Spaces:
Runtime error
Runtime error
init
Browse files- .gitignore +142 -0
- DIS/Inference.py +53 -0
- DIS/IsNetPipeLine.py +131 -0
- DIS/models/__init__.py +1 -0
- DIS/models/isnet.py +608 -0
- DIS/pytorch18.yml +92 -0
- app.py +36 -0
- requirements.txt +8 -0
.gitignore
ADDED
@@ -0,0 +1,142 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Byte-compiled / optimized / DLL files
|
2 |
+
__pycache__/
|
3 |
+
*.py[cod]
|
4 |
+
*$py.class
|
5 |
+
|
6 |
+
# C extensions
|
7 |
+
*.so
|
8 |
+
|
9 |
+
.idea
|
10 |
+
|
11 |
+
*.pth
|
12 |
+
|
13 |
+
# Distribution / packaging
|
14 |
+
.Python
|
15 |
+
build/
|
16 |
+
develop-eggs/
|
17 |
+
dist/
|
18 |
+
downloads/
|
19 |
+
eggs/
|
20 |
+
.eggs/
|
21 |
+
lib/
|
22 |
+
lib64/
|
23 |
+
parts/
|
24 |
+
sdist/
|
25 |
+
var/
|
26 |
+
wheels/
|
27 |
+
share/python-wheels/
|
28 |
+
*.egg-info/
|
29 |
+
.installed.cfg
|
30 |
+
*.egg
|
31 |
+
MANIFEST
|
32 |
+
|
33 |
+
# PyInstaller
|
34 |
+
# Usually these files are written by a python script from a template
|
35 |
+
# before PyInstaller builds the exe, so as to inject date/other infos into it.
|
36 |
+
*.manifest
|
37 |
+
*.spec
|
38 |
+
|
39 |
+
# Installer logs
|
40 |
+
pip-log.txt
|
41 |
+
pip-delete-this-directory.txt
|
42 |
+
|
43 |
+
# Unit test / coverage reports
|
44 |
+
htmlcov/
|
45 |
+
.tox/
|
46 |
+
.nox/
|
47 |
+
.coverage
|
48 |
+
.coverage.*
|
49 |
+
.cache
|
50 |
+
nosetests.xml
|
51 |
+
coverage.xml
|
52 |
+
*.cover
|
53 |
+
*.py,cover
|
54 |
+
.hypothesis/
|
55 |
+
.pytest_cache/
|
56 |
+
cover/
|
57 |
+
|
58 |
+
# Translations
|
59 |
+
*.mo
|
60 |
+
*.pot
|
61 |
+
|
62 |
+
# Django stuff:
|
63 |
+
*.log
|
64 |
+
local_settings.py
|
65 |
+
db.sqlite3
|
66 |
+
db.sqlite3-journal
|
67 |
+
|
68 |
+
# Flask stuff:
|
69 |
+
instance/
|
70 |
+
.webassets-cache
|
71 |
+
|
72 |
+
# Scrapy stuff:
|
73 |
+
.scrapy
|
74 |
+
|
75 |
+
# Sphinx documentation
|
76 |
+
docs/_build/
|
77 |
+
|
78 |
+
# PyBuilder
|
79 |
+
.pybuilder/
|
80 |
+
target/
|
81 |
+
|
82 |
+
# Jupyter Notebook
|
83 |
+
.ipynb_checkpoints
|
84 |
+
|
85 |
+
# IPython
|
86 |
+
profile_default/
|
87 |
+
ipython_config.py
|
88 |
+
|
89 |
+
# pyenv
|
90 |
+
# For a library or package, you might want to ignore these files since the code is
|
91 |
+
# intended to run in multiple environments; otherwise, check them in:
|
92 |
+
# .python-version
|
93 |
+
|
94 |
+
# pipenv
|
95 |
+
# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
|
96 |
+
# However, in case of collaboration, if having platform-specific dependencies or dependencies
|
97 |
+
# having no cross-platform support, pipenv may install dependencies that don't work, or not
|
98 |
+
# install all needed dependencies.
|
99 |
+
#Pipfile.lock
|
100 |
+
|
101 |
+
# PEP 582; used by e.g. github.com/David-OConnor/pyflow
|
102 |
+
__pypackages__/
|
103 |
+
|
104 |
+
# Celery stuff
|
105 |
+
celerybeat-schedule
|
106 |
+
celerybeat.pid
|
107 |
+
|
108 |
+
# SageMath parsed files
|
109 |
+
*.sage.py
|
110 |
+
|
111 |
+
# Environments
|
112 |
+
.env
|
113 |
+
.venv
|
114 |
+
env/
|
115 |
+
venv/
|
116 |
+
ENV/
|
117 |
+
env.bak/
|
118 |
+
venv.bak/
|
119 |
+
|
120 |
+
# Spyder project settings
|
121 |
+
.spyderproject
|
122 |
+
.spyproject
|
123 |
+
|
124 |
+
# Rope project settings
|
125 |
+
.ropeproject
|
126 |
+
|
127 |
+
# mkdocs documentation
|
128 |
+
/site
|
129 |
+
|
130 |
+
# mypy
|
131 |
+
.mypy_cache/
|
132 |
+
.dmypy.json
|
133 |
+
dmypy.json
|
134 |
+
|
135 |
+
# Pyre type checker
|
136 |
+
.pyre/
|
137 |
+
|
138 |
+
# pytype static type analyzer
|
139 |
+
.pytype/
|
140 |
+
|
141 |
+
# Cython debug symbols
|
142 |
+
cython_debug/
|
DIS/Inference.py
ADDED
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import time
|
3 |
+
import numpy as np
|
4 |
+
from skimage import io
|
5 |
+
import time
|
6 |
+
from glob import glob
|
7 |
+
from tqdm import tqdm
|
8 |
+
|
9 |
+
import torch, gc
|
10 |
+
import torch.nn as nn
|
11 |
+
from torch.autograd import Variable
|
12 |
+
import torch.optim as optim
|
13 |
+
import torch.nn.functional as F
|
14 |
+
from torchvision.transforms.functional import normalize
|
15 |
+
|
16 |
+
from models import *
|
17 |
+
|
18 |
+
|
19 |
+
if __name__ == "__main__":
|
20 |
+
dataset_path="../demo_datasets/your_dataset" #Your dataset path
|
21 |
+
model_path="../saved_models/IS-Net/isnet-general-use.pth" # the model path
|
22 |
+
result_path="../demo_datasets/your_dataset_result" #The folder path that you want to save the results
|
23 |
+
input_size=[1024,1024]
|
24 |
+
net=ISNetDIS()
|
25 |
+
|
26 |
+
if torch.cuda.is_available():
|
27 |
+
net.load_state_dict(torch.load(model_path))
|
28 |
+
net=net.cuda()
|
29 |
+
else:
|
30 |
+
net.load_state_dict(torch.load(model_path,map_location="cpu"))
|
31 |
+
net.eval()
|
32 |
+
im_list = glob(dataset_path+"/*.jpg")+glob(dataset_path+"/*.JPG")+glob(dataset_path+"/*.jpeg")+glob(dataset_path+"/*.JPEG")+glob(dataset_path+"/*.png")+glob(dataset_path+"/*.PNG")+glob(dataset_path+"/*.bmp")+glob(dataset_path+"/*.BMP")+glob(dataset_path+"/*.tiff")+glob(dataset_path+"/*.TIFF")
|
33 |
+
with torch.no_grad():
|
34 |
+
for i, im_path in tqdm(enumerate(im_list), total=len(im_list)):
|
35 |
+
print("im_path: ", im_path)
|
36 |
+
im = io.imread(im_path)
|
37 |
+
if len(im.shape) < 3:
|
38 |
+
im = im[:, :, np.newaxis]
|
39 |
+
im_shp=im.shape[0:2]
|
40 |
+
im_tensor = torch.tensor(im, dtype=torch.float32).permute(2,0,1)
|
41 |
+
im_tensor = F.upsample(torch.unsqueeze(im_tensor,0), input_size, mode="bilinear").type(torch.uint8)
|
42 |
+
image = torch.divide(im_tensor,255.0)
|
43 |
+
image = normalize(image,[0.5,0.5,0.5],[1.0,1.0,1.0])
|
44 |
+
|
45 |
+
if torch.cuda.is_available():
|
46 |
+
image=image.cuda()
|
47 |
+
result=net(image)
|
48 |
+
result=torch.squeeze(F.upsample(result[0][0],im_shp,mode='bilinear'),0)
|
49 |
+
ma = torch.max(result)
|
50 |
+
mi = torch.min(result)
|
51 |
+
result = (result-mi)/(ma-mi)
|
52 |
+
im_name=im_path.split('/')[-1].split('.')[0]
|
53 |
+
io.imsave(os.path.join(result_path,im_name+".png"),(result*255).permute(1,2,0).cpu().data.numpy().astype(np.uint8))
|
DIS/IsNetPipeLine.py
ADDED
@@ -0,0 +1,131 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
reference: https://github.com/xuebinqin/DIS
|
3 |
+
"""
|
4 |
+
|
5 |
+
import PIL.Image
|
6 |
+
import numpy as np
|
7 |
+
import torch
|
8 |
+
import torch.nn.functional as F
|
9 |
+
from PIL import Image
|
10 |
+
from torch import nn
|
11 |
+
from torch.autograd import Variable
|
12 |
+
from torchvision import transforms
|
13 |
+
from torchvision.transforms.functional import normalize
|
14 |
+
|
15 |
+
from .models import ISNetDIS
|
16 |
+
|
17 |
+
# Helpers
|
18 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
19 |
+
|
20 |
+
|
21 |
+
class GOSNormalize(object):
|
22 |
+
"""
|
23 |
+
Normalize the Image using torch.transforms
|
24 |
+
"""
|
25 |
+
|
26 |
+
def __init__(self, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]):
|
27 |
+
self.mean = mean
|
28 |
+
self.std = std
|
29 |
+
|
30 |
+
def __call__(self, image):
|
31 |
+
image = normalize(image, self.mean, self.std)
|
32 |
+
return image
|
33 |
+
|
34 |
+
|
35 |
+
def im_preprocess(im, size):
|
36 |
+
if len(im.shape) < 3:
|
37 |
+
im = im[:, :, np.newaxis]
|
38 |
+
if im.shape[2] == 1:
|
39 |
+
im = np.repeat(im, 3, axis=2)
|
40 |
+
im_tensor = torch.tensor(im.copy(), dtype=torch.float32)
|
41 |
+
im_tensor = torch.transpose(torch.transpose(im_tensor, 1, 2), 0, 1)
|
42 |
+
if len(size) < 2:
|
43 |
+
return im_tensor, im.shape[0:2]
|
44 |
+
else:
|
45 |
+
im_tensor = torch.unsqueeze(im_tensor, 0)
|
46 |
+
im_tensor = F.upsample(im_tensor, size, mode="bilinear")
|
47 |
+
im_tensor = torch.squeeze(im_tensor, 0)
|
48 |
+
|
49 |
+
return im_tensor.type(torch.uint8), im.shape[0:2]
|
50 |
+
|
51 |
+
|
52 |
+
class IsNetPipeLine:
|
53 |
+
def __init__(self, model_path=None, model_digit="full"):
|
54 |
+
self.model_digit = model_digit
|
55 |
+
self.model = ISNetDIS()
|
56 |
+
self.cache_size = [1024, 1024]
|
57 |
+
self.transform = transforms.Compose([
|
58 |
+
GOSNormalize([0.5, 0.5, 0.5], [1.0, 1.0, 1.0])
|
59 |
+
])
|
60 |
+
|
61 |
+
# Build Model
|
62 |
+
self.build_model(model_path)
|
63 |
+
|
64 |
+
def load_image(self, image: PIL.Image.Image):
|
65 |
+
im = np.array(image.convert("RGB"))
|
66 |
+
im, im_shp = im_preprocess(im, self.cache_size)
|
67 |
+
im = torch.divide(im, 255.0)
|
68 |
+
shape = torch.from_numpy(np.array(im_shp))
|
69 |
+
return self.transform(im).unsqueeze(0), shape.unsqueeze(0) # make a batch of image, shape
|
70 |
+
|
71 |
+
def build_model(self, model_path=None):
|
72 |
+
if model_path is not None:
|
73 |
+
self.model.load_state_dict(torch.load(model_path, map_location=device))
|
74 |
+
|
75 |
+
# convert to half precision
|
76 |
+
if self.model_digit == "half":
|
77 |
+
self.model.half()
|
78 |
+
for layer in self.model.modules():
|
79 |
+
if isinstance(layer, nn.BatchNorm2d):
|
80 |
+
layer.float()
|
81 |
+
self.model.to(device)
|
82 |
+
self.model.eval()
|
83 |
+
|
84 |
+
def __call__(self, image: PIL.Image.Image):
|
85 |
+
image_tensor, orig_size = self.load_image(image)
|
86 |
+
mask = self.predict(image_tensor, orig_size)
|
87 |
+
|
88 |
+
pil_mask = Image.fromarray(mask).convert('L')
|
89 |
+
im_rgb = image.convert("RGB")
|
90 |
+
|
91 |
+
im_rgba = im_rgb.copy()
|
92 |
+
im_rgba.putalpha(pil_mask)
|
93 |
+
|
94 |
+
return [im_rgba, pil_mask]
|
95 |
+
|
96 |
+
def predict(self, inputs_val: torch.Tensor, shapes_val):
|
97 |
+
"""
|
98 |
+
Given an Image, predict the mask
|
99 |
+
"""
|
100 |
+
|
101 |
+
if self.model_digit == "full":
|
102 |
+
inputs_val = inputs_val.type(torch.FloatTensor)
|
103 |
+
else:
|
104 |
+
inputs_val = inputs_val.type(torch.HalfTensor)
|
105 |
+
|
106 |
+
inputs_val_v = Variable(inputs_val, requires_grad=False).to(device) # wrap inputs in Variable
|
107 |
+
|
108 |
+
ds_val = self.model(inputs_val_v)[0] # list of 6 results
|
109 |
+
|
110 |
+
# B x 1 x H x W # we want the first one which is the most accurate prediction
|
111 |
+
pred_val = ds_val[0][0, :, :, :]
|
112 |
+
|
113 |
+
# recover the prediction spatial size to the orignal image size
|
114 |
+
pred_val = torch.squeeze(
|
115 |
+
F.upsample(torch.unsqueeze(pred_val, 0), (shapes_val[0][0], shapes_val[0][1]), mode='bilinear'))
|
116 |
+
|
117 |
+
ma = torch.max(pred_val)
|
118 |
+
mi = torch.min(pred_val)
|
119 |
+
pred_val = (pred_val - mi) / (ma - mi) # max = 1
|
120 |
+
|
121 |
+
if device == 'cuda':
|
122 |
+
torch.cuda.empty_cache()
|
123 |
+
return (pred_val.detach().cpu().numpy() * 255).astype(np.uint8) # it is the mask we need
|
124 |
+
|
125 |
+
|
126 |
+
# a = IsNetPipeLine(model_path="save_models/isnet.pth")
|
127 |
+
# input_image = Image.open("image_0mx.png")
|
128 |
+
# rgb, mask = a(input_image)
|
129 |
+
#
|
130 |
+
# rgb.save("rgb.png")
|
131 |
+
# mask.save("mask.png")
|
DIS/models/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
from .isnet import ISNetGTEncoder, ISNetDIS
|
DIS/models/isnet.py
ADDED
@@ -0,0 +1,608 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
|
5 |
+
bce_loss = nn.BCELoss(size_average=True)
|
6 |
+
|
7 |
+
|
8 |
+
def muti_loss_fusion(preds, target):
|
9 |
+
loss0 = 0.0
|
10 |
+
loss = 0.0
|
11 |
+
|
12 |
+
for i in range(0, len(preds)):
|
13 |
+
# print("i: ", i, preds[i].shape)
|
14 |
+
if (preds[i].shape[2] != target.shape[2] or preds[i].shape[3] != target.shape[3]):
|
15 |
+
# tmp_target = _upsample_like(target,preds[i])
|
16 |
+
tmp_target = F.interpolate(target, size=preds[i].size()[2:], mode='bilinear', align_corners=True)
|
17 |
+
loss = loss + bce_loss(preds[i], tmp_target)
|
18 |
+
else:
|
19 |
+
loss = loss + bce_loss(preds[i], target)
|
20 |
+
if (i == 0):
|
21 |
+
loss0 = loss
|
22 |
+
return loss0, loss
|
23 |
+
|
24 |
+
|
25 |
+
fea_loss = nn.MSELoss(size_average=True)
|
26 |
+
kl_loss = nn.KLDivLoss(size_average=True)
|
27 |
+
l1_loss = nn.L1Loss(size_average=True)
|
28 |
+
smooth_l1_loss = nn.SmoothL1Loss(size_average=True)
|
29 |
+
|
30 |
+
|
31 |
+
def muti_loss_fusion_kl(preds, target, dfs, fs, mode='MSE'):
|
32 |
+
loss0 = 0.0
|
33 |
+
loss = 0.0
|
34 |
+
|
35 |
+
for i in range(0, len(preds)):
|
36 |
+
# print("i: ", i, preds[i].shape)
|
37 |
+
if (preds[i].shape[2] != target.shape[2] or preds[i].shape[3] != target.shape[3]):
|
38 |
+
# tmp_target = _upsample_like(target,preds[i])
|
39 |
+
tmp_target = F.interpolate(target, size=preds[i].size()[2:], mode='bilinear', align_corners=True)
|
40 |
+
loss = loss + bce_loss(preds[i], tmp_target)
|
41 |
+
else:
|
42 |
+
loss = loss + bce_loss(preds[i], target)
|
43 |
+
if (i == 0):
|
44 |
+
loss0 = loss
|
45 |
+
|
46 |
+
for i in range(0, len(dfs)):
|
47 |
+
if (mode == 'MSE'):
|
48 |
+
loss = loss + fea_loss(dfs[i], fs[i]) ### add the mse loss of features as additional constraints
|
49 |
+
# print("fea_loss: ", fea_loss(dfs[i],fs[i]).item())
|
50 |
+
elif (mode == 'KL'):
|
51 |
+
loss = loss + kl_loss(F.log_softmax(dfs[i], dim=1), F.softmax(fs[i], dim=1))
|
52 |
+
# print("kl_loss: ", kl_loss(F.log_softmax(dfs[i],dim=1),F.softmax(fs[i],dim=1)).item())
|
53 |
+
elif (mode == 'MAE'):
|
54 |
+
loss = loss + l1_loss(dfs[i], fs[i])
|
55 |
+
# print("ls_loss: ", l1_loss(dfs[i],fs[i]))
|
56 |
+
elif (mode == 'SmoothL1'):
|
57 |
+
loss = loss + smooth_l1_loss(dfs[i], fs[i])
|
58 |
+
# print("SmoothL1: ", smooth_l1_loss(dfs[i],fs[i]).item())
|
59 |
+
|
60 |
+
return loss0, loss
|
61 |
+
|
62 |
+
|
63 |
+
class REBNCONV(nn.Module):
|
64 |
+
def __init__(self, in_ch=3, out_ch=3, dirate=1, stride=1):
|
65 |
+
super(REBNCONV, self).__init__()
|
66 |
+
|
67 |
+
self.conv_s1 = nn.Conv2d(in_ch, out_ch, 3, padding=1 * dirate, dilation=1 * dirate, stride=stride)
|
68 |
+
self.bn_s1 = nn.BatchNorm2d(out_ch)
|
69 |
+
self.relu_s1 = nn.ReLU(inplace=True)
|
70 |
+
|
71 |
+
def forward(self, x):
|
72 |
+
hx = x
|
73 |
+
xout = self.relu_s1(self.bn_s1(self.conv_s1(hx)))
|
74 |
+
|
75 |
+
return xout
|
76 |
+
|
77 |
+
|
78 |
+
## upsample tensor 'src' to have the same spatial size with tensor 'tar'
|
79 |
+
def _upsample_like(src, tar):
|
80 |
+
src = F.upsample(src, size=tar.shape[2:], mode='bilinear')
|
81 |
+
|
82 |
+
return src
|
83 |
+
|
84 |
+
|
85 |
+
### RSU-7 ###
|
86 |
+
class RSU7(nn.Module):
|
87 |
+
|
88 |
+
def __init__(self, in_ch=3, mid_ch=12, out_ch=3, img_size=512):
|
89 |
+
super(RSU7, self).__init__()
|
90 |
+
|
91 |
+
self.in_ch = in_ch
|
92 |
+
self.mid_ch = mid_ch
|
93 |
+
self.out_ch = out_ch
|
94 |
+
|
95 |
+
self.rebnconvin = REBNCONV(in_ch, out_ch, dirate=1) ## 1 -> 1/2
|
96 |
+
|
97 |
+
self.rebnconv1 = REBNCONV(out_ch, mid_ch, dirate=1)
|
98 |
+
self.pool1 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
99 |
+
|
100 |
+
self.rebnconv2 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
101 |
+
self.pool2 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
102 |
+
|
103 |
+
self.rebnconv3 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
104 |
+
self.pool3 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
105 |
+
|
106 |
+
self.rebnconv4 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
107 |
+
self.pool4 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
108 |
+
|
109 |
+
self.rebnconv5 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
110 |
+
self.pool5 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
111 |
+
|
112 |
+
self.rebnconv6 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
113 |
+
|
114 |
+
self.rebnconv7 = REBNCONV(mid_ch, mid_ch, dirate=2)
|
115 |
+
|
116 |
+
self.rebnconv6d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
117 |
+
self.rebnconv5d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
118 |
+
self.rebnconv4d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
119 |
+
self.rebnconv3d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
120 |
+
self.rebnconv2d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
121 |
+
self.rebnconv1d = REBNCONV(mid_ch * 2, out_ch, dirate=1)
|
122 |
+
|
123 |
+
def forward(self, x):
|
124 |
+
b, c, h, w = x.shape
|
125 |
+
|
126 |
+
hx = x
|
127 |
+
hxin = self.rebnconvin(hx)
|
128 |
+
|
129 |
+
hx1 = self.rebnconv1(hxin)
|
130 |
+
hx = self.pool1(hx1)
|
131 |
+
|
132 |
+
hx2 = self.rebnconv2(hx)
|
133 |
+
hx = self.pool2(hx2)
|
134 |
+
|
135 |
+
hx3 = self.rebnconv3(hx)
|
136 |
+
hx = self.pool3(hx3)
|
137 |
+
|
138 |
+
hx4 = self.rebnconv4(hx)
|
139 |
+
hx = self.pool4(hx4)
|
140 |
+
|
141 |
+
hx5 = self.rebnconv5(hx)
|
142 |
+
hx = self.pool5(hx5)
|
143 |
+
|
144 |
+
hx6 = self.rebnconv6(hx)
|
145 |
+
|
146 |
+
hx7 = self.rebnconv7(hx6)
|
147 |
+
|
148 |
+
hx6d = self.rebnconv6d(torch.cat((hx7, hx6), 1))
|
149 |
+
hx6dup = _upsample_like(hx6d, hx5)
|
150 |
+
|
151 |
+
hx5d = self.rebnconv5d(torch.cat((hx6dup, hx5), 1))
|
152 |
+
hx5dup = _upsample_like(hx5d, hx4)
|
153 |
+
|
154 |
+
hx4d = self.rebnconv4d(torch.cat((hx5dup, hx4), 1))
|
155 |
+
hx4dup = _upsample_like(hx4d, hx3)
|
156 |
+
|
157 |
+
hx3d = self.rebnconv3d(torch.cat((hx4dup, hx3), 1))
|
158 |
+
hx3dup = _upsample_like(hx3d, hx2)
|
159 |
+
|
160 |
+
hx2d = self.rebnconv2d(torch.cat((hx3dup, hx2), 1))
|
161 |
+
hx2dup = _upsample_like(hx2d, hx1)
|
162 |
+
|
163 |
+
hx1d = self.rebnconv1d(torch.cat((hx2dup, hx1), 1))
|
164 |
+
|
165 |
+
return hx1d + hxin
|
166 |
+
|
167 |
+
|
168 |
+
### RSU-6 ###
|
169 |
+
class RSU6(nn.Module):
|
170 |
+
|
171 |
+
def __init__(self, in_ch=3, mid_ch=12, out_ch=3):
|
172 |
+
super(RSU6, self).__init__()
|
173 |
+
|
174 |
+
self.rebnconvin = REBNCONV(in_ch, out_ch, dirate=1)
|
175 |
+
|
176 |
+
self.rebnconv1 = REBNCONV(out_ch, mid_ch, dirate=1)
|
177 |
+
self.pool1 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
178 |
+
|
179 |
+
self.rebnconv2 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
180 |
+
self.pool2 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
181 |
+
|
182 |
+
self.rebnconv3 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
183 |
+
self.pool3 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
184 |
+
|
185 |
+
self.rebnconv4 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
186 |
+
self.pool4 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
187 |
+
|
188 |
+
self.rebnconv5 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
189 |
+
|
190 |
+
self.rebnconv6 = REBNCONV(mid_ch, mid_ch, dirate=2)
|
191 |
+
|
192 |
+
self.rebnconv5d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
193 |
+
self.rebnconv4d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
194 |
+
self.rebnconv3d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
195 |
+
self.rebnconv2d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
196 |
+
self.rebnconv1d = REBNCONV(mid_ch * 2, out_ch, dirate=1)
|
197 |
+
|
198 |
+
def forward(self, x):
|
199 |
+
hx = x
|
200 |
+
|
201 |
+
hxin = self.rebnconvin(hx)
|
202 |
+
|
203 |
+
hx1 = self.rebnconv1(hxin)
|
204 |
+
hx = self.pool1(hx1)
|
205 |
+
|
206 |
+
hx2 = self.rebnconv2(hx)
|
207 |
+
hx = self.pool2(hx2)
|
208 |
+
|
209 |
+
hx3 = self.rebnconv3(hx)
|
210 |
+
hx = self.pool3(hx3)
|
211 |
+
|
212 |
+
hx4 = self.rebnconv4(hx)
|
213 |
+
hx = self.pool4(hx4)
|
214 |
+
|
215 |
+
hx5 = self.rebnconv5(hx)
|
216 |
+
|
217 |
+
hx6 = self.rebnconv6(hx5)
|
218 |
+
|
219 |
+
hx5d = self.rebnconv5d(torch.cat((hx6, hx5), 1))
|
220 |
+
hx5dup = _upsample_like(hx5d, hx4)
|
221 |
+
|
222 |
+
hx4d = self.rebnconv4d(torch.cat((hx5dup, hx4), 1))
|
223 |
+
hx4dup = _upsample_like(hx4d, hx3)
|
224 |
+
|
225 |
+
hx3d = self.rebnconv3d(torch.cat((hx4dup, hx3), 1))
|
226 |
+
hx3dup = _upsample_like(hx3d, hx2)
|
227 |
+
|
228 |
+
hx2d = self.rebnconv2d(torch.cat((hx3dup, hx2), 1))
|
229 |
+
hx2dup = _upsample_like(hx2d, hx1)
|
230 |
+
|
231 |
+
hx1d = self.rebnconv1d(torch.cat((hx2dup, hx1), 1))
|
232 |
+
|
233 |
+
return hx1d + hxin
|
234 |
+
|
235 |
+
|
236 |
+
### RSU-5 ###
|
237 |
+
class RSU5(nn.Module):
|
238 |
+
|
239 |
+
def __init__(self, in_ch=3, mid_ch=12, out_ch=3):
|
240 |
+
super(RSU5, self).__init__()
|
241 |
+
|
242 |
+
self.rebnconvin = REBNCONV(in_ch, out_ch, dirate=1)
|
243 |
+
|
244 |
+
self.rebnconv1 = REBNCONV(out_ch, mid_ch, dirate=1)
|
245 |
+
self.pool1 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
246 |
+
|
247 |
+
self.rebnconv2 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
248 |
+
self.pool2 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
249 |
+
|
250 |
+
self.rebnconv3 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
251 |
+
self.pool3 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
252 |
+
|
253 |
+
self.rebnconv4 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
254 |
+
|
255 |
+
self.rebnconv5 = REBNCONV(mid_ch, mid_ch, dirate=2)
|
256 |
+
|
257 |
+
self.rebnconv4d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
258 |
+
self.rebnconv3d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
259 |
+
self.rebnconv2d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
260 |
+
self.rebnconv1d = REBNCONV(mid_ch * 2, out_ch, dirate=1)
|
261 |
+
|
262 |
+
def forward(self, x):
|
263 |
+
hx = x
|
264 |
+
|
265 |
+
hxin = self.rebnconvin(hx)
|
266 |
+
|
267 |
+
hx1 = self.rebnconv1(hxin)
|
268 |
+
hx = self.pool1(hx1)
|
269 |
+
|
270 |
+
hx2 = self.rebnconv2(hx)
|
271 |
+
hx = self.pool2(hx2)
|
272 |
+
|
273 |
+
hx3 = self.rebnconv3(hx)
|
274 |
+
hx = self.pool3(hx3)
|
275 |
+
|
276 |
+
hx4 = self.rebnconv4(hx)
|
277 |
+
|
278 |
+
hx5 = self.rebnconv5(hx4)
|
279 |
+
|
280 |
+
hx4d = self.rebnconv4d(torch.cat((hx5, hx4), 1))
|
281 |
+
hx4dup = _upsample_like(hx4d, hx3)
|
282 |
+
|
283 |
+
hx3d = self.rebnconv3d(torch.cat((hx4dup, hx3), 1))
|
284 |
+
hx3dup = _upsample_like(hx3d, hx2)
|
285 |
+
|
286 |
+
hx2d = self.rebnconv2d(torch.cat((hx3dup, hx2), 1))
|
287 |
+
hx2dup = _upsample_like(hx2d, hx1)
|
288 |
+
|
289 |
+
hx1d = self.rebnconv1d(torch.cat((hx2dup, hx1), 1))
|
290 |
+
|
291 |
+
return hx1d + hxin
|
292 |
+
|
293 |
+
|
294 |
+
### RSU-4 ###
|
295 |
+
class RSU4(nn.Module):
|
296 |
+
|
297 |
+
def __init__(self, in_ch=3, mid_ch=12, out_ch=3):
|
298 |
+
super(RSU4, self).__init__()
|
299 |
+
|
300 |
+
self.rebnconvin = REBNCONV(in_ch, out_ch, dirate=1)
|
301 |
+
|
302 |
+
self.rebnconv1 = REBNCONV(out_ch, mid_ch, dirate=1)
|
303 |
+
self.pool1 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
304 |
+
|
305 |
+
self.rebnconv2 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
306 |
+
self.pool2 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
307 |
+
|
308 |
+
self.rebnconv3 = REBNCONV(mid_ch, mid_ch, dirate=1)
|
309 |
+
|
310 |
+
self.rebnconv4 = REBNCONV(mid_ch, mid_ch, dirate=2)
|
311 |
+
|
312 |
+
self.rebnconv3d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
313 |
+
self.rebnconv2d = REBNCONV(mid_ch * 2, mid_ch, dirate=1)
|
314 |
+
self.rebnconv1d = REBNCONV(mid_ch * 2, out_ch, dirate=1)
|
315 |
+
|
316 |
+
def forward(self, x):
|
317 |
+
hx = x
|
318 |
+
|
319 |
+
hxin = self.rebnconvin(hx)
|
320 |
+
|
321 |
+
hx1 = self.rebnconv1(hxin)
|
322 |
+
hx = self.pool1(hx1)
|
323 |
+
|
324 |
+
hx2 = self.rebnconv2(hx)
|
325 |
+
hx = self.pool2(hx2)
|
326 |
+
|
327 |
+
hx3 = self.rebnconv3(hx)
|
328 |
+
|
329 |
+
hx4 = self.rebnconv4(hx3)
|
330 |
+
|
331 |
+
hx3d = self.rebnconv3d(torch.cat((hx4, hx3), 1))
|
332 |
+
hx3dup = _upsample_like(hx3d, hx2)
|
333 |
+
|
334 |
+
hx2d = self.rebnconv2d(torch.cat((hx3dup, hx2), 1))
|
335 |
+
hx2dup = _upsample_like(hx2d, hx1)
|
336 |
+
|
337 |
+
hx1d = self.rebnconv1d(torch.cat((hx2dup, hx1), 1))
|
338 |
+
|
339 |
+
return hx1d + hxin
|
340 |
+
|
341 |
+
|
342 |
+
### RSU-4F ###
|
343 |
+
class RSU4F(nn.Module):
|
344 |
+
|
345 |
+
def __init__(self, in_ch=3, mid_ch=12, out_ch=3):
|
346 |
+
super(RSU4F, self).__init__()
|
347 |
+
|
348 |
+
self.rebnconvin = REBNCONV(in_ch, out_ch, dirate=1)
|
349 |
+
|
350 |
+
self.rebnconv1 = REBNCONV(out_ch, mid_ch, dirate=1)
|
351 |
+
self.rebnconv2 = REBNCONV(mid_ch, mid_ch, dirate=2)
|
352 |
+
self.rebnconv3 = REBNCONV(mid_ch, mid_ch, dirate=4)
|
353 |
+
|
354 |
+
self.rebnconv4 = REBNCONV(mid_ch, mid_ch, dirate=8)
|
355 |
+
|
356 |
+
self.rebnconv3d = REBNCONV(mid_ch * 2, mid_ch, dirate=4)
|
357 |
+
self.rebnconv2d = REBNCONV(mid_ch * 2, mid_ch, dirate=2)
|
358 |
+
self.rebnconv1d = REBNCONV(mid_ch * 2, out_ch, dirate=1)
|
359 |
+
|
360 |
+
def forward(self, x):
|
361 |
+
hx = x
|
362 |
+
|
363 |
+
hxin = self.rebnconvin(hx)
|
364 |
+
|
365 |
+
hx1 = self.rebnconv1(hxin)
|
366 |
+
hx2 = self.rebnconv2(hx1)
|
367 |
+
hx3 = self.rebnconv3(hx2)
|
368 |
+
|
369 |
+
hx4 = self.rebnconv4(hx3)
|
370 |
+
|
371 |
+
hx3d = self.rebnconv3d(torch.cat((hx4, hx3), 1))
|
372 |
+
hx2d = self.rebnconv2d(torch.cat((hx3d, hx2), 1))
|
373 |
+
hx1d = self.rebnconv1d(torch.cat((hx2d, hx1), 1))
|
374 |
+
|
375 |
+
return hx1d + hxin
|
376 |
+
|
377 |
+
|
378 |
+
class myrebnconv(nn.Module):
|
379 |
+
def __init__(self, in_ch=3,
|
380 |
+
out_ch=1,
|
381 |
+
kernel_size=3,
|
382 |
+
stride=1,
|
383 |
+
padding=1,
|
384 |
+
dilation=1,
|
385 |
+
groups=1):
|
386 |
+
super(myrebnconv, self).__init__()
|
387 |
+
|
388 |
+
self.conv = nn.Conv2d(in_ch,
|
389 |
+
out_ch,
|
390 |
+
kernel_size=kernel_size,
|
391 |
+
stride=stride,
|
392 |
+
padding=padding,
|
393 |
+
dilation=dilation,
|
394 |
+
groups=groups)
|
395 |
+
self.bn = nn.BatchNorm2d(out_ch)
|
396 |
+
self.rl = nn.ReLU(inplace=True)
|
397 |
+
|
398 |
+
def forward(self, x):
|
399 |
+
return self.rl(self.bn(self.conv(x)))
|
400 |
+
|
401 |
+
|
402 |
+
class ISNetGTEncoder(nn.Module):
|
403 |
+
|
404 |
+
def __init__(self, in_ch=1, out_ch=1):
|
405 |
+
super(ISNetGTEncoder, self).__init__()
|
406 |
+
|
407 |
+
self.conv_in = myrebnconv(in_ch, 16, 3, stride=2, padding=1) # nn.Conv2d(in_ch,64,3,stride=2,padding=1)
|
408 |
+
|
409 |
+
self.stage1 = RSU7(16, 16, 64)
|
410 |
+
self.pool12 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
411 |
+
|
412 |
+
self.stage2 = RSU6(64, 16, 64)
|
413 |
+
self.pool23 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
414 |
+
|
415 |
+
self.stage3 = RSU5(64, 32, 128)
|
416 |
+
self.pool34 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
417 |
+
|
418 |
+
self.stage4 = RSU4(128, 32, 256)
|
419 |
+
self.pool45 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
420 |
+
|
421 |
+
self.stage5 = RSU4F(256, 64, 512)
|
422 |
+
self.pool56 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
423 |
+
|
424 |
+
self.stage6 = RSU4F(512, 64, 512)
|
425 |
+
|
426 |
+
self.side1 = nn.Conv2d(64, out_ch, 3, padding=1)
|
427 |
+
self.side2 = nn.Conv2d(64, out_ch, 3, padding=1)
|
428 |
+
self.side3 = nn.Conv2d(128, out_ch, 3, padding=1)
|
429 |
+
self.side4 = nn.Conv2d(256, out_ch, 3, padding=1)
|
430 |
+
self.side5 = nn.Conv2d(512, out_ch, 3, padding=1)
|
431 |
+
self.side6 = nn.Conv2d(512, out_ch, 3, padding=1)
|
432 |
+
|
433 |
+
def compute_loss(self, preds, targets):
|
434 |
+
return muti_loss_fusion(preds, targets)
|
435 |
+
|
436 |
+
def forward(self, x):
|
437 |
+
hx = x
|
438 |
+
|
439 |
+
hxin = self.conv_in(hx)
|
440 |
+
# hx = self.pool_in(hxin)
|
441 |
+
|
442 |
+
# stage 1
|
443 |
+
hx1 = self.stage1(hxin)
|
444 |
+
hx = self.pool12(hx1)
|
445 |
+
|
446 |
+
# stage 2
|
447 |
+
hx2 = self.stage2(hx)
|
448 |
+
hx = self.pool23(hx2)
|
449 |
+
|
450 |
+
# stage 3
|
451 |
+
hx3 = self.stage3(hx)
|
452 |
+
hx = self.pool34(hx3)
|
453 |
+
|
454 |
+
# stage 4
|
455 |
+
hx4 = self.stage4(hx)
|
456 |
+
hx = self.pool45(hx4)
|
457 |
+
|
458 |
+
# stage 5
|
459 |
+
hx5 = self.stage5(hx)
|
460 |
+
hx = self.pool56(hx5)
|
461 |
+
|
462 |
+
# stage 6
|
463 |
+
hx6 = self.stage6(hx)
|
464 |
+
|
465 |
+
# side output
|
466 |
+
d1 = self.side1(hx1)
|
467 |
+
d1 = _upsample_like(d1, x)
|
468 |
+
|
469 |
+
d2 = self.side2(hx2)
|
470 |
+
d2 = _upsample_like(d2, x)
|
471 |
+
|
472 |
+
d3 = self.side3(hx3)
|
473 |
+
d3 = _upsample_like(d3, x)
|
474 |
+
|
475 |
+
d4 = self.side4(hx4)
|
476 |
+
d4 = _upsample_like(d4, x)
|
477 |
+
|
478 |
+
d5 = self.side5(hx5)
|
479 |
+
d5 = _upsample_like(d5, x)
|
480 |
+
|
481 |
+
d6 = self.side6(hx6)
|
482 |
+
d6 = _upsample_like(d6, x)
|
483 |
+
|
484 |
+
# d0 = self.outconv(torch.cat((d1,d2,d3,d4,d5,d6),1))
|
485 |
+
|
486 |
+
return [F.sigmoid(d1), F.sigmoid(d2), F.sigmoid(d3), F.sigmoid(d4), F.sigmoid(d5), F.sigmoid(d6)], [hx1, hx2,
|
487 |
+
hx3, hx4,
|
488 |
+
hx5, hx6]
|
489 |
+
|
490 |
+
|
491 |
+
class ISNetDIS(nn.Module):
|
492 |
+
|
493 |
+
def __init__(self, in_ch=3, out_ch=1):
|
494 |
+
super(ISNetDIS, self).__init__()
|
495 |
+
|
496 |
+
self.conv_in = nn.Conv2d(in_ch, 64, 3, stride=2, padding=1)
|
497 |
+
self.pool_in = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
498 |
+
|
499 |
+
self.stage1 = RSU7(64, 32, 64)
|
500 |
+
self.pool12 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
501 |
+
|
502 |
+
self.stage2 = RSU6(64, 32, 128)
|
503 |
+
self.pool23 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
504 |
+
|
505 |
+
self.stage3 = RSU5(128, 64, 256)
|
506 |
+
self.pool34 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
507 |
+
|
508 |
+
self.stage4 = RSU4(256, 128, 512)
|
509 |
+
self.pool45 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
510 |
+
|
511 |
+
self.stage5 = RSU4F(512, 256, 512)
|
512 |
+
self.pool56 = nn.MaxPool2d(2, stride=2, ceil_mode=True)
|
513 |
+
|
514 |
+
self.stage6 = RSU4F(512, 256, 512)
|
515 |
+
|
516 |
+
# decoder
|
517 |
+
self.stage5d = RSU4F(1024, 256, 512)
|
518 |
+
self.stage4d = RSU4(1024, 128, 256)
|
519 |
+
self.stage3d = RSU5(512, 64, 128)
|
520 |
+
self.stage2d = RSU6(256, 32, 64)
|
521 |
+
self.stage1d = RSU7(128, 16, 64)
|
522 |
+
|
523 |
+
self.side1 = nn.Conv2d(64, out_ch, 3, padding=1)
|
524 |
+
self.side2 = nn.Conv2d(64, out_ch, 3, padding=1)
|
525 |
+
self.side3 = nn.Conv2d(128, out_ch, 3, padding=1)
|
526 |
+
self.side4 = nn.Conv2d(256, out_ch, 3, padding=1)
|
527 |
+
self.side5 = nn.Conv2d(512, out_ch, 3, padding=1)
|
528 |
+
self.side6 = nn.Conv2d(512, out_ch, 3, padding=1)
|
529 |
+
|
530 |
+
# self.outconv = nn.Conv2d(6*out_ch,out_ch,1)
|
531 |
+
|
532 |
+
def compute_loss_kl(self, preds, targets, dfs, fs, mode='MSE'):
|
533 |
+
# return muti_loss_fusion(preds,targets)
|
534 |
+
return muti_loss_fusion_kl(preds, targets, dfs, fs, mode=mode)
|
535 |
+
|
536 |
+
def compute_loss(self, preds, targets):
|
537 |
+
# return muti_loss_fusion(preds,targets)
|
538 |
+
return muti_loss_fusion(preds, targets)
|
539 |
+
|
540 |
+
def forward(self, x):
|
541 |
+
hx = x
|
542 |
+
|
543 |
+
hxin = self.conv_in(hx)
|
544 |
+
# hx = self.pool_in(hxin)
|
545 |
+
|
546 |
+
# stage 1
|
547 |
+
hx1 = self.stage1(hxin)
|
548 |
+
hx = self.pool12(hx1)
|
549 |
+
|
550 |
+
# stage 2
|
551 |
+
hx2 = self.stage2(hx)
|
552 |
+
hx = self.pool23(hx2)
|
553 |
+
|
554 |
+
# stage 3
|
555 |
+
hx3 = self.stage3(hx)
|
556 |
+
hx = self.pool34(hx3)
|
557 |
+
|
558 |
+
# stage 4
|
559 |
+
hx4 = self.stage4(hx)
|
560 |
+
hx = self.pool45(hx4)
|
561 |
+
|
562 |
+
# stage 5
|
563 |
+
hx5 = self.stage5(hx)
|
564 |
+
hx = self.pool56(hx5)
|
565 |
+
|
566 |
+
# stage 6
|
567 |
+
hx6 = self.stage6(hx)
|
568 |
+
hx6up = _upsample_like(hx6, hx5)
|
569 |
+
|
570 |
+
# -------------------- decoder --------------------
|
571 |
+
hx5d = self.stage5d(torch.cat((hx6up, hx5), 1))
|
572 |
+
hx5dup = _upsample_like(hx5d, hx4)
|
573 |
+
|
574 |
+
hx4d = self.stage4d(torch.cat((hx5dup, hx4), 1))
|
575 |
+
hx4dup = _upsample_like(hx4d, hx3)
|
576 |
+
|
577 |
+
hx3d = self.stage3d(torch.cat((hx4dup, hx3), 1))
|
578 |
+
hx3dup = _upsample_like(hx3d, hx2)
|
579 |
+
|
580 |
+
hx2d = self.stage2d(torch.cat((hx3dup, hx2), 1))
|
581 |
+
hx2dup = _upsample_like(hx2d, hx1)
|
582 |
+
|
583 |
+
hx1d = self.stage1d(torch.cat((hx2dup, hx1), 1))
|
584 |
+
|
585 |
+
# side output
|
586 |
+
d1 = self.side1(hx1d)
|
587 |
+
d1 = _upsample_like(d1, x)
|
588 |
+
|
589 |
+
d2 = self.side2(hx2d)
|
590 |
+
d2 = _upsample_like(d2, x)
|
591 |
+
|
592 |
+
d3 = self.side3(hx3d)
|
593 |
+
d3 = _upsample_like(d3, x)
|
594 |
+
|
595 |
+
d4 = self.side4(hx4d)
|
596 |
+
d4 = _upsample_like(d4, x)
|
597 |
+
|
598 |
+
d5 = self.side5(hx5d)
|
599 |
+
d5 = _upsample_like(d5, x)
|
600 |
+
|
601 |
+
d6 = self.side6(hx6)
|
602 |
+
d6 = _upsample_like(d6, x)
|
603 |
+
|
604 |
+
# d0 = self.outconv(torch.cat((d1,d2,d3,d4,d5,d6),1))
|
605 |
+
|
606 |
+
return [F.sigmoid(d1), F.sigmoid(d2), F.sigmoid(d3), F.sigmoid(d4), F.sigmoid(d5), F.sigmoid(d6)], [hx1d, hx2d,
|
607 |
+
hx3d, hx4d,
|
608 |
+
hx5d, hx6]
|
DIS/pytorch18.yml
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
name: pytorch18
|
2 |
+
channels:
|
3 |
+
- conda-forge
|
4 |
+
- anaconda
|
5 |
+
- pytorch
|
6 |
+
- defaults
|
7 |
+
dependencies:
|
8 |
+
- _libgcc_mutex=0.1=main
|
9 |
+
- _openmp_mutex=4.5=1_gnu
|
10 |
+
- blas=1.0=mkl
|
11 |
+
- brotli=1.0.9=he6710b0_2
|
12 |
+
- bzip2=1.0.8=h7b6447c_0
|
13 |
+
- ca-certificates=2022.2.1=h06a4308_0
|
14 |
+
- certifi=2021.10.8=py37h06a4308_2
|
15 |
+
- cloudpickle=2.0.0=pyhd3eb1b0_0
|
16 |
+
- colorama=0.4.4=pyhd3eb1b0_0
|
17 |
+
- cudatoolkit=10.2.89=hfd86e86_1
|
18 |
+
- cycler=0.11.0=pyhd3eb1b0_0
|
19 |
+
- cytoolz=0.11.0=py37h7b6447c_0
|
20 |
+
- dask-core=2021.10.0=pyhd3eb1b0_0
|
21 |
+
- ffmpeg=4.3=hf484d3e_0
|
22 |
+
- fonttools=4.25.0=pyhd3eb1b0_0
|
23 |
+
- freetype=2.11.0=h70c0345_0
|
24 |
+
- fsspec=2022.2.0=pyhd3eb1b0_0
|
25 |
+
- gmp=6.2.1=h2531618_2
|
26 |
+
- gnutls=3.6.15=he1e5248_0
|
27 |
+
- imageio=2.9.0=pyhd3eb1b0_0
|
28 |
+
- intel-openmp=2021.4.0=h06a4308_3561
|
29 |
+
- jpeg=9b=h024ee3a_2
|
30 |
+
- kiwisolver=1.3.2=py37h295c915_0
|
31 |
+
- lame=3.100=h7b6447c_0
|
32 |
+
- lcms2=2.12=h3be6417_0
|
33 |
+
- ld_impl_linux-64=2.35.1=h7274673_9
|
34 |
+
- libffi=3.3=he6710b0_2
|
35 |
+
- libgcc-ng=9.3.0=h5101ec6_17
|
36 |
+
- libgfortran-ng=7.5.0=ha8ba4b0_17
|
37 |
+
- libgfortran4=7.5.0=ha8ba4b0_17
|
38 |
+
- libgomp=9.3.0=h5101ec6_17
|
39 |
+
- libiconv=1.15=h63c8f33_5
|
40 |
+
- libidn2=2.3.2=h7f8727e_0
|
41 |
+
- libpng=1.6.37=hbc83047_0
|
42 |
+
- libstdcxx-ng=9.3.0=hd4cf53a_17
|
43 |
+
- libtasn1=4.16.0=h27cfd23_0
|
44 |
+
- libtiff=4.2.0=h85742a9_0
|
45 |
+
- libunistring=0.9.10=h27cfd23_0
|
46 |
+
- libuv=1.40.0=h7b6447c_0
|
47 |
+
- libwebp-base=1.2.2=h7f8727e_0
|
48 |
+
- locket=0.2.1=py37h06a4308_2
|
49 |
+
- lz4-c=1.9.3=h295c915_1
|
50 |
+
- matplotlib-base=3.5.1=py37ha18d171_1
|
51 |
+
- mkl=2021.4.0=h06a4308_640
|
52 |
+
- mkl-service=2.4.0=py37h7f8727e_0
|
53 |
+
- mkl_fft=1.3.1=py37hd3c417c_0
|
54 |
+
- mkl_random=1.2.2=py37h51133e4_0
|
55 |
+
- munkres=1.1.4=py_0
|
56 |
+
- ncurses=6.3=h7f8727e_2
|
57 |
+
- nettle=3.7.3=hbbd107a_1
|
58 |
+
- networkx=2.6.3=pyhd3eb1b0_0
|
59 |
+
- ninja=1.10.2=py37hd09550d_3
|
60 |
+
- numpy=1.21.2=py37h20f2e39_0
|
61 |
+
- numpy-base=1.21.2=py37h79a1101_0
|
62 |
+
- olefile=0.46=py37_0
|
63 |
+
- openh264=2.1.1=h4ff587b_0
|
64 |
+
- openssl=1.1.1n=h7f8727e_0
|
65 |
+
- packaging=21.3=pyhd3eb1b0_0
|
66 |
+
- partd=1.2.0=pyhd3eb1b0_1
|
67 |
+
- pillow=8.0.0=py37h9a89aac_0
|
68 |
+
- pip=21.2.2=py37h06a4308_0
|
69 |
+
- pyparsing=3.0.4=pyhd3eb1b0_0
|
70 |
+
- python=3.7.11=h12debd9_0
|
71 |
+
- python-dateutil=2.8.2=pyhd3eb1b0_0
|
72 |
+
- pytorch=1.8.0=py3.7_cuda10.2_cudnn7.6.5_0
|
73 |
+
- pywavelets=1.1.1=py37h7b6447c_2
|
74 |
+
- pyyaml=6.0=py37h7f8727e_1
|
75 |
+
- readline=8.1.2=h7f8727e_1
|
76 |
+
- scikit-image=0.15.0=py37hb3f55d8_2
|
77 |
+
- scipy=1.7.3=py37hc147768_0
|
78 |
+
- setuptools=58.0.4=py37h06a4308_0
|
79 |
+
- six=1.16.0=pyhd3eb1b0_1
|
80 |
+
- sqlite=3.38.0=hc218d9a_0
|
81 |
+
- tk=8.6.11=h1ccaba5_0
|
82 |
+
- toolz=0.11.2=pyhd3eb1b0_0
|
83 |
+
- torchaudio=0.8.0=py37
|
84 |
+
- torchvision=0.9.0=py37_cu102
|
85 |
+
- tqdm=4.63.0=pyhd8ed1ab_0
|
86 |
+
- typing_extensions=3.10.0.2=pyh06a4308_0
|
87 |
+
- wheel=0.37.1=pyhd3eb1b0_0
|
88 |
+
- xz=5.2.5=h7b6447c_0
|
89 |
+
- yaml=0.2.5=h7b6447c_0
|
90 |
+
- zlib=1.2.11=h7f8727e_4
|
91 |
+
- zstd=1.4.9=haebb681_0
|
92 |
+
prefix: /home/solar/anaconda3/envs/pytorch18
|
app.py
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
reference: https://github.com/xuebinqin/DIS
|
3 |
+
"""
|
4 |
+
|
5 |
+
import os
|
6 |
+
|
7 |
+
import gdown
|
8 |
+
import gradio as gr
|
9 |
+
|
10 |
+
from DIS.IsNetPipeLine import IsNetPipeLine
|
11 |
+
|
12 |
+
save_model_path = "DIS/save_models"
|
13 |
+
model_name = os.path.join(save_model_path, "isnet.pth")
|
14 |
+
# Download official weights
|
15 |
+
if not os.path.exists(model_name):
|
16 |
+
if not os.path.exists(save_model_path):
|
17 |
+
os.mkdir(save_model_path)
|
18 |
+
MODEL_PATH_URL = "https://huggingface.co/Superlang/ImageProcess/resolve/main/isnet.pth"
|
19 |
+
gdown.download(MODEL_PATH_URL, model_name, use_cookies=False)
|
20 |
+
|
21 |
+
pipe = IsNetPipeLine(model_path=model_name)
|
22 |
+
|
23 |
+
|
24 |
+
def inference(image):
|
25 |
+
return pipe(image)
|
26 |
+
|
27 |
+
|
28 |
+
title = "remove background"
|
29 |
+
interface = gr.Interface(
|
30 |
+
fn=inference,
|
31 |
+
inputs=gr.Image(type='pil'),
|
32 |
+
outputs=["image", "image"],
|
33 |
+
title=title,
|
34 |
+
allow_flagging='never',
|
35 |
+
cache_examples=True,
|
36 |
+
).queue(concurrency_count=1, api_open=True).launch(show_api=True, show_error=True)
|
requirements.txt
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch~=2.0.0
|
2 |
+
numpy~=1.23.3
|
3 |
+
scikit-image~=0.19.2
|
4 |
+
tqdm~=4.65.0
|
5 |
+
torchvision~=0.15.1
|
6 |
+
Pillow~=9.4.0
|
7 |
+
gdown~=4.7.1
|
8 |
+
gradio~=3.23.0
|