Spaces:
Runtime error
Runtime error
Update Generater.py
Browse files- Generater.py +237 -39
Generater.py
CHANGED
@@ -1,42 +1,240 @@
|
|
1 |
-
|
2 |
import numpy as np
|
|
|
3 |
import paddle
|
|
|
|
|
|
|
|
|
|
|
4 |
import paddle.nn.functional as F
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
import numpy as np
|
3 |
+
# import os
|
4 |
import paddle
|
5 |
+
import paddle.optimizer
|
6 |
+
import paddle.nn as nn
|
7 |
+
# from tqdm import tqdm
|
8 |
+
# from paddle.io import Dataset
|
9 |
+
# from paddle.io import DataLoader
|
10 |
import paddle.nn.functional as F
|
11 |
+
# import paddle.tensor as tensor
|
12 |
+
|
13 |
+
class VGG19(nn.Layer):
|
14 |
+
cfg = [
|
15 |
+
64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512,'M', 512, 512, 512, 512, 'M']
|
16 |
+
|
17 |
+
def __init__(self, output_index: int = 26) -> None:
|
18 |
+
super().__init__()
|
19 |
+
# arch = 'caffevgg19'
|
20 |
+
# weights_path = get_path_from_url(model_urls[arch][0],
|
21 |
+
# model_urls[arch][1])
|
22 |
+
data_dict: dict = np.load("./vgg19_no_fc.npy",
|
23 |
+
encoding='latin1',
|
24 |
+
allow_pickle=True).item()
|
25 |
+
self.features = self.make_layers(self.cfg, data_dict)
|
26 |
+
del data_dict
|
27 |
+
self.features = nn.Sequential(*self.features.sublayers()[:output_index])
|
28 |
+
mean = paddle.to_tensor([103.939, 116.779, 123.68])
|
29 |
+
self.mean = mean.unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
|
30 |
+
|
31 |
+
def _process(self, x):
|
32 |
+
rgb = (x * 0.5 + 0.5) * 255 # value to 255
|
33 |
+
bgr = paddle.stack((rgb[:, 2, :, :], rgb[:, 1, :, :], rgb[:, 0, :, :]),
|
34 |
+
1) # rgb to bgr
|
35 |
+
return bgr - self.mean # vgg norm
|
36 |
+
|
37 |
+
def _forward_impl(self, x):
|
38 |
+
x = self._process(x)
|
39 |
+
# NOTE get output with out relu activation
|
40 |
+
x = self.features(x)
|
41 |
+
return x
|
42 |
+
|
43 |
+
def forward(self, x):
|
44 |
+
return self._forward_impl(x)
|
45 |
+
|
46 |
+
@staticmethod
|
47 |
+
def get_conv_filter(data_dict, name):
|
48 |
+
return data_dict[name][0]
|
49 |
+
|
50 |
+
@staticmethod
|
51 |
+
def get_bias(data_dict, name):
|
52 |
+
return data_dict[name][1]
|
53 |
+
|
54 |
+
@staticmethod
|
55 |
+
def get_fc_weight(data_dict, name):
|
56 |
+
return data_dict[name][0]
|
57 |
+
|
58 |
+
def make_layers(self, cfg, data_dict, batch_norm=False) -> nn.Sequential:
|
59 |
+
layers = []
|
60 |
+
in_channels = 3
|
61 |
+
block = 1
|
62 |
+
number = 1
|
63 |
+
for v in cfg:
|
64 |
+
if v == 'M':
|
65 |
+
layers += [nn.MaxPool2D(kernel_size=2, stride=2)]
|
66 |
+
block += 1
|
67 |
+
number = 1
|
68 |
+
else:
|
69 |
+
conv2d = nn.Conv2D(in_channels, v, kernel_size=3, padding=1)
|
70 |
+
""" set value """
|
71 |
+
weight = paddle.to_tensor(
|
72 |
+
self.get_conv_filter(data_dict, f'conv{block}_{number}'))
|
73 |
+
weight = weight.transpose((3, 2, 0, 1))
|
74 |
+
bias = paddle.to_tensor(
|
75 |
+
self.get_bias(data_dict, f'conv{block}_{number}'))
|
76 |
+
conv2d.weight.set_value(weight)
|
77 |
+
conv2d.bias.set_value(bias)
|
78 |
+
number += 1
|
79 |
+
if batch_norm:
|
80 |
+
layers += [conv2d, nn.BatchNorm2D(v), nn.ReLU()]
|
81 |
+
else:
|
82 |
+
layers += [conv2d, nn.ReLU()]
|
83 |
+
in_channels = v
|
84 |
+
# print("number",block)
|
85 |
+
return nn.Sequential(*layers)
|
86 |
+
|
87 |
+
|
88 |
+
class InvertedresBlock(nn.Layer):
|
89 |
+
def __init__(self,
|
90 |
+
in_channels: int,
|
91 |
+
expansion: float,
|
92 |
+
out_channels: int,
|
93 |
+
bias_attr=False):
|
94 |
+
super().__init__()
|
95 |
+
self.in_channels = in_channels
|
96 |
+
self.expansion = expansion
|
97 |
+
self.out_channels = out_channels
|
98 |
+
self.bottle_channels = round(self.expansion * self.out_channels)
|
99 |
+
self.body = nn.Sequential(
|
100 |
+
# pw
|
101 |
+
Conv2DNormLReLU(self.in_channels,
|
102 |
+
self.bottle_channels,
|
103 |
+
kernel_size=1,
|
104 |
+
bias_attr=bias_attr),
|
105 |
+
# dw
|
106 |
+
nn.Conv2D(self.bottle_channels,
|
107 |
+
self.bottle_channels,
|
108 |
+
kernel_size=3,
|
109 |
+
stride=1,
|
110 |
+
padding=0,
|
111 |
+
groups=self.bottle_channels,
|
112 |
+
bias_attr=True),
|
113 |
+
nn.GroupNorm(1, self.bottle_channels),
|
114 |
+
nn.LeakyReLU(0.2),
|
115 |
+
# pw & linear
|
116 |
+
nn.Conv2D(self.bottle_channels,
|
117 |
+
self.out_channels,
|
118 |
+
kernel_size=1,
|
119 |
+
padding=0,
|
120 |
+
bias_attr=False),
|
121 |
+
nn.GroupNorm(1, self.out_channels),
|
122 |
+
)
|
123 |
+
|
124 |
+
|
125 |
+
def forward(self, x0):
|
126 |
+
x = self.body(x0)
|
127 |
+
if self.in_channels == self.out_channels:
|
128 |
+
out = paddle.add(x0, x)
|
129 |
+
else:
|
130 |
+
out = x
|
131 |
+
return x
|
132 |
+
class Conv2DNormLReLU(nn.Layer):
|
133 |
+
def __init__(self,
|
134 |
+
in_channels: int,
|
135 |
+
out_channels: int,
|
136 |
+
kernel_size: int = 3,
|
137 |
+
stride: int = 1,
|
138 |
+
padding: int = 1,
|
139 |
+
bias_attr=False) -> None:
|
140 |
+
super().__init__()
|
141 |
+
self.conv = nn.Conv2D(in_channels,
|
142 |
+
out_channels,
|
143 |
+
kernel_size,
|
144 |
+
stride,
|
145 |
+
padding,
|
146 |
+
bias_attr=bias_attr)
|
147 |
+
# NOTE layer norm is crucial for animegan!
|
148 |
+
self.norm = nn.GroupNorm(1, out_channels)
|
149 |
+
self.lrelu = nn.LeakyReLU(0.2)
|
150 |
+
|
151 |
+
def forward(self, x):
|
152 |
+
x = self.conv(x)
|
153 |
+
x = self.norm(x)
|
154 |
+
x = self.lrelu(x)
|
155 |
+
return x
|
156 |
+
|
157 |
+
|
158 |
+
|
159 |
+
class Generater(nn.Layer):
|
160 |
+
def __init__(self):
|
161 |
+
super().__init__()
|
162 |
+
self.VGG = VGG19()
|
163 |
+
self.A = nn.Sequential(InvertedresBlock(512, 2, 256),
|
164 |
+
InvertedresBlock(256, 2, 256),
|
165 |
+
InvertedresBlock(256, 2, 256),
|
166 |
+
InvertedresBlock(256, 2, 256),
|
167 |
+
Conv2DNormLReLU(256, 128))
|
168 |
+
self.B = nn.Sequential(nn.Upsample(scale_factor=2, mode='bilinear'),
|
169 |
+
Conv2DNormLReLU(128, 128),
|
170 |
+
Conv2DNormLReLU(128, 128))
|
171 |
+
self.C = nn.Sequential(nn.Upsample(scale_factor=2, mode='bilinear'),
|
172 |
+
Conv2DNormLReLU(128, 128),
|
173 |
+
Conv2DNormLReLU(128, 128))
|
174 |
+
self.D = nn.Sequential(nn.Upsample(scale_factor=2, mode='bilinear'),
|
175 |
+
Conv2DNormLReLU(128, 64),
|
176 |
+
Conv2DNormLReLU(64, 64),
|
177 |
+
Conv2DNormLReLU(64, 32, 7, padding=3))
|
178 |
+
|
179 |
+
self.out = nn.Sequential(nn.Conv2D(32, 3, 1, bias_attr=False),
|
180 |
+
nn.Tanh())
|
181 |
+
# ,nn.Sigmoid())
|
182 |
+
def style_projection(self,content_feature,style_feature,alpha = 0.7):
|
183 |
+
def scatter_numpy(dim, index, src):
|
184 |
+
dst = src.copy()
|
185 |
+
idx_xsection_shape = index.shape[:dim] + index.shape[dim + 1:]
|
186 |
+
# print("idx_xsection_shape",idx_xsection_shape)#(b,c)
|
187 |
+
dst_xsection_shape = dst.shape[:dim] + dst.shape[dim + 1:]
|
188 |
+
def make_slice(arr, dim, i):
|
189 |
+
slc = [slice(None)] * arr.ndim
|
190 |
+
slc[dim] = i
|
191 |
+
return tuple(slc)
|
192 |
+
|
193 |
+
# We use index and dim parameters to create idx
|
194 |
+
# idx is in a form that can be used as a NumPy advanced index for scattering of src param.
|
195 |
+
idx = [[
|
196 |
+
*np.indices(idx_xsection_shape).reshape(index.ndim - 1, -1), index[make_slice(index, dim, i)].reshape(1, -1)[0]
|
197 |
+
] for i in range(index.shape[dim])]
|
198 |
+
idx = list(np.concatenate(idx, axis=1))
|
199 |
+
# print("idx",idx)
|
200 |
+
# idx.insert(dim, idx.pop())
|
201 |
+
|
202 |
+
if not np.isscalar(src):
|
203 |
+
src_idx = list(idx)#使idx和src_idx并不是同一个内存空间
|
204 |
+
src_idx.pop(dim)
|
205 |
+
src_idx.insert(dim, np.repeat(np.arange(index.shape[dim]), np.prod(idx_xsection_shape)))
|
206 |
+
dst[tuple(idx)] = src[tuple(src_idx)]
|
207 |
+
else:
|
208 |
+
dst[idx] = src
|
209 |
+
return dst
|
210 |
+
b,c,h,w = content_feature.shape
|
211 |
+
style_feature = F.interpolate(x=style_feature, size=content_feature.shape[-2:],mode="BILINEAR")
|
212 |
+
content_feat = content_feature.reshape([b,c,h*w]).numpy()
|
213 |
+
style_feat = style_feature.reshape([b,c,h*w]).numpy()
|
214 |
+
# print("content_feat",content_feat.shape,b,c)
|
215 |
+
# content_feat = np.reshape(content_feat, (b,c, -1))#(b,c,-1)
|
216 |
+
# style_feat = np.reshape(style_feat, (b,c, -1))#(b,c,-1)
|
217 |
+
# print(content_feat)
|
218 |
+
content_feat_index = np.argsort(content_feat, axis=2)
|
219 |
+
style_feat = np.sort(style_feat, axis=2)
|
220 |
+
# print("content_feat_index",content_feat_index)
|
221 |
+
# print("style_feat",style_feat)
|
222 |
+
fr_feat = scatter_numpy(dim=2, index=content_feat_index, src=style_feat)
|
223 |
+
fr_feat = fr_feat * alpha + content_feat * (1 - alpha)
|
224 |
+
fr_feat = np.reshape(fr_feat, (b,c,h,w))
|
225 |
+
fr_feat = paddle.to_tensor(fr_feat)
|
226 |
+
return fr_feat
|
227 |
+
# @paddle.jit.to_static
|
228 |
+
def forward(self,real_image,style_image,alpha):
|
229 |
+
alpha = alpha.numpy()[0]
|
230 |
+
# print("real_image",real_image.shape)
|
231 |
+
content_feature = self.VGG(real_image)
|
232 |
+
# print("content_feat",content_feature.shape)
|
233 |
+
style_feature = self.VGG(style_image)
|
234 |
+
fr_feat = self.style_projection(content_feature,style_feature,alpha)
|
235 |
+
a = self.A(fr_feat)
|
236 |
+
b = self.B(a)
|
237 |
+
c = self.C(b)
|
238 |
+
d = self.D(c)
|
239 |
+
out = self.out(d)
|
240 |
+
return out
|