Upload folder using huggingface_hub

README.md

@@ -1,12 +1,6 @@
 ---
-title: Demo MSE-CNN
-emoji: 📉
-colorFrom: yellow
-colorTo: pink
+title: Demo_MSE-CNN
+app_file: demo.py
 sdk: gradio
-sdk_version: 3.44.3
-app_file: app.py
-pinned: false
+sdk_version: 3.34.0
 ---
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo.py

@@ -0,0 +1,440 @@
+"""@package docstring 
+
+@file demo.py 
+
+@brief Demonstration of the application of the MSE-CNN 
+
+Note: In order to run this script, you have to do it inside the folder
+
+@section libraries_demo Libraries 
+- msecnn
+- train_model_utils
+- cv2
+- dataset_utils
+- re
+- sys
+- numpy
+- gradio
+- torch
+- custom_dataset
+- PIL
+ 
+@section classes_demo Classes 
+- None 
+
+@section functions_demo Functions 
+- setup_model()
+- int2label(split)
+- draw_partition(img, split, cu_pos, cu_size)
+- split_fm(cu, cu_pos, split)
+- partition_img(img, img_yuv)
+- pipeline(img, text)
+- main()
+
+@section global_vars_demo Global Variables 
+- PATH_TO_COEFFS = "../../../model_coefficients/best_coefficients"
+- LOAD_IMAGE_ERROR = "load_image_error.png"
+- EXAMPLE_IMGS = ["example_img_1.jpeg", "example_img_2.jpeg"]
+- CTU_SIZE = (128, 128)
+- FIRST_CU_POS = torch.tensor([0, 0]).reshape(shape=(-1, 2))
+- FIRST_CU_SIZE = torch.tensor([64, 64]).reshape(shape=(-1, 2))
+- DEV = "cuda" if torch.cuda.is_available() else "cpu"
+- QP = 32
+- model = None
+- COLOR = (0, 247, 255)
+- LINE_THICKNESS = 1
+- DEFAULT_TEXT_FOR_COORDS = "Insert CTU position in the image..."
+
+@section todo_demo TODO 
+- Instead of obtaining the best split, do the thresholding and then split it until you find the right type of split 
+
+@section license License 
+MIT License 
+Copyright (c) 2022 Raul Kevin do Espirito Santo Viana
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+@section author_demo Author(s)
+- Created by Raul Kevin Viana
+- Last time modified is 2023-09-10 21:00:10.225508
+"""
+
+
+# ==============================================================
+# Imports
+# ==============================================================
+
+import gradio as gr
+import cv2 as cv
+import sys
+import torch
+from PIL import Image
+import numpy as np
+import re
+
+sys.path.append("../")
+import msecnn
+import dataset_utils as du
+import custom_dataset as cd
+import train_model_utils as tmu
+
+
+# ==============================================================
+# Constants and Global Variables
+# ==============================================================
+
+PATH_TO_COEFFS = "../../../model_coefficients/best_coefficients"
+LOAD_IMAGE_ERROR = "load_image_error.png"
+EXAMPLE_IMGS = ["example_img_1.jpeg", "example_img_2.jpeg"]
+CTU_SIZE = (128, 128)
+FIRST_CU_POS = torch.tensor([0, 0]).reshape(shape=(-1, 2))
+FIRST_CU_SIZE = torch.tensor([64, 64]).reshape(shape=(-1, 2))
+DEV = "cuda" if torch.cuda.is_available() else "cpu"
+QP = 32
+model = None
+COLOR = (0, 247, 255)
+LINE_THICKNESS = 1
+DEFAULT_TEXT_FOR_COORDS = "Insert CTU position in the image..."
+
+# ==============================================================
+# Functions
+# ==============================================================
+
+def setup_model():
+    """!
+    @brief Initializes and load the parameters of the MSE-CNN
+    """
+    # Initialize model
+    stg1_2 = msecnn.MseCnnStg1(device=DEV, QP=QP).to(DEV)
+    stg3 = msecnn.MseCnnStgX(device=DEV, QP=QP).to(DEV)
+    stg4 = msecnn.MseCnnStgX(device=DEV, QP=QP).to(DEV)
+    stg5 = msecnn.MseCnnStgX(device=DEV, QP=QP).to(DEV)
+    stg6 = msecnn.MseCnnStgX(device=DEV, QP=QP).to(DEV)
+    model = (stg1_2, stg3, stg4, stg5, stg6)
+
+    # Load model coefficients
+    model = tmu.load_model_parameters_eval(model, PATH_TO_COEFFS, DEV)
+
+    return model
+
+def int2label(split):
+    """!
+    @brief Obtain the string that corresponds to an integer value of the split
+
+    @param [in] split: Integer number representing the split tht the model chose
+    @param [out] str_split: Name of the corresponding split
+    """
+    if split == 0:
+        return "Non-Split"
+    elif split == 1:
+        return "Quad-Tree"
+    elif split == 2:
+        return "Horizontal Binary Tree"
+    elif split == 3:
+        return "Vertical Binary Tree"
+    elif split == 4:
+        return "Horizontal Ternary Tree"
+    elif split == 5:
+        return "Vertical Ternary Tree"
+    else:
+        return "Something wrong happened!"
+    
+def draw_partition(img, split, cu_pos, cu_size):
+    """!
+    @brief Draw partition in image based in the split outputed by the model
+
+    @param [in] img: User's input image
+    @param [in] cu_pos: CU position
+    @param [in] cu_size: CU size
+    @param [in] split: Integer number representing the split that the model chose
+    @param [out] str_split: Name of the corresponding split
+    """
+    # Parameters to draw the lines
+    ver_line_length = cu_size[0]
+    hor_line_length = cu_size[1]
+
+    if split == 1:
+        line1_start = (cu_pos[0], cu_pos[1]+hor_line_length//2)
+        line1_end = (cu_pos[0]+ver_line_length, cu_pos[1]+hor_line_length//2)
+        line2_start = (cu_pos[0]+ver_line_length//2, cu_pos[1])
+        line2_end = (cu_pos[0]+ver_line_length//2, cu_pos[1]+hor_line_length)
+        img = cv.line(img, line1_start, line1_end, COLOR, LINE_THICKNESS)
+        img = cv.line(img, line2_start, line2_end, COLOR, LINE_THICKNESS)
+    elif split == 2:
+        line1_start = (cu_pos[0]+ver_line_length//2, cu_pos[1])
+        line1_end = (cu_pos[0]+ver_line_length//2, cu_pos[1]+hor_line_length)
+        # assert line1_start[0]-line1_end[0] == 0 or line1_start[1]-line1_end[1] == 0  # Make sure that the lines are either horizontal or vertical
+        img = cv.line(img, line1_start, line1_end, COLOR, LINE_THICKNESS)
+    elif split == 3:
+        line1_start = (cu_pos[0], cu_pos[1]+hor_line_length//2)
+        line1_end = (cu_pos[0]+ver_line_length, cu_pos[1]+hor_line_length//2)
+        img = cv.line(img, line1_start, line1_end, COLOR, LINE_THICKNESS)
+    elif split == 4:
+        line1_start = (cu_pos[0]+ver_line_length//3, cu_pos[1])
+        line1_end = (cu_pos[0]+ver_line_length//3, cu_pos[1]+hor_line_length)
+        line2_start = (cu_pos[0]+(ver_line_length*2)//3, cu_pos[1])
+        line2_end = (cu_pos[0]+(ver_line_length*2)//3, cu_pos[1]+hor_line_length)
+        img = cv.line(img, line1_start, line1_end, COLOR, LINE_THICKNESS)
+        img = cv.line(img, line2_start, line2_end, COLOR, LINE_THICKNESS)
+    elif split == 5:
+        line1_start = (cu_pos[0], cu_pos[1]+hor_line_length//3)
+        line1_end = (cu_pos[0]+ver_line_length, cu_pos[1]+hor_line_length//3)
+        line2_start = (cu_pos[0], cu_pos[1]+(hor_line_length*2)//3)
+        line2_end = (cu_pos[0]+ver_line_length, cu_pos[1]+(hor_line_length*2)//3)
+        img = cv.line(img, line1_start, line1_end, COLOR, LINE_THICKNESS)
+        img = cv.line(img, line2_start, line2_end, COLOR, LINE_THICKNESS)
+    else:
+        raise Exception("Something wrong happened!")
+    
+    return img
+
+
+def split_fm(cu, cu_pos, split):
+    """!
+    @brief Splits feature maps in specific way
+
+    @param [in] cu: Input to the model
+    @param [in] cu_pos: Coordinate of the CU
+    @param [in] split: Way to split CU
+    @param [out] cu_out: New Feature maps
+    @param [out] cu_pos_out: Position of the new CUs
+    """
+    # Initizalize list
+    if split == 0:  # Non-split
+        cu_out = cu
+        cu_pos = [cu_pos]
+    
+    elif split == 1:  # Quad-tree
+        # Split CU and add to list
+        cu_1 = torch.split(cu, cu.shape[-2]//2, -2)
+        cu_2 = torch.split(cu_1[1], cu_1[1].shape[-1]//2, -1)
+        cu_1 = torch.split(cu_1[0], cu_1[0].shape[-1]//2, -1)
+        cu_out = cu_1 + cu_2
+        cu_pos = [[cu_pos[0], cu_pos[1]], [cu_pos[0], cu_pos[1]+cu.shape[-1]//2],
+                  [cu_pos[0]+cu.shape[-2]//2, cu_pos[1]], [cu_pos[0]+cu.shape[-2]//2, cu_pos[1]+cu.shape[-1]//2]]
+
+    elif split == 2:  # HBT
+        # Split CU and add to list
+        cu_out = torch.split(cu, cu.shape[-2]//2, -2)
+        cu_pos = [[cu_pos[0], cu_pos[1]], [cu_pos[0]+cu.shape[-2]//2, cu_pos[1]]]
+    
+    elif split == 3:  # VBT
+        # Split CU and add to list
+        cu_out = torch.split(cu, cu.shape[-1]//2, -1)
+        cu_pos = [[cu_pos[0], cu_pos[1]], [cu_pos[0], cu_pos[1]+cu.shape[-1]//2]]
+                
+    elif split == 4:  # HTT
+        # Split CU and add to list
+        cu_out = torch.split(cu, cu.shape[-2]//3, -2)
+        cu_pos = [[cu_pos[0], cu_pos[1]], [cu_pos[0]+cu.shape[-2]//3, cu_pos[1]], [(2*cu.shape[-2])//3+cu_pos[0], cu_pos[1]]]
+
+    
+    elif split == 5:  # VTT
+        # Split CU and add to list
+        cu_out = torch.split(cu, cu.shape[-1]//3, -1)
+        cu_pos = [[cu_pos[0], cu_pos[1]], [cu_pos[0], cu_pos[1]+cu.shape[-1]//3], [(2*cu.shape[-1])//3+cu_pos[0], cu_pos[1]]]
+    
+    else:
+        raise Exception("This can't happen! Wrong split mode number: ", str(split))
+
+    if type(cu_out) is tuple:
+        if len(cu_out) != 1:
+            cu_out = torch.cat(cu_out)
+        else:
+            cu_out = cu_out[0]
+
+    return cu_out, cu_pos
+
+def partition_img(img, img_yuv):
+    """!
+    @brief Partitions a full 128x128 CTU and draws the partition in the original image
+
+    TODO: Instead of obtaining the best split, do the thresholding and then split it until you find the right type of split
+
+    @param [in] img: Image in BGR
+    @param [in] img_yuv: Image in YUV
+    @param [in] stg: Current stage being partitioned
+    @param [in] cu_pos: Current stage being partitioned
+    @param [in] cu_size: Current stage being partitioned
+    @param [out] img: Image in with partitions drawn to it
+    """
+    global model
+    # Stage 1 and 2
+    pos_1 = torch.tensor([[0, 0]])
+    pos_2 = torch.tensor([[0, 64]])
+    pos_3 = torch.tensor([[64, 0]])
+    pos_4 = torch.tensor([[64, 64]])
+    split_1, CUs_1, ap_1 = model[0](img_yuv, FIRST_CU_SIZE, pos_1)
+    split_2, CUs_2, ap_2 = model[0](img_yuv, FIRST_CU_SIZE, pos_2)
+    split_3, CUs_3, ap_3 = model[0](img_yuv, FIRST_CU_SIZE, pos_3)
+    split_4, CUs_4, ap_4 = model[0](img_yuv, FIRST_CU_SIZE, pos_4)
+    all_cus_stg1 = [(split_1, CUs_1, ap_1, (0, 0)), (split_2, CUs_2, ap_2, (0, 64)),
+                      (split_3, CUs_3, ap_3, (64, 0)), (split_4, CUs_4, ap_4, (64, 64))]
+    img = draw_partition(img, 1, (0, 0), (128, 128))
+
+    # Stage 2: spliting
+    for cus_stg1 in all_cus_stg1:
+        split_stg1, cu_stg1, ap_stg1, pos_stg1 = cus_stg1
+        split_stg1 = tmu.obtain_mode(split_stg1)
+        if split_stg1 == 0:
+            continue
+        # compute new cus
+        try:
+            cu_out_2, cu_pos_2 = split_fm(cu_stg1, pos_stg1, split_stg1)
+        except RuntimeError:
+            # Weird partition happened
+            continue
+        # draw partition to original image
+        img = draw_partition(img, split_stg1, pos_stg1, (cu_stg1.shape[-2], cu_stg1.shape[-1]))
+
+        all_cus_stg2 = [(cu_out_2[idx, :, :, :].unsqueeze(0), ap_stg1, cu_pos_2[idx]) for idx in range(cu_out_2.shape[0])]
+        
+        # Stage 3
+        for cus_stg2 in all_cus_stg2:
+            cu_stg2, ap_stg2, pos_stg2 = cus_stg2
+            pred_stg3, cu_stg3, ap_stg3 = model[1](cu_stg2, ap_stg2)
+            pred_stg3 = tmu.obtain_mode(pred_stg3)
+            # ap_stg3 = ap_stg3.item()
+            if pred_stg3 == 0:
+                continue
+            # compute new cus
+            try:
+                cu_out_3, cu_pos_3 = split_fm(cu_stg3, pos_stg2, pred_stg3)
+            except RuntimeError:
+                # Weird partition happened; skip
+                continue
+            # draw partition to original image
+            img = draw_partition(img, pred_stg3, pos_stg2, (cu_stg3.shape[-2], cu_stg3.shape[-1]))
+            
+            all_cus_stg3 = [(cu_out_3[idx, :, :, :].unsqueeze(0), ap_stg3, cu_pos_3[idx]) for idx in range(cu_out_3.shape[0])]
+
+            # Stage 4
+            for cus_stg3 in all_cus_stg3:
+                cu_stg3, ap_stg3, pos_stg3 = cus_stg3
+                pred_stg4, cu_stg4, ap_stg4 = model[2](cu_stg3, ap_stg3)
+                pred_stg4 = tmu.obtain_mode(pred_stg4)
+                # ap_stg4 = ap_stg4.item()
+                if pred_stg4 == 0:
+                    continue
+                
+                # compute new cus
+                try:
+                   cu_out_4, cu_pos_4 = split_fm(cu_stg4, pos_stg3, pred_stg4)
+                except RuntimeError:
+                    # Weird partition happened; skip
+                    continue
+                # draw partition to original image
+                img = draw_partition(img, pred_stg4, pos_stg3, (cu_stg4.shape[-2], cu_stg4.shape[-1]))
+                all_cus_stg4 = [(cu_out_4[idx, :, :, :].unsqueeze(0), ap_stg4, cu_pos_4[idx]) for idx in range(cu_out_4.shape[0])]
+
+                # Stage 5
+                for cus_stg4 in all_cus_stg4:
+                    cu_stg4, ap_stg4, pos_stg4 = cus_stg4
+                    pred_stg5, cu_stg5, ap_stg5 = model[3](cu_stg4, ap_stg4)
+                    pred_stg5 = tmu.obtain_mode(pred_stg5)
+                    # ap_stg5 = ap_stg5.item()
+                    if pred_stg5 == 0:
+                        continue
+                    # compute new cus
+                    try:
+                        cu_out_5, cu_pos_5 = split_fm(cu_stg5, pos_stg4, pred_stg5)
+                    except RuntimeError:
+                        # Weird partition happened; skip
+                        continue
+                    # draw partition to original image
+                    img = draw_partition(img, pred_stg5, pos_stg4, (cu_stg5.shape[-2], cu_stg5.shape[-1]))
+
+                    all_cus_stg5 = [(cu_out_5[idx, :, :, :].unsqueeze(0), ap_stg5, cu_pos_5[idx]) for idx in range(cu_out_5.shape[0])]
+
+                    # Stage 6
+                    for cus_stg5 in all_cus_stg5:
+                        cu_stg5, ap_stg5, pos_stg5 = cus_stg5
+                        pred_stg6, cu_stg6, ap_stg6 = model[4](cu_stg5, ap_stg5)
+                        pred_stg6 = tmu.obtain_mode(pred_stg6)
+                        # ap_stg6 = ap_stg6.item()
+                        if pred_stg6 == 0:
+                            continue
+                        # draw partition to original image
+                        img = draw_partition(img, pred_stg6, pos_stg5, (cu_stg6.shape[-2], cu_stg6.shape[-1]))
+    
+    return img
+
+def pipeline(img, text):
+    """!
+    @brief Pipeline to implement the functionalities to demonstrate the potential of the MSE-CNN
+
+    @param [in] img: Image in RGB
+    @param [out] mod_img: Modified image with drawings into it in RGB
+    @param [out] best_split: Best split (BTV, BTH, TTV, TTH, Non-split, QT)
+    """
+    global model
+
+    # Obtain coordinates of the CTU
+    coords = re.findall(r"\d+", text)
+    coords = list(map(lambda x: int(x), coords))
+
+    # In case nothing is submitted, return a default image and text
+    if type(img) is type(None) or type(text) is type(None) or text == DEFAULT_TEXT_FOR_COORDS:
+        img_error = Image.open(LOAD_IMAGE_ERROR)  # Replace with the path to your first image file
+        img_error = np.array(img_error)
+        return img_error, "Load the image first and also make sure you specify the position of the CTU!"
+    
+    # Convert image to appropriate size
+    img = img[coords[0]:coords[0]+128, coords[1]:coords[1]+128, :]
+    if img.shape[0] % 2 != 0:
+        img = img[: img.shape[0]-1, :, :]
+    if img.shape[1] % 2 != 0:
+        img = img[:, :img.shape[1]-1, :]
+
+    # convert to yuv
+    img_yuv = cv.cvtColor(img, cv.COLOR_RGB2YUV_I420)
+    # convert to pytorch tensor
+    img_yuv = torch.from_numpy(img_yuv)
+    # obtain luma channel
+    _, ctu_y, _, _ = cd.get_cu_v2(img_yuv, CTU_SIZE, (0, 0), CTU_SIZE)
+    # change shape
+    ctu_y = torch.reshape(ctu_y, (1, 1, 128, 128)).to(DEV).float()
+
+    # Load model
+    model = setup_model()
+
+    # Partition Image
+    img = partition_img(img, ctu_y)
+
+    return img, "Partitioned Image"
+
+def main():
+    with open("description.md", encoding="utf-8") as f:
+        description = f.read()
+
+    in_text_box = gr.Textbox(value=DEFAULT_TEXT_FOR_COORDS, label="Coordinates of CTU", info="You have to provide two numbers indicating the position of the CTU in the image")
+    in_image = gr.Image(label="Input image", info="Either use the example image or an image of your choosing")
+    out_text_box = gr.Textbox(label="Completion Message")
+    out_image = gr.Image(label="Partitioned CTU", info="Result of partitioning using MSE-CNN")
+
+    demo = gr.Interface(fn=pipeline, inputs=[in_image, in_text_box], examples=[[EXAMPLE_IMGS[0], "300, 800"], [EXAMPLE_IMGS[0], "100, 200"], [EXAMPLE_IMGS[1], "600, 925"], [EXAMPLE_IMGS[1], "450, 1600"]], thumbnail="msecnn_model.png",
+                        outputs=[out_image, out_text_box], description=description, debug=True, # inbrowser=True,
+                        title="MSE-CNN Demo", image="msecnn_model.png")
+
+    demo.launch()
+
+
+# ==============================================================
+# Main
+# ==============================================================
+
+if __name__ == "__main__":
+    main()

description.md

@@ -0,0 +1,14 @@
+With this demo you will be able to understand better how the MSE-CNN works and its goal! :D
+
+<center>
+<img src="file/msecnn_model.png" width=500 />
+</center>
+
+## Tutorial
+To use this demo follow these steps ;)
+
+1. Load an image from your PC. Since the model needs to be fed 128x128 images, if your image is larger than that, then the 128 by 128 area within your image top left will be loaded. Additionally, you can tell the app the coordinates of the area of the snapshot you want to see. The coordinates are related to the image's upper-left corner, therefore position 0,0 refers to that particular area. Furthermore, you must first specify the height position (y axis) and then the width position (x axis).
+2. Click "Submit" to pass the image through the model. 
+3. After the previous step, an image will be displayed with the best way, according with the model, to partition that section of the image. The possible ways to partition an block (coding unit, CU) in VVC is in Quartenary Tree (QT), Binary Tree Horizontal or Vertical (BTH or BTV), Ternary Tree Horizontal or Vertical (TTH, TTV) and Non-split.
+
+**Note**: This demo implementation has some limitations, such as the fact that the model occasionally makes illogical predictions. For instance, splitting a 16x32 CU using VTT is incorrect. This occurs as a result of the model's inherent constraints as well as the fact that only the best split is being chosen. One way to reduce this behaviour is to evaluate not only the optimal split but also alternate splits. This can be accomplished by, for instance, applying the multi-thresholding method to the model's forecasts and determining the splits that are most likely to occur. Additionally, when unreasonable splits are predicted, the code immediately halts the partitioning of that particular block.