Spaces:
Sleeping
Sleeping
初次提交
Browse files- .gitignore +2 -0
- app.py +62 -0
- gitattributes.txt +27 -0
- packages.txt +1 -0
- requirements.txt +5 -0
- test-img.jpg +0 -0
- upcunet_v3.py +714 -0
- weights_v3/up2x-latest-conservative.pth +3 -0
- weights_v3/up2x-latest-denoise1x.pth +3 -0
- weights_v3/up2x-latest-denoise2x.pth +3 -0
- weights_v3/up2x-latest-denoise3x.pth +3 -0
- weights_v3/up2x-latest-no-denoise.pth +3 -0
- weights_v3/up3x-latest-conservative.pth +3 -0
- weights_v3/up3x-latest-denoise3x.pth +3 -0
- weights_v3/up3x-latest-no-denoise.pth +3 -0
- weights_v3/up4x-latest-conservative.pth +3 -0
- weights_v3/up4x-latest-denoise3x.pth +3 -0
- weights_v3/up4x-latest-no-denoise.pth +3 -0
.gitignore
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
|
2 |
+
__pycache__/
|
app.py
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from upcunet_v3 import RealWaifuUpScaler
|
2 |
+
import gradio as gr
|
3 |
+
import time
|
4 |
+
import logging
|
5 |
+
import os
|
6 |
+
from PIL import ImageOps
|
7 |
+
import numpy as np
|
8 |
+
import math
|
9 |
+
|
10 |
+
|
11 |
+
def greet(input_img, input_model_name, input_tile_mode):
|
12 |
+
# if input_img.size[0] * input_img.size[1] > 256 * 256:
|
13 |
+
# y = int(math.sqrt(256*256/input_img.size[0]*input_img.size[1]))
|
14 |
+
# x = int(input_img.size[0]/input_img.size[1]*y)
|
15 |
+
# input_img = ImageOps.fit(input_img, (x, y))
|
16 |
+
input_img = np.array(input_img)
|
17 |
+
if input_model_name not in model_cache:
|
18 |
+
t1 = time.time()
|
19 |
+
upscaler = RealWaifuUpScaler(input_model_name[2], ModelPath + input_model_name, half=False, device="cpu")
|
20 |
+
t2 = time.time()
|
21 |
+
logger.info(f'load model time, {t2 - t1}')
|
22 |
+
model_cache[input_model_name] = upscaler
|
23 |
+
else:
|
24 |
+
upscaler = model_cache[input_model_name]
|
25 |
+
logger.info(f'load model from cache')
|
26 |
+
|
27 |
+
start = time.time()
|
28 |
+
result = upscaler(input_img, tile_mode=input_tile_mode)
|
29 |
+
end = time.time()
|
30 |
+
logger.info(f'input_model_name, {input_model_name}')
|
31 |
+
logger.info(f'input_tile_mode, {input_tile_mode}')
|
32 |
+
logger.info(f'input shape, {input_img.shape}')
|
33 |
+
logger.info(f'output shape, {result.shape}')
|
34 |
+
logger.info(f'speed time, {end - start}')
|
35 |
+
return result
|
36 |
+
|
37 |
+
|
38 |
+
if __name__ == '__main__':
|
39 |
+
logging.basicConfig(level=logging.INFO, format="[%(asctime)s] [%(process)d] [%(levelname)s] %(message)s")
|
40 |
+
logger = logging.getLogger()
|
41 |
+
|
42 |
+
ModelPath = "weights_v3/"
|
43 |
+
model_cache = {}
|
44 |
+
|
45 |
+
input_model_name = gr.inputs.Dropdown(os.listdir(ModelPath), default="up2x-latest-denoise2x.pth", label='选择model')
|
46 |
+
input_tile_mode = gr.inputs.Dropdown([0, 1, 2, 3, 4], default=2, label='选择tile_mode')
|
47 |
+
input_img = gr.inputs.Image(label='image', type='pil')
|
48 |
+
|
49 |
+
inputs = [input_img, input_model_name, input_tile_mode]
|
50 |
+
outputs = "image"
|
51 |
+
iface = gr.Interface(fn=greet,
|
52 |
+
inputs=inputs,
|
53 |
+
outputs=outputs,
|
54 |
+
allow_screenshot=False,
|
55 |
+
allow_flagging='never',
|
56 |
+
examples=[['test-img.jpg', "up2x-latest-denoise2x.pth", 2]],
|
57 |
+
article='[https://github.com/bilibili/ailab/tree/main/Real-CUGAN](https://github.com/bilibili/ailab/tree/main/Real-CUGAN)<br>'
|
58 |
+
'感谢b站开源的项目,图片过大会导致内存不足,所有我将图片裁剪小,想体验大图片的效果请自行前往上面的链接。<br>'
|
59 |
+
'修改bbb'
|
60 |
+
'The large image will lead to memory limit exceeded. So I crop and resize image. '
|
61 |
+
'If you want to experience the large image, please go to the link above.')
|
62 |
+
iface.launch()
|
gitattributes.txt
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
3 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
4 |
+
*.bin.* filter=lfs diff=lfs merge=lfs -text
|
5 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
6 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
7 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
8 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
9 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
10 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
11 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
12 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
13 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
14 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
15 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
16 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
17 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
18 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
19 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
20 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
21 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
22 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
23 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
24 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
25 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
26 |
+
*.zstandard filter=lfs diff=lfs merge=lfs -text
|
27 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
packages.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
python3-opencv
|
requirements.txt
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch
|
2 |
+
opencv-python
|
3 |
+
numpy
|
4 |
+
gradio
|
5 |
+
jinja2
|
test-img.jpg
ADDED
upcunet_v3.py
ADDED
@@ -0,0 +1,714 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch import nn as nn
|
3 |
+
from torch.nn import functional as F
|
4 |
+
import os, sys
|
5 |
+
import numpy as np
|
6 |
+
|
7 |
+
root_path = os.path.abspath('.')
|
8 |
+
sys.path.append(root_path)
|
9 |
+
|
10 |
+
|
11 |
+
class SEBlock(nn.Module):
|
12 |
+
def __init__(self, in_channels, reduction=8, bias=False):
|
13 |
+
super(SEBlock, self).__init__()
|
14 |
+
self.conv1 = nn.Conv2d(in_channels, in_channels // reduction, 1, 1, 0, bias=bias)
|
15 |
+
self.conv2 = nn.Conv2d(in_channels // reduction, in_channels, 1, 1, 0, bias=bias)
|
16 |
+
|
17 |
+
def forward(self, x):
|
18 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
19 |
+
x0 = torch.mean(x.float(), dim=(2, 3), keepdim=True).half()
|
20 |
+
else:
|
21 |
+
x0 = torch.mean(x, dim=(2, 3), keepdim=True)
|
22 |
+
x0 = self.conv1(x0)
|
23 |
+
x0 = F.relu(x0, inplace=True)
|
24 |
+
x0 = self.conv2(x0)
|
25 |
+
x0 = torch.sigmoid(x0)
|
26 |
+
x = torch.mul(x, x0)
|
27 |
+
return x
|
28 |
+
|
29 |
+
def forward_mean(self, x, x0):
|
30 |
+
x0 = self.conv1(x0)
|
31 |
+
x0 = F.relu(x0, inplace=True)
|
32 |
+
x0 = self.conv2(x0)
|
33 |
+
x0 = torch.sigmoid(x0)
|
34 |
+
x = torch.mul(x, x0)
|
35 |
+
return x
|
36 |
+
|
37 |
+
|
38 |
+
class UNetConv(nn.Module):
|
39 |
+
def __init__(self, in_channels, mid_channels, out_channels, se):
|
40 |
+
super(UNetConv, self).__init__()
|
41 |
+
self.conv = nn.Sequential(
|
42 |
+
nn.Conv2d(in_channels, mid_channels, 3, 1, 0),
|
43 |
+
nn.LeakyReLU(0.1, inplace=True),
|
44 |
+
nn.Conv2d(mid_channels, out_channels, 3, 1, 0),
|
45 |
+
nn.LeakyReLU(0.1, inplace=True),
|
46 |
+
)
|
47 |
+
if se:
|
48 |
+
self.seblock = SEBlock(out_channels, reduction=8, bias=True)
|
49 |
+
else:
|
50 |
+
self.seblock = None
|
51 |
+
|
52 |
+
def forward(self, x):
|
53 |
+
z = self.conv(x)
|
54 |
+
if self.seblock is not None:
|
55 |
+
z = self.seblock(z)
|
56 |
+
return z
|
57 |
+
|
58 |
+
|
59 |
+
class UNet1(nn.Module):
|
60 |
+
def __init__(self, in_channels, out_channels, deconv):
|
61 |
+
super(UNet1, self).__init__()
|
62 |
+
self.conv1 = UNetConv(in_channels, 32, 64, se=False)
|
63 |
+
self.conv1_down = nn.Conv2d(64, 64, 2, 2, 0)
|
64 |
+
self.conv2 = UNetConv(64, 128, 64, se=True)
|
65 |
+
self.conv2_up = nn.ConvTranspose2d(64, 64, 2, 2, 0)
|
66 |
+
self.conv3 = nn.Conv2d(64, 64, 3, 1, 0)
|
67 |
+
|
68 |
+
if deconv:
|
69 |
+
self.conv_bottom = nn.ConvTranspose2d(64, out_channels, 4, 2, 3)
|
70 |
+
else:
|
71 |
+
self.conv_bottom = nn.Conv2d(64, out_channels, 3, 1, 0)
|
72 |
+
|
73 |
+
for m in self.modules():
|
74 |
+
if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
|
75 |
+
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
|
76 |
+
elif isinstance(m, nn.Linear):
|
77 |
+
nn.init.normal_(m.weight, 0, 0.01)
|
78 |
+
if m.bias is not None:
|
79 |
+
nn.init.constant_(m.bias, 0)
|
80 |
+
|
81 |
+
def forward(self, x):
|
82 |
+
x1 = self.conv1(x)
|
83 |
+
x2 = self.conv1_down(x1)
|
84 |
+
x2 = F.leaky_relu(x2, 0.1, inplace=True)
|
85 |
+
x2 = self.conv2(x2)
|
86 |
+
x2 = self.conv2_up(x2)
|
87 |
+
x2 = F.leaky_relu(x2, 0.1, inplace=True)
|
88 |
+
|
89 |
+
x1 = F.pad(x1, (-4, -4, -4, -4))
|
90 |
+
x3 = self.conv3(x1 + x2)
|
91 |
+
x3 = F.leaky_relu(x3, 0.1, inplace=True)
|
92 |
+
z = self.conv_bottom(x3)
|
93 |
+
return z
|
94 |
+
|
95 |
+
def forward_a(self, x):
|
96 |
+
x1 = self.conv1(x)
|
97 |
+
x2 = self.conv1_down(x1)
|
98 |
+
x2 = F.leaky_relu(x2, 0.1, inplace=True)
|
99 |
+
x2 = self.conv2.conv(x2)
|
100 |
+
return x1, x2
|
101 |
+
|
102 |
+
def forward_b(self, x1, x2):
|
103 |
+
x2 = self.conv2_up(x2)
|
104 |
+
x2 = F.leaky_relu(x2, 0.1, inplace=True)
|
105 |
+
|
106 |
+
x1 = F.pad(x1, (-4, -4, -4, -4))
|
107 |
+
x3 = self.conv3(x1 + x2)
|
108 |
+
x3 = F.leaky_relu(x3, 0.1, inplace=True)
|
109 |
+
z = self.conv_bottom(x3)
|
110 |
+
return z
|
111 |
+
|
112 |
+
|
113 |
+
class UNet1x3(nn.Module):
|
114 |
+
def __init__(self, in_channels, out_channels, deconv):
|
115 |
+
super(UNet1x3, self).__init__()
|
116 |
+
self.conv1 = UNetConv(in_channels, 32, 64, se=False)
|
117 |
+
self.conv1_down = nn.Conv2d(64, 64, 2, 2, 0)
|
118 |
+
self.conv2 = UNetConv(64, 128, 64, se=True)
|
119 |
+
self.conv2_up = nn.ConvTranspose2d(64, 64, 2, 2, 0)
|
120 |
+
self.conv3 = nn.Conv2d(64, 64, 3, 1, 0)
|
121 |
+
|
122 |
+
if deconv:
|
123 |
+
self.conv_bottom = nn.ConvTranspose2d(64, out_channels, 5, 3, 2)
|
124 |
+
else:
|
125 |
+
self.conv_bottom = nn.Conv2d(64, out_channels, 3, 1, 0)
|
126 |
+
|
127 |
+
for m in self.modules():
|
128 |
+
if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
|
129 |
+
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
|
130 |
+
elif isinstance(m, nn.Linear):
|
131 |
+
nn.init.normal_(m.weight, 0, 0.01)
|
132 |
+
if m.bias is not None:
|
133 |
+
nn.init.constant_(m.bias, 0)
|
134 |
+
|
135 |
+
def forward(self, x):
|
136 |
+
x1 = self.conv1(x)
|
137 |
+
x2 = self.conv1_down(x1)
|
138 |
+
x2 = F.leaky_relu(x2, 0.1, inplace=True)
|
139 |
+
x2 = self.conv2(x2)
|
140 |
+
x2 = self.conv2_up(x2)
|
141 |
+
x2 = F.leaky_relu(x2, 0.1, inplace=True)
|
142 |
+
|
143 |
+
x1 = F.pad(x1, (-4, -4, -4, -4))
|
144 |
+
x3 = self.conv3(x1 + x2)
|
145 |
+
x3 = F.leaky_relu(x3, 0.1, inplace=True)
|
146 |
+
z = self.conv_bottom(x3)
|
147 |
+
return z
|
148 |
+
|
149 |
+
def forward_a(self, x):
|
150 |
+
x1 = self.conv1(x)
|
151 |
+
x2 = self.conv1_down(x1)
|
152 |
+
x2 = F.leaky_relu(x2, 0.1, inplace=True)
|
153 |
+
x2 = self.conv2.conv(x2)
|
154 |
+
return x1, x2
|
155 |
+
|
156 |
+
def forward_b(self, x1, x2):
|
157 |
+
x2 = self.conv2_up(x2)
|
158 |
+
x2 = F.leaky_relu(x2, 0.1, inplace=True)
|
159 |
+
|
160 |
+
x1 = F.pad(x1, (-4, -4, -4, -4))
|
161 |
+
x3 = self.conv3(x1 + x2)
|
162 |
+
x3 = F.leaky_relu(x3, 0.1, inplace=True)
|
163 |
+
z = self.conv_bottom(x3)
|
164 |
+
return z
|
165 |
+
|
166 |
+
|
167 |
+
class UNet2(nn.Module):
|
168 |
+
def __init__(self, in_channels, out_channels, deconv):
|
169 |
+
super(UNet2, self).__init__()
|
170 |
+
|
171 |
+
self.conv1 = UNetConv(in_channels, 32, 64, se=False)
|
172 |
+
self.conv1_down = nn.Conv2d(64, 64, 2, 2, 0)
|
173 |
+
self.conv2 = UNetConv(64, 64, 128, se=True)
|
174 |
+
self.conv2_down = nn.Conv2d(128, 128, 2, 2, 0)
|
175 |
+
self.conv3 = UNetConv(128, 256, 128, se=True)
|
176 |
+
self.conv3_up = nn.ConvTranspose2d(128, 128, 2, 2, 0)
|
177 |
+
self.conv4 = UNetConv(128, 64, 64, se=True)
|
178 |
+
self.conv4_up = nn.ConvTranspose2d(64, 64, 2, 2, 0)
|
179 |
+
self.conv5 = nn.Conv2d(64, 64, 3, 1, 0)
|
180 |
+
|
181 |
+
if deconv:
|
182 |
+
self.conv_bottom = nn.ConvTranspose2d(64, out_channels, 4, 2, 3)
|
183 |
+
else:
|
184 |
+
self.conv_bottom = nn.Conv2d(64, out_channels, 3, 1, 0)
|
185 |
+
|
186 |
+
for m in self.modules():
|
187 |
+
if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
|
188 |
+
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
|
189 |
+
elif isinstance(m, nn.Linear):
|
190 |
+
nn.init.normal_(m.weight, 0, 0.01)
|
191 |
+
if m.bias is not None:
|
192 |
+
nn.init.constant_(m.bias, 0)
|
193 |
+
|
194 |
+
def forward(self, x):
|
195 |
+
x1 = self.conv1(x)
|
196 |
+
x2 = self.conv1_down(x1)
|
197 |
+
x2 = F.leaky_relu(x2, 0.1, inplace=True)
|
198 |
+
x2 = self.conv2(x2)
|
199 |
+
|
200 |
+
x3 = self.conv2_down(x2)
|
201 |
+
x3 = F.leaky_relu(x3, 0.1, inplace=True)
|
202 |
+
x3 = self.conv3(x3)
|
203 |
+
x3 = self.conv3_up(x3)
|
204 |
+
x3 = F.leaky_relu(x3, 0.1, inplace=True)
|
205 |
+
|
206 |
+
x2 = F.pad(x2, (-4, -4, -4, -4))
|
207 |
+
x4 = self.conv4(x2 + x3)
|
208 |
+
x4 = self.conv4_up(x4)
|
209 |
+
x4 = F.leaky_relu(x4, 0.1, inplace=True)
|
210 |
+
|
211 |
+
x1 = F.pad(x1, (-16, -16, -16, -16))
|
212 |
+
x5 = self.conv5(x1 + x4)
|
213 |
+
x5 = F.leaky_relu(x5, 0.1, inplace=True)
|
214 |
+
|
215 |
+
z = self.conv_bottom(x5)
|
216 |
+
return z
|
217 |
+
|
218 |
+
def forward_a(self, x): # conv234结尾有se
|
219 |
+
x1 = self.conv1(x)
|
220 |
+
x2 = self.conv1_down(x1)
|
221 |
+
x2 = F.leaky_relu(x2, 0.1, inplace=True)
|
222 |
+
x2 = self.conv2.conv(x2)
|
223 |
+
return x1, x2
|
224 |
+
|
225 |
+
def forward_b(self, x2): # conv234结尾有se
|
226 |
+
x3 = self.conv2_down(x2)
|
227 |
+
x3 = F.leaky_relu(x3, 0.1, inplace=True)
|
228 |
+
x3 = self.conv3.conv(x3)
|
229 |
+
return x3
|
230 |
+
|
231 |
+
def forward_c(self, x2, x3): # conv234结尾有se
|
232 |
+
x3 = self.conv3_up(x3)
|
233 |
+
x3 = F.leaky_relu(x3, 0.1, inplace=True)
|
234 |
+
|
235 |
+
x2 = F.pad(x2, (-4, -4, -4, -4))
|
236 |
+
x4 = self.conv4.conv(x2 + x3)
|
237 |
+
return x4
|
238 |
+
|
239 |
+
def forward_d(self, x1, x4): # conv234结尾有se
|
240 |
+
x4 = self.conv4_up(x4)
|
241 |
+
x4 = F.leaky_relu(x4, 0.1, inplace=True)
|
242 |
+
|
243 |
+
x1 = F.pad(x1, (-16, -16, -16, -16))
|
244 |
+
x5 = self.conv5(x1 + x4)
|
245 |
+
x5 = F.leaky_relu(x5, 0.1, inplace=True)
|
246 |
+
|
247 |
+
z = self.conv_bottom(x5)
|
248 |
+
return z
|
249 |
+
|
250 |
+
|
251 |
+
class UpCunet2x(nn.Module): # 完美tile,全程无损
|
252 |
+
def __init__(self, in_channels=3, out_channels=3):
|
253 |
+
super(UpCunet2x, self).__init__()
|
254 |
+
self.unet1 = UNet1(in_channels, out_channels, deconv=True)
|
255 |
+
self.unet2 = UNet2(in_channels, out_channels, deconv=False)
|
256 |
+
|
257 |
+
def forward(self, x, tile_mode): # 1.7G
|
258 |
+
n, c, h0, w0 = x.shape
|
259 |
+
if (tile_mode == 0): # 不tile
|
260 |
+
ph = ((h0 - 1) // 2 + 1) * 2
|
261 |
+
pw = ((w0 - 1) // 2 + 1) * 2
|
262 |
+
x = F.pad(x, (18, 18 + pw - w0, 18, 18 + ph - h0), 'reflect') # 需要保证被2整除
|
263 |
+
x = self.unet1.forward(x)
|
264 |
+
x0 = self.unet2.forward(x)
|
265 |
+
x1 = F.pad(x, (-20, -20, -20, -20))
|
266 |
+
x = torch.add(x0, x1)
|
267 |
+
if (w0 != pw or h0 != ph): x = x[:, :, :h0 * 2, :w0 * 2]
|
268 |
+
return x
|
269 |
+
elif (tile_mode == 1): # 对长边减半
|
270 |
+
if (w0 >= h0):
|
271 |
+
crop_size_w = ((w0 - 1) // 4 * 4 + 4) // 2 # 减半后能被2整除,所以要先被4整除
|
272 |
+
crop_size_h = (h0 - 1) // 2 * 2 + 2 # 能被2整除
|
273 |
+
else:
|
274 |
+
crop_size_h = ((h0 - 1) // 4 * 4 + 4) // 2 # 减半后能被2整除,所以要先被4整除
|
275 |
+
crop_size_w = (w0 - 1) // 2 * 2 + 2 # 能被2整除
|
276 |
+
crop_size = (crop_size_h, crop_size_w) # 6.6G
|
277 |
+
elif (tile_mode == 2): # hw都减半
|
278 |
+
crop_size = (((h0 - 1) // 4 * 4 + 4) // 2, ((w0 - 1) // 4 * 4 + 4) // 2) # 5.6G
|
279 |
+
elif (tile_mode == 3): # hw都三分之一
|
280 |
+
crop_size = (((h0 - 1) // 6 * 6 + 6) // 3, ((w0 - 1) // 6 * 6 + 6) // 3) # 4.2G
|
281 |
+
elif (tile_mode == 4): # hw都四分���一
|
282 |
+
crop_size = (((h0 - 1) // 8 * 8 + 8) // 4, ((w0 - 1) // 8 * 8 + 8) // 4) # 3.7G
|
283 |
+
ph = ((h0 - 1) // crop_size[0] + 1) * crop_size[0]
|
284 |
+
pw = ((w0 - 1) // crop_size[1] + 1) * crop_size[1]
|
285 |
+
x = F.pad(x, (18, 18 + pw - w0, 18, 18 + ph - h0), 'reflect')
|
286 |
+
n, c, h, w = x.shape
|
287 |
+
se_mean0 = torch.zeros((n, 64, 1, 1)).to(x.device)
|
288 |
+
if ("Half" in x.type()):
|
289 |
+
se_mean0 = se_mean0.half()
|
290 |
+
n_patch = 0
|
291 |
+
tmp_dict = {}
|
292 |
+
opt_res_dict = {}
|
293 |
+
for i in range(0, h - 36, crop_size[0]):
|
294 |
+
tmp_dict[i] = {}
|
295 |
+
for j in range(0, w - 36, crop_size[1]):
|
296 |
+
x_crop = x[:, :, i:i + crop_size[0] + 36, j:j + crop_size[1] + 36]
|
297 |
+
n, c1, h1, w1 = x_crop.shape
|
298 |
+
tmp0, x_crop = self.unet1.forward_a(x_crop)
|
299 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
300 |
+
tmp_se_mean = torch.mean(x_crop.float(), dim=(2, 3), keepdim=True).half()
|
301 |
+
else:
|
302 |
+
tmp_se_mean = torch.mean(x_crop, dim=(2, 3), keepdim=True)
|
303 |
+
se_mean0 += tmp_se_mean
|
304 |
+
n_patch += 1
|
305 |
+
tmp_dict[i][j] = (tmp0, x_crop)
|
306 |
+
se_mean0 /= n_patch
|
307 |
+
se_mean1 = torch.zeros((n, 128, 1, 1)).to(x.device) # 64#128#128#64
|
308 |
+
if ("Half" in x.type()):
|
309 |
+
se_mean1 = se_mean1.half()
|
310 |
+
for i in range(0, h - 36, crop_size[0]):
|
311 |
+
for j in range(0, w - 36, crop_size[1]):
|
312 |
+
tmp0, x_crop = tmp_dict[i][j]
|
313 |
+
x_crop = self.unet1.conv2.seblock.forward_mean(x_crop, se_mean0)
|
314 |
+
opt_unet1 = self.unet1.forward_b(tmp0, x_crop)
|
315 |
+
tmp_x1, tmp_x2 = self.unet2.forward_a(opt_unet1)
|
316 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
317 |
+
tmp_se_mean = torch.mean(tmp_x2.float(), dim=(2, 3), keepdim=True).half()
|
318 |
+
else:
|
319 |
+
tmp_se_mean = torch.mean(tmp_x2, dim=(2, 3), keepdim=True)
|
320 |
+
se_mean1 += tmp_se_mean
|
321 |
+
tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2)
|
322 |
+
se_mean1 /= n_patch
|
323 |
+
se_mean0 = torch.zeros((n, 128, 1, 1)).to(x.device) # 64#128#128#64
|
324 |
+
if ("Half" in x.type()):
|
325 |
+
se_mean0 = se_mean0.half()
|
326 |
+
for i in range(0, h - 36, crop_size[0]):
|
327 |
+
for j in range(0, w - 36, crop_size[1]):
|
328 |
+
opt_unet1, tmp_x1, tmp_x2 = tmp_dict[i][j]
|
329 |
+
tmp_x2 = self.unet2.conv2.seblock.forward_mean(tmp_x2, se_mean1)
|
330 |
+
tmp_x3 = self.unet2.forward_b(tmp_x2)
|
331 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
332 |
+
tmp_se_mean = torch.mean(tmp_x3.float(), dim=(2, 3), keepdim=True).half()
|
333 |
+
else:
|
334 |
+
tmp_se_mean = torch.mean(tmp_x3, dim=(2, 3), keepdim=True)
|
335 |
+
se_mean0 += tmp_se_mean
|
336 |
+
tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2, tmp_x3)
|
337 |
+
se_mean0 /= n_patch
|
338 |
+
se_mean1 = torch.zeros((n, 64, 1, 1)).to(x.device) # 64#128#128#64
|
339 |
+
if ("Half" in x.type()):
|
340 |
+
se_mean1 = se_mean1.half()
|
341 |
+
for i in range(0, h - 36, crop_size[0]):
|
342 |
+
for j in range(0, w - 36, crop_size[1]):
|
343 |
+
opt_unet1, tmp_x1, tmp_x2, tmp_x3 = tmp_dict[i][j]
|
344 |
+
tmp_x3 = self.unet2.conv3.seblock.forward_mean(tmp_x3, se_mean0)
|
345 |
+
tmp_x4 = self.unet2.forward_c(tmp_x2, tmp_x3)
|
346 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
347 |
+
tmp_se_mean = torch.mean(tmp_x4.float(), dim=(2, 3), keepdim=True).half()
|
348 |
+
else:
|
349 |
+
tmp_se_mean = torch.mean(tmp_x4, dim=(2, 3), keepdim=True)
|
350 |
+
se_mean1 += tmp_se_mean
|
351 |
+
tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x4)
|
352 |
+
se_mean1 /= n_patch
|
353 |
+
for i in range(0, h - 36, crop_size[0]):
|
354 |
+
opt_res_dict[i] = {}
|
355 |
+
for j in range(0, w - 36, crop_size[1]):
|
356 |
+
opt_unet1, tmp_x1, tmp_x4 = tmp_dict[i][j]
|
357 |
+
tmp_x4 = self.unet2.conv4.seblock.forward_mean(tmp_x4, se_mean1)
|
358 |
+
x0 = self.unet2.forward_d(tmp_x1, tmp_x4)
|
359 |
+
x1 = F.pad(opt_unet1, (-20, -20, -20, -20))
|
360 |
+
x_crop = torch.add(x0, x1) # x0是unet2的最终输出
|
361 |
+
opt_res_dict[i][j] = x_crop
|
362 |
+
del tmp_dict
|
363 |
+
torch.cuda.empty_cache()
|
364 |
+
res = torch.zeros((n, c, h * 2 - 72, w * 2 - 72)).to(x.device)
|
365 |
+
if ("Half" in x.type()):
|
366 |
+
res = res.half()
|
367 |
+
for i in range(0, h - 36, crop_size[0]):
|
368 |
+
for j in range(0, w - 36, crop_size[1]):
|
369 |
+
res[:, :, i * 2:i * 2 + h1 * 2 - 72, j * 2:j * 2 + w1 * 2 - 72] = opt_res_dict[i][j]
|
370 |
+
del opt_res_dict
|
371 |
+
torch.cuda.empty_cache()
|
372 |
+
if (w0 != pw or h0 != ph): res = res[:, :, :h0 * 2, :w0 * 2]
|
373 |
+
return res #
|
374 |
+
|
375 |
+
|
376 |
+
class UpCunet3x(nn.Module): # 完美tile,全程无损
|
377 |
+
def __init__(self, in_channels=3, out_channels=3):
|
378 |
+
super(UpCunet3x, self).__init__()
|
379 |
+
self.unet1 = UNet1x3(in_channels, out_channels, deconv=True)
|
380 |
+
self.unet2 = UNet2(in_channels, out_channels, deconv=False)
|
381 |
+
|
382 |
+
def forward(self, x, tile_mode): # 1.7G
|
383 |
+
n, c, h0, w0 = x.shape
|
384 |
+
if (tile_mode == 0): # 不tile
|
385 |
+
ph = ((h0 - 1) // 4 + 1) * 4
|
386 |
+
pw = ((w0 - 1) // 4 + 1) * 4
|
387 |
+
x = F.pad(x, (14, 14 + pw - w0, 14, 14 + ph - h0), 'reflect') # 需要保证被2整除
|
388 |
+
x = self.unet1.forward(x)
|
389 |
+
x0 = self.unet2.forward(x)
|
390 |
+
x1 = F.pad(x, (-20, -20, -20, -20))
|
391 |
+
x = torch.add(x0, x1)
|
392 |
+
if (w0 != pw or h0 != ph): x = x[:, :, :h0 * 3, :w0 * 3]
|
393 |
+
return x
|
394 |
+
elif (tile_mode == 1): # 对长边减半
|
395 |
+
if (w0 >= h0):
|
396 |
+
crop_size_w = ((w0 - 1) // 8 * 8 + 8) // 2 # 减半后能被4整除,所以要先被8整除
|
397 |
+
crop_size_h = (h0 - 1) // 4 * 4 + 4 # 能被4整除
|
398 |
+
else:
|
399 |
+
crop_size_h = ((h0 - 1) // 8 * 8 + 8) // 2 # 减半后能被4整除,所以要先被8整除
|
400 |
+
crop_size_w = (w0 - 1) // 4 * 4 + 4 # 能被4整除
|
401 |
+
crop_size = (crop_size_h, crop_size_w) # 6.6G
|
402 |
+
elif (tile_mode == 2): # hw都减半
|
403 |
+
crop_size = (((h0 - 1) // 8 * 8 + 8) // 2, ((w0 - 1) // 8 * 8 + 8) // 2) # 5.6G
|
404 |
+
elif (tile_mode == 3): # hw都三分之一
|
405 |
+
crop_size = (((h0 - 1) // 12 * 12 + 12) // 3, ((w0 - 1) // 12 * 12 + 12) // 3) # 4.2G
|
406 |
+
elif (tile_mode == 4): # hw都四分之一
|
407 |
+
crop_size = (((h0 - 1) // 16 * 16 + 16) // 4, ((w0 - 1) // 16 * 16 + 16) // 4) # 3.7G
|
408 |
+
ph = ((h0 - 1) // crop_size[0] + 1) * crop_size[0]
|
409 |
+
pw = ((w0 - 1) // crop_size[1] + 1) * crop_size[1]
|
410 |
+
x = F.pad(x, (14, 14 + pw - w0, 14, 14 + ph - h0), 'reflect')
|
411 |
+
n, c, h, w = x.shape
|
412 |
+
se_mean0 = torch.zeros((n, 64, 1, 1)).to(x.device)
|
413 |
+
if ("Half" in x.type()):
|
414 |
+
se_mean0 = se_mean0.half()
|
415 |
+
n_patch = 0
|
416 |
+
tmp_dict = {}
|
417 |
+
opt_res_dict = {}
|
418 |
+
for i in range(0, h - 28, crop_size[0]):
|
419 |
+
tmp_dict[i] = {}
|
420 |
+
for j in range(0, w - 28, crop_size[1]):
|
421 |
+
x_crop = x[:, :, i:i + crop_size[0] + 28, j:j + crop_size[1] + 28]
|
422 |
+
n, c1, h1, w1 = x_crop.shape
|
423 |
+
tmp0, x_crop = self.unet1.forward_a(x_crop)
|
424 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
425 |
+
tmp_se_mean = torch.mean(x_crop.float(), dim=(2, 3), keepdim=True).half()
|
426 |
+
else:
|
427 |
+
tmp_se_mean = torch.mean(x_crop, dim=(2, 3), keepdim=True)
|
428 |
+
se_mean0 += tmp_se_mean
|
429 |
+
n_patch += 1
|
430 |
+
tmp_dict[i][j] = (tmp0, x_crop)
|
431 |
+
se_mean0 /= n_patch
|
432 |
+
se_mean1 = torch.zeros((n, 128, 1, 1)).to(x.device) # 64#128#128#64
|
433 |
+
if ("Half" in x.type()):
|
434 |
+
se_mean1 = se_mean1.half()
|
435 |
+
for i in range(0, h - 28, crop_size[0]):
|
436 |
+
for j in range(0, w - 28, crop_size[1]):
|
437 |
+
tmp0, x_crop = tmp_dict[i][j]
|
438 |
+
x_crop = self.unet1.conv2.seblock.forward_mean(x_crop, se_mean0)
|
439 |
+
opt_unet1 = self.unet1.forward_b(tmp0, x_crop)
|
440 |
+
tmp_x1, tmp_x2 = self.unet2.forward_a(opt_unet1)
|
441 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
442 |
+
tmp_se_mean = torch.mean(tmp_x2.float(), dim=(2, 3), keepdim=True).half()
|
443 |
+
else:
|
444 |
+
tmp_se_mean = torch.mean(tmp_x2, dim=(2, 3), keepdim=True)
|
445 |
+
se_mean1 += tmp_se_mean
|
446 |
+
tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2)
|
447 |
+
se_mean1 /= n_patch
|
448 |
+
se_mean0 = torch.zeros((n, 128, 1, 1)).to(x.device) # 64#128#128#64
|
449 |
+
if ("Half" in x.type()):
|
450 |
+
se_mean0 = se_mean0.half()
|
451 |
+
for i in range(0, h - 28, crop_size[0]):
|
452 |
+
for j in range(0, w - 28, crop_size[1]):
|
453 |
+
opt_unet1, tmp_x1, tmp_x2 = tmp_dict[i][j]
|
454 |
+
tmp_x2 = self.unet2.conv2.seblock.forward_mean(tmp_x2, se_mean1)
|
455 |
+
tmp_x3 = self.unet2.forward_b(tmp_x2)
|
456 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
457 |
+
tmp_se_mean = torch.mean(tmp_x3.float(), dim=(2, 3), keepdim=True).half()
|
458 |
+
else:
|
459 |
+
tmp_se_mean = torch.mean(tmp_x3, dim=(2, 3), keepdim=True)
|
460 |
+
se_mean0 += tmp_se_mean
|
461 |
+
tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2, tmp_x3)
|
462 |
+
se_mean0 /= n_patch
|
463 |
+
se_mean1 = torch.zeros((n, 64, 1, 1)).to(x.device) # 64#128#128#64
|
464 |
+
if ("Half" in x.type()):
|
465 |
+
se_mean1 = se_mean1.half()
|
466 |
+
for i in range(0, h - 28, crop_size[0]):
|
467 |
+
for j in range(0, w - 28, crop_size[1]):
|
468 |
+
opt_unet1, tmp_x1, tmp_x2, tmp_x3 = tmp_dict[i][j]
|
469 |
+
tmp_x3 = self.unet2.conv3.seblock.forward_mean(tmp_x3, se_mean0)
|
470 |
+
tmp_x4 = self.unet2.forward_c(tmp_x2, tmp_x3)
|
471 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
472 |
+
tmp_se_mean = torch.mean(tmp_x4.float(), dim=(2, 3), keepdim=True).half()
|
473 |
+
else:
|
474 |
+
tmp_se_mean = torch.mean(tmp_x4, dim=(2, 3), keepdim=True)
|
475 |
+
se_mean1 += tmp_se_mean
|
476 |
+
tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x4)
|
477 |
+
se_mean1 /= n_patch
|
478 |
+
for i in range(0, h - 28, crop_size[0]):
|
479 |
+
opt_res_dict[i] = {}
|
480 |
+
for j in range(0, w - 28, crop_size[1]):
|
481 |
+
opt_unet1, tmp_x1, tmp_x4 = tmp_dict[i][j]
|
482 |
+
tmp_x4 = self.unet2.conv4.seblock.forward_mean(tmp_x4, se_mean1)
|
483 |
+
x0 = self.unet2.forward_d(tmp_x1, tmp_x4)
|
484 |
+
x1 = F.pad(opt_unet1, (-20, -20, -20, -20))
|
485 |
+
x_crop = torch.add(x0, x1) # x0是unet2的最终输出
|
486 |
+
opt_res_dict[i][j] = x_crop #
|
487 |
+
del tmp_dict
|
488 |
+
torch.cuda.empty_cache()
|
489 |
+
res = torch.zeros((n, c, h * 3 - 84, w * 3 - 84)).to(x.device)
|
490 |
+
if ("Half" in x.type()):
|
491 |
+
res = res.half()
|
492 |
+
for i in range(0, h - 28, crop_size[0]):
|
493 |
+
for j in range(0, w - 28, crop_size[1]):
|
494 |
+
res[:, :, i * 3:i * 3 + h1 * 3 - 84, j * 3:j * 3 + w1 * 3 - 84] = opt_res_dict[i][j]
|
495 |
+
del opt_res_dict
|
496 |
+
torch.cuda.empty_cache()
|
497 |
+
if (w0 != pw or h0 != ph): res = res[:, :, :h0 * 3, :w0 * 3]
|
498 |
+
return res
|
499 |
+
|
500 |
+
|
501 |
+
class UpCunet4x(nn.Module): # 完美tile,全程无损
|
502 |
+
def __init__(self, in_channels=3, out_channels=3):
|
503 |
+
super(UpCunet4x, self).__init__()
|
504 |
+
self.unet1 = UNet1(in_channels, 64, deconv=True)
|
505 |
+
self.unet2 = UNet2(64, 64, deconv=False)
|
506 |
+
self.ps = nn.PixelShuffle(2)
|
507 |
+
self.conv_final = nn.Conv2d(64, 12, 3, 1, padding=0, bias=True)
|
508 |
+
|
509 |
+
def forward(self, x, tile_mode):
|
510 |
+
n, c, h0, w0 = x.shape
|
511 |
+
x00 = x
|
512 |
+
if (tile_mode == 0): # 不tile
|
513 |
+
ph = ((h0 - 1) // 2 + 1) * 2
|
514 |
+
pw = ((w0 - 1) // 2 + 1) * 2
|
515 |
+
x = F.pad(x, (19, 19 + pw - w0, 19, 19 + ph - h0), 'reflect') # 需要保证被2整除
|
516 |
+
x = self.unet1.forward(x)
|
517 |
+
x0 = self.unet2.forward(x)
|
518 |
+
x1 = F.pad(x, (-20, -20, -20, -20))
|
519 |
+
x = torch.add(x0, x1)
|
520 |
+
x = self.conv_final(x)
|
521 |
+
x = F.pad(x, (-1, -1, -1, -1))
|
522 |
+
x = self.ps(x)
|
523 |
+
if (w0 != pw or h0 != ph): x = x[:, :, :h0 * 4, :w0 * 4]
|
524 |
+
x += F.interpolate(x00, scale_factor=4, mode='nearest')
|
525 |
+
return x
|
526 |
+
elif (tile_mode == 1): # 对长边减半
|
527 |
+
if (w0 >= h0):
|
528 |
+
crop_size_w = ((w0 - 1) // 4 * 4 + 4) // 2 # 减半后能被2整除,所以要先被4整除
|
529 |
+
crop_size_h = (h0 - 1) // 2 * 2 + 2 # 能被2整除
|
530 |
+
else:
|
531 |
+
crop_size_h = ((h0 - 1) // 4 * 4 + 4) // 2 # 减半后能被2整除,所以要先被4整除
|
532 |
+
crop_size_w = (w0 - 1) // 2 * 2 + 2 # 能被2整除
|
533 |
+
crop_size = (crop_size_h, crop_size_w) # 6.6G
|
534 |
+
elif (tile_mode == 2): # hw都减半
|
535 |
+
crop_size = (((h0 - 1) // 4 * 4 + 4) // 2, ((w0 - 1) // 4 * 4 + 4) // 2) # 5.6G
|
536 |
+
elif (tile_mode == 3): # hw都三分之一
|
537 |
+
crop_size = (((h0 - 1) // 6 * 6 + 6) // 3, ((w0 - 1) // 6 * 6 + 6) // 3) # 4.1G
|
538 |
+
elif (tile_mode == 4): # hw都四分之一
|
539 |
+
crop_size = (((h0 - 1) // 8 * 8 + 8) // 4, ((w0 - 1) // 8 * 8 + 8) // 4) # 3.7G
|
540 |
+
ph = ((h0 - 1) // crop_size[0] + 1) * crop_size[0]
|
541 |
+
pw = ((w0 - 1) // crop_size[1] + 1) * crop_size[1]
|
542 |
+
x = F.pad(x, (19, 19 + pw - w0, 19, 19 + ph - h0), 'reflect')
|
543 |
+
n, c, h, w = x.shape
|
544 |
+
se_mean0 = torch.zeros((n, 64, 1, 1)).to(x.device)
|
545 |
+
if ("Half" in x.type()):
|
546 |
+
se_mean0 = se_mean0.half()
|
547 |
+
n_patch = 0
|
548 |
+
tmp_dict = {}
|
549 |
+
opt_res_dict = {}
|
550 |
+
for i in range(0, h - 38, crop_size[0]):
|
551 |
+
tmp_dict[i] = {}
|
552 |
+
for j in range(0, w - 38, crop_size[1]):
|
553 |
+
x_crop = x[:, :, i:i + crop_size[0] + 38, j:j + crop_size[1] + 38]
|
554 |
+
n, c1, h1, w1 = x_crop.shape
|
555 |
+
tmp0, x_crop = self.unet1.forward_a(x_crop)
|
556 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
557 |
+
tmp_se_mean = torch.mean(x_crop.float(), dim=(2, 3), keepdim=True).half()
|
558 |
+
else:
|
559 |
+
tmp_se_mean = torch.mean(x_crop, dim=(2, 3), keepdim=True)
|
560 |
+
se_mean0 += tmp_se_mean
|
561 |
+
n_patch += 1
|
562 |
+
tmp_dict[i][j] = (tmp0, x_crop)
|
563 |
+
se_mean0 /= n_patch
|
564 |
+
se_mean1 = torch.zeros((n, 128, 1, 1)).to(x.device) # 64#128#128#64
|
565 |
+
if ("Half" in x.type()):
|
566 |
+
se_mean1 = se_mean1.half()
|
567 |
+
for i in range(0, h - 38, crop_size[0]):
|
568 |
+
for j in range(0, w - 38, crop_size[1]):
|
569 |
+
tmp0, x_crop = tmp_dict[i][j]
|
570 |
+
x_crop = self.unet1.conv2.seblock.forward_mean(x_crop, se_mean0)
|
571 |
+
opt_unet1 = self.unet1.forward_b(tmp0, x_crop)
|
572 |
+
tmp_x1, tmp_x2 = self.unet2.forward_a(opt_unet1)
|
573 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
574 |
+
tmp_se_mean = torch.mean(tmp_x2.float(), dim=(2, 3), keepdim=True).half()
|
575 |
+
else:
|
576 |
+
tmp_se_mean = torch.mean(tmp_x2, dim=(2, 3), keepdim=True)
|
577 |
+
se_mean1 += tmp_se_mean
|
578 |
+
tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2)
|
579 |
+
se_mean1 /= n_patch
|
580 |
+
se_mean0 = torch.zeros((n, 128, 1, 1)).to(x.device) # 64#128#128#64
|
581 |
+
if ("Half" in x.type()):
|
582 |
+
se_mean0 = se_mean0.half()
|
583 |
+
for i in range(0, h - 38, crop_size[0]):
|
584 |
+
for j in range(0, w - 38, crop_size[1]):
|
585 |
+
opt_unet1, tmp_x1, tmp_x2 = tmp_dict[i][j]
|
586 |
+
tmp_x2 = self.unet2.conv2.seblock.forward_mean(tmp_x2, se_mean1)
|
587 |
+
tmp_x3 = self.unet2.forward_b(tmp_x2)
|
588 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
589 |
+
tmp_se_mean = torch.mean(tmp_x3.float(), dim=(2, 3), keepdim=True).half()
|
590 |
+
else:
|
591 |
+
tmp_se_mean = torch.mean(tmp_x3, dim=(2, 3), keepdim=True)
|
592 |
+
se_mean0 += tmp_se_mean
|
593 |
+
tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2, tmp_x3)
|
594 |
+
se_mean0 /= n_patch
|
595 |
+
se_mean1 = torch.zeros((n, 64, 1, 1)).to(x.device) # 64#128#128#64
|
596 |
+
if ("Half" in x.type()):
|
597 |
+
se_mean1 = se_mean1.half()
|
598 |
+
for i in range(0, h - 38, crop_size[0]):
|
599 |
+
for j in range(0, w - 38, crop_size[1]):
|
600 |
+
opt_unet1, tmp_x1, tmp_x2, tmp_x3 = tmp_dict[i][j]
|
601 |
+
tmp_x3 = self.unet2.conv3.seblock.forward_mean(tmp_x3, se_mean0)
|
602 |
+
tmp_x4 = self.unet2.forward_c(tmp_x2, tmp_x3)
|
603 |
+
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
|
604 |
+
tmp_se_mean = torch.mean(tmp_x4.float(), dim=(2, 3), keepdim=True).half()
|
605 |
+
else:
|
606 |
+
tmp_se_mean = torch.mean(tmp_x4, dim=(2, 3), keepdim=True)
|
607 |
+
se_mean1 += tmp_se_mean
|
608 |
+
tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x4)
|
609 |
+
se_mean1 /= n_patch
|
610 |
+
for i in range(0, h - 38, crop_size[0]):
|
611 |
+
opt_res_dict[i] = {}
|
612 |
+
for j in range(0, w - 38, crop_size[1]):
|
613 |
+
opt_unet1, tmp_x1, tmp_x4 = tmp_dict[i][j]
|
614 |
+
tmp_x4 = self.unet2.conv4.seblock.forward_mean(tmp_x4, se_mean1)
|
615 |
+
x0 = self.unet2.forward_d(tmp_x1, tmp_x4)
|
616 |
+
x1 = F.pad(opt_unet1, (-20, -20, -20, -20))
|
617 |
+
x_crop = torch.add(x0, x1) # x0是unet2的最终输出
|
618 |
+
x_crop = self.conv_final(x_crop)
|
619 |
+
x_crop = F.pad(x_crop, (-1, -1, -1, -1))
|
620 |
+
x_crop = self.ps(x_crop)
|
621 |
+
opt_res_dict[i][j] = x_crop
|
622 |
+
del tmp_dict
|
623 |
+
torch.cuda.empty_cache()
|
624 |
+
res = torch.zeros((n, c, h * 4 - 152, w * 4 - 152)).to(x.device)
|
625 |
+
if ("Half" in x.type()):
|
626 |
+
res = res.half()
|
627 |
+
for i in range(0, h - 38, crop_size[0]):
|
628 |
+
for j in range(0, w - 38, crop_size[1]):
|
629 |
+
# print(opt_res_dict[i][j].shape,res[:, :, i * 4:i * 4 + h1 * 4 - 144, j * 4:j * 4 + w1 * 4 - 144].shape)
|
630 |
+
res[:, :, i * 4:i * 4 + h1 * 4 - 152, j * 4:j * 4 + w1 * 4 - 152] = opt_res_dict[i][j]
|
631 |
+
del opt_res_dict
|
632 |
+
torch.cuda.empty_cache()
|
633 |
+
if (w0 != pw or h0 != ph): res = res[:, :, :h0 * 4, :w0 * 4]
|
634 |
+
res += F.interpolate(x00, scale_factor=4, mode='nearest')
|
635 |
+
return res #
|
636 |
+
|
637 |
+
|
638 |
+
class RealWaifuUpScaler(object):
|
639 |
+
def __init__(self, scale, weight_path, half, device):
|
640 |
+
weight = torch.load(weight_path, map_location="cpu")
|
641 |
+
self.model = eval("UpCunet%sx" % scale)()
|
642 |
+
if (half == True):
|
643 |
+
self.model = self.model.half().to(device)
|
644 |
+
else:
|
645 |
+
self.model = self.model.to(device)
|
646 |
+
self.model.load_state_dict(weight, strict=True)
|
647 |
+
self.model.eval()
|
648 |
+
self.half = half
|
649 |
+
self.device = device
|
650 |
+
|
651 |
+
def np2tensor(self, np_frame):
|
652 |
+
if (self.half == False):
|
653 |
+
return torch.from_numpy(np.transpose(np_frame, (2, 0, 1))).unsqueeze(0).to(self.device).float() / 255
|
654 |
+
else:
|
655 |
+
return torch.from_numpy(np.transpose(np_frame, (2, 0, 1))).unsqueeze(0).to(self.device).half() / 255
|
656 |
+
|
657 |
+
def tensor2np(self, tensor):
|
658 |
+
if (self.half == False):
|
659 |
+
return (
|
660 |
+
np.transpose((tensor.data.squeeze() * 255.0).round().clamp_(0, 255).byte().cpu().numpy(), (1, 2, 0)))
|
661 |
+
else:
|
662 |
+
return (np.transpose((tensor.data.squeeze().float() * 255.0).round().clamp_(0, 255).byte().cpu().numpy(),
|
663 |
+
(1, 2, 0)))
|
664 |
+
|
665 |
+
def __call__(self, frame, tile_mode):
|
666 |
+
with torch.no_grad():
|
667 |
+
tensor = self.np2tensor(frame)
|
668 |
+
result = self.tensor2np(self.model(tensor, tile_mode))
|
669 |
+
return result
|
670 |
+
|
671 |
+
|
672 |
+
if __name__ == "__main__":
|
673 |
+
###########inference_img
|
674 |
+
import time, cv2, sys
|
675 |
+
from time import time as ttime
|
676 |
+
|
677 |
+
for weight_path, scale in [("weights_v3/up2x-latest-denoise3x.pth", 2), ("weights_v3/up3x-latest-denoise3x.pth", 3),
|
678 |
+
("weights_v3/up4x-latest-denoise3x.pth", 4)]:
|
679 |
+
for tile_mode in [0, 1, 2, 3, 4]:
|
680 |
+
upscaler2x = RealWaifuUpScaler(scale, weight_path, half=True, device="cuda:0")
|
681 |
+
input_dir = "%s/input_dir1" % root_path
|
682 |
+
output_dir = "%s/opt-dir-all-test" % root_path
|
683 |
+
os.makedirs(output_dir, exist_ok=True)
|
684 |
+
for name in os.listdir(input_dir):
|
685 |
+
print(name)
|
686 |
+
tmp = name.split(".")
|
687 |
+
inp_path = os.path.join(input_dir, name)
|
688 |
+
suffix = tmp[-1]
|
689 |
+
prefix = ".".join(tmp[:-1])
|
690 |
+
tmp_path = os.path.join(root_path, "tmp", "%s.%s" % (int(time.time() * 1000000), suffix))
|
691 |
+
print(inp_path, tmp_path)
|
692 |
+
# 支持中文路径
|
693 |
+
# os.link(inp_path, tmp_path)#win用硬链接
|
694 |
+
os.symlink(inp_path, tmp_path) # linux用软链接
|
695 |
+
frame = cv2.imread(tmp_path)[:, :, [2, 1, 0]]
|
696 |
+
t0 = ttime()
|
697 |
+
result = upscaler2x(frame, tile_mode=tile_mode)[:, :, ::-1]
|
698 |
+
t1 = ttime()
|
699 |
+
print(prefix, "done", t1 - t0)
|
700 |
+
tmp_opt_path = os.path.join(root_path, "tmp", "%s.%s" % (int(time.time() * 1000000), suffix))
|
701 |
+
cv2.imwrite(tmp_opt_path, result)
|
702 |
+
n = 0
|
703 |
+
while (1):
|
704 |
+
if (n == 0):
|
705 |
+
suffix = "_%sx_tile%s.png" % (scale, tile_mode)
|
706 |
+
else:
|
707 |
+
suffix = "_%sx_tile%s_%s.png" % (scale, tile_mode, n) #
|
708 |
+
if (os.path.exists(os.path.join(output_dir, prefix + suffix)) == False):
|
709 |
+
break
|
710 |
+
else:
|
711 |
+
n += 1
|
712 |
+
final_opt_path = os.path.join(output_dir, prefix + suffix)
|
713 |
+
os.rename(tmp_opt_path, final_opt_path)
|
714 |
+
os.remove(tmp_path)
|
weights_v3/up2x-latest-conservative.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:6cfe3b23687915d08ba96010f25198d9cfe8a683aa4131f1acf7eaa58ee1de93
|
3 |
+
size 5147249
|
weights_v3/up2x-latest-denoise1x.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:2e783c39da6a6394fbc250fdd069c55eaedc43971c4f2405322f18949ce38573
|
3 |
+
size 5147249
|
weights_v3/up2x-latest-denoise2x.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8188b3faef4258cf748c59360cbc8086ebedf4a63eb9d5d6637d45f819d32496
|
3 |
+
size 5147249
|
weights_v3/up2x-latest-denoise3x.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:0a14739f3f5fcbd74ec3ce2806d13a47916c916b20afe4a39d95f6df4ca6abd8
|
3 |
+
size 5147249
|
weights_v3/up2x-latest-no-denoise.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f491f9ecf6964ead9f3a36bf03e83527f32c6a341b683f7378ac6c1e2a5f0d16
|
3 |
+
size 5147249
|
weights_v3/up3x-latest-conservative.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f6ea5fd20380413beb2701182483fd80c2e86f3b3f08053eb3df4975184aefe3
|
3 |
+
size 5154161
|
weights_v3/up3x-latest-denoise3x.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:39f1e6e90d50e5528a63f4ba1866bad23365a737cbea22a80769b2ec4c1c3285
|
3 |
+
size 5154161
|
weights_v3/up3x-latest-no-denoise.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:763f0a87e70d744673f1a41db5396d5f334d22de97fff68ffc40deb91404a584
|
3 |
+
size 5154161
|
weights_v3/up4x-latest-conservative.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a8c8185def699b0883662a02df0ef2e6db3b0275170b6cc0d28089b64b273427
|
3 |
+
size 5636403
|
weights_v3/up4x-latest-denoise3x.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:42bd8fcdae37c12c5b25ed59625266bfa65780071a8d38192d83756cb85e98dd
|
3 |
+
size 5636403
|
weights_v3/up4x-latest-no-denoise.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:aaf3ef78a488cce5d3842154925eb70ff8423b8298e2cd189ec66eb7f6f66fae
|
3 |
+
size 5636403
|