (X*D/2) + (threadIdx.x % Q);   // index of u_Y or u_X
+
+    // grab the cos,sin appropriate for me
+    const float freq = pos[blockIdx.x*2+X] * shared_inv_freq[threadIdx.x % Q];
+    const float cos = cosf(freq);
+    const float sin = sinf(freq);
+    /*
+    float* shared_cos_sin = shared + D + D/4;
+    if ((threadIdx.x % (D/2)) < Q)
+        shared_cos_sin[m+0] = cosf(freq);
+    else
+        shared_cos_sin[m+Q] = sinf(freq);
+    __syncthreads();
+    const float cos = shared_cos_sin[m+0];
+    const float sin = shared_cos_sin[m+Q];
+    */
+
+    for (int h = 0; h < H; h++)
+    {
+        // then, load all the token for this head in shared memory
+        shared[threadIdx.x] = tokens[b][n][h][threadIdx.x];
+        __syncthreads();
+
+        const float u = shared[m];
+        const float v = shared[m+Q];
+        
+        // write output
+        if ((threadIdx.x % (D/2)) < Q)
+            tokens[b][n][h][threadIdx.x] = u*cos - v*sin;
+        else
+            tokens[b][n][h][threadIdx.x] = v*cos + u*sin;
+    }
+}
+
+void rope_2d_cuda( torch::Tensor tokens, const torch::Tensor pos, const float base, const float fwd ) 
+{
+    const int B = tokens.size(0); // batch size
+    const int N = tokens.size(1); // sequence length
+    const int H = tokens.size(2); // number of heads
+    const int D = tokens.size(3); // dimension per head
+
+    TORCH_CHECK(tokens.stride(3) == 1 && tokens.stride(2) == D, "tokens are not contiguous");
+    TORCH_CHECK(pos.is_contiguous(), "positions are not contiguous");
+    TORCH_CHECK(pos.size(0) == B && pos.size(1) == N && pos.size(2) == 2, "bad pos.shape");
+    TORCH_CHECK(D % 4 == 0, "token dim must be multiple of 4");
+
+    // one block for each layer, one thread per local-max
+    const int THREADS_PER_BLOCK = D;
+    const int N_BLOCKS = B * N; // each block takes care of H*D values
+    const int SHARED_MEM = sizeof(float) * (D + D/4);
+
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(tokens.type(), "rope_2d_cuda", ([&] {
+        rope_2d_cuda_kernel<scalar_t> <<<N_BLOCKS, THREADS_PER_BLOCK, SHARED_MEM>>> (
+            //tokens.data_ptr<scalar_t>(), 
+            tokens.packed_accessor32<scalar_t,4,torch::RestrictPtrTraits>(),
+            pos.data_ptr<int64_t>(), 
+            base, fwd); //, N, H, D );
+    }));
+}

third_party/dust3r/croco/models/curope/setup.py

@@ -0,0 +1,34 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+from setuptools import setup
+from torch import cuda
+from torch.utils.cpp_extension import BuildExtension, CUDAExtension
+
+# compile for all possible CUDA architectures
+all_cuda_archs = cuda.get_gencode_flags().replace('compute=','arch=').split()
+# alternatively, you can list cuda archs that you want, eg:
+# all_cuda_archs = [
+    # '-gencode', 'arch=compute_70,code=sm_70',
+    # '-gencode', 'arch=compute_75,code=sm_75',
+    # '-gencode', 'arch=compute_80,code=sm_80',
+    # '-gencode', 'arch=compute_86,code=sm_86'
+# ]
+
+setup(
+    name = 'curope',
+    ext_modules = [
+        CUDAExtension(
+                name='curope',
+                sources=[
+                    "curope.cpp",
+                    "kernels.cu",
+                ],
+                extra_compile_args = dict(
+                    nvcc=['-O3','--ptxas-options=-v',"--use_fast_math"]+all_cuda_archs, 
+                    cxx=['-O3'])
+                )
+    ],
+    cmdclass = {
+        'build_ext': BuildExtension
+    })

third_party/dust3r/croco/models/dpt_block.py

@@ -0,0 +1,450 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+# --------------------------------------------------------
+# DPT head for ViTs
+# --------------------------------------------------------
+# References: 
+# https://github.com/isl-org/DPT
+# https://github.com/EPFL-VILAB/MultiMAE/blob/main/multimae/output_adapters.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from typing import Union, Tuple, Iterable, List, Optional, Dict
+
+def pair(t):
+    return t if isinstance(t, tuple) else (t, t)
+
+def make_scratch(in_shape, out_shape, groups=1, expand=False):
+    scratch = nn.Module()
+
+    out_shape1 = out_shape
+    out_shape2 = out_shape
+    out_shape3 = out_shape
+    out_shape4 = out_shape
+    if expand == True:
+        out_shape1 = out_shape
+        out_shape2 = out_shape * 2
+        out_shape3 = out_shape * 4
+        out_shape4 = out_shape * 8
+
+    scratch.layer1_rn = nn.Conv2d(
+        in_shape[0],
+        out_shape1,
+        kernel_size=3,
+        stride=1,
+        padding=1,
+        bias=False,
+        groups=groups,
+    )
+    scratch.layer2_rn = nn.Conv2d(
+        in_shape[1],
+        out_shape2,
+        kernel_size=3,
+        stride=1,
+        padding=1,
+        bias=False,
+        groups=groups,
+    )
+    scratch.layer3_rn = nn.Conv2d(
+        in_shape[2],
+        out_shape3,
+        kernel_size=3,
+        stride=1,
+        padding=1,
+        bias=False,
+        groups=groups,
+    )
+    scratch.layer4_rn = nn.Conv2d(
+        in_shape[3],
+        out_shape4,
+        kernel_size=3,
+        stride=1,
+        padding=1,
+        bias=False,
+        groups=groups,
+    )
+
+    scratch.layer_rn = nn.ModuleList([
+        scratch.layer1_rn,
+        scratch.layer2_rn,
+        scratch.layer3_rn,
+        scratch.layer4_rn,
+    ])
+
+    return scratch
+
+class ResidualConvUnit_custom(nn.Module):
+    """Residual convolution module."""
+
+    def __init__(self, features, activation, bn):
+        """Init.
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.bn = bn
+
+        self.groups = 1
+
+        self.conv1 = nn.Conv2d(
+            features,
+            features,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            bias=not self.bn,
+            groups=self.groups,
+        )
+
+        self.conv2 = nn.Conv2d(
+            features,
+            features,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            bias=not self.bn,
+            groups=self.groups,
+        )
+
+        if self.bn == True:
+            self.bn1 = nn.BatchNorm2d(features)
+            self.bn2 = nn.BatchNorm2d(features)
+
+        self.activation = activation
+
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, x):
+        """Forward pass.
+        Args:
+            x (tensor): input
+        Returns:
+            tensor: output
+        """
+
+        out = self.activation(x)
+        out = self.conv1(out)
+        if self.bn == True:
+            out = self.bn1(out)
+
+        out = self.activation(out)
+        out = self.conv2(out)
+        if self.bn == True:
+            out = self.bn2(out)
+
+        if self.groups > 1:
+            out = self.conv_merge(out)
+
+        return self.skip_add.add(out, x)
+
+class FeatureFusionBlock_custom(nn.Module):
+    """Feature fusion block."""
+
+    def __init__(
+        self,
+        features,
+        activation,
+        deconv=False,
+        bn=False,
+        expand=False,
+        align_corners=True,
+        width_ratio=1,
+    ):
+        """Init.
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock_custom, self).__init__()
+        self.width_ratio = width_ratio
+
+        self.deconv = deconv
+        self.align_corners = align_corners
+
+        self.groups = 1
+
+        self.expand = expand
+        out_features = features
+        if self.expand == True:
+            out_features = features // 2
+
+        self.out_conv = nn.Conv2d(
+            features,
+            out_features,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=True,
+            groups=1,
+        )
+
+        self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn)
+        self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn)
+
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, *xs):
+        """Forward pass.
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if len(xs) == 2:
+            res = self.resConfUnit1(xs[1])
+            if self.width_ratio != 1:
+                res = F.interpolate(res, size=(output.shape[2], output.shape[3]), mode='bilinear')
+
+            output = self.skip_add.add(output, res)
+            # output += res
+
+        output = self.resConfUnit2(output)
+
+        if self.width_ratio != 1:
+            # and output.shape[3] < self.width_ratio * output.shape[2]
+            #size=(image.shape[])
+            if (output.shape[3] / output.shape[2]) < (2 / 3) * self.width_ratio:
+                shape = 3 * output.shape[3]
+            else:
+                shape = int(self.width_ratio * 2 * output.shape[2])
+            output  = F.interpolate(output, size=(2* output.shape[2], shape), mode='bilinear')
+        else:
+            output = nn.functional.interpolate(output, scale_factor=2,
+                    mode="bilinear", align_corners=self.align_corners)
+        output = self.out_conv(output)
+        return output
+
+def make_fusion_block(features, use_bn, width_ratio=1):
+    return FeatureFusionBlock_custom(
+        features,
+        nn.ReLU(False),
+        deconv=False,
+        bn=use_bn,
+        expand=False,
+        align_corners=True,
+        width_ratio=width_ratio,
+    )
+
+class Interpolate(nn.Module):
+    """Interpolation module."""
+
+    def __init__(self, scale_factor, mode, align_corners=False):
+        """Init.
+        Args:
+            scale_factor (float): scaling
+            mode (str): interpolation mode
+        """
+        super(Interpolate, self).__init__()
+
+        self.interp = nn.functional.interpolate
+        self.scale_factor = scale_factor
+        self.mode = mode
+        self.align_corners = align_corners
+
+    def forward(self, x):
+        """Forward pass.
+        Args:
+            x (tensor): input
+        Returns:
+            tensor: interpolated data
+        """
+
+        x = self.interp(
+            x,
+            scale_factor=self.scale_factor,
+            mode=self.mode,
+            align_corners=self.align_corners,
+        )
+
+        return x
+
+class DPTOutputAdapter(nn.Module):
+    """DPT output adapter.
+
+    :param num_cahnnels: Number of output channels
+    :param stride_level: tride level compared to the full-sized image.
+        E.g. 4 for 1/4th the size of the image.
+    :param patch_size_full: Int or tuple of the patch size over the full image size.
+        Patch size for smaller inputs will be computed accordingly.
+    :param hooks: Index of intermediate layers
+    :param layer_dims: Dimension of intermediate layers
+    :param feature_dim: Feature dimension
+    :param last_dim: out_channels/in_channels for the last two Conv2d when head_type == regression
+    :param use_bn: If set to True, activates batch norm
+    :param dim_tokens_enc:  Dimension of tokens coming from encoder
+    """
+
+    def __init__(self,
+                 num_channels: int = 1,
+                 stride_level: int = 1,
+                 patch_size: Union[int, Tuple[int, int]] = 16,
+                 main_tasks: Iterable[str] = ('rgb',),
+                 hooks: List[int] = [2, 5, 8, 11],
+                 layer_dims: List[int] = [96, 192, 384, 768],
+                 feature_dim: int = 256,
+                 last_dim: int = 32,
+                 use_bn: bool = False,
+                 dim_tokens_enc: Optional[int] = None,
+                 head_type: str = 'regression',
+                 output_width_ratio=1,
+                 **kwargs):
+        super().__init__()
+        self.num_channels = num_channels
+        self.stride_level = stride_level
+        self.patch_size = pair(patch_size)
+        self.main_tasks = main_tasks
+        self.hooks = hooks
+        self.layer_dims = layer_dims
+        self.feature_dim = feature_dim
+        self.dim_tokens_enc = dim_tokens_enc * len(self.main_tasks) if dim_tokens_enc is not None else None
+        self.head_type = head_type
+
+        # Actual patch height and width, taking into account stride of input
+        self.P_H = max(1, self.patch_size[0] // stride_level)
+        self.P_W = max(1, self.patch_size[1] // stride_level)
+
+        self.scratch = make_scratch(layer_dims, feature_dim, groups=1, expand=False)
+
+        self.scratch.refinenet1 = make_fusion_block(feature_dim, use_bn, output_width_ratio)
+        self.scratch.refinenet2 = make_fusion_block(feature_dim, use_bn, output_width_ratio)
+        self.scratch.refinenet3 = make_fusion_block(feature_dim, use_bn, output_width_ratio)
+        self.scratch.refinenet4 = make_fusion_block(feature_dim, use_bn, output_width_ratio)
+
+        if self.head_type == 'regression':
+            # The "DPTDepthModel" head
+            self.head = nn.Sequential(
+                nn.Conv2d(feature_dim, feature_dim // 2, kernel_size=3, stride=1, padding=1),
+                Interpolate(scale_factor=2, mode="bilinear", align_corners=True),
+                nn.Conv2d(feature_dim // 2, last_dim, kernel_size=3, stride=1, padding=1),
+                nn.ReLU(True),
+                nn.Conv2d(last_dim, self.num_channels, kernel_size=1, stride=1, padding=0)
+            )
+        elif self.head_type == 'semseg':
+            # The "DPTSegmentationModel" head
+            self.head = nn.Sequential(
+                nn.Conv2d(feature_dim, feature_dim, kernel_size=3, padding=1, bias=False),
+                nn.BatchNorm2d(feature_dim) if use_bn else nn.Identity(),
+                nn.ReLU(True),
+                nn.Dropout(0.1, False),
+                nn.Conv2d(feature_dim, self.num_channels, kernel_size=1),
+                Interpolate(scale_factor=2, mode="bilinear", align_corners=True),
+            )
+        else:
+            raise ValueError('DPT head_type must be "regression" or "semseg".')
+
+        if self.dim_tokens_enc is not None:
+            self.init(dim_tokens_enc=dim_tokens_enc)
+
+    def init(self, dim_tokens_enc=768):
+        """
+        Initialize parts of decoder that are dependent on dimension of encoder tokens.
+        Should be called when setting up MultiMAE.
+
+        :param dim_tokens_enc: Dimension of tokens coming from encoder
+        """
+        #print(dim_tokens_enc)
+
+        # Set up activation postprocessing layers
+        if isinstance(dim_tokens_enc, int):
+            dim_tokens_enc = 4 * [dim_tokens_enc]
+
+        self.dim_tokens_enc = [dt * len(self.main_tasks) for dt in dim_tokens_enc]
+
+        self.act_1_postprocess = nn.Sequential(
+            nn.Conv2d(
+                in_channels=self.dim_tokens_enc[0],
+                out_channels=self.layer_dims[0],
+                kernel_size=1, stride=1, padding=0,
+            ),
+            nn.ConvTranspose2d(
+                in_channels=self.layer_dims[0],
+                out_channels=self.layer_dims[0],
+                kernel_size=4, stride=4, padding=0,
+                bias=True, dilation=1, groups=1,
+            )
+        )
+
+        self.act_2_postprocess = nn.Sequential(
+            nn.Conv2d(
+                in_channels=self.dim_tokens_enc[1],
+                out_channels=self.layer_dims[1],
+                kernel_size=1, stride=1, padding=0,
+            ),
+            nn.ConvTranspose2d(
+                in_channels=self.layer_dims[1],
+                out_channels=self.layer_dims[1],
+                kernel_size=2, stride=2, padding=0,
+                bias=True, dilation=1, groups=1,
+            )
+        )
+
+        self.act_3_postprocess = nn.Sequential(
+            nn.Conv2d(
+                in_channels=self.dim_tokens_enc[2],
+                out_channels=self.layer_dims[2],
+                kernel_size=1, stride=1, padding=0,
+            )
+        )
+
+        self.act_4_postprocess = nn.Sequential(
+            nn.Conv2d(
+                in_channels=self.dim_tokens_enc[3],
+                out_channels=self.layer_dims[3],
+                kernel_size=1, stride=1, padding=0,
+            ),
+            nn.Conv2d(
+                in_channels=self.layer_dims[3],
+                out_channels=self.layer_dims[3],
+                kernel_size=3, stride=2, padding=1,
+            )
+        )
+
+        self.act_postprocess = nn.ModuleList([
+            self.act_1_postprocess,
+            self.act_2_postprocess,
+            self.act_3_postprocess,
+            self.act_4_postprocess
+        ])
+
+    def adapt_tokens(self, encoder_tokens):
+        # Adapt tokens
+        x = []
+        x.append(encoder_tokens[:, :])
+        x = torch.cat(x, dim=-1)
+        return x
+
+    def forward(self, encoder_tokens: List[torch.Tensor], image_size):
+            #input_info: Dict):
+        assert self.dim_tokens_enc is not None, 'Need to call init(dim_tokens_enc) function first'
+        H, W = image_size
+        
+        # Number of patches in height and width
+        N_H = H // (self.stride_level * self.P_H)
+        N_W = W // (self.stride_level * self.P_W)
+
+        # Hook decoder onto 4 layers from specified ViT layers
+        layers = [encoder_tokens[hook] for hook in self.hooks]
+
+        # Extract only task-relevant tokens and ignore global tokens.
+        layers = [self.adapt_tokens(l) for l in layers]
+
+        # Reshape tokens to spatial representation
+        layers = [rearrange(l, 'b (nh nw) c -> b c nh nw', nh=N_H, nw=N_W) for l in layers]
+
+        layers = [self.act_postprocess[idx](l) for idx, l in enumerate(layers)]
+        # Project layers to chosen feature dim
+        layers = [self.scratch.layer_rn[idx](l) for idx, l in enumerate(layers)]
+
+        # Fuse layers using refinement stages
+        path_4 = self.scratch.refinenet4(layers[3])
+        path_3 = self.scratch.refinenet3(path_4, layers[2])
+        path_2 = self.scratch.refinenet2(path_3, layers[1])
+        path_1 = self.scratch.refinenet1(path_2, layers[0])
+
+        # Output head
+        out = self.head(path_1)
+
+        return out

third_party/dust3r/croco/models/head_downstream.py

@@ -0,0 +1,58 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+# --------------------------------------------------------
+# Heads for downstream tasks
+# --------------------------------------------------------
+
+"""
+A head is a module where the __init__ defines only the head hyperparameters.
+A method setup(croconet) takes a CroCoNet and set all layers according to the head and croconet attributes.
+The forward takes the features as well as a dictionary img_info containing the keys 'width' and 'height'
+"""
+
+import torch
+import torch.nn as nn
+from .dpt_block import DPTOutputAdapter
+
+
+class PixelwiseTaskWithDPT(nn.Module):
+    """ DPT module for CroCo.
+    by default, hooks_idx will be equal to:
+    * for encoder-only: 4 equally spread layers
+    * for encoder+decoder: last encoder + 3 equally spread layers of the decoder 
+    """
+
+    def __init__(self, *, hooks_idx=None, layer_dims=[96,192,384,768],
+                 output_width_ratio=1, num_channels=1, postprocess=None, **kwargs):
+        super(PixelwiseTaskWithDPT, self).__init__()
+        self.return_all_blocks = True # backbone needs to return all layers 
+        self.postprocess = postprocess
+        self.output_width_ratio = output_width_ratio
+        self.num_channels = num_channels
+        self.hooks_idx = hooks_idx
+        self.layer_dims = layer_dims
+    
+    def setup(self, croconet):
+        dpt_args = {'output_width_ratio': self.output_width_ratio, 'num_channels': self.num_channels}
+        if self.hooks_idx is None:
+            if hasattr(croconet, 'dec_blocks'): # encoder + decoder 
+                step = {8: 3, 12: 4, 24: 8}[croconet.dec_depth]
+                hooks_idx = [croconet.dec_depth+croconet.enc_depth-1-i*step for i in range(3,-1,-1)]
+            else: # encoder only
+                step = croconet.enc_depth//4
+                hooks_idx = [croconet.enc_depth-1-i*step for i in range(3,-1,-1)]
+            self.hooks_idx = hooks_idx
+            print(f'  PixelwiseTaskWithDPT: automatically setting hook_idxs={self.hooks_idx}')
+        dpt_args['hooks'] = self.hooks_idx
+        dpt_args['layer_dims'] = self.layer_dims
+        self.dpt = DPTOutputAdapter(**dpt_args)
+        dim_tokens = [croconet.enc_embed_dim if hook<croconet.enc_depth else croconet.dec_embed_dim for hook in self.hooks_idx]
+        dpt_init_args = {'dim_tokens_enc': dim_tokens}
+        self.dpt.init(**dpt_init_args)
+
+
+    def forward(self, x, img_info):
+        out = self.dpt(x, image_size=(img_info['height'],img_info['width']))
+        if self.postprocess: out = self.postprocess(out)
+        return out

third_party/dust3r/croco/models/masking.py

@@ -0,0 +1,25 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+
+# --------------------------------------------------------
+# Masking utils
+# --------------------------------------------------------
+
+import torch
+import torch.nn as nn    
+    
+class RandomMask(nn.Module):
+    """
+    random masking
+    """
+
+    def __init__(self, num_patches, mask_ratio):
+        super().__init__()
+        self.num_patches = num_patches
+        self.num_mask = int(mask_ratio * self.num_patches)
+    
+    def __call__(self, x):
+        noise = torch.rand(x.size(0), self.num_patches, device=x.device) 
+        argsort = torch.argsort(noise, dim=1) 
+        return argsort < self.num_mask

third_party/dust3r/croco/models/pos_embed.py

@@ -0,0 +1,159 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+
+# --------------------------------------------------------
+# Position embedding utils
+# --------------------------------------------------------
+
+
+
+import numpy as np
+
+import torch
+
+# --------------------------------------------------------
+# 2D sine-cosine position embedding
+# References:
+# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py
+# Transformer: https://github.com/tensorflow/models/blob/master/official/nlp/transformer/model_utils.py
+# MoCo v3: https://github.com/facebookresearch/moco-v3
+# --------------------------------------------------------
+def get_2d_sincos_pos_embed(embed_dim, grid_size, n_cls_token=0):
+    """
+    grid_size: int of the grid height and width
+    return:
+    pos_embed: [grid_size*grid_size, embed_dim] or [n_cls_token+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+    """
+    grid_h = np.arange(grid_size, dtype=np.float32)
+    grid_w = np.arange(grid_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, grid_size, grid_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if n_cls_token>0:
+        pos_embed = np.concatenate([np.zeros([n_cls_token, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=float)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out) # (M, D/2)
+    emb_cos = np.cos(out) # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+
+# --------------------------------------------------------
+# Interpolate position embeddings for high-resolution
+# References:
+# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py
+# DeiT: https://github.com/facebookresearch/deit
+# --------------------------------------------------------
+def interpolate_pos_embed(model, checkpoint_model):
+    if 'pos_embed' in checkpoint_model:
+        pos_embed_checkpoint = checkpoint_model['pos_embed']
+        embedding_size = pos_embed_checkpoint.shape[-1]
+        num_patches = model.patch_embed.num_patches
+        num_extra_tokens = model.pos_embed.shape[-2] - num_patches
+        # height (== width) for the checkpoint position embedding
+        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
+        # height (== width) for the new position embedding
+        new_size = int(num_patches ** 0.5)
+        # class_token and dist_token are kept unchanged
+        if orig_size != new_size:
+            print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
+            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
+            # only the position tokens are interpolated
+            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
+            pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
+            pos_tokens = torch.nn.functional.interpolate(
+                pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
+            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
+            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
+            checkpoint_model['pos_embed'] = new_pos_embed
+
+
+#----------------------------------------------------------
+# RoPE2D: RoPE implementation in 2D
+#----------------------------------------------------------
+
+try:
+    from models.curope import cuRoPE2D
+    RoPE2D = cuRoPE2D
+except ImportError:
+    print('Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead')
+
+    class RoPE2D(torch.nn.Module):
+        
+        def __init__(self, freq=100.0, F0=1.0):
+            super().__init__()
+            self.base = freq 
+            self.F0 = F0
+            self.cache = {}
+
+        def get_cos_sin(self, D, seq_len, device, dtype):
+            if (D,seq_len,device,dtype) not in self.cache:
+                inv_freq = 1.0 / (self.base ** (torch.arange(0, D, 2).float().to(device) / D))
+                t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)
+                freqs = torch.einsum("i,j->ij", t, inv_freq).to(dtype)
+                freqs = torch.cat((freqs, freqs), dim=-1)
+                cos = freqs.cos() # (Seq, Dim)
+                sin = freqs.sin()
+                self.cache[D,seq_len,device,dtype] = (cos,sin)
+            return self.cache[D,seq_len,device,dtype]
+            
+        @staticmethod
+        def rotate_half(x):
+            x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
+            return torch.cat((-x2, x1), dim=-1)
+            
+        def apply_rope1d(self, tokens, pos1d, cos, sin):
+            assert pos1d.ndim==2
+            cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :]
+            sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :]
+            return (tokens * cos) + (self.rotate_half(tokens) * sin)
+            
+        def forward(self, tokens, positions):
+            """
+            input:
+                * tokens: batch_size x nheads x ntokens x dim
+                * positions: batch_size x ntokens x 2 (y and x position of each token)
+            output:
+                * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim)
+            """
+            assert tokens.size(3)%2==0, "number of dimensions should be a multiple of two"
+            D = tokens.size(3) // 2
+            assert positions.ndim==3 and positions.shape[-1] == 2 # Batch, Seq, 2
+            cos, sin = self.get_cos_sin(D, int(positions.max())+1, tokens.device, tokens.dtype)
+            # split features into two along the feature dimension, and apply rope1d on each half
+            y, x = tokens.chunk(2, dim=-1)
+            y = self.apply_rope1d(y, positions[:,:,0], cos, sin)
+            x = self.apply_rope1d(x, positions[:,:,1], cos, sin)
+            tokens = torch.cat((y, x), dim=-1)
+            return tokens

third_party/dust3r/croco/pretrain.py

@@ -0,0 +1,254 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+# 
+# --------------------------------------------------------
+# Pre-training CroCo 
+# --------------------------------------------------------
+# References:
+# MAE: https://github.com/facebookresearch/mae
+# DeiT: https://github.com/facebookresearch/deit
+# BEiT: https://github.com/microsoft/unilm/tree/master/beit
+# --------------------------------------------------------
+import argparse
+import datetime
+import json
+import numpy as np
+import os
+import sys
+import time
+import math 
+from pathlib import Path
+from typing import Iterable
+
+import torch
+import torch.distributed as dist
+import torch.backends.cudnn as cudnn
+from torch.utils.tensorboard import SummaryWriter
+import torchvision.transforms as transforms
+import torchvision.datasets as datasets
+
+import utils.misc as misc
+from utils.misc import NativeScalerWithGradNormCount as NativeScaler
+from models.croco import CroCoNet
+from models.criterion import MaskedMSE
+from datasets.pairs_dataset import PairsDataset
+
+
+def get_args_parser():
+    parser = argparse.ArgumentParser('CroCo pre-training', add_help=False)
+    # model and criterion
+    parser.add_argument('--model', default='CroCoNet()', type=str, help="string containing the model to build")
+    parser.add_argument('--norm_pix_loss', default=1, choices=[0,1], help="apply per-patch mean/std normalization before applying the loss")
+    # dataset 
+    parser.add_argument('--dataset', default='habitat_release', type=str, help="training set")
+    parser.add_argument('--transforms', default='crop224+acolor', type=str, help="transforms to apply") # in the paper, we also use some homography and rotation, but find later that they were not useful or even harmful
+    # training 
+    parser.add_argument('--seed', default=0, type=int, help="Random seed")
+    parser.add_argument('--batch_size', default=64, type=int, help="Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus")
+    parser.add_argument('--epochs', default=800, type=int, help="Maximum number of epochs for the scheduler")
+    parser.add_argument('--max_epoch', default=400, type=int, help="Stop training at this epoch")
+    parser.add_argument('--accum_iter', default=1, type=int, help="Accumulate gradient iterations (for increasing the effective batch size under memory constraints)")
+    parser.add_argument('--weight_decay', type=float, default=0.05, help="weight decay (default: 0.05)")
+    parser.add_argument('--lr', type=float, default=None, metavar='LR', help='learning rate (absolute lr)')
+    parser.add_argument('--blr', type=float, default=1.5e-4, metavar='LR', help='base learning rate: absolute_lr = base_lr * total_batch_size / 256')
+    parser.add_argument('--min_lr', type=float, default=0., metavar='LR', help='lower lr bound for cyclic schedulers that hit 0')
+    parser.add_argument('--warmup_epochs', type=int, default=40, metavar='N', help='epochs to warmup LR')              
+    parser.add_argument('--amp', type=int, default=1, choices=[0,1], help="Use Automatic Mixed Precision for pretraining")
+    # others 
+    parser.add_argument('--num_workers', default=8, type=int)
+    parser.add_argument('--world_size', default=1, type=int, help='number of distributed processes')
+    parser.add_argument('--local_rank', default=-1, type=int)
+    parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training')
+    parser.add_argument('--save_freq', default=1, type=int, help='frequence (number of epochs) to save checkpoint in checkpoint-last.pth')
+    parser.add_argument('--keep_freq', default=20, type=int, help='frequence (number of epochs) to save checkpoint in checkpoint-%d.pth')
+    parser.add_argument('--print_freq', default=20, type=int, help='frequence (number of iterations) to print infos while training')
+    # paths 
+    parser.add_argument('--output_dir', default='./output/', type=str, help="path where to save the output")
+    parser.add_argument('--data_dir', default='./data/', type=str, help="path where data are stored")
+    return parser
+
+
+
+        
+def main(args):
+    misc.init_distributed_mode(args)
+    global_rank = misc.get_rank()
+    world_size = misc.get_world_size()
+    
+    print("output_dir: "+args.output_dir)
+    if args.output_dir:
+        Path(args.output_dir).mkdir(parents=True, exist_ok=True)                         
+
+    # auto resume 
+    last_ckpt_fname = os.path.join(args.output_dir, f'checkpoint-last.pth')
+    args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None
+
+    print('job dir: {}'.format(os.path.dirname(os.path.realpath(__file__))))
+    print("{}".format(args).replace(', ', ',\n'))
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    device = torch.device(device)
+
+    # fix the seed
+    seed = args.seed + misc.get_rank()
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+
+    cudnn.benchmark = True
+
+    ## training dataset and loader 
+    print('Building dataset for {:s} with transforms {:s}'.format(args.dataset, args.transforms))
+    dataset = PairsDataset(args.dataset, trfs=args.transforms, data_dir=args.data_dir)
+    if world_size>1:
+        sampler_train = torch.utils.data.DistributedSampler(
+            dataset, num_replicas=world_size, rank=global_rank, shuffle=True
+        )
+        print("Sampler_train = %s" % str(sampler_train))
+    else:
+        sampler_train = torch.utils.data.RandomSampler(dataset)
+    data_loader_train = torch.utils.data.DataLoader(
+        dataset, sampler=sampler_train,
+        batch_size=args.batch_size,
+        num_workers=args.num_workers,
+        pin_memory=True,
+        drop_last=True,
+    )
+   
+    ## model 
+    print('Loading model: {:s}'.format(args.model))
+    model = eval(args.model)
+    print('Loading criterion: MaskedMSE(norm_pix_loss={:s})'.format(str(bool(args.norm_pix_loss))))
+    criterion = MaskedMSE(norm_pix_loss=bool(args.norm_pix_loss))
+   
+    model.to(device)
+    model_without_ddp = model
+    print("Model = %s" % str(model_without_ddp))
+
+    eff_batch_size = args.batch_size * args.accum_iter * misc.get_world_size()
+    if args.lr is None:  # only base_lr is specified
+        args.lr = args.blr * eff_batch_size / 256
+    print("base lr: %.2e" % (args.lr * 256 / eff_batch_size))
+    print("actual lr: %.2e" % args.lr)
+    print("accumulate grad iterations: %d" % args.accum_iter)
+    print("effective batch size: %d" % eff_batch_size)
+
+    if args.distributed:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], find_unused_parameters=True, static_graph=True)
+        model_without_ddp = model.module
+    
+    param_groups = misc.get_parameter_groups(model_without_ddp, args.weight_decay) # following timm: set wd as 0 for bias and norm layers
+    optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))
+    print(optimizer)
+    loss_scaler = NativeScaler()
+
+    misc.load_model(args=args, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler)
+
+    if global_rank == 0 and args.output_dir is not None:
+        log_writer = SummaryWriter(log_dir=args.output_dir)
+    else:
+        log_writer = None
+
+    print(f"Start training until {args.max_epoch} epochs")
+    start_time = time.time()
+    for epoch in range(args.start_epoch, args.max_epoch):
+        if world_size>1:
+            data_loader_train.sampler.set_epoch(epoch)
+            
+        train_stats = train_one_epoch(
+            model, criterion, data_loader_train,
+            optimizer, device, epoch, loss_scaler,
+            log_writer=log_writer,
+            args=args
+        )
+        
+        if args.output_dir and epoch % args.save_freq == 0 :
+            misc.save_model(
+                args=args, model_without_ddp=model_without_ddp, optimizer=optimizer,
+                loss_scaler=loss_scaler, epoch=epoch, fname='last')
+                
+        if args.output_dir and (epoch % args.keep_freq == 0 or epoch + 1 == args.max_epoch) and (epoch>0 or args.max_epoch==1):
+            misc.save_model(
+                args=args, model_without_ddp=model_without_ddp, optimizer=optimizer,
+                loss_scaler=loss_scaler, epoch=epoch)
+
+        log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
+                        'epoch': epoch,}
+
+        if args.output_dir and misc.is_main_process():
+            if log_writer is not None:
+                log_writer.flush()
+            with open(os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8") as f:
+                f.write(json.dumps(log_stats) + "\n")
+
+    total_time = time.time() - start_time
+    total_time_str = str(datetime.timedelta(seconds=int(total_time)))
+    print('Training time {}'.format(total_time_str))
+
+
+
+
+def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module,
+                    data_loader: Iterable, optimizer: torch.optim.Optimizer,
+                    device: torch.device, epoch: int, loss_scaler,
+                    log_writer=None,
+                    args=None):
+    model.train(True)
+    metric_logger = misc.MetricLogger(delimiter="  ")
+    metric_logger.add_meter('lr', misc.SmoothedValue(window_size=1, fmt='{value:.6f}'))
+    header = 'Epoch: [{}]'.format(epoch)
+    accum_iter = args.accum_iter
+
+    optimizer.zero_grad()
+
+    if log_writer is not None:
+        print('log_dir: {}'.format(log_writer.log_dir))
+
+    for data_iter_step, (image1, image2) in enumerate(metric_logger.log_every(data_loader, args.print_freq, header)):
+
+        # we use a per iteration  lr scheduler
+        if data_iter_step % accum_iter == 0:
+            misc.adjust_learning_rate(optimizer, data_iter_step / len(data_loader) + epoch, args)
+
+        image1 = image1.to(device, non_blocking=True) 
+        image2 = image2.to(device, non_blocking=True)
+        with torch.cuda.amp.autocast(enabled=bool(args.amp)):
+            out, mask, target = model(image1, image2)
+            loss = criterion(out, mask, target)
+
+        loss_value = loss.item()
+
+        if not math.isfinite(loss_value):
+            print("Loss is {}, stopping training".format(loss_value))
+            sys.exit(1)
+
+        loss /= accum_iter
+        loss_scaler(loss, optimizer, parameters=model.parameters(),
+                    update_grad=(data_iter_step + 1) % accum_iter == 0)
+        if (data_iter_step + 1) % accum_iter == 0:
+            optimizer.zero_grad()
+
+        torch.cuda.synchronize()
+
+        metric_logger.update(loss=loss_value)
+
+        lr = optimizer.param_groups[0]["lr"]
+        metric_logger.update(lr=lr)
+
+        loss_value_reduce = misc.all_reduce_mean(loss_value)
+        if log_writer is not None and ((data_iter_step + 1) % (accum_iter*args.print_freq)) == 0:
+            # x-axis is based on epoch_1000x in the tensorboard, calibrating differences curves when batch size changes 
+            epoch_1000x = int((data_iter_step / len(data_loader) + epoch) * 1000)
+            log_writer.add_scalar('train_loss', loss_value_reduce, epoch_1000x)
+            log_writer.add_scalar('lr', lr, epoch_1000x)
+
+    # gather the stats from all processes
+    metric_logger.synchronize_between_processes()
+    print("Averaged stats:", metric_logger)
+    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
+    
+    
+
+if __name__ == '__main__':
+    args = get_args_parser()
+    args = args.parse_args()
+    main(args)

third_party/dust3r/croco/stereoflow/README.MD

@@ -0,0 +1,318 @@
+## CroCo-Stereo and CroCo-Flow
+
+This README explains how to use CroCo-Stereo and CroCo-Flow as well as how they were trained.
+All commands should be launched from the root directory.
+
+### Simple inference example
+
+We provide a simple inference exemple for CroCo-Stereo and CroCo-Flow in the Totebook `croco-stereo-flow-demo.ipynb`.
+Before running it, please download the trained models with:
+```
+bash stereoflow/download_model.sh crocostereo.pth
+bash stereoflow/download_model.sh crocoflow.pth
+```
+
+### Prepare data for training or evaluation
+
+Put the datasets used for training/evaluation in `./data/stereoflow` (or update the paths at the top of `stereoflow/datasets_stereo.py` and `stereoflow/datasets_flow.py`).
+Please find below on the file structure should look for each dataset:
+<details>
+<summary>FlyingChairs</summary>
+
+```
+./data/stereoflow/FlyingChairs/
+└───chairs_split.txt
+└───data/
+    └─── ...
+```
+</details>
+
+<details>
+<summary>MPI-Sintel</summary>
+
+```
+./data/stereoflow/MPI-Sintel/
+└───training/
+│   └───clean/
+│   └───final/
+│   └───flow/
+└───test/
+    └───clean/
+    └───final/
+```
+</details>
+
+<details>
+<summary>SceneFlow (including FlyingThings)</summary>
+
+```
+./data/stereoflow/SceneFlow/
+└───Driving/
+│   └───disparity/
+│   └───frames_cleanpass/
+│   └───frames_finalpass/
+└───FlyingThings/
+│   └───disparity/
+│   └───frames_cleanpass/
+│   └───frames_finalpass/
+│   └───optical_flow/
+└───Monkaa/
+    └───disparity/
+    └───frames_cleanpass/
+    └───frames_finalpass/
+```
+</details>
+
+<details>
+<summary>TartanAir</summary>
+
+```
+./data/stereoflow/TartanAir/
+└───abandonedfactory/
+│   └───.../
+└───abandonedfactory_night/
+│   └───.../
+└───.../
+```
+</details>
+
+<details>
+<summary>Booster</summary>
+
+```
+./data/stereoflow/booster_gt/
+└───train/
+    └───balanced/
+        └───Bathroom/
+        └───Bedroom/
+        └───...
+```
+</details>
+
+<details>
+<summary>CREStereo</summary>
+
+```
+./data/stereoflow/crenet_stereo_trainset/
+└───stereo_trainset/
+    └───crestereo/
+        └───hole/
+        └───reflective/
+        └───shapenet/
+        └───tree/
+```
+</details>
+
+<details>
+<summary>ETH3D Two-view Low-res</summary>
+
+```
+./data/stereoflow/eth3d_lowres/
+└───test/
+│   └───lakeside_1l/
+│   └───...
+└───train/
+│   └───delivery_area_1l/
+│   └───...
+└───train_gt/
+    └───delivery_area_1l/
+    └───...
+```
+</details>
+
+<details>
+<summary>KITTI 2012</summary>
+
+```
+./data/stereoflow/kitti-stereo-2012/
+└───testing/
+│   └───colored_0/
+│   └───colored_1/
+└───training/
+    └───colored_0/
+    └───colored_1/
+    └───disp_occ/
+    └───flow_occ/
+```
+</details>
+
+<details>
+<summary>KITTI 2015</summary>
+
+```
+./data/stereoflow/kitti-stereo-2015/
+└───testing/
+│   └───image_2/
+│   └───image_3/
+└───training/
+    └───image_2/
+    └───image_3/
+    └───disp_occ_0/
+    └───flow_occ/
+```
+</details>
+
+<details>
+<summary>Middlebury</summary>
+
+```
+./data/stereoflow/middlebury
+└───2005/
+│   └───train/
+│       └───Art/
+│       └───...
+└───2006/
+│   └───Aloe/
+│   └───Baby1/
+│   └───...
+└───2014/
+│   └───Adirondack-imperfect/
+│   └───Adirondack-perfect/
+│   └───...
+└───2021/
+│   └───data/
+│       └───artroom1/
+│       └───artroom2/
+│       └───...
+└───MiddEval3_F/
+    └───test/
+    │   └───Australia/
+    │   └───...
+    └───train/
+        └───Adirondack/
+        └───...
+```
+</details>
+
+<details>
+<summary>Spring</summary>
+
+```
+./data/stereoflow/spring/
+└───test/
+│   └───0003/
+│   └───...
+└───train/
+    └───0001/
+    └───...
+```
+</details>
+
+
+### CroCo-Stereo
+
+##### Main model 
+
+The main training of CroCo-Stereo was performed on a series of datasets, and it was used as it for Middlebury v3 benchmark.
+
+```
+# Download the model
+bash stereoflow/download_model.sh crocostereo.pth
+# Middlebury v3 submission
+python stereoflow/test.py --model stereoflow_models/crocostereo.pth --dataset "MdEval3('all_full')" --save submission --tile_overlap 0.9
+# Training command that was used, using checkpoint-last.pth
+python -u stereoflow/train.py stereo --criterion "LaplacianLossBounded2()" --dataset "CREStereo('train')+SceneFlow('train_allpass')+30*ETH3DLowRes('train')+50*Md05('train')+50*Md06('train')+50*Md14('train')+50*Md21('train')+50*MdEval3('train_full')+Booster('train_balanced')" --val_dataset "SceneFlow('test1of100_finalpass')+SceneFlow('test1of100_cleanpass')+ETH3DLowRes('subval')+Md05('subval')+Md06('subval')+Md14('subval')+Md21('subval')+MdEval3('subval_full')+Booster('subval_balanced')" --lr 3e-5 --batch_size 6 --epochs 32 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --output_dir xps/crocostereo/main/
+# or it can be launched on multiple gpus (while maintaining the effective batch size), e.g. on 3 gpus:
+torchrun --nproc_per_node 3 stereoflow/train.py stereo --criterion "LaplacianLossBounded2()" --dataset "CREStereo('train')+SceneFlow('train_allpass')+30*ETH3DLowRes('train')+50*Md05('train')+50*Md06('train')+50*Md14('train')+50*Md21('train')+50*MdEval3('train_full')+Booster('train_balanced')" --val_dataset "SceneFlow('test1of100_finalpass')+SceneFlow('test1of100_cleanpass')+ETH3DLowRes('subval')+Md05('subval')+Md06('subval')+Md14('subval')+Md21('subval')+MdEval3('subval_full')+Booster('subval_balanced')" --lr 3e-5 --batch_size 2 --epochs 32 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --output_dir xps/crocostereo/main/
+```
+
+For evaluation of validation set, we also provide the model trained on the `subtrain` subset of the training sets.
+
+```
+# Download the model
+bash stereoflow/download_model.sh crocostereo_subtrain.pth
+# Evaluation on validation sets 
+python stereoflow/test.py --model stereoflow_models/crocostereo_subtrain.pth --dataset "MdEval3('subval_full')+ETH3DLowRes('subval')+SceneFlow('test_finalpass')+SceneFlow('test_cleanpass')" --save metrics --tile_overlap 0.9
+# Training command that was used (same as above but on subtrain, using checkpoint-best.pth), can also be launched on multiple gpus
+python -u stereoflow/train.py stereo --criterion "LaplacianLossBounded2()" --dataset "CREStereo('train')+SceneFlow('train_allpass')+30*ETH3DLowRes('subtrain')+50*Md05('subtrain')+50*Md06('subtrain')+50*Md14('subtrain')+50*Md21('subtrain')+50*MdEval3('subtrain_full')+Booster('subtrain_balanced')" --val_dataset "SceneFlow('test1of100_finalpass')+SceneFlow('test1of100_cleanpass')+ETH3DLowRes('subval')+Md05('subval')+Md06('subval')+Md14('subval')+Md21('subval')+MdEval3('subval_full')+Booster('subval_balanced')" --lr 3e-5 --batch_size 6 --epochs 32 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --output_dir xps/crocostereo/main_subtrain/
+```
+
+##### Other models 
+
+<details>
+	<summary>Model for ETH3D</summary> 
+	The model used for the submission on ETH3D is trained with the same command but using an unbounded Laplacian loss.
+	
+	# Download the model
+	bash stereoflow/download_model.sh crocostereo_eth3d.pth
+	# ETH3D submission
+	python stereoflow/test.py --model stereoflow_models/crocostereo_eth3d.pth --dataset "ETH3DLowRes('all')" --save submission --tile_overlap 0.9
+	# Training command that was used
+	python -u stereoflow/train.py stereo --criterion "LaplacianLoss()" --tile_conf_mode conf_expbeta3 --dataset "CREStereo('train')+SceneFlow('train_allpass')+30*ETH3DLowRes('train')+50*Md05('train')+50*Md06('train')+50*Md14('train')+50*Md21('train')+50*MdEval3('train_full')+Booster('train_balanced')" --val_dataset "SceneFlow('test1of100_finalpass')+SceneFlow('test1of100_cleanpass')+ETH3DLowRes('subval')+Md05('subval')+Md06('subval')+Md14('subval')+Md21('subval')+MdEval3('subval_full')+Booster('subval_balanced')" --lr 3e-5 --batch_size 6 --epochs 32 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --output_dir xps/crocostereo/main_eth3d/
+	
+</details>
+
+<details>
+	<summary>Main model finetuned on Kitti</summary>
+	
+	# Download the model
+	bash stereoflow/download_model.sh crocostereo_finetune_kitti.pth
+	# Kitti submission 
+	python stereoflow/test.py --model stereoflow_models/crocostereo_finetune_kitti.pth --dataset "Kitti15('test')" --save submission --tile_overlap 0.9
+	# Training that was used
+	python -u stereoflow/train.py stereo --crop 352 1216 --criterion "LaplacianLossBounded2()" --dataset "Kitti12('train')+Kitti15('train')" --lr 3e-5 --batch_size 1 --accum_iter 6 --epochs 20 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --start_from stereoflow_models/crocostereo.pth --output_dir xps/crocostereo/finetune_kitti/ --save_every 5
+</details>
+
+<details>
+	<summary>Main model finetuned on Spring</summary>
+	
+	# Download the model
+	bash stereoflow/download_model.sh crocostereo_finetune_spring.pth
+	# Spring submission 
+	python stereoflow/test.py --model stereoflow_models/crocostereo_finetune_spring.pth --dataset "Spring('test')" --save submission --tile_overlap 0.9
+	# Training command that was used
+	python -u stereoflow/train.py stereo --criterion "LaplacianLossBounded2()" --dataset "Spring('train')" --lr 3e-5 --batch_size 6 --epochs 8 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --start_from stereoflow_models/crocostereo.pth --output_dir xps/crocostereo/finetune_spring/
+</details>
+
+<details>
+	<summary>Smaller models</summary>
+	To train CroCo-Stereo with smaller CroCo pretrained models, simply replace the <code>--pretrained</code> argument. To download the smaller CroCo-Stereo models based on CroCo v2 pretraining with ViT-Base encoder and Small encoder, use <code>bash stereoflow/download_model.sh crocostereo_subtrain_vitb_smalldecoder.pth</code>, and for the model with a ViT-Base encoder and a Base decoder, use <code>bash stereoflow/download_model.sh crocostereo_subtrain_vitb_basedecoder.pth</code>.
+</details>
+	
+
+### CroCo-Flow
+
+##### Main model
+
+The main training of CroCo-Flow was performed on the FlyingThings, FlyingChairs, MPI-Sintel and TartanAir datasets.
+It was used for our submission to the MPI-Sintel benchmark.
+
+```
+# Download the model 
+bash stereoflow/download_model.sh crocoflow.pth
+# Evaluation 
+python stereoflow/test.py --model stereoflow_models/crocoflow.pth --dataset "MPISintel('subval_cleanpass')+MPISintel('subval_finalpass')" --save metrics --tile_overlap 0.9
+# Sintel submission
+python stereoflow/test.py --model stereoflow_models/crocoflow.pth --dataset "MPISintel('test_allpass')" --save submission --tile_overlap 0.9
+# Training command that was used, with checkpoint-best.pth
+python -u stereoflow/train.py flow --criterion "LaplacianLossBounded()" --dataset "40*MPISintel('subtrain_cleanpass')+40*MPISintel('subtrain_finalpass')+4*FlyingThings('train_allpass')+4*FlyingChairs('train')+TartanAir('train')" --val_dataset "MPISintel('subval_cleanpass')+MPISintel('subval_finalpass')" --lr 2e-5 --batch_size 8 --epochs 240 --img_per_epoch 30000 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --output_dir xps/crocoflow/main/
+```
+
+##### Other models 
+
+<details>
+	<summary>Main model finetuned on Kitti</summary>
+	
+	# Download the model
+	bash stereoflow/download_model.sh crocoflow_finetune_kitti.pth
+	# Kitti submission 
+	python stereoflow/test.py --model stereoflow_models/crocoflow_finetune_kitti.pth --dataset "Kitti15('test')" --save submission --tile_overlap 0.99
+	# Training that was used, with checkpoint-last.pth
+	python -u stereoflow/train.py flow --crop 352 1216 --criterion "LaplacianLossBounded()" --dataset "Kitti15('train')+Kitti12('train')" --lr 2e-5 --batch_size 1 --accum_iter 8 --epochs 150 --save_every 5 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --start_from stereoflow_models/crocoflow.pth --output_dir xps/crocoflow/finetune_kitti/
+</details>
+
+<details>
+	<summary>Main model finetuned on Spring</summary>
+	
+	# Download the model
+	bash stereoflow/download_model.sh crocoflow_finetune_spring.pth
+	# Spring submission 
+	python stereoflow/test.py --model stereoflow_models/crocoflow_finetune_spring.pth --dataset "Spring('test')" --save submission --tile_overlap 0.9
+	# Training command that was used, with checkpoint-last.pth
+	python -u stereoflow/train.py flow --criterion "LaplacianLossBounded()" --dataset "Spring('train')" --lr 2e-5 --batch_size 8 --epochs 12 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --start_from stereoflow_models/crocoflow.pth --output_dir xps/crocoflow/finetune_spring/
+</details>
+
+<details>
+	<summary>Smaller models</summary>
+	To train CroCo-Flow with smaller CroCo pretrained models, simply replace the <code>--pretrained</code> argument. To download the smaller CroCo-Flow models based on CroCo v2 pretraining with ViT-Base encoder and Small encoder, use <code>bash stereoflow/download_model.sh crocoflow_vitb_smalldecoder.pth</code>, and for the model with a ViT-Base encoder and a Base decoder, use <code>bash stereoflow/download_model.sh crocoflow_vitb_basedecoder.pth</code>.
+</details>

third_party/dust3r/croco/stereoflow/augmentor.py

@@ -0,0 +1,290 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+# --------------------------------------------------------
+# Data augmentation for training stereo and flow
+# --------------------------------------------------------
+
+# References
+# https://github.com/autonomousvision/unimatch/blob/master/dataloader/stereo/transforms.py
+# https://github.com/autonomousvision/unimatch/blob/master/dataloader/flow/transforms.py
+
+
+import numpy as np
+import random
+from PIL import Image
+
+import cv2
+cv2.setNumThreads(0)
+cv2.ocl.setUseOpenCL(False)
+
+import torch
+from torchvision.transforms import ColorJitter
+import torchvision.transforms.functional as FF
+
+class StereoAugmentor(object):
+
+    def __init__(self, crop_size, scale_prob=0.5, scale_xonly=True, lhth=800., lminscale=0.0, lmaxscale=1.0, hminscale=-0.2, hmaxscale=0.4, scale_interp_nearest=True, rightjitterprob=0.5, v_flip_prob=0.5, color_aug_asym=True, color_choice_prob=0.5):
+        self.crop_size = crop_size
+        self.scale_prob = scale_prob
+        self.scale_xonly = scale_xonly
+        self.lhth = lhth
+        self.lminscale = lminscale
+        self.lmaxscale = lmaxscale
+        self.hminscale = hminscale
+        self.hmaxscale = hmaxscale
+        self.scale_interp_nearest = scale_interp_nearest
+        self.rightjitterprob = rightjitterprob
+        self.v_flip_prob = v_flip_prob
+        self.color_aug_asym = color_aug_asym
+        self.color_choice_prob = color_choice_prob
+        
+    def _random_scale(self, img1, img2, disp):
+        ch,cw = self.crop_size
+        h,w = img1.shape[:2]
+        if self.scale_prob>0. and np.random.rand()<self.scale_prob:
+            min_scale, max_scale = (self.lminscale,self.lmaxscale) if min(h,w) < self.lhth else (self.hminscale,self.hmaxscale)
+            scale_x = 2. ** np.random.uniform(min_scale, max_scale)
+            scale_x = np.clip(scale_x, (cw+8) / float(w), None)
+            scale_y = 1.
+            if not self.scale_xonly:
+                scale_y = scale_x
+                scale_y = np.clip(scale_y, (ch+8) / float(h), None)
+            img1 = cv2.resize(img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR)
+            img2 = cv2.resize(img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR)
+            disp = cv2.resize(disp, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR if not self.scale_interp_nearest else cv2.INTER_NEAREST) * scale_x
+        else: # check if we need to resize to be able to crop 
+            h,w = img1.shape[:2]
+            clip_scale = (cw+8) / float(w)
+            if clip_scale>1.:
+                scale_x = clip_scale
+                scale_y = scale_x if not self.scale_xonly else 1.0
+                img1 = cv2.resize(img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR)
+                img2 = cv2.resize(img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR)
+                disp = cv2.resize(disp, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR if not self.scale_interp_nearest else cv2.INTER_NEAREST) * scale_x
+        return img1, img2, disp 
+                
+    def _random_crop(self, img1, img2, disp): 
+        h,w = img1.shape[:2]
+        ch,cw = self.crop_size
+        assert ch<=h and cw<=w, (img1.shape, h,w,ch,cw)
+        offset_x = np.random.randint(w - cw + 1)
+        offset_y = np.random.randint(h - ch + 1)
+        img1 = img1[offset_y:offset_y+ch,offset_x:offset_x+cw]
+        img2 = img2[offset_y:offset_y+ch,offset_x:offset_x+cw]
+        disp = disp[offset_y:offset_y+ch,offset_x:offset_x+cw]
+        return img1, img2, disp
+    
+    def _random_vflip(self, img1, img2, disp):
+        # vertical flip
+        if self.v_flip_prob>0 and np.random.rand() < self.v_flip_prob:
+            img1 = np.copy(np.flipud(img1))
+            img2 = np.copy(np.flipud(img2))
+            disp = np.copy(np.flipud(disp))
+        return img1, img2, disp
+        
+    def _random_rotate_shift_right(self, img2):
+        if self.rightjitterprob>0. and np.random.rand()<self.rightjitterprob:
+            angle, pixel = 0.1, 2
+            px = np.random.uniform(-pixel, pixel)
+            ag = np.random.uniform(-angle, angle)
+            image_center = (np.random.uniform(0, img2.shape[0]), np.random.uniform(0, img2.shape[1])  )
+            rot_mat = cv2.getRotationMatrix2D(image_center, ag, 1.0)
+            img2 = cv2.warpAffine(img2, rot_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR)
+            trans_mat = np.float32([[1, 0, 0], [0, 1, px]])
+            img2 = cv2.warpAffine(img2, trans_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR)
+        return img2
+            
+    def _random_color_contrast(self, img1, img2):
+        if np.random.random() < 0.5:
+            contrast_factor = np.random.uniform(0.8, 1.2)
+            img1 = FF.adjust_contrast(img1, contrast_factor)
+            if self.color_aug_asym and np.random.random() < 0.5: contrast_factor = np.random.uniform(0.8, 1.2)
+            img2 = FF.adjust_contrast(img2, contrast_factor)
+        return img1, img2
+    def _random_color_gamma(self, img1, img2):
+        if np.random.random() < 0.5:
+            gamma = np.random.uniform(0.7, 1.5)
+            img1 = FF.adjust_gamma(img1, gamma)
+            if self.color_aug_asym and np.random.random() < 0.5: gamma = np.random.uniform(0.7, 1.5)
+            img2 = FF.adjust_gamma(img2, gamma)
+        return img1, img2
+    def _random_color_brightness(self, img1, img2):
+        if np.random.random() < 0.5:
+            brightness = np.random.uniform(0.5, 2.0)
+            img1 = FF.adjust_brightness(img1, brightness)
+            if self.color_aug_asym and np.random.random() < 0.5: brightness = np.random.uniform(0.5, 2.0)
+            img2 = FF.adjust_brightness(img2, brightness)
+        return img1, img2
+    def _random_color_hue(self, img1, img2):
+        if np.random.random() < 0.5:
+            hue = np.random.uniform(-0.1, 0.1)
+            img1 = FF.adjust_hue(img1, hue)
+            if self.color_aug_asym and np.random.random() < 0.5: hue = np.random.uniform(-0.1, 0.1)
+            img2 = FF.adjust_hue(img2, hue)
+        return img1, img2
+    def _random_color_saturation(self, img1, img2):
+        if np.random.random() < 0.5:
+            saturation = np.random.uniform(0.8, 1.2)
+            img1 = FF.adjust_saturation(img1, saturation)
+            if self.color_aug_asym and np.random.random() < 0.5: saturation = np.random.uniform(-0.8,1.2)
+            img2 = FF.adjust_saturation(img2, saturation)
+        return img1, img2   
+    def _random_color(self, img1, img2):
+        trfs = [self._random_color_contrast,self._random_color_gamma,self._random_color_brightness,self._random_color_hue,self._random_color_saturation]
+        img1 = Image.fromarray(img1.astype('uint8'))
+        img2 = Image.fromarray(img2.astype('uint8'))
+        if np.random.random() < self.color_choice_prob:
+            # A single transform
+            t = random.choice(trfs)
+            img1, img2 = t(img1, img2)
+        else:
+            # Combination of trfs
+            # Random order
+            random.shuffle(trfs)
+            for t in trfs:
+                img1, img2 = t(img1, img2)
+        img1 = np.array(img1).astype(np.float32)
+        img2 = np.array(img2).astype(np.float32)
+        return img1, img2
+                    
+    def __call__(self, img1, img2, disp, dataset_name):
+        img1, img2, disp = self._random_scale(img1, img2, disp)
+        img1, img2, disp = self._random_crop(img1, img2, disp)
+        img1, img2, disp = self._random_vflip(img1, img2, disp)
+        img2 = self._random_rotate_shift_right(img2)
+        img1, img2 = self._random_color(img1, img2)
+        return img1, img2, disp
+
+
+
+class FlowAugmentor:
+
+    def __init__(self, crop_size, min_scale=-0.2, max_scale=0.5, spatial_aug_prob=0.8, stretch_prob=0.8, max_stretch=0.2, h_flip_prob=0.5, v_flip_prob=0.1, asymmetric_color_aug_prob=0.2):
+    
+        # spatial augmentation params
+        self.crop_size = crop_size
+        self.min_scale = min_scale
+        self.max_scale = max_scale
+        self.spatial_aug_prob = spatial_aug_prob
+        self.stretch_prob = stretch_prob
+        self.max_stretch = max_stretch
+
+        # flip augmentation params
+        self.h_flip_prob = h_flip_prob
+        self.v_flip_prob = v_flip_prob
+
+        # photometric augmentation params
+        self.photo_aug = ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4, hue=0.5 / 3.14)
+
+        self.asymmetric_color_aug_prob = asymmetric_color_aug_prob
+        
+    def color_transform(self, img1, img2):
+        """ Photometric augmentation """
+
+        # asymmetric
+        if np.random.rand() < self.asymmetric_color_aug_prob:
+            img1 = np.array(self.photo_aug(Image.fromarray(img1)), dtype=np.uint8)
+            img2 = np.array(self.photo_aug(Image.fromarray(img2)), dtype=np.uint8)
+
+        # symmetric
+        else:
+            image_stack = np.concatenate([img1, img2], axis=0)
+            image_stack = np.array(self.photo_aug(Image.fromarray(image_stack)), dtype=np.uint8)
+            img1, img2 = np.split(image_stack, 2, axis=0)
+
+        return img1, img2
+
+    def _resize_flow(self, flow, scale_x, scale_y, factor=1.0):
+        if np.all(np.isfinite(flow)):
+            flow = cv2.resize(flow, None, fx=scale_x/factor, fy=scale_y/factor, interpolation=cv2.INTER_LINEAR)
+            flow = flow * [scale_x, scale_y]
+        else: # sparse version
+            fx, fy = scale_x, scale_y
+            ht, wd = flow.shape[:2]
+            coords = np.meshgrid(np.arange(wd), np.arange(ht))
+            coords = np.stack(coords, axis=-1)
+
+            coords = coords.reshape(-1, 2).astype(np.float32)
+            flow = flow.reshape(-1, 2).astype(np.float32)
+            valid = np.isfinite(flow[:,0])
+
+            coords0 = coords[valid]
+            flow0 = flow[valid]
+
+            ht1 = int(round(ht * fy/factor))
+            wd1 = int(round(wd * fx/factor))
+            
+            rescale = np.expand_dims(np.array([fx, fy]), axis=0)
+            coords1 = coords0 * rescale / factor
+            flow1 = flow0 * rescale
+
+            xx = np.round(coords1[:, 0]).astype(np.int32)
+            yy = np.round(coords1[:, 1]).astype(np.int32)
+
+            v = (xx > 0) & (xx < wd1) & (yy > 0) & (yy < ht1)
+            xx = xx[v]
+            yy = yy[v]
+            flow1 = flow1[v]
+
+            flow = np.inf * np.ones([ht1, wd1, 2], dtype=np.float32) # invalid value every where, before we fill it with the correct ones
+            flow[yy, xx] = flow1
+        return flow
+        
+    def spatial_transform(self, img1, img2, flow, dname):
+    
+        if np.random.rand() < self.spatial_aug_prob:
+            # randomly sample scale
+            ht, wd = img1.shape[:2]
+            clip_min_scale = np.maximum(
+                (self.crop_size[0] + 8) / float(ht),
+                (self.crop_size[1] + 8) / float(wd))
+            min_scale, max_scale = self.min_scale, self.max_scale
+            scale = 2 ** np.random.uniform(self.min_scale, self.max_scale)
+            scale_x = scale
+            scale_y = scale
+            if np.random.rand() < self.stretch_prob:
+                scale_x *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch)
+                scale_y *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch)
+            scale_x = np.clip(scale_x, clip_min_scale, None)
+            scale_y = np.clip(scale_y, clip_min_scale, None)
+            # rescale the images
+            img1 = cv2.resize(img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR)
+            img2 = cv2.resize(img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR)
+            flow = self._resize_flow(flow, scale_x, scale_y, factor=2.0 if dname=='Spring' else 1.0)
+        elif dname=="Spring":
+            flow = self._resize_flow(flow, 1.0, 1.0, factor=2.0)
+
+        if self.h_flip_prob>0. and np.random.rand() < self.h_flip_prob:  # h-flip
+            img1 = img1[:, ::-1]
+            img2 = img2[:, ::-1]
+            flow = flow[:, ::-1] * [-1.0, 1.0]
+
+        if self.v_flip_prob>0. and np.random.rand() < self.v_flip_prob:  # v-flip
+            img1 = img1[::-1, :]
+            img2 = img2[::-1, :]
+            flow = flow[::-1, :] * [1.0, -1.0]
+                
+        # In case no cropping
+        if img1.shape[0] - self.crop_size[0] > 0:
+            y0 = np.random.randint(0, img1.shape[0] - self.crop_size[0])
+        else:
+            y0 = 0
+        if img1.shape[1] - self.crop_size[1] > 0:
+            x0 = np.random.randint(0, img1.shape[1] - self.crop_size[1])
+        else:
+            x0 = 0
+
+        img1 = img1[y0:y0 + self.crop_size[0], x0:x0 + self.crop_size[1]]
+        img2 = img2[y0:y0 + self.crop_size[0], x0:x0 + self.crop_size[1]]
+        flow = flow[y0:y0 + self.crop_size[0], x0:x0 + self.crop_size[1]]
+
+        return img1, img2, flow
+
+    def __call__(self, img1, img2, flow, dname):
+        img1, img2, flow = self.spatial_transform(img1, img2, flow, dname)
+        img1, img2 = self.color_transform(img1, img2)
+        img1 = np.ascontiguousarray(img1)
+        img2 = np.ascontiguousarray(img2)
+        flow = np.ascontiguousarray(flow)
+        return img1, img2, flow

third_party/dust3r/croco/stereoflow/criterion.py

@@ -0,0 +1,251 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+# --------------------------------------------------------
+# Losses, metrics per batch, metrics per dataset 
+# --------------------------------------------------------
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+def _get_gtnorm(gt):
+    if gt.size(1)==1: # stereo
+        return gt
+    # flow 
+    return torch.sqrt(torch.sum(gt**2, dim=1, keepdims=True)) # Bx1xHxW
+
+############ losses without confidence
+
+class L1Loss(nn.Module):
+    
+    def __init__(self, max_gtnorm=None):
+        super().__init__()
+        self.max_gtnorm = max_gtnorm
+        self.with_conf = False 
+    
+    def _error(self, gt, predictions):
+        return torch.abs(gt-predictions)
+    
+    def forward(self, predictions, gt, inspect=False):
+        mask = torch.isfinite(gt)
+        if self.max_gtnorm is not None: 
+            mask *= _get_gtnorm(gt).expand(-1,gt.size(1),-1,-1)<self.max_gtnorm
+        if inspect:
+            return self._error(gt, predictions)
+        return self._error(gt[mask],predictions[mask]).mean()
+
+############## losses with confience
+## there are several parametrizations
+
+class LaplacianLoss(nn.Module): # used for CroCo-Stereo on ETH3D, d'=exp(d)
+    
+    def __init__(self, max_gtnorm=None):
+        super().__init__()
+        self.max_gtnorm = max_gtnorm
+        self.with_conf = True
+        
+    def forward(self, predictions, gt, conf):
+        mask = torch.isfinite(gt)
+        mask = mask[:,0,:,:]
+        if self.max_gtnorm is not None: mask *= _get_gtnorm(gt)[:,0,:,:]<self.max_gtnorm
+        conf = conf.squeeze(1)
+        return ( torch.abs(gt-predictions).sum(dim=1)[mask] / torch.exp(conf[mask]) + conf[mask] ).mean()# + torch.log(2) => which is a constant
+
+
+class LaplacianLossBounded(nn.Module): # used for CroCo-Flow ; in the equation of the paper, we have a=1/b
+    def __init__(self, max_gtnorm=10000., a=0.25, b=4.):
+        super().__init__()
+        self.max_gtnorm = max_gtnorm
+        self.with_conf = True
+        self.a, self.b = a, b
+        
+    def forward(self, predictions, gt, conf):
+        mask = torch.isfinite(gt)
+        mask = mask[:,0,:,:]
+        if self.max_gtnorm is not None: mask *= _get_gtnorm(gt)[:,0,:,:]<self.max_gtnorm
+        conf = conf.squeeze(1)
+        conf = (self.b - self.a) * torch.sigmoid(conf) + self.a
+        return ( torch.abs(gt-predictions).sum(dim=1)[mask] / conf[mask] + torch.log(conf)[mask] ).mean()# + torch.log(2) => which is a constant
+
+class LaplacianLossBounded2(nn.Module): # used for CroCo-Stereo (except for ETH3D) ; in the equation of the paper, we have a=b
+    def __init__(self, max_gtnorm=None, a=3.0, b=3.0):
+        super().__init__()
+        self.max_gtnorm = max_gtnorm
+        self.with_conf = True
+        self.a, self.b = a, b
+        
+    def forward(self, predictions, gt, conf):
+        mask = torch.isfinite(gt)
+        mask = mask[:,0,:,:]
+        if self.max_gtnorm is not None: mask *= _get_gtnorm(gt)[:,0,:,:]<self.max_gtnorm
+        conf = conf.squeeze(1)
+        conf = 2 * self.a * (torch.sigmoid(conf / self.b) - 0.5 )
+        return ( torch.abs(gt-predictions).sum(dim=1)[mask] / torch.exp(conf[mask]) + conf[mask] ).mean()# + torch.log(2) => which is a constant
+        
+############## metrics per batch 
+
+class StereoMetrics(nn.Module):
+
+    def __init__(self, do_quantile=False):
+        super().__init__()
+        self.bad_ths = [0.5,1,2,3]
+        self.do_quantile = do_quantile
+        
+    def forward(self, predictions, gt):
+        B = predictions.size(0)
+        metrics = {}
+        gtcopy = gt.clone() 
+        mask = torch.isfinite(gtcopy)
+        gtcopy[~mask] = 999999.0 # we make a copy and put a non-infinite value, such that it does not become nan once multiplied by the mask value 0
+        Npx = mask.view(B,-1).sum(dim=1)
+        L1error = (torch.abs(gtcopy-predictions)*mask).view(B,-1)
+        L2error = (torch.square(gtcopy-predictions)*mask).view(B,-1)
+        # avgerr
+        metrics['avgerr'] = torch.mean(L1error.sum(dim=1)/Npx )
+        # rmse
+        metrics['rmse'] = torch.sqrt(L2error.sum(dim=1)/Npx).mean(dim=0)
+        # err > t for t in [0.5,1,2,3]
+        for ths in self.bad_ths:
+            metrics['bad@{:.1f}'.format(ths)] = (((L1error>ths)* mask.view(B,-1)).sum(dim=1)/Npx).mean(dim=0) * 100
+        return metrics
+        
+class FlowMetrics(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.bad_ths = [1,3,5]
+        
+    def forward(self, predictions, gt):
+        B = predictions.size(0)        
+        metrics = {}
+        mask = torch.isfinite(gt[:,0,:,:]) # both x and y would be infinite
+        Npx = mask.view(B,-1).sum(dim=1)
+        gtcopy = gt.clone() # to compute L1/L2 error, we need to have non-infinite value, the error computed at this locations will be ignored
+        gtcopy[:,0,:,:][~mask] = 999999.0
+        gtcopy[:,1,:,:][~mask] = 999999.0
+        L1error = (torch.abs(gtcopy-predictions).sum(dim=1)*mask).view(B,-1)
+        L2error = (torch.sqrt(torch.sum(torch.square(gtcopy-predictions),dim=1))*mask).view(B,-1)
+        metrics['L1err'] = torch.mean(L1error.sum(dim=1)/Npx )
+        metrics['EPE'] = torch.mean(L2error.sum(dim=1)/Npx )
+        for ths in self.bad_ths:
+            metrics['bad@{:.1f}'.format(ths)] = (((L2error>ths)* mask.view(B,-1)).sum(dim=1)/Npx).mean(dim=0) * 100
+        return metrics
+        
+############## metrics per dataset
+## we update the average and maintain the number of pixels while adding data batch per batch 
+## at the beggining, call reset()
+## after each batch, call add_batch(...)
+## at the end: call get_results()
+
+class StereoDatasetMetrics(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+        self.bad_ths = [0.5,1,2,3]
+        
+    def reset(self):
+        self.agg_N = 0 # number of pixels so far 
+        self.agg_L1err = torch.tensor(0.0) # L1 error so far 
+        self.agg_Nbad = [0 for _ in self.bad_ths] # counter of bad pixels 
+        self._metrics = None
+                
+    def add_batch(self, predictions, gt):
+        assert predictions.size(1)==1, predictions.size()
+        assert gt.size(1)==1, gt.size()
+        if gt.size(2)==predictions.size(2)*2 and gt.size(3)==predictions.size(3)*2: # special case for Spring ...
+            L1err = torch.minimum( torch.minimum( torch.minimum(
+                torch.sum(torch.abs(gt[:,:,0::2,0::2]-predictions),dim=1),
+                torch.sum(torch.abs(gt[:,:,1::2,0::2]-predictions),dim=1)),
+                torch.sum(torch.abs(gt[:,:,0::2,1::2]-predictions),dim=1)),
+                torch.sum(torch.abs(gt[:,:,1::2,1::2]-predictions),dim=1))
+            valid = torch.isfinite(L1err)
+        else:
+            valid = torch.isfinite(gt[:,0,:,:]) # both x and y would be infinite
+            L1err = torch.sum(torch.abs(gt-predictions),dim=1)
+        N = valid.sum()
+        Nnew = self.agg_N + N
+        self.agg_L1err = float(self.agg_N)/Nnew * self.agg_L1err + L1err[valid].mean().cpu() * float(N)/Nnew
+        self.agg_N = Nnew
+        for i,th in enumerate(self.bad_ths):
+            self.agg_Nbad[i] += (L1err[valid]>th).sum().cpu()
+   
+    def _compute_metrics(self):
+        if self._metrics is not None: return
+        out = {}
+        out['L1err'] = self.agg_L1err.item()
+        for i,th in enumerate(self.bad_ths):
+            out['bad@{:.1f}'.format(th)] = (float(self.agg_Nbad[i]) / self.agg_N).item() * 100.0
+        self._metrics = out
+        
+    def get_results(self): 
+        self._compute_metrics() # to avoid recompute them multiple times
+        return self._metrics
+
+class FlowDatasetMetrics(nn.Module):
+    
+    def __init__(self):
+        super().__init__()
+        self.bad_ths = [0.5,1,3,5]
+        self.speed_ths = [(0,10),(10,40),(40,torch.inf)]
+    
+    def reset(self):
+        self.agg_N = 0 # number of pixels so far 
+        self.agg_L1err = torch.tensor(0.0) # L1 error so far 
+        self.agg_L2err = torch.tensor(0.0) # L2 (=EPE) error so far 
+        self.agg_Nbad = [0 for _ in self.bad_ths] # counter of bad pixels 
+        self.agg_EPEspeed = [torch.tensor(0.0) for _ in self.speed_ths] # EPE per speed bin so far 
+        self.agg_Nspeed = [0 for _ in self.speed_ths] # N pixels per speed bin so far
+        self._metrics = None
+        self.pairname_results = {}
+
+    def add_batch(self, predictions, gt):
+        assert predictions.size(1)==2, predictions.size()
+        assert gt.size(1)==2, gt.size()
+        if gt.size(2)==predictions.size(2)*2 and gt.size(3)==predictions.size(3)*2: # special case for Spring ...
+            L1err = torch.minimum( torch.minimum( torch.minimum(
+                torch.sum(torch.abs(gt[:,:,0::2,0::2]-predictions),dim=1),
+                torch.sum(torch.abs(gt[:,:,1::2,0::2]-predictions),dim=1)),
+                torch.sum(torch.abs(gt[:,:,0::2,1::2]-predictions),dim=1)),
+                torch.sum(torch.abs(gt[:,:,1::2,1::2]-predictions),dim=1))
+            L2err = torch.minimum( torch.minimum( torch.minimum(
+                torch.sqrt(torch.sum(torch.square(gt[:,:,0::2,0::2]-predictions),dim=1)),
+                torch.sqrt(torch.sum(torch.square(gt[:,:,1::2,0::2]-predictions),dim=1))),
+                torch.sqrt(torch.sum(torch.square(gt[:,:,0::2,1::2]-predictions),dim=1))),
+                torch.sqrt(torch.sum(torch.square(gt[:,:,1::2,1::2]-predictions),dim=1)))
+            valid = torch.isfinite(L1err)
+            gtspeed = (torch.sqrt(torch.sum(torch.square(gt[:,:,0::2,0::2]),dim=1)) + torch.sqrt(torch.sum(torch.square(gt[:,:,0::2,1::2]),dim=1)) +\
+                       torch.sqrt(torch.sum(torch.square(gt[:,:,1::2,0::2]),dim=1)) + torch.sqrt(torch.sum(torch.square(gt[:,:,1::2,1::2]),dim=1)) ) / 4.0 # let's just average them
+        else:
+            valid = torch.isfinite(gt[:,0,:,:]) # both x and y would be infinite
+            L1err = torch.sum(torch.abs(gt-predictions),dim=1)
+            L2err = torch.sqrt(torch.sum(torch.square(gt-predictions),dim=1))
+            gtspeed = torch.sqrt(torch.sum(torch.square(gt),dim=1))
+        N = valid.sum()
+        Nnew = self.agg_N + N
+        self.agg_L1err = float(self.agg_N)/Nnew * self.agg_L1err + L1err[valid].mean().cpu() * float(N)/Nnew
+        self.agg_L2err = float(self.agg_N)/Nnew * self.agg_L2err + L2err[valid].mean().cpu() * float(N)/Nnew
+        self.agg_N = Nnew
+        for i,th in enumerate(self.bad_ths):
+            self.agg_Nbad[i] += (L2err[valid]>th).sum().cpu()
+        for i,(th1,th2) in enumerate(self.speed_ths):
+            vv = (gtspeed[valid]>=th1) * (gtspeed[valid]<th2)
+            iNspeed = vv.sum()
+            if iNspeed==0: continue
+            iNnew = self.agg_Nspeed[i] + iNspeed
+            self.agg_EPEspeed[i] = float(self.agg_Nspeed[i]) / iNnew * self.agg_EPEspeed[i] + float(iNspeed) / iNnew * L2err[valid][vv].mean().cpu()
+            self.agg_Nspeed[i] = iNnew
+
+    def _compute_metrics(self):
+        if self._metrics is not None: return
+        out = {}
+        out['L1err'] = self.agg_L1err.item()
+        out['EPE']  = self.agg_L2err.item()
+        for i,th in enumerate(self.bad_ths):
+            out['bad@{:.1f}'.format(th)] = (float(self.agg_Nbad[i]) / self.agg_N).item() * 100.0
+        for i,(th1,th2) in enumerate(self.speed_ths):
+            out['s{:d}{:s}'.format(th1, '-'+str(th2) if th2<torch.inf else '+')] = self.agg_EPEspeed[i].item()
+        self._metrics = out
+    
+    def get_results(self): 
+        self._compute_metrics() # to avoid recompute them multiple times
+        return self._metrics

third_party/dust3r/croco/stereoflow/datasets_flow.py

@@ -0,0 +1,630 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+# --------------------------------------------------------
+# Dataset structure for flow
+# --------------------------------------------------------
+
+import os
+import os.path as osp
+import pickle
+import numpy as np
+import struct
+from PIL import Image 
+import json
+import h5py
+import torch
+from torch.utils import data
+
+from .augmentor import FlowAugmentor
+from .datasets_stereo import _read_img, img_to_tensor, dataset_to_root, _read_pfm
+from copy import deepcopy
+dataset_to_root = deepcopy(dataset_to_root)
+
+dataset_to_root.update(**{
+    'TartanAir': './data/stereoflow/TartanAir',
+    'FlyingChairs': './data/stereoflow/FlyingChairs/',
+    'FlyingThings': osp.join(dataset_to_root['SceneFlow'],'FlyingThings')+'/',
+    'MPISintel': './data/stereoflow//MPI-Sintel/'+'/',
+})
+cache_dir = "./data/stereoflow/datasets_flow_cache/"
+
+
+def flow_to_tensor(disp):
+    return torch.from_numpy(disp).float().permute(2, 0, 1)
+
+class FlowDataset(data.Dataset):
+    
+    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):
+        self.split = split
+        if not augmentor: assert crop_size is None 
+        if crop_size is not None: assert augmentor
+        self.crop_size = crop_size
+        self.augmentor_str = augmentor
+        self.augmentor = FlowAugmentor(crop_size) if augmentor else None
+        self.totensor = totensor
+        self.rmul = 1 # keep track of rmul
+        self.has_constant_resolution = True # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)
+        self._prepare_data()
+        self._load_or_build_cache()
+        
+    def prepare_data(self):
+        """
+        to be defined for each dataset 
+        """
+        raise NotImplementedError 
+        
+    def __len__(self):
+        return len(self.pairnames) # each pairname is typically of the form (str, int1, int2) 
+        
+    def __getitem__(self, index):
+        pairname = self.pairnames[index]
+        
+        # get filenames 
+        img1name = self.pairname_to_img1name(pairname)
+        img2name = self.pairname_to_img2name(pairname)
+        flowname = self.pairname_to_flowname(pairname) if self.pairname_to_flowname is not None else None
+        
+        # load images and disparities
+        img1 = _read_img(img1name)
+        img2 = _read_img(img2name)
+        flow = self.load_flow(flowname) if flowname is not None else None
+
+        # apply augmentations
+        if self.augmentor is not None:
+            img1, img2, flow = self.augmentor(img1, img2, flow, self.name)
+        
+        if self.totensor:
+            img1 = img_to_tensor(img1)
+            img2 = img_to_tensor(img2)
+            if flow is not None: 
+                flow = flow_to_tensor(flow)
+            else:
+                flow = torch.tensor([]) # to allow dataloader batching with default collate_gn
+            pairname = str(pairname) # transform potential tuple to str to be able to batch it
+
+        return img1, img2, flow, pairname
+        
+    def __rmul__(self, v):
+        self.rmul *= v
+        self.pairnames = v * self.pairnames
+        return self
+        
+    def __str__(self):
+        return f'{self.__class__.__name__}_{self.split}'
+        
+    def __repr__(self):
+        s = f'{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})'
+        if self.rmul==1:
+            s+=f'\n\tnum pairs: {len(self.pairnames)}'
+        else:
+            s+=f'\n\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})'
+        return s
+
+    def _set_root(self):
+        self.root = dataset_to_root[self.name]
+        assert os.path.isdir(self.root), f"could not find root directory for dataset {self.name}: {self.root}"       
+
+    def _load_or_build_cache(self):
+        cache_file = osp.join(cache_dir, self.name+'.pkl')
+        if osp.isfile(cache_file):
+            with open(cache_file, 'rb') as fid:
+                self.pairnames = pickle.load(fid)[self.split]
+        else:
+            tosave = self._build_cache()
+            os.makedirs(cache_dir, exist_ok=True)
+            with open(cache_file, 'wb') as fid:
+                pickle.dump(tosave, fid)
+            self.pairnames = tosave[self.split]
+
+class TartanAirDataset(FlowDataset):
+
+    def _prepare_data(self):
+        self.name = "TartanAir"
+        self._set_root()
+        assert self.split in ['train']
+        self.pairname_to_img1name = lambda pairname: osp.join(self.root, pairname[0], 'image_left/{:06d}_left.png'.format(pairname[1]))
+        self.pairname_to_img2name = lambda pairname: osp.join(self.root, pairname[0], 'image_left/{:06d}_left.png'.format(pairname[2]))
+        self.pairname_to_flowname = lambda pairname: osp.join(self.root, pairname[0], 'flow/{:06d}_{:06d}_flow.npy'.format(pairname[1],pairname[2]))
+        self.pairname_to_str = lambda pairname: os.path.join(pairname[0][pairname[0].find('/')+1:], '{:06d}_{:06d}'.format(pairname[1], pairname[2]))
+        self.load_flow = _read_numpy_flow
+        
+    def _build_cache(self):
+        seqs = sorted(os.listdir(self.root))
+        pairs = [(osp.join(s,s,difficulty,Pxxx),int(a[:6]),int(a[:6])+1) for s in seqs for difficulty in ['Easy','Hard'] for Pxxx in sorted(os.listdir(osp.join(self.root,s,s,difficulty))) for a in sorted(os.listdir(osp.join(self.root,s,s,difficulty,Pxxx,'image_left/')))[:-1]]
+        assert len(pairs)==306268, "incorrect parsing of pairs in TartanAir"
+        tosave = {'train': pairs}
+        return tosave
+        
+class FlyingChairsDataset(FlowDataset):
+
+    def _prepare_data(self):
+        self.name = "FlyingChairs"
+        self._set_root()
+        assert self.split in ['train','val']
+        self.pairname_to_img1name = lambda pairname: osp.join(self.root, 'data', pairname+'_img1.ppm')
+        self.pairname_to_img2name = lambda pairname: osp.join(self.root, 'data', pairname+'_img2.ppm')
+        self.pairname_to_flowname = lambda pairname: osp.join(self.root, 'data', pairname+'_flow.flo')
+        self.pairname_to_str = lambda pairname: pairname
+        self.load_flow = _read_flo_file
+        
+    def _build_cache(self):
+        split_file = osp.join(self.root, 'chairs_split.txt')
+        split_list = np.loadtxt(split_file, dtype=np.int32)
+        trainpairs = ['{:05d}'.format(i) for i in np.where(split_list==1)[0]+1]
+        valpairs = ['{:05d}'.format(i) for i in np.where(split_list==2)[0]+1]
+        assert len(trainpairs)==22232 and len(valpairs)==640, "incorrect parsing of pairs in MPI-Sintel"
+        tosave = {'train': trainpairs, 'val': valpairs}
+        return tosave
+        
+class FlyingThingsDataset(FlowDataset):
+    
+    def _prepare_data(self):
+        self.name = "FlyingThings"
+        self._set_root()
+        assert self.split in [f'{set_}_{pass_}pass{camstr}' for set_ in ['train','test','test1024'] for camstr in ['','_rightcam'] for pass_ in ['clean','final','all']]
+        self.pairname_to_img1name = lambda pairname: osp.join(self.root, f'frames_{pairname[3]}pass', pairname[0].replace('into_future','').replace('into_past',''), '{:04d}.png'.format(pairname[1]))
+        self.pairname_to_img2name = lambda pairname: osp.join(self.root, f'frames_{pairname[3]}pass', pairname[0].replace('into_future','').replace('into_past',''), '{:04d}.png'.format(pairname[2]))
+        self.pairname_to_flowname = lambda pairname: osp.join(self.root, 'optical_flow', pairname[0], 'OpticalFlowInto{f:s}_{i:04d}_{c:s}.pfm'.format(f='Future' if 'future' in pairname[0] else 'Past', i=pairname[1], c='L' if 'left' in pairname[0] else 'R' ))
+        self.pairname_to_str = lambda pairname: os.path.join(pairname[3]+'pass', pairname[0], 'Into{f:s}_{i:04d}_{c:s}'.format(f='Future' if 'future' in pairname[0] else 'Past',  i=pairname[1], c='L' if 'left' in pairname[0] else 'R' ))
+        self.load_flow = _read_pfm_flow
+        
+    def _build_cache(self):
+        tosave = {}
+        # train and test splits for the different passes 
+        for set_ in ['train', 'test']:
+            sroot = osp.join(self.root, 'optical_flow', set_.upper())
+            fname_to_i = lambda f: int(f[len('OpticalFlowIntoFuture_'):-len('_L.pfm')])
+            pp = [(osp.join(set_.upper(), d, s, 'into_future/left'),fname_to_i(fname)) for d in sorted(os.listdir(sroot)) for s in sorted(os.listdir(osp.join(sroot,d))) for fname in sorted(os.listdir(osp.join(sroot,d, s, 'into_future/left')))[:-1]]
+            pairs  = [(a,i,i+1) for a,i in pp]
+            pairs += [(a.replace('into_future','into_past'),i+1,i) for a,i in pp]
+            assert len(pairs)=={'train': 40302, 'test': 7866}[set_], "incorrect parsing of pairs Flying Things"
+            for cam in ['left','right']:
+                camstr = '' if cam=='left' else f'_{cam}cam'
+                for pass_ in ['final', 'clean']:
+                    tosave[f'{set_}_{pass_}pass{camstr}'] = [(a.replace('left',cam),i,j,pass_) for a,i,j in pairs]
+                tosave[f'{set_}_allpass{camstr}'] = tosave[f'{set_}_cleanpass{camstr}'] + tosave[f'{set_}_finalpass{camstr}']
+        # test1024: this is the same split as unimatch 'validation' split
+        # see https://github.com/autonomousvision/unimatch/blob/master/dataloader/flow/datasets.py#L229
+        test1024_nsamples = 1024
+        alltest_nsamples = len(tosave['test_cleanpass'])  # 7866
+        stride = alltest_nsamples // test1024_nsamples
+        remove = alltest_nsamples % test1024_nsamples
+        for cam in ['left','right']:
+            camstr = '' if cam=='left' else f'_{cam}cam'
+            for pass_ in ['final','clean']:
+                tosave[f'test1024_{pass_}pass{camstr}'] = sorted(tosave[f'test_{pass_}pass{camstr}'])[:-remove][::stride] # warning, it was not sorted before
+            assert len(tosave['test1024_cleanpass'])==1024, "incorrect parsing of pairs in Flying Things"
+            tosave[f'test1024_allpass{camstr}'] = tosave[f'test1024_cleanpass{camstr}'] + tosave[f'test1024_finalpass{camstr}']
+        return tosave
+        
+      
+class MPISintelDataset(FlowDataset):
+    
+    def _prepare_data(self):
+        self.name = "MPISintel"
+        self._set_root()
+        assert self.split in [s+'_'+p for s in ['train','test','subval','subtrain'] for p in ['cleanpass','finalpass','allpass']]
+        self.pairname_to_img1name = lambda pairname: osp.join(self.root, pairname[0], 'frame_{:04d}.png'.format(pairname[1]))
+        self.pairname_to_img2name = lambda pairname: osp.join(self.root, pairname[0], 'frame_{:04d}.png'.format(pairname[1]+1))
+        self.pairname_to_flowname = lambda pairname: None if pairname[0].startswith('test/') else osp.join(self.root, pairname[0].replace('/clean/','/flow/').replace('/final/','/flow/'), 'frame_{:04d}.flo'.format(pairname[1]))
+        self.pairname_to_str = lambda pairname: osp.join(pairname[0], 'frame_{:04d}'.format(pairname[1]))
+        self.load_flow = _read_flo_file
+        
+    def _build_cache(self):
+        trainseqs = sorted(os.listdir(self.root+'training/clean'))
+        trainpairs = [ (osp.join('training/clean', s),i) for s in trainseqs for i in range(1, len(os.listdir(self.root+'training/clean/'+s)))]
+        subvalseqs = ['temple_2','temple_3']
+        subtrainseqs = [s for s in trainseqs if s not in subvalseqs]
+        subvalpairs = [ (p,i) for p,i in trainpairs if any(s in p for s in subvalseqs)]
+        subtrainpairs = [ (p,i) for p,i in trainpairs if any(s in p for s in subtrainseqs)]
+        testseqs = sorted(os.listdir(self.root+'test/clean'))
+        testpairs = [ (osp.join('test/clean', s),i) for s in testseqs for i in range(1, len(os.listdir(self.root+'test/clean/'+s)))]
+        assert len(trainpairs)==1041 and len(testpairs)==552 and len(subvalpairs)==98 and len(subtrainpairs)==943, "incorrect parsing of pairs in MPI-Sintel"
+        tosave = {}
+        tosave['train_cleanpass'] = trainpairs
+        tosave['test_cleanpass'] = testpairs
+        tosave['subval_cleanpass'] = subvalpairs
+        tosave['subtrain_cleanpass'] = subtrainpairs         
+        for t in ['train','test','subval','subtrain']: 
+            tosave[t+'_finalpass'] = [(p.replace('/clean/','/final/'),i) for p,i in tosave[t+'_cleanpass']]
+            tosave[t+'_allpass'] = tosave[t+'_cleanpass'] + tosave[t+'_finalpass'] 
+        return tosave
+        
+    def submission_save_pairname(self, pairname, prediction, outdir, _time):
+        assert prediction.shape[2]==2
+        outfile = os.path.join(outdir, 'submission', self.pairname_to_str(pairname)+'.flo')
+        os.makedirs( os.path.dirname(outfile), exist_ok=True)
+        writeFlowFile(prediction, outfile)
+        
+    def finalize_submission(self, outdir):
+        assert self.split == 'test_allpass'
+        bundle_exe = "/nfs/data/ffs-3d/datasets/StereoFlow/MPI-Sintel/bundler/linux-x64/bundler" # eg <bundle_exe> <path_to_results_for_clean> <path_to_results_for_final> <output/bundled.lzma>
+        if os.path.isfile(bundle_exe):
+            cmd = f'{bundle_exe} "{outdir}/submission/test/clean/" "{outdir}/submission/test/final" "{outdir}/submission/bundled.lzma"'
+            print(cmd)
+            os.system(cmd)
+            print(f'Done. Submission file at: "{outdir}/submission/bundled.lzma"')
+        else:
+            print('Could not find bundler executable for submission.')
+            print('Please download it and run:')
+            print(f'<bundle_exe> "{outdir}/submission/test/clean/" "{outdir}/submission/test/final" "{outdir}/submission/bundled.lzma"')
+        
+class SpringDataset(FlowDataset):
+
+    def _prepare_data(self):
+        self.name = "Spring"
+        self._set_root()
+        assert self.split in ['train','test','subtrain','subval']
+        self.pairname_to_img1name = lambda pairname: osp.join(self.root, pairname[0], pairname[1], 'frame_'+pairname[3], 'frame_{:s}_{:04d}.png'.format(pairname[3], pairname[4]))
+        self.pairname_to_img2name = lambda pairname: osp.join(self.root, pairname[0], pairname[1], 'frame_'+pairname[3], 'frame_{:s}_{:04d}.png'.format(pairname[3], pairname[4]+(1 if pairname[2]=='FW' else -1)))
+        self.pairname_to_flowname = lambda pairname: None if pairname[0]=='test' else osp.join(self.root, pairname[0], pairname[1], f'flow_{pairname[2]}_{pairname[3]}', f'flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}.flo5')
+        self.pairname_to_str = lambda pairname: osp.join(pairname[0], pairname[1], f'flow_{pairname[2]}_{pairname[3]}', f'flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}')
+        self.load_flow = _read_hdf5_flow
+
+    def _build_cache(self):
+        # train 
+        trainseqs = sorted(os.listdir( osp.join(self.root,'train')))
+        trainpairs = []
+        for leftright in ['left','right']:
+            for fwbw in ['FW','BW']:
+                trainpairs += [('train',s,fwbw,leftright,int(f[len(f'flow_{fwbw}_{leftright}_'):-len('.flo5')])) for s in trainseqs for f in sorted(os.listdir(osp.join(self.root,'train',s,f'flow_{fwbw}_{leftright}')))]
+        # test 
+        testseqs = sorted(os.listdir( osp.join(self.root,'test')))
+        testpairs = []
+        for leftright in ['left','right']:
+            testpairs += [('test',s,'FW',leftright,int(f[len(f'frame_{leftright}_'):-len('.png')])) for s in testseqs for f in sorted(os.listdir(osp.join(self.root,'test',s,f'frame_{leftright}')))[:-1]]
+            testpairs += [('test',s,'BW',leftright,int(f[len(f'frame_{leftright}_'):-len('.png')])+1) for s in testseqs for f in sorted(os.listdir(osp.join(self.root,'test',s,f'frame_{leftright}')))[:-1]]
+        # subtrain / subval
+        subtrainpairs = [p for p in trainpairs if p[1]!='0041']
+        subvalpairs = [p for p in trainpairs if p[1]=='0041']
+        assert len(trainpairs)==19852 and len(testpairs)==3960 and len(subtrainpairs)==19472 and len(subvalpairs)==380, "incorrect parsing of pairs in Spring"
+        tosave = {'train': trainpairs, 'test': testpairs, 'subtrain': subtrainpairs, 'subval': subvalpairs}
+        return tosave
+        
+    def submission_save_pairname(self, pairname, prediction, outdir, time):
+        assert prediction.ndim==3
+        assert prediction.shape[2]==2
+        assert prediction.dtype==np.float32
+        outfile = osp.join(outdir, pairname[0], pairname[1], f'flow_{pairname[2]}_{pairname[3]}', f'flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}.flo5')
+        os.makedirs( os.path.dirname(outfile), exist_ok=True)
+        writeFlo5File(prediction, outfile)
+        
+    def finalize_submission(self, outdir):
+        assert self.split=='test'
+        exe = "{self.root}/flow_subsampling"
+        if os.path.isfile(exe):
+            cmd = f'cd "{outdir}/test"; {exe} .'
+            print(cmd)
+            os.system(cmd)
+            print(f'Done. Submission file at {outdir}/test/flow_submission.hdf5')
+        else:
+            print('Could not find flow_subsampling executable for submission.')
+            print('Please download it and run:')
+            print(f'cd "{outdir}/test"; <flow_subsampling_exe> .')
+
+        
+class Kitti12Dataset(FlowDataset):
+
+    def _prepare_data(self):
+        self.name = "Kitti12"
+        self._set_root()
+        assert self.split in ['train','test']
+        self.pairname_to_img1name = lambda pairname: osp.join(self.root, pairname+'_10.png')
+        self.pairname_to_img2name = lambda pairname: osp.join(self.root, pairname+'_11.png')
+        self.pairname_to_flowname = None if self.split=='test' else lambda pairname: osp.join(self.root, pairname.replace('/colored_0/','/flow_occ/')+'_10.png')
+        self.pairname_to_str = lambda pairname: pairname.replace('/colored_0/','/')
+        self.load_flow = _read_kitti_flow
+        
+    def _build_cache(self):
+        trainseqs = ["training/colored_0/%06d"%(i) for i in range(194)]
+        testseqs = ["testing/colored_0/%06d"%(i) for i in range(195)]
+        assert len(trainseqs)==194 and len(testseqs)==195, "incorrect parsing of pairs in Kitti12"
+        tosave = {'train': trainseqs, 'test': testseqs}
+        return tosave 
+
+    def submission_save_pairname(self, pairname, prediction, outdir, time):
+        assert prediction.ndim==3
+        assert prediction.shape[2]==2
+        outfile = os.path.join(outdir, pairname.split('/')[-1]+'_10.png')
+        os.makedirs( os.path.dirname(outfile), exist_ok=True)
+        writeFlowKitti(outfile, prediction)
+
+    def finalize_submission(self, outdir):
+        assert self.split=='test'
+        cmd = f'cd {outdir}/; zip -r "kitti12_flow_results.zip" .'
+        print(cmd)
+        os.system(cmd)
+        print(f'Done. Submission file at {outdir}/kitti12_flow_results.zip')
+
+
+class Kitti15Dataset(FlowDataset):
+    
+    def _prepare_data(self):
+        self.name = "Kitti15"
+        self._set_root()
+        assert self.split in ['train','subtrain','subval','test']
+        self.pairname_to_img1name = lambda pairname: osp.join(self.root, pairname+'_10.png')
+        self.pairname_to_img2name = lambda pairname: osp.join(self.root, pairname+'_11.png')
+        self.pairname_to_flowname = None if self.split=='test' else lambda pairname: osp.join(self.root, pairname.replace('/image_2/','/flow_occ/')+'_10.png')
+        self.pairname_to_str = lambda pairname: pairname.replace('/image_2/','/')
+        self.load_flow = _read_kitti_flow
+        
+    def _build_cache(self):
+        trainseqs = ["training/image_2/%06d"%(i) for i in range(200)]
+        subtrainseqs = trainseqs[:-10]
+        subvalseqs = trainseqs[-10:]
+        testseqs = ["testing/image_2/%06d"%(i) for i in range(200)]
+        assert len(trainseqs)==200 and len(subtrainseqs)==190 and len(subvalseqs)==10 and len(testseqs)==200, "incorrect parsing of pairs in Kitti15"
+        tosave = {'train': trainseqs, 'subtrain': subtrainseqs, 'subval': subvalseqs, 'test': testseqs}
+        return tosave 
+
+    def submission_save_pairname(self, pairname, prediction, outdir, time):
+        assert prediction.ndim==3
+        assert prediction.shape[2]==2
+        outfile = os.path.join(outdir, 'flow', pairname.split('/')[-1]+'_10.png')
+        os.makedirs( os.path.dirname(outfile), exist_ok=True)
+        writeFlowKitti(outfile, prediction)
+
+    def finalize_submission(self, outdir):
+        assert self.split=='test'
+        cmd = f'cd {outdir}/; zip -r "kitti15_flow_results.zip" flow'
+        print(cmd)
+        os.system(cmd)
+        print(f'Done. Submission file at {outdir}/kitti15_flow_results.zip')
+
+
+import cv2
+def _read_numpy_flow(filename): 
+    return np.load(filename)
+    
+def _read_pfm_flow(filename):
+    f, _ = _read_pfm(filename)
+    assert np.all(f[:,:,2]==0.0)
+    return np.ascontiguousarray(f[:,:,:2])
+
+TAG_FLOAT = 202021.25 # tag to check the sanity of the file
+TAG_STRING = 'PIEH'   # string containing the tag
+MIN_WIDTH = 1
+MAX_WIDTH = 99999
+MIN_HEIGHT = 1
+MAX_HEIGHT = 99999
+def readFlowFile(filename):
+    """
+    readFlowFile(<FILENAME>) reads a flow file <FILENAME> into a 2-band np.array.
+    if <FILENAME> does not exist, an IOError is raised.
+    if <FILENAME> does not finish by '.flo' or the tag, the width, the height or the file's size is illegal, an Expcetion is raised.
+    ---- PARAMETERS ----
+        filename: string containg the name of the file to read a flow
+    ---- OUTPUTS ----
+        a np.array of dimension (height x width x 2) containing the flow of type 'float32'
+    """
+        
+    # check filename
+    if not filename.endswith(".flo"):
+        raise Exception("readFlowFile({:s}): filename must finish with '.flo'".format(filename))
+    
+    # open the file and read it
+    with open(filename,'rb') as f:
+        # check tag
+        tag = struct.unpack('f',f.read(4))[0]
+        if tag != TAG_FLOAT:
+            raise Exception("flow_utils.readFlowFile({:s}): wrong tag".format(filename))
+        # read dimension
+        w,h = struct.unpack('ii',f.read(8))
+        if w < MIN_WIDTH or w > MAX_WIDTH:
+            raise Exception("flow_utils.readFlowFile({:s}: illegal width {:d}".format(filename,w))
+        if h < MIN_HEIGHT or h > MAX_HEIGHT:
+            raise Exception("flow_utils.readFlowFile({:s}: illegal height {:d}".format(filename,h))
+        flow = np.fromfile(f,'float32')
+        if not flow.shape == (h*w*2,):
+            raise Exception("flow_utils.readFlowFile({:s}: illegal size of the file".format(filename))
+        flow.shape = (h,w,2)
+        return flow
+
+def writeFlowFile(flow,filename):
+    """
+    writeFlowFile(flow,<FILENAME>) write flow to the file <FILENAME>.
+    if <FILENAME> does not exist, an IOError is raised.
+    if <FILENAME> does not finish with '.flo' or the flow has not 2 bands, an Exception is raised.
+    ---- PARAMETERS ----
+        flow: np.array of dimension (height x width x 2) containing the flow to write
+        filename: string containg the name of the file to write a flow
+    """
+    
+    # check filename
+    if not filename.endswith(".flo"):
+        raise Exception("flow_utils.writeFlowFile(<flow>,{:s}): filename must finish with '.flo'".format(filename))
+    
+    if not flow.shape[2:] == (2,):
+        raise Exception("flow_utils.writeFlowFile(<flow>,{:s}): <flow> must have 2 bands".format(filename))
+
+
+    # open the file and write it
+    with open(filename,'wb') as f:
+        # write TAG
+        f.write( TAG_STRING.encode('utf-8') )
+        # write dimension
+        f.write( struct.pack('ii',flow.shape[1],flow.shape[0]) )
+        # write the flow
+        
+        flow.astype(np.float32).tofile(f)
+        
+_read_flo_file = readFlowFile
+
+def _read_kitti_flow(filename):
+    flow = cv2.imread(filename, cv2.IMREAD_ANYDEPTH | cv2.IMREAD_COLOR)
+    flow = flow[:, :, ::-1].astype(np.float32)
+    valid = flow[:, :, 2]>0
+    flow = flow[:, :, :2]
+    flow = (flow - 2 ** 15) / 64.0
+    flow[~valid,0] = np.inf
+    flow[~valid,1] = np.inf
+    return flow
+_read_hd1k_flow = _read_kitti_flow
+    
+        
+def writeFlowKitti(filename, uv):
+    uv = 64.0 * uv + 2 ** 15
+    valid = np.ones([uv.shape[0], uv.shape[1], 1])
+    uv = np.concatenate([uv, valid], axis=-1).astype(np.uint16)
+    cv2.imwrite(filename, uv[..., ::-1])
+
+def writeFlo5File(flow, filename):
+    with h5py.File(filename, "w") as f:
+        f.create_dataset("flow", data=flow, compression="gzip", compression_opts=5)
+    
+def _read_hdf5_flow(filename):
+    flow = np.asarray(h5py.File(filename)['flow'])
+    flow[np.isnan(flow)] = np.inf # make invalid values as +inf
+    return flow.astype(np.float32)
+
+# flow visualization
+RY = 15
+YG = 6
+GC = 4
+CB = 11
+BM = 13
+MR = 6
+UNKNOWN_THRESH = 1e9
+
+def colorTest():
+    """
+    flow_utils.colorTest(): display an example of image showing the color encoding scheme
+    """
+    import matplotlib.pylab as plt
+    truerange = 1
+    h,w = 151,151
+    trange = truerange*1.04
+    s2 = round(h/2)
+    x,y = np.meshgrid(range(w),range(h))
+    u = x*trange/s2-trange
+    v = y*trange/s2-trange
+    img = _computeColor(np.concatenate((u[:,:,np.newaxis],v[:,:,np.newaxis]),2)/trange/np.sqrt(2))
+    plt.imshow(img)
+    plt.axis('off')
+    plt.axhline(round(h/2),color='k')
+    plt.axvline(round(w/2),color='k')
+    
+def flowToColor(flow, maxflow=None, maxmaxflow=None, saturate=False):
+    """
+    flow_utils.flowToColor(flow): return a color code flow field, normalized based on the maximum l2-norm of the flow
+    flow_utils.flowToColor(flow,maxflow): return a color code flow field, normalized by maxflow
+    ---- PARAMETERS ----
+        flow: flow to display of shape (height x width x 2)
+        maxflow (default:None): if given, normalize the flow by its value, otherwise by the flow norm
+        maxmaxflow (default:None): if given, normalize the flow by the max of its value and the flow norm
+    ---- OUTPUT ----
+        an np.array of shape (height x width x 3) of type uint8 containing a color code of the flow
+    """
+    h,w,n = flow.shape
+    # check size of flow
+    assert n == 2, "flow_utils.flowToColor(flow): flow must have 2 bands"
+    # fix unknown flow
+    unknown_idx = np.max(np.abs(flow),2)>UNKNOWN_THRESH
+    flow[unknown_idx] = 0.0
+    # compute max flow if needed
+    if maxflow is None:
+        maxflow = flowMaxNorm(flow)
+    if maxmaxflow is not None:
+        maxflow = min(maxmaxflow, maxflow)
+    # normalize flow
+    eps = np.spacing(1) # minimum positive float value to avoid division by 0
+    # compute the flow
+    img = _computeColor(flow/(maxflow+eps), saturate=saturate)
+    # put black pixels in unknown location
+    img[ np.tile( unknown_idx[:,:,np.newaxis],[1,1,3]) ] = 0.0 
+    return img
+
+def flowMaxNorm(flow):
+    """
+    flow_utils.flowMaxNorm(flow): return the maximum of the l2-norm of the given flow
+    ---- PARAMETERS ----
+        flow: the flow
+        
+    ---- OUTPUT ----
+        a float containing the maximum of the l2-norm of the flow
+    """
+    return np.max( np.sqrt( np.sum( np.square( flow ) , 2) ) )
+
+def _computeColor(flow, saturate=True):
+    """
+    flow_utils._computeColor(flow): compute color codes for the flow field flow
+    
+    ---- PARAMETERS ----
+        flow: np.array of dimension (height x width x 2) containing the flow to display
+    ---- OUTPUTS ----
+        an np.array of dimension (height x width x 3) containing the color conversion of the flow
+    """
+    # set nan to 0
+    nanidx = np.isnan(flow[:,:,0])
+    flow[nanidx] = 0.0
+    
+    # colorwheel
+    ncols = RY + YG + GC + CB + BM + MR
+    nchans = 3
+    colorwheel = np.zeros((ncols,nchans),'uint8')
+    col = 0;
+    #RY
+    colorwheel[:RY,0] = 255
+    colorwheel[:RY,1] = [(255*i) // RY for i in range(RY)]
+    col += RY
+    # YG    
+    colorwheel[col:col+YG,0] = [255 - (255*i) // YG for i in range(YG)]
+    colorwheel[col:col+YG,1] = 255
+    col += YG
+    # GC
+    colorwheel[col:col+GC,1] = 255
+    colorwheel[col:col+GC,2] = [(255*i) // GC for i in range(GC)]
+    col += GC
+    # CB
+    colorwheel[col:col+CB,1] = [255 - (255*i) // CB for i in range(CB)]
+    colorwheel[col:col+CB,2] = 255
+    col += CB
+    # BM
+    colorwheel[col:col+BM,0] = [(255*i) // BM for i in range(BM)]
+    colorwheel[col:col+BM,2] = 255
+    col += BM
+    # MR
+    colorwheel[col:col+MR,0] = 255
+    colorwheel[col:col+MR,2] = [255 - (255*i) // MR for i in range(MR)]
+
+    # compute utility variables
+    rad = np.sqrt( np.sum( np.square(flow) , 2) ) # magnitude
+    a = np.arctan2( -flow[:,:,1] , -flow[:,:,0]) / np.pi # angle
+    fk = (a+1)/2 * (ncols-1) # map [-1,1] to [0,ncols-1]
+    k0 = np.floor(fk).astype('int')
+    k1 = k0+1
+    k1[k1==ncols] = 0
+    f = fk-k0
+
+    if not saturate:
+        rad = np.minimum(rad,1)
+
+    # compute the image
+    img = np.zeros( (flow.shape[0],flow.shape[1],nchans), 'uint8' )
+    for i in range(nchans):
+        tmp = colorwheel[:,i].astype('float')
+        col0 = tmp[k0]/255
+        col1 = tmp[k1]/255
+        col = (1-f)*col0 + f*col1
+        idx = (rad <= 1)
+        col[idx] = 1-rad[idx]*(1-col[idx]) # increase saturation with radius
+        col[~idx] *= 0.75 # out of range
+        img[:,:,i] = (255*col*(1-nanidx.astype('float'))).astype('uint8')
+
+    return img
+    
+# flow dataset getter 
+    
+def get_train_dataset_flow(dataset_str, augmentor=True, crop_size=None):
+    dataset_str = dataset_str.replace('(','Dataset(')
+    if augmentor:
+        dataset_str = dataset_str.replace(')',', augmentor=True)')
+    if crop_size is not None:
+        dataset_str = dataset_str.replace(')',', crop_size={:s})'.format(str(crop_size)))
+    return eval(dataset_str)
+    
+def get_test_datasets_flow(dataset_str):
+    dataset_str = dataset_str.replace('(','Dataset(')
+    return [eval(s) for s in dataset_str.split('+')]

third_party/dust3r/croco/stereoflow/datasets_stereo.py

@@ -0,0 +1,674 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+# --------------------------------------------------------
+# Dataset structure for stereo
+# --------------------------------------------------------
+
+import sys, os
+import os.path as osp
+import pickle
+import numpy as np
+from PIL import Image
+import json
+import h5py
+from glob import glob
+import cv2
+
+import torch
+from torch.utils import data
+
+from .augmentor import StereoAugmentor
+
+
+
+dataset_to_root = {
+    'CREStereo': './data/stereoflow//crenet_stereo_trainset/stereo_trainset/crestereo/',
+    'SceneFlow': './data/stereoflow//SceneFlow/',
+    'ETH3DLowRes': './data/stereoflow/eth3d_lowres/',
+    'Booster': './data/stereoflow/booster_gt/',
+    'Middlebury2021': './data/stereoflow/middlebury/2021/data/',
+    'Middlebury2014': './data/stereoflow/middlebury/2014/',
+    'Middlebury2006': './data/stereoflow/middlebury/2006/',
+    'Middlebury2005': './data/stereoflow/middlebury/2005/train/',
+    'MiddleburyEval3':  './data/stereoflow/middlebury/MiddEval3/',
+    'Spring': './data/stereoflow/spring/', 
+    'Kitti15': './data/stereoflow/kitti-stereo-2015/',
+    'Kitti12': './data/stereoflow/kitti-stereo-2012/',
+}
+cache_dir = "./data/stereoflow/datasets_stereo_cache/"
+
+
+in1k_mean = torch.tensor([0.485, 0.456, 0.406]).view(3,1,1)
+in1k_std =  torch.tensor([0.229, 0.224, 0.225]).view(3,1,1)
+def img_to_tensor(img):
+    img = torch.from_numpy(img).permute(2, 0, 1).float() / 255.
+    img = (img-in1k_mean)/in1k_std
+    return img
+def disp_to_tensor(disp):
+    return torch.from_numpy(disp)[None,:,:]
+
+class StereoDataset(data.Dataset):
+    
+    def __init__(self, split, augmentor=False, crop_size=None, totensor=True):
+        self.split = split
+        if not augmentor: assert crop_size is None 
+        if crop_size: assert augmentor
+        self.crop_size = crop_size
+        self.augmentor_str = augmentor
+        self.augmentor = StereoAugmentor(crop_size) if augmentor else None
+        self.totensor = totensor
+        self.rmul = 1 # keep track of rmul
+        self.has_constant_resolution = True # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time)
+        self._prepare_data()
+        self._load_or_build_cache()
+        
+    def prepare_data(self):
+        """
+        to be defined for each dataset 
+        """
+        raise NotImplementedError 
+        
+    def __len__(self):
+        return len(self.pairnames)
+        
+    def __getitem__(self, index):
+        pairname = self.pairnames[index]
+        
+        # get filenames 
+        Limgname = self.pairname_to_Limgname(pairname)
+        Rimgname = self.pairname_to_Rimgname(pairname)
+        Ldispname = self.pairname_to_Ldispname(pairname) if self.pairname_to_Ldispname is not None else None
+        
+        # load images and disparities
+        Limg = _read_img(Limgname)
+        Rimg = _read_img(Rimgname)
+        disp = self.load_disparity(Ldispname) if Ldispname is not None else None
+        
+        # sanity check
+        if disp is not None: assert np.all(disp>0) or self.name=="Spring", (self.name, pairname, Ldispname)
+        
+        # apply augmentations
+        if self.augmentor is not None:
+            Limg, Rimg, disp = self.augmentor(Limg, Rimg, disp, self.name)
+        
+        if self.totensor:
+            Limg = img_to_tensor(Limg)
+            Rimg = img_to_tensor(Rimg)
+            if disp is None:
+                disp = torch.tensor([]) # to allow dataloader batching with default collate_gn
+            else:
+                disp = disp_to_tensor(disp)
+        
+        return Limg, Rimg, disp, str(pairname)
+        
+    def __rmul__(self, v):
+        self.rmul *= v
+        self.pairnames = v * self.pairnames
+        return self
+        
+    def __str__(self):
+        return f'{self.__class__.__name__}_{self.split}'
+        
+    def __repr__(self):
+        s = f'{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})'
+        if self.rmul==1:
+            s+=f'\n\tnum pairs: {len(self.pairnames)}'
+        else:
+            s+=f'\n\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})'
+        return s
+
+    def _set_root(self):
+        self.root = dataset_to_root[self.name]
+        assert os.path.isdir(self.root), f"could not find root directory for dataset {self.name}: {self.root}"       
+
+    def _load_or_build_cache(self):
+        cache_file = osp.join(cache_dir, self.name+'.pkl')
+        if osp.isfile(cache_file):
+            with open(cache_file, 'rb') as fid:
+                self.pairnames = pickle.load(fid)[self.split]
+        else:
+            tosave = self._build_cache()
+            os.makedirs(cache_dir, exist_ok=True)
+            with open(cache_file, 'wb') as fid:
+                pickle.dump(tosave, fid)
+            self.pairnames = tosave[self.split]
+        
+class CREStereoDataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = 'CREStereo'
+        self._set_root()
+        assert self.split in ['train']
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname+'_left.jpg')
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname+'_right.jpg')
+        self.pairname_to_Ldispname = lambda pairname: osp.join(self.root, pairname+'_left.disp.png')
+        self.pairname_to_str = lambda pairname: pairname
+        self.load_disparity = _read_crestereo_disp
+        
+    
+    def _build_cache(self):
+        allpairs = [s+'/'+f[:-len('_left.jpg')] for s in sorted(os.listdir(self.root)) for f in sorted(os.listdir(self.root+'/'+s)) if f.endswith('_left.jpg')]
+        assert len(allpairs)==200000, "incorrect parsing of pairs in CreStereo"
+        tosave = {'train': allpairs}
+        return tosave
+        
+class SceneFlowDataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "SceneFlow"
+        self._set_root()
+        assert self.split in ['train_finalpass','train_cleanpass','train_allpass','test_finalpass','test_cleanpass','test_allpass','test1of100_cleanpass','test1of100_finalpass']
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname).replace('/left/','/right/')
+        self.pairname_to_Ldispname = lambda pairname: osp.join(self.root, pairname).replace('/frames_finalpass/','/disparity/').replace('/frames_cleanpass/','/disparity/')[:-4]+'.pfm'
+        self.pairname_to_str = lambda pairname: pairname[:-4]
+        self.load_disparity = _read_sceneflow_disp
+        
+    def _build_cache(self):
+        trainpairs = []
+        # driving
+        pairs = sorted(glob(self.root+'Driving/frames_finalpass/*/*/*/left/*.png'))
+        pairs = list(map(lambda x: x[len(self.root):], pairs))
+        assert len(pairs) == 4400, "incorrect parsing of pairs in SceneFlow"
+        trainpairs += pairs
+        # monkaa
+        pairs = sorted(glob(self.root+'Monkaa/frames_finalpass/*/left/*.png'))
+        pairs = list(map(lambda x: x[len(self.root):], pairs))
+        assert len(pairs) == 8664, "incorrect parsing of pairs in SceneFlow"
+        trainpairs += pairs
+        # flyingthings
+        pairs = sorted(glob(self.root+'FlyingThings/frames_finalpass/TRAIN/*/*/left/*.png'))
+        pairs = list(map(lambda x: x[len(self.root):], pairs))
+        assert len(pairs) == 22390, "incorrect parsing of pairs in SceneFlow"
+        trainpairs += pairs
+        assert len(trainpairs) == 35454, "incorrect parsing of pairs in SceneFlow"
+        testpairs = sorted(glob(self.root+'FlyingThings/frames_finalpass/TEST/*/*/left/*.png'))
+        testpairs = list(map(lambda x: x[len(self.root):], testpairs))
+        assert len(testpairs) == 4370, "incorrect parsing of pairs in SceneFlow"
+        test1of100pairs = testpairs[::100]
+        assert len(test1of100pairs) == 44, "incorrect parsing of pairs in SceneFlow"
+        # all 
+        tosave = {'train_finalpass': trainpairs,
+                  'train_cleanpass': list(map(lambda x: x.replace('frames_finalpass','frames_cleanpass'), trainpairs)),
+                  'test_finalpass': testpairs,
+                  'test_cleanpass': list(map(lambda x: x.replace('frames_finalpass','frames_cleanpass'), testpairs)),
+                  'test1of100_finalpass': test1of100pairs,
+                  'test1of100_cleanpass': list(map(lambda x: x.replace('frames_finalpass','frames_cleanpass'), test1of100pairs)),
+                 }
+        tosave['train_allpass'] = tosave['train_finalpass']+tosave['train_cleanpass']
+        tosave['test_allpass'] = tosave['test_finalpass']+tosave['test_cleanpass']
+        return tosave
+   
+class Md21Dataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "Middlebury2021"
+        self._set_root()
+        assert self.split in ['train','subtrain','subval']
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname.replace('/im0','/im1'))
+        self.pairname_to_Ldispname = lambda pairname: osp.join(self.root, pairname.split('/')[0], 'disp0.pfm')
+        self.pairname_to_str = lambda pairname: pairname[:-4]
+        self.load_disparity = _read_middlebury_disp
+        
+    def _build_cache(self):
+        seqs = sorted(os.listdir(self.root))
+        trainpairs = []
+        for s in seqs:
+            #trainpairs += [s+'/im0.png'] # we should remove it, it is included as such in other lightings
+            trainpairs += [s+'/ambient/'+b+'/'+a for b in sorted(os.listdir(osp.join(self.root,s,'ambient'))) for a in sorted(os.listdir(osp.join(self.root,s,'ambient',b))) if a.startswith('im0')]
+        assert len(trainpairs)==355
+        subtrainpairs = [p for p in trainpairs if any(p.startswith(s+'/') for s in seqs[:-2])]
+        subvalpairs = [p for p in trainpairs if any(p.startswith(s+'/') for s in seqs[-2:])]
+        assert len(subtrainpairs)==335 and len(subvalpairs)==20, "incorrect parsing of pairs in Middlebury 2021"
+        tosave = {'train': trainpairs, 'subtrain': subtrainpairs, 'subval': subvalpairs}
+        return tosave 
+
+class Md14Dataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "Middlebury2014"
+        self._set_root()
+        assert self.split in ['train','subtrain','subval']
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, osp.dirname(pairname), 'im0.png')
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname)
+        self.pairname_to_Ldispname = lambda pairname: osp.join(self.root, osp.dirname(pairname), 'disp0.pfm')
+        self.pairname_to_str = lambda pairname: pairname[:-4]
+        self.load_disparity = _read_middlebury_disp
+        self.has_constant_resolution = False
+        
+    def _build_cache(self):
+        seqs = sorted(os.listdir(self.root))
+        trainpairs = []
+        for s in seqs:
+            trainpairs += [s+'/im1.png',s+'/im1E.png',s+'/im1L.png']
+        assert len(trainpairs)==138
+        valseqs = ['Umbrella-imperfect','Vintage-perfect']
+        assert all(s in seqs for s in valseqs)
+        subtrainpairs = [p for p in trainpairs if not any(p.startswith(s+'/') for s in valseqs)]
+        subvalpairs = [p for p in trainpairs if any(p.startswith(s+'/') for s in valseqs)]
+        assert len(subtrainpairs)==132 and len(subvalpairs)==6, "incorrect parsing of pairs in Middlebury 2014"
+        tosave = {'train': trainpairs, 'subtrain': subtrainpairs, 'subval': subvalpairs}
+        return tosave 
+
+class Md06Dataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "Middlebury2006"
+        self._set_root()
+        assert self.split in ['train','subtrain','subval']
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, osp.dirname(pairname), 'view5.png')
+        self.pairname_to_Ldispname = lambda pairname: osp.join(self.root, pairname.split('/')[0], 'disp1.png')
+        self.load_disparity = _read_middlebury20052006_disp
+        self.has_constant_resolution = False
+        
+    def _build_cache(self):
+        seqs = sorted(os.listdir(self.root))
+        trainpairs = []
+        for s in seqs:
+            for i in ['Illum1','Illum2','Illum3']:
+                for e in ['Exp0','Exp1','Exp2']:
+                    trainpairs.append(osp.join(s,i,e,'view1.png'))
+        assert len(trainpairs)==189
+        valseqs = ['Rocks1','Wood2']
+        assert all(s in seqs for s in valseqs)
+        subtrainpairs = [p for p in trainpairs if not any(p.startswith(s+'/') for s in valseqs)]
+        subvalpairs = [p for p in trainpairs if any(p.startswith(s+'/') for s in valseqs)]
+        assert len(subtrainpairs)==171 and len(subvalpairs)==18, "incorrect parsing of pairs in Middlebury 2006"
+        tosave = {'train': trainpairs, 'subtrain': subtrainpairs, 'subval': subvalpairs}
+        return tosave
+
+class Md05Dataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "Middlebury2005"
+        self._set_root()
+        assert self.split in ['train','subtrain','subval']
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, osp.dirname(pairname), 'view5.png')
+        self.pairname_to_Ldispname = lambda pairname: osp.join(self.root, pairname.split('/')[0], 'disp1.png')
+        self.pairname_to_str = lambda pairname: pairname[:-4]
+        self.load_disparity = _read_middlebury20052006_disp
+        
+    def _build_cache(self):
+        seqs = sorted(os.listdir(self.root))
+        trainpairs = []
+        for s in seqs:
+            for i in ['Illum1','Illum2','Illum3']:
+                for e in ['Exp0','Exp1','Exp2']:
+                    trainpairs.append(osp.join(s,i,e,'view1.png'))
+        assert len(trainpairs)==54, "incorrect parsing of pairs in Middlebury 2005"
+        valseqs = ['Reindeer']
+        assert all(s in seqs for s in valseqs)
+        subtrainpairs = [p for p in trainpairs if not any(p.startswith(s+'/') for s in valseqs)]
+        subvalpairs = [p for p in trainpairs if any(p.startswith(s+'/') for s in valseqs)]
+        assert len(subtrainpairs)==45 and len(subvalpairs)==9, "incorrect parsing of pairs in Middlebury 2005"
+        tosave = {'train': trainpairs, 'subtrain': subtrainpairs, 'subval': subvalpairs}
+        return tosave
+        
+class MdEval3Dataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "MiddleburyEval3"
+        self._set_root()
+        assert self.split in [s+'_'+r for s in ['train','subtrain','subval','test','all'] for r in ['full','half','quarter']]
+        if self.split.endswith('_full'):
+            self.root = self.root.replace('/MiddEval3','/MiddEval3_F')
+        elif self.split.endswith('_half'):        
+            self.root = self.root.replace('/MiddEval3','/MiddEval3_H')
+        else:
+            assert self.split.endswith('_quarter')
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname, 'im0.png')
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname, 'im1.png')
+        self.pairname_to_Ldispname = lambda pairname: None if pairname.startswith('test') else osp.join(self.root, pairname, 'disp0GT.pfm')
+        self.pairname_to_str = lambda pairname: pairname
+        self.load_disparity = _read_middlebury_disp
+        # for submission only
+        self.submission_methodname = "CroCo-Stereo"
+        self.submission_sresolution = 'F' if self.split.endswith('_full') else ('H' if self.split.endswith('_half') else 'Q')
+        
+    def _build_cache(self):
+        trainpairs = ['train/'+s for s in sorted(os.listdir(self.root+'train/'))]
+        testpairs = ['test/'+s for s in sorted(os.listdir(self.root+'test/'))]
+        subvalpairs = trainpairs[-1:]
+        subtrainpairs = trainpairs[:-1]
+        allpairs = trainpairs+testpairs
+        assert len(trainpairs)==15 and len(testpairs)==15 and len(subvalpairs)==1 and len(subtrainpairs)==14 and len(allpairs)==30, "incorrect parsing of pairs in Middlebury Eval v3"
+        tosave = {}
+        for r in ['full','half','quarter']:
+            tosave.update(**{'train_'+r: trainpairs, 'subtrain_'+r: subtrainpairs, 'subval_'+r: subvalpairs, 'test_'+r: testpairs, 'all_'+r: allpairs})
+        return tosave
+        
+    def submission_save_pairname(self, pairname, prediction, outdir, time):
+        assert prediction.ndim==2
+        assert prediction.dtype==np.float32
+        outfile = os.path.join(outdir, pairname.split('/')[0].replace('train','training')+self.submission_sresolution, pairname.split('/')[1], 'disp0'+self.submission_methodname+'.pfm')
+        os.makedirs( os.path.dirname(outfile), exist_ok=True)
+        writePFM(outfile, prediction)
+        timefile = os.path.join( os.path.dirname(outfile), "time"+self.submission_methodname+'.txt')
+        with open(timefile, 'w') as fid:
+            fid.write(str(time))
+
+    def finalize_submission(self, outdir):
+        cmd = f'cd {outdir}/; zip -r "{self.submission_methodname}.zip" .'
+        print(cmd)
+        os.system(cmd)
+        print(f'Done. Submission file at {outdir}/{self.submission_methodname}.zip')
+
+class ETH3DLowResDataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "ETH3DLowRes"
+        self._set_root()
+        assert self.split in ['train','test','subtrain','subval','all']
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname, 'im0.png')
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname, 'im1.png')
+        self.pairname_to_Ldispname = None if self.split=='test' else lambda pairname: None if pairname.startswith('test/') else osp.join(self.root, pairname.replace('train/','train_gt/'), 'disp0GT.pfm')
+        self.pairname_to_str = lambda pairname: pairname
+        self.load_disparity = _read_eth3d_disp
+        self.has_constant_resolution = False
+        
+    def _build_cache(self):
+        trainpairs = ['train/' + s for s in sorted(os.listdir(self.root+'train/'))]
+        testpairs = ['test/' + s for s in sorted(os.listdir(self.root+'test/'))]
+        assert len(trainpairs) == 27 and len(testpairs) == 20, "incorrect parsing of pairs in ETH3D Low Res"
+        subvalpairs = ['train/delivery_area_3s','train/electro_3l','train/playground_3l']
+        assert all(p in trainpairs for p in subvalpairs)
+        subtrainpairs = [p for p in trainpairs if not p in subvalpairs]
+        assert len(subvalpairs)==3 and len(subtrainpairs)==24, "incorrect parsing of pairs in ETH3D Low Res"
+        tosave = {'train': trainpairs, 'test': testpairs, 'subtrain': subtrainpairs, 'subval': subvalpairs, 'all': trainpairs+testpairs}
+        return tosave
+
+    def submission_save_pairname(self, pairname, prediction, outdir, time):
+        assert prediction.ndim==2
+        assert prediction.dtype==np.float32
+        outfile = os.path.join(outdir, 'low_res_two_view', pairname.split('/')[1]+'.pfm')
+        os.makedirs( os.path.dirname(outfile), exist_ok=True)
+        writePFM(outfile, prediction)
+        timefile = outfile[:-4]+'.txt'
+        with open(timefile, 'w') as fid:
+            fid.write('runtime '+str(time))
+
+    def finalize_submission(self, outdir):
+        cmd = f'cd {outdir}/; zip -r "eth3d_low_res_two_view_results.zip" low_res_two_view'
+        print(cmd)
+        os.system(cmd)
+        print(f'Done. Submission file at {outdir}/eth3d_low_res_two_view_results.zip')
+
+class BoosterDataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "Booster"
+        self._set_root()
+        assert self.split in ['train_balanced','test_balanced','subtrain_balanced','subval_balanced'] # we use only the balanced version
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname)
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname).replace('/camera_00/','/camera_02/')
+        self.pairname_to_Ldispname = lambda pairname: osp.join(self.root, osp.dirname(pairname), '../disp_00.npy') # same images with different colors, same gt per sequence
+        self.pairname_to_str = lambda pairname: pairname[:-4].replace('/camera_00/','/')
+        self.load_disparity = _read_booster_disp
+        
+        
+    def _build_cache(self):
+        trainseqs = sorted(os.listdir(self.root+'train/balanced'))
+        trainpairs = ['train/balanced/'+s+'/camera_00/'+imname for s in trainseqs for imname in sorted(os.listdir(self.root+'train/balanced/'+s+'/camera_00/'))]
+        testpairs = ['test/balanced/'+s+'/camera_00/'+imname for s in sorted(os.listdir(self.root+'test/balanced')) for imname in sorted(os.listdir(self.root+'test/balanced/'+s+'/camera_00/'))]
+        assert len(trainpairs) == 228 and len(testpairs) == 191
+        subtrainpairs = [p for p in trainpairs if any(s in p for s in trainseqs[:-2])]
+        subvalpairs = [p for p in trainpairs if any(s in p for s in trainseqs[-2:])]
+        # warning: if we do validation split, we should split scenes!!!
+        tosave = {'train_balanced': trainpairs, 'test_balanced': testpairs, 'subtrain_balanced': subtrainpairs, 'subval_balanced': subvalpairs,}
+        return tosave
+        
+class SpringDataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "Spring"
+        self._set_root()
+        assert self.split in ['train', 'test', 'subtrain', 'subval']
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname+'.png')
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname+'.png').replace('frame_right','<frame_right>').replace('frame_left','frame_right').replace('<frame_right>','frame_left')
+        self.pairname_to_Ldispname = lambda pairname: None if pairname.startswith('test') else osp.join(self.root, pairname+'.dsp5').replace('frame_left','disp1_left').replace('frame_right','disp1_right')
+        self.pairname_to_str = lambda pairname: pairname
+        self.load_disparity = _read_hdf5_disp        
+        
+    def _build_cache(self):
+        trainseqs = sorted(os.listdir( osp.join(self.root,'train')))
+        trainpairs = [osp.join('train',s,'frame_left',f[:-4]) for s in trainseqs for f in sorted(os.listdir(osp.join(self.root,'train',s,'frame_left')))]
+        testseqs = sorted(os.listdir( osp.join(self.root,'test')))
+        testpairs = [osp.join('test',s,'frame_left',f[:-4]) for s in testseqs for f in sorted(os.listdir(osp.join(self.root,'test',s,'frame_left')))]
+        testpairs += [p.replace('frame_left','frame_right') for p in testpairs]
+        """maxnorm = {'0001': 32.88, '0002': 228.5, '0004': 298.2, '0005': 142.5, '0006': 113.6, '0007': 27.3, '0008': 554.5, '0009': 155.6, '0010': 126.1, '0011': 87.6, '0012': 303.2, '0013': 24.14, '0014': 82.56, '0015': 98.44, '0016': 156.9, '0017': 28.17, '0018': 21.03, '0020': 178.0, '0021': 58.06, '0022': 354.2, '0023': 8.79, '0024': 97.06, '0025': 55.16, '0026': 91.9, '0027': 156.6, '0030': 200.4, '0032': 58.66, '0033': 373.5, '0036': 149.4, '0037': 5.625, '0038': 37.0, '0039': 12.2, '0041': 453.5, '0043': 457.0, '0044': 379.5, '0045': 161.8, '0047': 105.44} # => let'use 0041"""
+        subtrainpairs = [p for p in trainpairs if p.split('/')[1]!='0041']
+        subvalpairs = [p for p in trainpairs if p.split('/')[1]=='0041']
+        assert len(trainpairs)==5000 and len(testpairs)==2000 and len(subtrainpairs)==4904 and len(subvalpairs)==96, "incorrect parsing of pairs in Spring"
+        tosave = {'train': trainpairs, 'test': testpairs, 'subtrain': subtrainpairs, 'subval': subvalpairs}
+        return tosave
+        
+    def submission_save_pairname(self, pairname, prediction, outdir, time):
+        assert prediction.ndim==2
+        assert prediction.dtype==np.float32
+        outfile = os.path.join(outdir, pairname+'.dsp5').replace('frame_left','disp1_left').replace('frame_right','disp1_right')
+        os.makedirs( os.path.dirname(outfile), exist_ok=True)
+        writeDsp5File(prediction, outfile)
+        
+    def finalize_submission(self, outdir):
+        assert self.split=='test'
+        exe = "{self.root}/disp1_subsampling"
+        if os.path.isfile(exe):
+            cmd = f'cd "{outdir}/test"; {exe} .'
+            print(cmd)
+            os.system(cmd)
+        else:
+            print('Could not find disp1_subsampling executable for submission.')
+            print('Please download it and run:')
+            print(f'cd "{outdir}/test"; <disp1_subsampling_exe> .')
+
+class Kitti12Dataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "Kitti12"
+        self._set_root()
+        assert self.split in ['train','test']
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname+'_10.png')
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname.replace('/colored_0/','/colored_1/')+'_10.png')
+        self.pairname_to_Ldispname = None if self.split=='test' else lambda pairname: osp.join(self.root, pairname.replace('/colored_0/','/disp_occ/')+'_10.png')
+        self.pairname_to_str = lambda pairname: pairname.replace('/colored_0/','/')
+        self.load_disparity = _read_kitti_disp
+        
+    def _build_cache(self):
+        trainseqs = ["training/colored_0/%06d"%(i) for i in range(194)]
+        testseqs = ["testing/colored_0/%06d"%(i) for i in range(195)]
+        assert len(trainseqs)==194 and len(testseqs)==195, "incorrect parsing of pairs in Kitti12"
+        tosave = {'train': trainseqs, 'test': testseqs}
+        return tosave 
+
+    def submission_save_pairname(self, pairname, prediction, outdir, time):
+        assert prediction.ndim==2
+        assert prediction.dtype==np.float32
+        outfile = os.path.join(outdir, pairname.split('/')[-1]+'_10.png')
+        os.makedirs( os.path.dirname(outfile), exist_ok=True)
+        img = (prediction * 256).astype('uint16')
+        Image.fromarray(img).save(outfile)
+
+    def finalize_submission(self, outdir):
+        assert self.split=='test'
+        cmd = f'cd {outdir}/; zip -r "kitti12_results.zip" .'
+        print(cmd)
+        os.system(cmd)
+        print(f'Done. Submission file at {outdir}/kitti12_results.zip')
+
+class Kitti15Dataset(StereoDataset):
+
+    def _prepare_data(self):
+        self.name = "Kitti15"
+        self._set_root()
+        assert self.split in ['train','subtrain','subval','test']
+        self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname+'_10.png')
+        self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname.replace('/image_2/','/image_3/')+'_10.png')
+        self.pairname_to_Ldispname = None if self.split=='test' else lambda pairname: osp.join(self.root, pairname.replace('/image_2/','/disp_occ_0/')+'_10.png')
+        self.pairname_to_str = lambda pairname: pairname.replace('/image_2/','/')
+        self.load_disparity = _read_kitti_disp
+        
+    def _build_cache(self):
+        trainseqs = ["training/image_2/%06d"%(i) for i in range(200)]
+        subtrainseqs = trainseqs[:-5]
+        subvalseqs = trainseqs[-5:]
+        testseqs = ["testing/image_2/%06d"%(i) for i in range(200)]
+        assert len(trainseqs)==200 and len(subtrainseqs)==195 and len(subvalseqs)==5 and len(testseqs)==200, "incorrect parsing of pairs in Kitti15"
+        tosave = {'train': trainseqs, 'subtrain': subtrainseqs, 'subval': subvalseqs, 'test': testseqs}
+        return tosave 
+
+    def submission_save_pairname(self, pairname, prediction, outdir, time):
+        assert prediction.ndim==2
+        assert prediction.dtype==np.float32
+        outfile = os.path.join(outdir, 'disp_0', pairname.split('/')[-1]+'_10.png')
+        os.makedirs( os.path.dirname(outfile), exist_ok=True)
+        img = (prediction * 256).astype('uint16')
+        Image.fromarray(img).save(outfile)
+
+    def finalize_submission(self, outdir):
+        assert self.split=='test'
+        cmd = f'cd {outdir}/; zip -r "kitti15_results.zip" disp_0'
+        print(cmd)
+        os.system(cmd)
+        print(f'Done. Submission file at {outdir}/kitti15_results.zip')
+
+
+### auxiliary functions
+
+def _read_img(filename):
+    # convert to RGB for scene flow finalpass data
+    img = np.asarray(Image.open(filename).convert('RGB'))
+    return img
+
+def _read_booster_disp(filename):
+    disp = np.load(filename)
+    disp[disp==0.0] = np.inf
+    return disp
+
+def _read_png_disp(filename, coef=1.0):
+    disp = np.asarray(Image.open(filename))
+    disp = disp.astype(np.float32) / coef
+    disp[disp==0.0] = np.inf
+    return disp 
+
+def _read_pfm_disp(filename):
+    disp = np.ascontiguousarray(_read_pfm(filename)[0])
+    disp[disp<=0] = np.inf # eg /nfs/data/ffs-3d/datasets/middlebury/2014/Shopvac-imperfect/disp0.pfm
+    return disp
+
+def _read_npy_disp(filename):
+    return np.load(filename)
+
+def _read_crestereo_disp(filename): return _read_png_disp(filename, coef=32.0)
+def _read_middlebury20052006_disp(filename): return _read_png_disp(filename, coef=1.0)
+def _read_kitti_disp(filename): return _read_png_disp(filename, coef=256.0)
+_read_sceneflow_disp = _read_pfm_disp
+_read_eth3d_disp = _read_pfm_disp
+_read_middlebury_disp = _read_pfm_disp
+_read_carla_disp = _read_pfm_disp
+_read_tartanair_disp = _read_npy_disp
+    
+def _read_hdf5_disp(filename):
+    disp = np.asarray(h5py.File(filename)['disparity'])
+    disp[np.isnan(disp)] = np.inf # make invalid values as +inf
+    #disp[disp==0.0] = np.inf # make invalid values as +inf
+    return disp.astype(np.float32)
+    
+import re
+def _read_pfm(file):
+    file = open(file, 'rb')
+
+    color = None
+    width = None
+    height = None
+    scale = None
+    endian = None
+
+    header = file.readline().rstrip()
+    if header.decode("ascii") == 'PF':
+        color = True
+    elif header.decode("ascii") == 'Pf':
+        color = False
+    else:
+        raise Exception('Not a PFM file.')
+
+    dim_match = re.match(r'^(\d+)\s(\d+)\s$', file.readline().decode("ascii"))
+    if dim_match:
+        width, height = list(map(int, dim_match.groups()))
+    else:
+        raise Exception('Malformed PFM header.')
+
+    scale = float(file.readline().decode("ascii").rstrip())
+    if scale < 0:  # little-endian
+        endian = '<'
+        scale = -scale
+    else:
+        endian = '>'  # big-endian
+
+    data = np.fromfile(file, endian + 'f')
+    shape = (height, width, 3) if color else (height, width)
+
+    data = np.reshape(data, shape)
+    data = np.flipud(data)
+    return data, scale
+
+def writePFM(file, image, scale=1):
+    file = open(file, 'wb')
+
+    color = None
+
+    if image.dtype.name != 'float32':
+        raise Exception('Image dtype must be float32.')
+
+    image = np.flipud(image)
+
+    if len(image.shape) == 3 and image.shape[2] == 3:  # color image
+        color = True
+    elif len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1:  # greyscale
+        color = False
+    else:
+        raise Exception('Image must have H x W x 3, H x W x 1 or H x W dimensions.')
+
+    file.write('PF\n' if color else 'Pf\n'.encode())
+    file.write('%d %d\n'.encode() % (image.shape[1], image.shape[0]))
+
+    endian = image.dtype.byteorder
+
+    if endian == '<' or endian == '=' and sys.byteorder == 'little':
+        scale = -scale
+
+    file.write('%f\n'.encode() % scale)
+
+    image.tofile(file)
+
+def writeDsp5File(disp, filename):
+    with h5py.File(filename, "w") as f:
+        f.create_dataset("disparity", data=disp, compression="gzip", compression_opts=5)
+
+
+# disp visualization
+
+def vis_disparity(disp, m=None, M=None):
+    if m is None: m = disp.min()
+    if M is None: M = disp.max()
+    disp_vis = (disp - m) / (M-m) * 255.0
+    disp_vis = disp_vis.astype("uint8")
+    disp_vis = cv2.applyColorMap(disp_vis, cv2.COLORMAP_INFERNO)
+    return disp_vis
+
+# dataset getter 
+    
+def get_train_dataset_stereo(dataset_str, augmentor=True, crop_size=None):
+    dataset_str = dataset_str.replace('(','Dataset(')
+    if augmentor:
+        dataset_str = dataset_str.replace(')',', augmentor=True)')
+    if crop_size is not None:
+        dataset_str = dataset_str.replace(')',', crop_size={:s})'.format(str(crop_size)))
+    return eval(dataset_str)
+    
+def get_test_datasets_stereo(dataset_str):
+    dataset_str = dataset_str.replace('(','Dataset(')
+    return [eval(s) for s in dataset_str.split('+')]

third_party/dust3r/croco/stereoflow/download_model.sh

@@ -0,0 +1,12 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+model=$1
+outfile="stereoflow_models/${model}"
+if [[ ! -f $outfile ]]
+then
+	mkdir -p stereoflow_models/;
+	wget https://download.europe.naverlabs.com/ComputerVision/CroCo/StereoFlow_models/$1 -P stereoflow_models/;
+else
+	echo "Model ${model} already downloaded in ${outfile}."
+fi

third_party/dust3r/croco/stereoflow/engine.py

@@ -0,0 +1,280 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+# --------------------------------------------------------
+# Main function for training one epoch or testing
+# --------------------------------------------------------
+
+import math
+import sys
+from typing import Iterable
+import numpy as np
+import torch
+import torchvision
+
+from utils import misc as misc
+
+
+def split_prediction_conf(predictions, with_conf=False):
+    if not with_conf:
+        return predictions, None
+    conf = predictions[:,-1:,:,:]
+    predictions = predictions[:,:-1,:,:]
+    return predictions, conf
+
+def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module, metrics: torch.nn.Module,
+                    data_loader: Iterable, optimizer: torch.optim.Optimizer,
+                    device: torch.device, epoch: int, loss_scaler,
+                    log_writer=None, print_freq = 20,
+                    args=None):
+    model.train(True)
+    metric_logger = misc.MetricLogger(delimiter="  ")
+    metric_logger.add_meter('lr', misc.SmoothedValue(window_size=1, fmt='{value:.6f}'))
+    header = 'Epoch: [{}]'.format(epoch)
+
+    accum_iter = args.accum_iter
+
+    optimizer.zero_grad()
+
+    details = {}
+
+    if log_writer is not None:
+        print('log_dir: {}'.format(log_writer.log_dir))
+
+    if args.img_per_epoch:
+        iter_per_epoch = args.img_per_epoch // args.batch_size + int(args.img_per_epoch % args.batch_size > 0)
+        assert len(data_loader) >= iter_per_epoch, 'Dataset is too small for so many iterations'
+        len_data_loader = iter_per_epoch
+    else:
+        len_data_loader, iter_per_epoch = len(data_loader), None
+
+    for data_iter_step, (image1, image2, gt, pairname) in enumerate(metric_logger.log_every(data_loader, print_freq, header, max_iter=iter_per_epoch)):
+        
+        image1 = image1.to(device, non_blocking=True)
+        image2 = image2.to(device, non_blocking=True)
+        gt = gt.to(device, non_blocking=True)
+        
+        # we use a per iteration (instead of per epoch) lr scheduler
+        if data_iter_step % accum_iter == 0:
+            misc.adjust_learning_rate(optimizer, data_iter_step / len_data_loader + epoch, args)
+
+        with torch.cuda.amp.autocast(enabled=bool(args.amp)):
+            prediction = model(image1, image2)
+            prediction, conf = split_prediction_conf(prediction, criterion.with_conf)
+            batch_metrics = metrics(prediction.detach(), gt)
+            loss = criterion(prediction, gt) if conf is None else criterion(prediction, gt, conf)
+            
+        loss_value = loss.item()
+        if not math.isfinite(loss_value):
+            print("Loss is {}, stopping training".format(loss_value))
+            sys.exit(1)
+
+        loss /= accum_iter
+        loss_scaler(loss, optimizer, parameters=model.parameters(),
+                    update_grad=(data_iter_step + 1) % accum_iter == 0)
+        if (data_iter_step + 1) % accum_iter == 0:
+            optimizer.zero_grad()
+
+        torch.cuda.synchronize()
+        
+        metric_logger.update(loss=loss_value)
+        for k,v in batch_metrics.items():
+            metric_logger.update(**{k: v.item()})
+        lr = optimizer.param_groups[0]["lr"]
+        metric_logger.update(lr=lr)
+
+        #if args.dsitributed: loss_value_reduce = misc.all_reduce_mean(loss_value)
+        time_to_log = ((data_iter_step + 1) % (args.tboard_log_step * accum_iter) == 0 or data_iter_step == len_data_loader-1)
+        loss_value_reduce = misc.all_reduce_mean(loss_value)
+        if log_writer is not None and time_to_log:
+            epoch_1000x = int((data_iter_step / len_data_loader + epoch) * 1000)
+            # We use epoch_1000x as the x-axis in tensorboard. This calibrates different curves when batch size changes.
+            log_writer.add_scalar('train/loss', loss_value_reduce, epoch_1000x)
+            log_writer.add_scalar('lr', lr, epoch_1000x)
+            for k,v in batch_metrics.items():
+                log_writer.add_scalar('train/'+k, v.item(), epoch_1000x)
+
+    # gather the stats from all processes
+    #if args.distributed: metric_logger.synchronize_between_processes()
+    print("Averaged stats:", metric_logger)
+    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
+
+
+@torch.no_grad()
+def validate_one_epoch(model: torch.nn.Module,
+                   criterion: torch.nn.Module,
+                   metrics: torch.nn.Module,
+                   data_loaders: list[Iterable],
+                   device: torch.device,
+                   epoch: int,
+                   log_writer=None,
+                   args=None):
+
+    model.eval()
+    metric_loggers = []
+    header = 'Epoch: [{}]'.format(epoch)
+    print_freq = 20
+
+    conf_mode = args.tile_conf_mode
+    crop = args.crop
+    
+    if log_writer is not None:
+        print('log_dir: {}'.format(log_writer.log_dir))
+
+    results = {}
+    dnames = []
+    image1, image2, gt, prediction = None, None, None, None
+    for didx, data_loader in enumerate(data_loaders):
+        dname = str(data_loader.dataset)
+        dnames.append(dname)
+        metric_loggers.append(misc.MetricLogger(delimiter="  "))
+        for data_iter_step, (image1, image2, gt, pairname) in enumerate(metric_loggers[didx].log_every(data_loader, print_freq, header)):
+            image1 = image1.to(device, non_blocking=True)
+            image2 = image2.to(device, non_blocking=True)
+            gt = gt.to(device, non_blocking=True)
+            if dname.startswith('Spring'):
+                assert gt.size(2)==image1.size(2)*2 and gt.size(3)==image1.size(3)*2
+                gt = (gt[:,:,0::2,0::2] + gt[:,:,0::2,1::2] + gt[:,:,1::2,0::2] + gt[:,:,1::2,1::2] ) / 4.0 # we approximate the gt based on the 2x upsampled ones
+
+            with torch.inference_mode():
+                prediction, tiled_loss, c = tiled_pred(model, criterion, image1, image2, gt, conf_mode=conf_mode, overlap=args.val_overlap, crop=crop, with_conf=criterion.with_conf)
+                batch_metrics = metrics(prediction.detach(), gt)
+                loss = criterion(prediction.detach(), gt) if not criterion.with_conf else criterion(prediction.detach(), gt, c)
+                loss_value = loss.item()
+                metric_loggers[didx].update(loss_tiled=tiled_loss.item())
+                metric_loggers[didx].update(**{f'loss': loss_value})
+                for k,v in batch_metrics.items():
+                    metric_loggers[didx].update(**{dname+'_' + k: v.item()})
+        
+    results = {k: meter.global_avg for ml in metric_loggers for k, meter in ml.meters.items()}
+    if len(dnames)>1:
+        for k in batch_metrics.keys():
+            results['AVG_'+k] = sum(results[dname+'_'+k] for dname in dnames) / len(dnames)
+            
+    if log_writer is not None :
+        epoch_1000x = int((1 + epoch) * 1000)
+        for k,v in results.items():
+            log_writer.add_scalar('val/'+k, v, epoch_1000x)
+
+    print("Averaged stats:", results)
+    return results
+
+import torch.nn.functional as F
+def _resize_img(img, new_size):
+    return F.interpolate(img, size=new_size, mode='bicubic', align_corners=False)
+def _resize_stereo_or_flow(data, new_size):
+    assert data.ndim==4
+    assert data.size(1) in [1,2]
+    scale_x = new_size[1]/float(data.size(3))
+    out = F.interpolate(data, size=new_size, mode='bicubic', align_corners=False)
+    out[:,0,:,:] *= scale_x
+    if out.size(1)==2:
+        scale_y = new_size[0]/float(data.size(2))        
+        out[:,1,:,:] *= scale_y
+        print(scale_x, new_size, data.shape)
+    return out
+    
+
+@torch.no_grad()
+def tiled_pred(model, criterion, img1, img2, gt,
+               overlap=0.5, bad_crop_thr=0.05,
+               downscale=False, crop=512, ret='loss',
+               conf_mode='conf_expsigmoid_10_5', with_conf=False, 
+               return_time=False):
+                     
+    # for each image, we are going to run inference on many overlapping patches
+    # then, all predictions will be weighted-averaged
+    if gt is not None:
+        B, C, H, W = gt.shape
+    else:
+        B, _, H, W = img1.shape
+        C = model.head.num_channels-int(with_conf)
+    win_height, win_width = crop[0], crop[1]
+    
+    # upscale to be larger than the crop
+    do_change_scale =  H<win_height or W<win_width
+    if do_change_scale: 
+        upscale_factor = max(win_width/W, win_height/W)
+        original_size = (H,W)
+        new_size = (round(H*upscale_factor),round(W*upscale_factor))
+        img1 = _resize_img(img1, new_size)
+        img2 = _resize_img(img2, new_size)
+        # resize gt just for the computation of tiled losses
+        if gt is not None: gt = _resize_stereo_or_flow(gt, new_size)
+        H,W = img1.shape[2:4]
+        
+    if conf_mode.startswith('conf_expsigmoid_'): # conf_expsigmoid_30_10
+        beta, betasigmoid = map(float, conf_mode[len('conf_expsigmoid_'):].split('_'))
+    elif conf_mode.startswith('conf_expbeta'): # conf_expbeta3
+        beta = float(conf_mode[len('conf_expbeta'):])
+    else:
+        raise NotImplementedError(f"conf_mode {conf_mode} is not implemented")
+
+    def crop_generator():
+        for sy in _overlapping(H, win_height, overlap):
+          for sx in _overlapping(W, win_width, overlap):
+            yield sy, sx, sy, sx, True
+
+    # keep track of weighted sum of prediction*weights and weights
+    accu_pred = img1.new_zeros((B, C, H, W)) # accumulate the weighted sum of predictions 
+    accu_conf = img1.new_zeros((B, H, W)) + 1e-16 # accumulate the weights 
+    accu_c = img1.new_zeros((B, H, W)) # accumulate the weighted sum of confidences ; not so useful except for computing some losses
+
+    tiled_losses = []
+    
+    if return_time:
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+        start.record()
+
+    for sy1, sx1, sy2, sx2, aligned in crop_generator():
+        # compute optical flow there
+        pred =  model(_crop(img1,sy1,sx1), _crop(img2,sy2,sx2))
+        pred, predconf = split_prediction_conf(pred, with_conf=with_conf)
+        
+        if gt is not None: gtcrop = _crop(gt,sy1,sx1)
+        if criterion is not None and gt is not None: 
+            tiled_losses.append( criterion(pred, gtcrop).item() if predconf is None else criterion(pred, gtcrop, predconf).item() )
+        
+        if conf_mode.startswith('conf_expsigmoid_'):
+            conf = torch.exp(- beta * 2 * (torch.sigmoid(predconf / betasigmoid) - 0.5)).view(B,win_height,win_width)
+        elif conf_mode.startswith('conf_expbeta'):
+            conf = torch.exp(- beta * predconf).view(B,win_height,win_width)
+        else:
+            raise NotImplementedError
+                        
+        accu_pred[...,sy1,sx1] += pred * conf[:,None,:,:]
+        accu_conf[...,sy1,sx1] += conf
+        accu_c[...,sy1,sx1] += predconf.view(B,win_height,win_width) * conf 
+        
+    pred = accu_pred / accu_conf[:, None,:,:]
+    c = accu_c / accu_conf
+    assert not torch.any(torch.isnan(pred))
+
+    if return_time:
+        end.record()
+        torch.cuda.synchronize()
+        time = start.elapsed_time(end)/1000.0 # this was in milliseconds
+
+    if do_change_scale:
+        pred = _resize_stereo_or_flow(pred, original_size)
+    
+    if return_time:
+        return pred, torch.mean(torch.tensor(tiled_losses)), c, time
+    return pred, torch.mean(torch.tensor(tiled_losses)), c
+
+
+def _overlapping(total, window, overlap=0.5):
+    assert total >= window and 0 <= overlap < 1, (total, window, overlap)
+    num_windows = 1 + int(np.ceil( (total - window) / ((1-overlap) * window) ))
+    offsets = np.linspace(0, total-window, num_windows).round().astype(int)
+    yield from (slice(x, x+window) for x in offsets)
+
+def _crop(img, sy, sx):
+    B, THREE, H, W = img.shape
+    if 0 <= sy.start and sy.stop <= H and 0 <= sx.start and sx.stop <= W:
+        return img[:,:,sy,sx]
+    l, r = max(0,-sx.start), max(0,sx.stop-W)
+    t, b = max(0,-sy.start), max(0,sy.stop-H)
+    img = torch.nn.functional.pad(img, (l,r,t,b), mode='constant')
+    return img[:, :, slice(sy.start+t,sy.stop+t), slice(sx.start+l,sx.stop+l)]

third_party/dust3r/croco/stereoflow/test.py

@@ -0,0 +1,216 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+# --------------------------------------------------------
+# Main test function
+# --------------------------------------------------------
+
+import os
+import argparse
+import pickle
+from PIL import Image
+import numpy as np
+from tqdm import tqdm
+
+import torch
+from torch.utils.data import DataLoader
+
+import utils.misc as misc
+from models.croco_downstream import CroCoDownstreamBinocular
+from models.head_downstream import PixelwiseTaskWithDPT
+
+from stereoflow.criterion import *
+from stereoflow.datasets_stereo import get_test_datasets_stereo
+from stereoflow.datasets_flow import get_test_datasets_flow
+from stereoflow.engine import tiled_pred
+
+from stereoflow.datasets_stereo import vis_disparity
+from stereoflow.datasets_flow import flowToColor
+
+def get_args_parser():
+    parser = argparse.ArgumentParser('Test CroCo models on stereo/flow', add_help=False)
+    # important argument 
+    parser.add_argument('--model', required=True, type=str, help='Path to the model to evaluate')
+    parser.add_argument('--dataset', required=True, type=str, help="test dataset (there can be multiple dataset separated by a +)")
+    # tiling 
+    parser.add_argument('--tile_conf_mode', type=str, default='', help='Weights for the tiling aggregation based on confidence (empty means use the formula from the loaded checkpoint')
+    parser.add_argument('--tile_overlap', type=float, default=0.7, help='overlap between tiles')
+    # save (it will automatically go to <model_path>_<dataset_str>/<tile_str>_<save>)
+    parser.add_argument('--save', type=str, nargs='+', default=[], 
+                        help='what to save: \
+                              metrics (pickle file), \
+                              pred (raw prediction save as torch tensor), \
+                              visu (visualization in png of each prediction), \
+                              err10 (visualization in png of the error clamp at 10 for each prediction), \
+                              submission (submission file)')
+    # other (no impact)
+    parser.add_argument('--num_workers', default=4, type=int)
+    return parser
+    
+    
+def _load_model_and_criterion(model_path, do_load_metrics, device):
+    print('loading model from', model_path)
+    assert os.path.isfile(model_path)
+    ckpt = torch.load(model_path, 'cpu')
+    
+    ckpt_args = ckpt['args']
+    task = ckpt_args.task
+    tile_conf_mode = ckpt_args.tile_conf_mode
+    num_channels = {'stereo': 1, 'flow': 2}[task]
+    with_conf =  eval(ckpt_args.criterion).with_conf
+    if with_conf: num_channels += 1
+    print('head: PixelwiseTaskWithDPT()')
+    head = PixelwiseTaskWithDPT()
+    head.num_channels = num_channels
+    print('croco_args:', ckpt_args.croco_args)
+    model = CroCoDownstreamBinocular(head, **ckpt_args.croco_args)
+    msg = model.load_state_dict(ckpt['model'], strict=True)
+    model.eval()
+    model = model.to(device)
+    
+    if do_load_metrics:
+        if task=='stereo':
+            metrics = StereoDatasetMetrics().to(device)
+        else:
+            metrics = FlowDatasetMetrics().to(device)
+    else:
+        metrics = None
+    
+    return model, metrics, ckpt_args.crop, with_conf, task, tile_conf_mode
+    
+    
+def _save_batch(pred, gt, pairnames, dataset, task, save, outdir, time, submission_dir=None):
+
+    for i in range(len(pairnames)):
+        
+        pairname = eval(pairnames[i]) if pairnames[i].startswith('(') else pairnames[i] # unbatch pairname 
+        fname = os.path.join(outdir, dataset.pairname_to_str(pairname))
+        os.makedirs(os.path.dirname(fname), exist_ok=True)
+        
+        predi = pred[i,...]
+        if gt is not None: gti = gt[i,...]
+        
+        if 'pred' in save:
+            torch.save(predi.squeeze(0).cpu(), fname+'_pred.pth')
+            
+        if 'visu' in save:
+            if task=='stereo':
+                disparity = predi.permute((1,2,0)).squeeze(2).cpu().numpy()
+                m,M = None
+                if gt is not None:
+                    mask = torch.isfinite(gti)
+                    m = gt[mask].min()
+                    M = gt[mask].max()
+                img_disparity = vis_disparity(disparity, m=m, M=M)
+                Image.fromarray(img_disparity).save(fname+'_pred.png')
+            else:
+                # normalize flowToColor according to the maxnorm of gt (or prediction if not available)
+                flowNorm = torch.sqrt(torch.sum( (gti if gt is not None else predi)**2, dim=0)).max().item()
+                imgflow = flowToColor(predi.permute((1,2,0)).cpu().numpy(), maxflow=flowNorm)
+                Image.fromarray(imgflow).save(fname+'_pred.png')
+                
+        if 'err10' in save:
+            assert gt is not None
+            L2err = torch.sqrt(torch.sum( (gti-predi)**2, dim=0))
+            valid = torch.isfinite(gti[0,:,:])
+            L2err[~valid] = 0.0
+            L2err = torch.clamp(L2err, max=10.0)
+            red = (L2err*255.0/10.0).to(dtype=torch.uint8)[:,:,None]
+            zer = torch.zeros_like(red)
+            imgerr = torch.cat( (red,zer,zer), dim=2).cpu().numpy()
+            Image.fromarray(imgerr).save(fname+'_err10.png')
+            
+        if 'submission' in save:
+            assert submission_dir is not None
+            predi_np = predi.permute(1,2,0).squeeze(2).cpu().numpy() # transform into HxWx2 for flow or HxW for stereo
+            dataset.submission_save_pairname(pairname, predi_np, submission_dir, time)
+
+def main(args):
+        
+    # load the pretrained model and metrics
+    device = torch.device('cuda:0') if torch.cuda.is_available() else torch.device('cpu')
+    model, metrics, cropsize, with_conf, task, tile_conf_mode = _load_model_and_criterion(args.model, 'metrics' in args.save, device)
+    if args.tile_conf_mode=='': args.tile_conf_mode = tile_conf_mode
+    
+    # load the datasets 
+    datasets = (get_test_datasets_stereo if task=='stereo' else get_test_datasets_flow)(args.dataset)
+    dataloaders = [DataLoader(dataset, batch_size=1, shuffle=False, num_workers=args.num_workers, pin_memory=True, drop_last=False) for dataset in datasets]    
+       
+    # run
+    for i,dataloader in enumerate(dataloaders):
+        dataset = datasets[i]
+        dstr = args.dataset.split('+')[i]
+        
+        outdir = args.model+'_'+misc.filename(dstr)
+        if 'metrics' in args.save and len(args.save)==1:
+            fname = os.path.join(outdir, f'conf_{args.tile_conf_mode}_overlap_{args.tile_overlap}.pkl')
+            if os.path.isfile(fname) and len(args.save)==1:
+                print('  metrics already compute in '+fname)
+                with open(fname, 'rb') as fid:
+                    results = pickle.load(fid)
+                for k,v in results.items():
+                    print('{:s}: {:.3f}'.format(k, v))
+                continue
+                        
+        if 'submission' in args.save:
+            dirname = f'submission_conf_{args.tile_conf_mode}_overlap_{args.tile_overlap}'
+            submission_dir = os.path.join(outdir, dirname)
+        else:
+            submission_dir = None
+           
+        print('')
+        print('saving {:s} in {:s}'.format('+'.join(args.save), outdir))
+        print(repr(dataset))
+    
+        if metrics is not None: 
+            metrics.reset()
+                
+        for data_iter_step, (image1, image2, gt, pairnames) in enumerate(tqdm(dataloader)):
+        
+            do_flip = (task=='stereo' and dstr.startswith('Spring') and any("right" in p for p in pairnames)) # we flip the images and will flip the prediction after as we assume img1 is on the left 
+            
+            image1 = image1.to(device, non_blocking=True)
+            image2 = image2.to(device, non_blocking=True)
+            gt = gt.to(device, non_blocking=True) if gt.numel()>0 else None # special case for test time
+            if do_flip:
+                assert all("right" in p for p in pairnames) 
+                image1 = image1.flip(dims=[3]) # this is already the right frame, let's flip it
+                image2 = image2.flip(dims=[3])
+                gt = gt # that is ok
+                        
+            with torch.inference_mode():
+                pred, _, _, time = tiled_pred(model, None, image1, image2, None if dataset.name=='Spring' else gt, conf_mode=args.tile_conf_mode, overlap=args.tile_overlap, crop=cropsize, with_conf=with_conf, return_time=True)
+
+                if do_flip:
+                    pred = pred.flip(dims=[3])
+                
+                if metrics is not None: 
+                    metrics.add_batch(pred, gt)
+                
+                if any(k in args.save for k in ['pred','visu','err10','submission']):
+                    _save_batch(pred, gt, pairnames, dataset, task, args.save, outdir, time, submission_dir=submission_dir)                
+            
+
+        # print 
+        if metrics is not None: 
+            results = metrics.get_results()
+            for k,v in results.items():
+                print('{:s}: {:.3f}'.format(k, v))
+                
+        # save if needed
+        if 'metrics' in args.save:
+            os.makedirs(os.path.dirname(fname), exist_ok=True)
+            with open(fname, 'wb') as fid:
+                pickle.dump(results, fid)
+            print('metrics saved in', fname)
+            
+        # finalize submission if needed
+        if 'submission' in args.save:
+            dataset.finalize_submission(submission_dir)
+                
+        
+            
+if __name__ == '__main__':
+    args = get_args_parser()
+    args = args.parse_args()
+    main(args)

third_party/dust3r/croco/stereoflow/train.py

@@ -0,0 +1,253 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+
+# --------------------------------------------------------
+# Main training function
+# --------------------------------------------------------
+
+import argparse
+import datetime
+import json
+import numpy as np
+import os
+import sys
+import time
+
+import torch
+import torch.distributed as dist
+import torch.backends.cudnn as cudnn
+from torch.utils.tensorboard import SummaryWriter
+import torchvision.transforms as transforms
+import torchvision.datasets as datasets
+from torch.utils.data import DataLoader
+
+import utils
+import utils.misc as misc
+from utils.misc import NativeScalerWithGradNormCount as NativeScaler
+from models.croco_downstream import CroCoDownstreamBinocular, croco_args_from_ckpt
+from models.pos_embed import interpolate_pos_embed
+from models.head_downstream import PixelwiseTaskWithDPT
+
+from stereoflow.datasets_stereo import get_train_dataset_stereo, get_test_datasets_stereo
+from stereoflow.datasets_flow import get_train_dataset_flow, get_test_datasets_flow
+from stereoflow.engine import train_one_epoch, validate_one_epoch
+from stereoflow.criterion import *
+
+
+def get_args_parser():
+    # prepare subparsers 
+    parser = argparse.ArgumentParser('Finetuning CroCo models on stereo or flow', add_help=False)
+    subparsers = parser.add_subparsers(title="Task (stereo or flow)", dest="task", required=True)
+    parser_stereo = subparsers.add_parser('stereo', help='Training stereo model')
+    parser_flow = subparsers.add_parser('flow', help='Training flow model')
+    def add_arg(name_or_flags, default=None, default_stereo=None, default_flow=None, **kwargs):
+        if default is not None: assert default_stereo is None and default_flow is None, "setting default makes default_stereo and default_flow disabled"
+        parser_stereo.add_argument(name_or_flags, default=default if default is not None else default_stereo, **kwargs)
+        parser_flow.add_argument(name_or_flags, default=default if default is not None else default_flow, **kwargs)
+    # output dir 
+    add_arg('--output_dir', required=True, type=str, help='path where to save, if empty, automatically created')
+    # model
+    add_arg('--crop', type=int, nargs = '+', default_stereo=[352, 704], default_flow=[320, 384], help = "size of the random image crops used during training.")
+    add_arg('--pretrained', required=True, type=str, help="Load pretrained model (required as croco arguments come from there)")
+    # criterion  
+    add_arg('--criterion', default_stereo='LaplacianLossBounded2()', default_flow='LaplacianLossBounded()', type=str, help='string to evaluate to get criterion')
+    add_arg('--bestmetric', default_stereo='avgerr', default_flow='EPE', type=str)
+    # dataset 
+    add_arg('--dataset', type=str, required=True, help="training set")
+    # training 
+    add_arg('--seed', default=0, type=int, help='seed')
+    add_arg('--batch_size', default_stereo=6, default_flow=8, type=int, help='Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus')
+    add_arg('--epochs', default=32, type=int, help='number of training epochs')
+    add_arg('--img_per_epoch', type=int, default=None, help='Fix the number of images seen in an epoch (None means use all training pairs)')
+    add_arg('--accum_iter', default=1, type=int, help='Accumulate gradient iterations (for increasing the effective batch size under memory constraints)')
+    add_arg('--weight_decay', type=float, default=0.05, help='weight decay (default: 0.05)')
+    add_arg('--lr', type=float, default_stereo=3e-5, default_flow=2e-5, metavar='LR', help='learning rate (absolute lr)')
+    add_arg('--min_lr', type=float, default=0., metavar='LR', help='lower lr bound for cyclic schedulers that hit 0')
+    add_arg('--warmup_epochs', type=int, default=1, metavar='N', help='epochs to warmup LR')
+    add_arg('--optimizer', default='AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))', type=str,
+                        help="Optimizer from torch.optim [ default: AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95)) ]")
+    add_arg('--amp', default=0, type=int, choices=[0,1], help='enable automatic mixed precision training')
+    # validation
+    add_arg('--val_dataset', type=str, default='', help="Validation sets, multiple separated by + (empty string means that no validation is performed)")
+    add_arg('--tile_conf_mode', type=str, default_stereo='conf_expsigmoid_15_3', default_flow='conf_expsigmoid_10_5', help='Weights for tile aggregation')
+    add_arg('--val_overlap', default=0.7, type=float, help='Overlap value for the tiling')
+    # others
+    add_arg('--num_workers', default=8, type=int)
+    add_arg('--eval_every', type=int, default=1, help='Val loss evaluation frequency')
+    add_arg('--save_every', type=int, default=1, help='Save checkpoint frequency')
+    add_arg('--start_from', type=str, default=None, help='Start training using weights from an other model (eg for finetuning)')
+    add_arg('--tboard_log_step', type=int, default=100, help='Log to tboard every so many steps')
+    add_arg('--dist_url', default='env://', help='url used to set up distributed training')
+
+    return parser
+    
+        
+def main(args):
+    misc.init_distributed_mode(args)
+    global_rank = misc.get_rank()
+    num_tasks = misc.get_world_size()
+
+    assert os.path.isfile(args.pretrained)
+    print("output_dir: "+args.output_dir)
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    # fix the seed for reproducibility
+    seed = args.seed + misc.get_rank()
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+    cudnn.benchmark = True
+
+    # Metrics / criterion 
+    device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+    metrics = (StereoMetrics if args.task=='stereo' else FlowMetrics)().to(device)
+    criterion = eval(args.criterion).to(device)
+    print('Criterion: ', args.criterion)
+
+    # Prepare model
+    assert os.path.isfile(args.pretrained)
+    ckpt = torch.load(args.pretrained, 'cpu')
+    croco_args = croco_args_from_ckpt(ckpt)
+    croco_args['img_size'] = (args.crop[0], args.crop[1])
+    print('Croco args: '+str(croco_args))
+    args.croco_args = croco_args # saved for test time 
+    # prepare head 
+    num_channels = {'stereo': 1, 'flow': 2}[args.task]
+    if criterion.with_conf: num_channels += 1
+    print(f'Building head PixelwiseTaskWithDPT() with {num_channels} channel(s)')
+    head = PixelwiseTaskWithDPT()
+    head.num_channels = num_channels
+    # build model and load pretrained weights
+    model = CroCoDownstreamBinocular(head, **croco_args)
+    interpolate_pos_embed(model, ckpt['model'])
+    msg = model.load_state_dict(ckpt['model'], strict=False)
+    print(msg)
+
+    total_params = sum(p.numel() for p in model.parameters())
+    total_params_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    print(f"Total params: {total_params}")
+    print(f"Total params trainable: {total_params_trainable}")
+    model_without_ddp = model.to(device)
+
+    eff_batch_size = args.batch_size * args.accum_iter * misc.get_world_size()
+    print("lr: %.2e" % args.lr)
+    print("accumulate grad iterations: %d" % args.accum_iter)
+    print("effective batch size: %d" % eff_batch_size)
+
+    if args.distributed:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], static_graph=True)
+        model_without_ddp = model.module
+    
+    # following timm: set wd as 0 for bias and norm layers   
+    param_groups = misc.get_parameter_groups(model_without_ddp, args.weight_decay)
+    optimizer = eval(f"torch.optim.{args.optimizer}")
+    print(optimizer)
+    loss_scaler = NativeScaler()
+    
+    # automatic restart
+    last_ckpt_fname = os.path.join(args.output_dir, f'checkpoint-last.pth')
+    args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None
+
+    if not args.resume and args.start_from:
+        print(f"Starting from an other model's weights: {args.start_from}")
+        best_so_far = None
+        args.start_epoch = 0
+        ckpt = torch.load(args.start_from, 'cpu')
+        msg = model_without_ddp.load_state_dict(ckpt['model'], strict=False)
+        print(msg)
+    else:
+        best_so_far = misc.load_model(args=args, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler)
+
+    if best_so_far is None: best_so_far = np.inf
+    
+    # tensorboard
+    log_writer = None
+    if global_rank == 0 and args.output_dir is not None:
+        log_writer = SummaryWriter(log_dir=args.output_dir, purge_step=args.start_epoch*1000)
+
+    #  dataset and loader 
+    print('Building Train Data loader for dataset: ', args.dataset)
+    train_dataset = (get_train_dataset_stereo if args.task=='stereo' else get_train_dataset_flow)(args.dataset, crop_size=args.crop)
+    def _print_repr_dataset(d):
+        if isinstance(d, torch.utils.data.dataset.ConcatDataset):
+            for dd in d.datasets:
+                _print_repr_dataset(dd)
+        else:
+            print(repr(d))
+    _print_repr_dataset(train_dataset)
+    print('  total length:', len(train_dataset))
+    if args.distributed:
+        sampler_train = torch.utils.data.DistributedSampler(
+            train_dataset, num_replicas=num_tasks, rank=global_rank, shuffle=True
+        )
+    else:
+        sampler_train = torch.utils.data.RandomSampler(train_dataset)
+    data_loader_train = torch.utils.data.DataLoader(
+        train_dataset, sampler=sampler_train,
+        batch_size=args.batch_size,
+        num_workers=args.num_workers,
+        pin_memory=True,
+        drop_last=True,
+    )
+    if args.val_dataset=='':
+        data_loaders_val = None
+    else:
+        print('Building Val Data loader for datasets: ', args.val_dataset)
+        val_datasets = (get_test_datasets_stereo if args.task=='stereo' else get_test_datasets_flow)(args.val_dataset)
+        for val_dataset in val_datasets: print(repr(val_dataset))
+        data_loaders_val = [DataLoader(val_dataset, batch_size=1, shuffle=False, num_workers=args.num_workers, pin_memory=True, drop_last=False) for val_dataset in val_datasets]
+        bestmetric = ("AVG_" if len(data_loaders_val)>1 else str(data_loaders_val[0].dataset)+'_')+args.bestmetric
+       
+    print(f"Start training for {args.epochs} epochs")
+    start_time = time.time()
+    # Training Loop
+    for epoch in range(args.start_epoch, args.epochs):
+
+        if args.distributed: data_loader_train.sampler.set_epoch(epoch)
+            
+        # Train
+        epoch_start = time.time()
+        train_stats = train_one_epoch(model, criterion, metrics, data_loader_train, optimizer, device, epoch, loss_scaler, log_writer=log_writer, args=args)
+        epoch_time = time.time() - epoch_start
+
+        if args.distributed: dist.barrier()
+
+        # Validation (current naive implementation runs the validation on every gpu ... not smart ...)
+        if data_loaders_val is not None and args.eval_every > 0 and (epoch+1) % args.eval_every == 0:
+            val_epoch_start = time.time()
+            val_stats = validate_one_epoch(model, criterion, metrics, data_loaders_val, device, epoch, log_writer=log_writer, args=args)
+            val_epoch_time = time.time() - val_epoch_start
+
+            val_best = val_stats[bestmetric]
+            
+            # Save best of all
+            if val_best <= best_so_far:
+                best_so_far = val_best
+                misc.save_model(args=args, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler, epoch=epoch, best_so_far=best_so_far, fname='best')
+        
+            log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
+                         'epoch': epoch,
+                         **{f'val_{k}': v for k, v in val_stats.items()}}
+        else:
+            log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
+                         'epoch': epoch,}
+                             
+        if args.distributed: dist.barrier()
+        
+        # Save stuff
+        if args.output_dir and ((epoch+1) % args.save_every == 0 or epoch + 1 == args.epochs):
+            misc.save_model(args=args, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler, epoch=epoch, best_so_far=best_so_far, fname='last')
+
+        if args.output_dir:
+            if log_writer is not None:
+                log_writer.flush()
+            with open(os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8") as f:
+                f.write(json.dumps(log_stats) + "\n")
+        
+    total_time = time.time() - start_time
+    total_time_str = str(datetime.timedelta(seconds=int(total_time)))
+    print('Training time {}'.format(total_time_str))
+    
+if __name__ == '__main__':
+    args = get_args_parser()
+    args = args.parse_args()
+    main(args)

third_party/dust3r/croco/utils/misc.py

@@ -0,0 +1,463 @@
+# Copyright (C) 2022-present Naver Corporation. All rights reserved.
+# Licensed under CC BY-NC-SA 4.0 (non-commercial use only).
+# 
+# --------------------------------------------------------
+# utilitary functions for CroCo
+# --------------------------------------------------------
+# References:
+# MAE: https://github.com/facebookresearch/mae
+# DeiT: https://github.com/facebookresearch/deit
+# BEiT: https://github.com/microsoft/unilm/tree/master/beit
+# --------------------------------------------------------
+
+import builtins
+import datetime
+import os
+import time
+import math
+import json
+from collections import defaultdict, deque
+from pathlib import Path
+import numpy as np
+
+import torch
+import torch.distributed as dist
+from torch import inf
+
+class SmoothedValue(object):
+    """Track a series of values and provide access to smoothed values over a
+    window or the global series average.
+    """
+
+    def __init__(self, window_size=20, fmt=None):
+        if fmt is None:
+            fmt = "{median:.4f} ({global_avg:.4f})"
+        self.deque = deque(maxlen=window_size)
+        self.total = 0.0
+        self.count = 0
+        self.fmt = fmt
+
+    def update(self, value, n=1):
+        self.deque.append(value)
+        self.count += n
+        self.total += value * n
+
+    def synchronize_between_processes(self):
+        """
+        Warning: does not synchronize the deque!
+        """
+        if not is_dist_avail_and_initialized():
+            return
+        t = torch.tensor([self.count, self.total], dtype=torch.float64, device='cuda')
+        dist.barrier()
+        dist.all_reduce(t)
+        t = t.tolist()
+        self.count = int(t[0])
+        self.total = t[1]
+
+    @property
+    def median(self):
+        d = torch.tensor(list(self.deque))
+        return d.median().item()
+
+    @property
+    def avg(self):
+        d = torch.tensor(list(self.deque), dtype=torch.float32)
+        return d.mean().item()
+
+    @property
+    def global_avg(self):
+        return self.total / self.count
+
+    @property
+    def max(self):
+        return max(self.deque)
+
+    @property
+    def value(self):
+        return self.deque[-1]
+
+    def __str__(self):
+        return self.fmt.format(
+            median=self.median,
+            avg=self.avg,
+            global_avg=self.global_avg,
+            max=self.max,
+            value=self.value)
+
+
+class MetricLogger(object):
+    def __init__(self, delimiter="\t"):
+        self.meters = defaultdict(SmoothedValue)
+        self.delimiter = delimiter
+
+    def update(self, **kwargs):
+        for k, v in kwargs.items():
+            if v is None:
+                continue
+            if isinstance(v, torch.Tensor):
+                v = v.item()
+            assert isinstance(v, (float, int))
+            self.meters[k].update(v)
+
+    def __getattr__(self, attr):
+        if attr in self.meters:
+            return self.meters[attr]
+        if attr in self.__dict__:
+            return self.__dict__[attr]
+        raise AttributeError("'{}' object has no attribute '{}'".format(
+            type(self).__name__, attr))
+
+    def __str__(self):
+        loss_str = []
+        for name, meter in self.meters.items():
+            loss_str.append(
+                "{}: {}".format(name, str(meter))
+            )
+        return self.delimiter.join(loss_str)
+
+    def synchronize_between_processes(self):
+        for meter in self.meters.values():
+            meter.synchronize_between_processes()
+
+    def add_meter(self, name, meter):
+        self.meters[name] = meter
+
+    def log_every(self, iterable, print_freq, header=None, max_iter=None):
+        i = 0
+        if not header:
+            header = ''
+        start_time = time.time()
+        end = time.time()
+        iter_time = SmoothedValue(fmt='{avg:.4f}')
+        data_time = SmoothedValue(fmt='{avg:.4f}')
+        len_iterable = min(len(iterable), max_iter) if max_iter else len(iterable)
+        space_fmt = ':' + str(len(str(len_iterable))) + 'd'
+        log_msg = [
+            header,
+            '[{0' + space_fmt + '}/{1}]',
+            'eta: {eta}',
+            '{meters}',
+            'time: {time}',
+            'data: {data}'
+        ]
+        if torch.cuda.is_available():
+            log_msg.append('max mem: {memory:.0f}')
+        log_msg = self.delimiter.join(log_msg)
+        MB = 1024.0 * 1024.0
+        for it,obj in enumerate(iterable):
+            data_time.update(time.time() - end)
+            yield obj
+            iter_time.update(time.time() - end)
+            if i % print_freq == 0 or i == len_iterable - 1:
+                eta_seconds = iter_time.global_avg * (len_iterable - i)
+                eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
+                if torch.cuda.is_available():
+                    print(log_msg.format(
+                        i, len_iterable, eta=eta_string,
+                        meters=str(self),
+                        time=str(iter_time), data=str(data_time),
+                        memory=torch.cuda.max_memory_allocated() / MB))
+                else:
+                    print(log_msg.format(
+                        i, len_iterable, eta=eta_string,
+                        meters=str(self),
+                        time=str(iter_time), data=str(data_time)))
+            i += 1
+            end = time.time()
+            if max_iter and it >= max_iter:
+                break
+        total_time = time.time() - start_time
+        total_time_str = str(datetime.timedelta(seconds=int(total_time)))
+        print('{} Total time: {} ({:.4f} s / it)'.format(
+            header, total_time_str, total_time / len_iterable))
+
+
+def setup_for_distributed(is_master):
+    """
+    This function disables printing when not in master process
+    """
+    builtin_print = builtins.print
+
+    def print(*args, **kwargs):
+        force = kwargs.pop('force', False)
+        force = force or (get_world_size() > 8)
+        if is_master or force:
+            now = datetime.datetime.now().time()
+            builtin_print('[{}] '.format(now), end='')  # print with time stamp
+            builtin_print(*args, **kwargs)
+
+    builtins.print = print
+
+
+def is_dist_avail_and_initialized():
+    if not dist.is_available():
+        return False
+    if not dist.is_initialized():
+        return False
+    return True
+
+
+def get_world_size():
+    if not is_dist_avail_and_initialized():
+        return 1
+    return dist.get_world_size()
+
+
+def get_rank():
+    if not is_dist_avail_and_initialized():
+        return 0
+    return dist.get_rank()
+
+
+def is_main_process():
+    return get_rank() == 0
+
+
+def save_on_master(*args, **kwargs):
+    if is_main_process():
+        torch.save(*args, **kwargs)
+
+
+def init_distributed_mode(args):
+    nodist = args.nodist if hasattr(args,'nodist') else False 
+    if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ and not nodist:
+        args.rank = int(os.environ["RANK"])
+        args.world_size = int(os.environ['WORLD_SIZE'])
+        args.gpu = int(os.environ['LOCAL_RANK'])
+    else:
+        print('Not using distributed mode')
+        setup_for_distributed(is_master=True)  # hack
+        args.distributed = False
+        return
+
+    args.distributed = True
+
+    torch.cuda.set_device(args.gpu)
+    args.dist_backend = 'nccl'
+    print('| distributed init (rank {}): {}, gpu {}'.format(
+        args.rank, args.dist_url, args.gpu), flush=True)
+    torch.distributed.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
+                                         world_size=args.world_size, rank=args.rank)
+    torch.distributed.barrier()
+    setup_for_distributed(args.rank == 0)
+
+
+class NativeScalerWithGradNormCount:
+    state_dict_key = "amp_scaler"
+
+    def __init__(self, enabled=True):
+        self._scaler = torch.cuda.amp.GradScaler(enabled=enabled)
+
+    def __call__(self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True):
+        self._scaler.scale(loss).backward(create_graph=create_graph)
+        if update_grad:
+            if clip_grad is not None:
+                assert parameters is not None
+                self._scaler.unscale_(optimizer)  # unscale the gradients of optimizer's assigned params in-place
+                norm = torch.nn.utils.clip_grad_norm_(parameters, clip_grad)
+            else:
+                self._scaler.unscale_(optimizer)
+                norm = get_grad_norm_(parameters)
+            self._scaler.step(optimizer)
+            self._scaler.update()
+        else:
+            norm = None
+        return norm
+
+    def state_dict(self):
+        return self._scaler.state_dict()
+
+    def load_state_dict(self, state_dict):
+        self._scaler.load_state_dict(state_dict)
+
+
+def get_grad_norm_(parameters, norm_type: float = 2.0) -> torch.Tensor:
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+    parameters = [p for p in parameters if p.grad is not None]
+    norm_type = float(norm_type)
+    if len(parameters) == 0:
+        return torch.tensor(0.)
+    device = parameters[0].grad.device
+    if norm_type == inf:
+        total_norm = max(p.grad.detach().abs().max().to(device) for p in parameters)
+    else:
+        total_norm = torch.norm(torch.stack([torch.norm(p.grad.detach(), norm_type).to(device) for p in parameters]), norm_type)
+    return total_norm
+
+
+
+
+def save_model(args, epoch, model_without_ddp, optimizer, loss_scaler, fname=None, best_so_far=None):
+    output_dir = Path(args.output_dir)
+    if fname is None: fname = str(epoch)
+    checkpoint_path = output_dir / ('checkpoint-%s.pth' % fname)
+    to_save = {
+        'model': model_without_ddp.state_dict(),
+        'optimizer': optimizer.state_dict(),
+        'scaler': loss_scaler.state_dict(),
+        'args': args,
+        'epoch': epoch,
+    }
+    if best_so_far is not None: to_save['best_so_far'] = best_so_far
+    print(f'>> Saving model to {checkpoint_path} ...')
+    save_on_master(to_save, checkpoint_path)
+
+
+def load_model(args, model_without_ddp, optimizer, loss_scaler):
+    args.start_epoch = 0
+    best_so_far = None
+    if args.resume is not None:
+        if args.resume.startswith('https'):
+            checkpoint = torch.hub.load_state_dict_from_url(
+                args.resume, map_location='cpu', check_hash=True)
+        else:
+            checkpoint = torch.load(args.resume, map_location='cpu')
+        print("Resume checkpoint %s" % args.resume)
+        model_without_ddp.load_state_dict(checkpoint['model'], strict=False)
+        args.start_epoch = checkpoint['epoch'] + 1
+        optimizer.load_state_dict(checkpoint['optimizer'])
+        if 'scaler' in checkpoint:
+            loss_scaler.load_state_dict(checkpoint['scaler'])
+        if 'best_so_far' in checkpoint:
+            best_so_far = checkpoint['best_so_far']
+            print(" & best_so_far={:g}".format(best_so_far))
+        else:
+            print("")
+        print("With optim & sched! start_epoch={:d}".format(args.start_epoch), end='')
+    return best_so_far
+
+def all_reduce_mean(x):
+    world_size = get_world_size()
+    if world_size > 1:
+        x_reduce = torch.tensor(x).cuda()
+        dist.all_reduce(x_reduce)
+        x_reduce /= world_size
+        return x_reduce.item()
+    else:
+        return x
+
+def _replace(text, src, tgt, rm=''):
+    """ Advanced string replacement.
+    Given a text:
+    - replace all elements in src by the corresponding element in tgt
+    - remove all elements in rm
+    """
+    if len(tgt) == 1: 
+        tgt = tgt * len(src)
+    assert len(src) == len(tgt), f"'{src}' and '{tgt}' should have the same len"
+    for s,t in zip(src, tgt):
+        text = text.replace(s,t)
+    for c in rm:
+        text = text.replace(c,'')
+    return text
+    
+def filename( obj ):
+    """ transform a python obj or cmd into a proper filename. 
+     - \1 gets replaced by slash '/'
+     - \2 gets replaced by comma ','
+    """
+    if not isinstance(obj, str): 
+        obj = repr(obj)
+    obj = str(obj).replace('()','')
+    obj = _replace(obj, '_,(*/\1\2','-__x%/,', rm=' )\'"')
+    assert all(len(s) < 256 for s in obj.split(os.sep)), 'filename too long (>256 characters):\n'+obj
+    return obj
+
+def _get_num_layer_for_vit(var_name, enc_depth, dec_depth):
+    if var_name in ("cls_token", "mask_token", "pos_embed", "global_tokens"):
+        return 0
+    elif var_name.startswith("patch_embed"):
+        return 0
+    elif var_name.startswith("enc_blocks"):
+        layer_id = int(var_name.split('.')[1])
+        return layer_id + 1
+    elif var_name.startswith('decoder_embed') or var_name.startswith('enc_norm'): # part of the last black
+        return enc_depth
+    elif var_name.startswith('dec_blocks'):
+        layer_id = int(var_name.split('.')[1])
+        return enc_depth + layer_id + 1
+    elif var_name.startswith('dec_norm'): # part of the last block
+        return enc_depth + dec_depth
+    elif any(var_name.startswith(k) for k in ['head','prediction_head']):
+        return enc_depth + dec_depth + 1
+    else:
+        raise NotImplementedError(var_name)
+
+def get_parameter_groups(model, weight_decay, layer_decay=1.0, skip_list=(), no_lr_scale_list=[]):
+    parameter_group_names = {}
+    parameter_group_vars = {}
+    enc_depth, dec_depth = None, None
+    # prepare layer decay values 
+    assert layer_decay==1.0 or 0.<layer_decay<1.
+    if layer_decay<1.:
+        enc_depth = model.enc_depth
+        dec_depth = model.dec_depth if hasattr(model, 'dec_blocks') else 0
+        num_layers = enc_depth+dec_depth
+        layer_decay_values = list(layer_decay ** (num_layers + 1 - i) for i in range(num_layers + 2))
+        
+    for name, param in model.named_parameters():
+        if not param.requires_grad:
+            continue  # frozen weights
+
+        # Assign weight decay values
+        if len(param.shape) == 1 or name.endswith(".bias") or name in skip_list:
+            group_name = "no_decay"
+            this_weight_decay = 0.
+        else:
+            group_name = "decay"
+            this_weight_decay = weight_decay
+
+        # Assign layer ID for LR scaling
+        if layer_decay<1.:
+            skip_scale = False
+            layer_id = _get_num_layer_for_vit(name, enc_depth, dec_depth)
+            group_name = "layer_%d_%s" % (layer_id, group_name)
+            if name in no_lr_scale_list:
+                skip_scale = True
+                group_name = f'{group_name}_no_lr_scale'
+        else:
+            layer_id = 0
+            skip_scale = True
+
+        if group_name not in parameter_group_names:
+            if not skip_scale:
+                scale = layer_decay_values[layer_id]
+            else:
+                scale = 1.
+
+            parameter_group_names[group_name] = {
+                "weight_decay": this_weight_decay,
+                "params": [],
+                "lr_scale": scale
+            }
+            parameter_group_vars[group_name] = {
+                "weight_decay": this_weight_decay,
+                "params": [],
+                "lr_scale": scale
+            }
+
+        parameter_group_vars[group_name]["params"].append(param)
+        parameter_group_names[group_name]["params"].append(name)
+    print("Param groups = %s" % json.dumps(parameter_group_names, indent=2))
+    return list(parameter_group_vars.values())
+
+
+
+def adjust_learning_rate(optimizer, epoch, args):
+    """Decay the learning rate with half-cycle cosine after warmup"""
+    
+    if epoch < args.warmup_epochs:
+        lr = args.lr * epoch / args.warmup_epochs 
+    else:
+        lr = args.min_lr + (args.lr - args.min_lr) * 0.5 * \
+            (1. + math.cos(math.pi * (epoch - args.warmup_epochs) / (args.epochs - args.warmup_epochs)))
+            
+    for param_group in optimizer.param_groups:
+        if "lr_scale" in param_group:
+            param_group["lr"] = lr * param_group["lr_scale"]
+        else:
+            param_group["lr"] = lr
+            
+    return lr

third_party/dust3r/datasets_preprocess/habitat/README.md

@@ -0,0 +1,66 @@
+## Steps to reproduce synthetic training data using the Habitat-Sim simulator
+
+### Create a conda environment
+```bash
+conda create -n habitat python=3.8 habitat-sim=0.2.1 headless=2.0 -c aihabitat -c conda-forge
+conda active habitat
+conda install pytorch -c pytorch
+pip install opencv-python tqdm
+```
+
+or (if you get the error `For headless systems, compile with --headless for EGL support`)
+```
+git clone --branch stable https://github.com/facebookresearch/habitat-sim.git
+cd habitat-sim
+
+conda create -n habitat python=3.9 cmake=3.14.0
+conda activate habitat
+pip install . -v
+conda install pytorch -c pytorch
+pip install opencv-python tqdm
+```
+
+### Download Habitat-Sim scenes
+Download Habitat-Sim scenes:
+- Download links can be found here: https://github.com/facebookresearch/habitat-sim/blob/main/DATASETS.md
+- We used scenes from the HM3D, habitat-test-scenes, ReplicaCad and ScanNet datasets.
+- Please put the scenes in a directory `$SCENES_DIR` following the structure below:
+(Note: the habitat-sim dataset installer may install an incompatible version for ReplicaCAD backed lighting.
+The correct scene dataset can be dowloaded from Huggingface: `git clone git@hf.co:datasets/ai-habitat/ReplicaCAD_baked_lighting`).
+```
+$SCENES_DIR/
+├──hm3d/
+├──gibson/
+├──habitat-test-scenes/
+├──ReplicaCAD_baked_lighting/
+└──scannet/
+```
+
+### Download renderings metadata 
+
+Download metadata corresponding to each scene and extract them into a directory `$METADATA_DIR`
+```bash
+wget https://download.europe.naverlabs.com/ComputerVision/DUSt3R/habitat_5views_v1_512x512_metadata.tar.gz
+tar -xvzf habitat_5views_v1_512x512_metadata.tar.gz
+```
+
+### Render the scenes
+
+Render the scenes in an output directory `$OUTPUT_DIR`
+```bash
+export METADATA_DIR="/path/to/habitat/5views_v1_512x512_metadata"
+export SCENES_DIR="/path/to/habitat/data/scene_datasets/"
+export OUTPUT_DIR="data/habitat_processed"
+cd datasets_preprocess/habitat/
+export PYTHONPATH=$(pwd)
+# Print commandlines to generate images corresponding to each scene
+python preprocess_habitat.py --scenes_dir=$SCENES_DIR --metadata_dir=$METADATA_DIR --output_dir=$OUTPUT_DIR
+# Launch these commandlines in parallel e.g. using GNU-Parallel as follows:
+python preprocess_habitat.py --scenes_dir=$SCENES_DIR --metadata_dir=$METADATA_DIR --output_dir=$OUTPUT_DIR | parallel -j 16
+```
+
+### Make a list of scenes
+
+```bash
+python find_scenes.py --root $OUTPUT_DIR
+```