PelosiFilippo commited on
Commit
d8eb455
1 Parent(s): c6cf61e

First commit

Browse files
Files changed (7) hide show
  1. app.py +91 -0
  2. models/network_swinir.py +867 -0
  3. packages.txt +1 -0
  4. render0001.png +0 -0
  5. render1546.png +0 -0
  6. render1682.png +0 -0
  7. requirements.txt +11 -0
app.py ADDED
@@ -0,0 +1,91 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ from PIL import Image
3
+
4
+ import torch
5
+ import numpy as np
6
+ from models.network_swinir import SwinIR as net
7
+
8
+ # model load
9
+ param_key_g = 'params_ema'
10
+ device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
11
+
12
+ super_res_model = net(upscale=4, in_chans=3, img_size=64, window_size=8,
13
+ img_range=1., depths=[6, 6, 6, 6, 6, 6, 6, 6, 6], embed_dim=240,
14
+ num_heads=[8, 8, 8, 8, 8, 8, 8, 8, 8],
15
+ mlp_ratio=2, upsampler='nearest+conv', resi_connection='3conv')
16
+ super_res_pretrained_model = torch.load("model_zoo/003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR-L_x4_PSNR.pth")
17
+ super_res_model.load_state_dict(super_res_pretrained_model[param_key_g] if param_key_g in super_res_pretrained_model.keys() else super_res_pretrained_model, strict=True)
18
+ super_res_model.eval()
19
+
20
+ def predict(input_img):
21
+ out = None
22
+
23
+ # preprocess input
24
+ if(input_img is not None):
25
+ # model predict
26
+ img_lq = input_img.astype(np.float32) / 255
27
+ img_lq = np.transpose(img_lq if img_lq.shape[2] == 1 else img_lq[:, :, [2, 1, 0]], (2, 0, 1)) # HCW-BGR to CHW-RGB
28
+ img_lq = torch.from_numpy(img_lq).float().unsqueeze(0).to(device) # CHW-RGB to NCHW-RGB
29
+
30
+ # inference
31
+ window_size = 8
32
+ model = super_res_model.to(device)
33
+
34
+ with torch.no_grad():
35
+ # pad input image to be a multiple of window_size
36
+ _, _, h_old, w_old = img_lq.size()
37
+ h_pad = (h_old // window_size + 1) * window_size - h_old
38
+ w_pad = (w_old // window_size + 1) * window_size - w_old
39
+ img_lq = torch.cat([img_lq, torch.flip(img_lq, [2])], 2)[:, :, :h_old + h_pad, :]
40
+ img_lq = torch.cat([img_lq, torch.flip(img_lq, [3])], 3)[:, :, :, :w_old + w_pad]
41
+ output = test(model, img_lq)
42
+ output = output[..., :h_old * 4, :w_old * 4]
43
+
44
+ # process image
45
+ output = output.data.squeeze().float().cpu().clamp_(0, 1).numpy()
46
+ if output.ndim == 3:
47
+ output = np.transpose(output[[2, 1, 0], :, :], (1, 2, 0)) # CHW-RGB to HCW-BGR
48
+ output = (output * 255.0).round().astype(np.uint8) # float32 to uint8
49
+
50
+ # convert to pil image
51
+ out = Image.fromarray(output)
52
+ return out
53
+
54
+ def test(model, img_lq):
55
+ # test the image tile by tile
56
+ b, c, h, w = img_lq.size()
57
+ tile = min(800, h, w)
58
+ tile_overlap = 32
59
+ sf = 4
60
+
61
+ stride = tile - tile_overlap
62
+ h_idx_list = list(range(0, h-tile, stride)) + [h-tile]
63
+ w_idx_list = list(range(0, w-tile, stride)) + [w-tile]
64
+ E = torch.zeros(b, c, h*sf, w*sf).type_as(img_lq)
65
+ W = torch.zeros_like(E)
66
+
67
+ for h_idx in h_idx_list:
68
+ for w_idx in w_idx_list:
69
+ in_patch = img_lq[..., h_idx:h_idx+tile, w_idx:w_idx+tile]
70
+ out_patch = model(in_patch)
71
+ out_patch_mask = torch.ones_like(out_patch)
72
+
73
+ E[..., h_idx*sf:(h_idx+tile)*sf, w_idx*sf:(w_idx+tile)*sf].add_(out_patch)
74
+ W[..., h_idx*sf:(h_idx+tile)*sf, w_idx*sf:(w_idx+tile)*sf].add_(out_patch_mask)
75
+ output = E.div_(W)
76
+ return output
77
+
78
+ gr.Interface(
79
+ fn=predict,
80
+ inputs=[
81
+ gr.inputs.Image()
82
+ ],
83
+ outputs=[
84
+ gr.inputs.Image()
85
+ ],
86
+ title="SwinIR moon super resolution",
87
+ description="Description of the app",
88
+ examples=[
89
+ "render0001.png", "render1546.png", "render1682.png"
90
+ ]
91
+ ).launch()
models/network_swinir.py ADDED
@@ -0,0 +1,867 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -----------------------------------------------------------------------------------
2
+ # SwinIR: Image Restoration Using Swin Transformer, https://arxiv.org/abs/2108.10257
3
+ # Originally Written by Ze Liu, Modified by Jingyun Liang.
4
+ # -----------------------------------------------------------------------------------
5
+
6
+ import math
7
+ import torch
8
+ import torch.nn as nn
9
+ import torch.nn.functional as F
10
+ import torch.utils.checkpoint as checkpoint
11
+ from timm.models.layers import DropPath, to_2tuple, trunc_normal_
12
+
13
+
14
+ class Mlp(nn.Module):
15
+ def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
16
+ super().__init__()
17
+ out_features = out_features or in_features
18
+ hidden_features = hidden_features or in_features
19
+ self.fc1 = nn.Linear(in_features, hidden_features)
20
+ self.act = act_layer()
21
+ self.fc2 = nn.Linear(hidden_features, out_features)
22
+ self.drop = nn.Dropout(drop)
23
+
24
+ def forward(self, x):
25
+ x = self.fc1(x)
26
+ x = self.act(x)
27
+ x = self.drop(x)
28
+ x = self.fc2(x)
29
+ x = self.drop(x)
30
+ return x
31
+
32
+
33
+ def window_partition(x, window_size):
34
+ """
35
+ Args:
36
+ x: (B, H, W, C)
37
+ window_size (int): window size
38
+
39
+ Returns:
40
+ windows: (num_windows*B, window_size, window_size, C)
41
+ """
42
+ B, H, W, C = x.shape
43
+ x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
44
+ windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
45
+ return windows
46
+
47
+
48
+ def window_reverse(windows, window_size, H, W):
49
+ """
50
+ Args:
51
+ windows: (num_windows*B, window_size, window_size, C)
52
+ window_size (int): Window size
53
+ H (int): Height of image
54
+ W (int): Width of image
55
+
56
+ Returns:
57
+ x: (B, H, W, C)
58
+ """
59
+ B = int(windows.shape[0] / (H * W / window_size / window_size))
60
+ x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
61
+ x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
62
+ return x
63
+
64
+
65
+ class WindowAttention(nn.Module):
66
+ r""" Window based multi-head self attention (W-MSA) module with relative position bias.
67
+ It supports both of shifted and non-shifted window.
68
+
69
+ Args:
70
+ dim (int): Number of input channels.
71
+ window_size (tuple[int]): The height and width of the window.
72
+ num_heads (int): Number of attention heads.
73
+ qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
74
+ qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
75
+ attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
76
+ proj_drop (float, optional): Dropout ratio of output. Default: 0.0
77
+ """
78
+
79
+ def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
80
+
81
+ super().__init__()
82
+ self.dim = dim
83
+ self.window_size = window_size # Wh, Ww
84
+ self.num_heads = num_heads
85
+ head_dim = dim // num_heads
86
+ self.scale = qk_scale or head_dim ** -0.5
87
+
88
+ # define a parameter table of relative position bias
89
+ self.relative_position_bias_table = nn.Parameter(
90
+ torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH
91
+
92
+ # get pair-wise relative position index for each token inside the window
93
+ coords_h = torch.arange(self.window_size[0])
94
+ coords_w = torch.arange(self.window_size[1])
95
+ coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
96
+ coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
97
+ relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww
98
+ relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
99
+ relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
100
+ relative_coords[:, :, 1] += self.window_size[1] - 1
101
+ relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
102
+ relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
103
+ self.register_buffer("relative_position_index", relative_position_index)
104
+
105
+ self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
106
+ self.attn_drop = nn.Dropout(attn_drop)
107
+ self.proj = nn.Linear(dim, dim)
108
+
109
+ self.proj_drop = nn.Dropout(proj_drop)
110
+
111
+ trunc_normal_(self.relative_position_bias_table, std=.02)
112
+ self.softmax = nn.Softmax(dim=-1)
113
+
114
+ def forward(self, x, mask=None):
115
+ """
116
+ Args:
117
+ x: input features with shape of (num_windows*B, N, C)
118
+ mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
119
+ """
120
+ B_, N, C = x.shape
121
+ qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
122
+ q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
123
+
124
+ q = q * self.scale
125
+ attn = (q @ k.transpose(-2, -1))
126
+
127
+ relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
128
+ self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH
129
+ relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
130
+ attn = attn + relative_position_bias.unsqueeze(0)
131
+
132
+ if mask is not None:
133
+ nW = mask.shape[0]
134
+ attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
135
+ attn = attn.view(-1, self.num_heads, N, N)
136
+ attn = self.softmax(attn)
137
+ else:
138
+ attn = self.softmax(attn)
139
+
140
+ attn = self.attn_drop(attn)
141
+
142
+ x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
143
+ x = self.proj(x)
144
+ x = self.proj_drop(x)
145
+ return x
146
+
147
+ def extra_repr(self) -> str:
148
+ return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
149
+
150
+ def flops(self, N):
151
+ # calculate flops for 1 window with token length of N
152
+ flops = 0
153
+ # qkv = self.qkv(x)
154
+ flops += N * self.dim * 3 * self.dim
155
+ # attn = (q @ k.transpose(-2, -1))
156
+ flops += self.num_heads * N * (self.dim // self.num_heads) * N
157
+ # x = (attn @ v)
158
+ flops += self.num_heads * N * N * (self.dim // self.num_heads)
159
+ # x = self.proj(x)
160
+ flops += N * self.dim * self.dim
161
+ return flops
162
+
163
+
164
+ class SwinTransformerBlock(nn.Module):
165
+ r""" Swin Transformer Block.
166
+
167
+ Args:
168
+ dim (int): Number of input channels.
169
+ input_resolution (tuple[int]): Input resulotion.
170
+ num_heads (int): Number of attention heads.
171
+ window_size (int): Window size.
172
+ shift_size (int): Shift size for SW-MSA.
173
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
174
+ qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
175
+ qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
176
+ drop (float, optional): Dropout rate. Default: 0.0
177
+ attn_drop (float, optional): Attention dropout rate. Default: 0.0
178
+ drop_path (float, optional): Stochastic depth rate. Default: 0.0
179
+ act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
180
+ norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
181
+ """
182
+
183
+ def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
184
+ mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
185
+ act_layer=nn.GELU, norm_layer=nn.LayerNorm):
186
+ super().__init__()
187
+ self.dim = dim
188
+ self.input_resolution = input_resolution
189
+ self.num_heads = num_heads
190
+ self.window_size = window_size
191
+ self.shift_size = shift_size
192
+ self.mlp_ratio = mlp_ratio
193
+ if min(self.input_resolution) <= self.window_size:
194
+ # if window size is larger than input resolution, we don't partition windows
195
+ self.shift_size = 0
196
+ self.window_size = min(self.input_resolution)
197
+ assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
198
+
199
+ self.norm1 = norm_layer(dim)
200
+ self.attn = WindowAttention(
201
+ dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
202
+ qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
203
+
204
+ self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
205
+ self.norm2 = norm_layer(dim)
206
+ mlp_hidden_dim = int(dim * mlp_ratio)
207
+ self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
208
+
209
+ if self.shift_size > 0:
210
+ attn_mask = self.calculate_mask(self.input_resolution)
211
+ else:
212
+ attn_mask = None
213
+
214
+ self.register_buffer("attn_mask", attn_mask)
215
+
216
+ def calculate_mask(self, x_size):
217
+ # calculate attention mask for SW-MSA
218
+ H, W = x_size
219
+ img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1
220
+ h_slices = (slice(0, -self.window_size),
221
+ slice(-self.window_size, -self.shift_size),
222
+ slice(-self.shift_size, None))
223
+ w_slices = (slice(0, -self.window_size),
224
+ slice(-self.window_size, -self.shift_size),
225
+ slice(-self.shift_size, None))
226
+ cnt = 0
227
+ for h in h_slices:
228
+ for w in w_slices:
229
+ img_mask[:, h, w, :] = cnt
230
+ cnt += 1
231
+
232
+ mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
233
+ mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
234
+ attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
235
+ attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
236
+
237
+ return attn_mask
238
+
239
+ def forward(self, x, x_size):
240
+ H, W = x_size
241
+ B, L, C = x.shape
242
+ # assert L == H * W, "input feature has wrong size"
243
+
244
+ shortcut = x
245
+ x = self.norm1(x)
246
+ x = x.view(B, H, W, C)
247
+
248
+ # cyclic shift
249
+ if self.shift_size > 0:
250
+ shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
251
+ else:
252
+ shifted_x = x
253
+
254
+ # partition windows
255
+ x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
256
+ x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C
257
+
258
+ # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
259
+ if self.input_resolution == x_size:
260
+ attn_windows = self.attn(x_windows, mask=self.attn_mask) # nW*B, window_size*window_size, C
261
+ else:
262
+ attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))
263
+
264
+ # merge windows
265
+ attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
266
+ shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C
267
+
268
+ # reverse cyclic shift
269
+ if self.shift_size > 0:
270
+ x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
271
+ else:
272
+ x = shifted_x
273
+ x = x.view(B, H * W, C)
274
+
275
+ # FFN
276
+ x = shortcut + self.drop_path(x)
277
+ x = x + self.drop_path(self.mlp(self.norm2(x)))
278
+
279
+ return x
280
+
281
+ def extra_repr(self) -> str:
282
+ return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
283
+ f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
284
+
285
+ def flops(self):
286
+ flops = 0
287
+ H, W = self.input_resolution
288
+ # norm1
289
+ flops += self.dim * H * W
290
+ # W-MSA/SW-MSA
291
+ nW = H * W / self.window_size / self.window_size
292
+ flops += nW * self.attn.flops(self.window_size * self.window_size)
293
+ # mlp
294
+ flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
295
+ # norm2
296
+ flops += self.dim * H * W
297
+ return flops
298
+
299
+
300
+ class PatchMerging(nn.Module):
301
+ r""" Patch Merging Layer.
302
+
303
+ Args:
304
+ input_resolution (tuple[int]): Resolution of input feature.
305
+ dim (int): Number of input channels.
306
+ norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
307
+ """
308
+
309
+ def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
310
+ super().__init__()
311
+ self.input_resolution = input_resolution
312
+ self.dim = dim
313
+ self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
314
+ self.norm = norm_layer(4 * dim)
315
+
316
+ def forward(self, x):
317
+ """
318
+ x: B, H*W, C
319
+ """
320
+ H, W = self.input_resolution
321
+ B, L, C = x.shape
322
+ assert L == H * W, "input feature has wrong size"
323
+ assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
324
+
325
+ x = x.view(B, H, W, C)
326
+
327
+ x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C
328
+ x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C
329
+ x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C
330
+ x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C
331
+ x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C
332
+ x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C
333
+
334
+ x = self.norm(x)
335
+ x = self.reduction(x)
336
+
337
+ return x
338
+
339
+ def extra_repr(self) -> str:
340
+ return f"input_resolution={self.input_resolution}, dim={self.dim}"
341
+
342
+ def flops(self):
343
+ H, W = self.input_resolution
344
+ flops = H * W * self.dim
345
+ flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
346
+ return flops
347
+
348
+
349
+ class BasicLayer(nn.Module):
350
+ """ A basic Swin Transformer layer for one stage.
351
+
352
+ Args:
353
+ dim (int): Number of input channels.
354
+ input_resolution (tuple[int]): Input resolution.
355
+ depth (int): Number of blocks.
356
+ num_heads (int): Number of attention heads.
357
+ window_size (int): Local window size.
358
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
359
+ qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
360
+ qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
361
+ drop (float, optional): Dropout rate. Default: 0.0
362
+ attn_drop (float, optional): Attention dropout rate. Default: 0.0
363
+ drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
364
+ norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
365
+ downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
366
+ use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
367
+ """
368
+
369
+ def __init__(self, dim, input_resolution, depth, num_heads, window_size,
370
+ mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
371
+ drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False):
372
+
373
+ super().__init__()
374
+ self.dim = dim
375
+ self.input_resolution = input_resolution
376
+ self.depth = depth
377
+ self.use_checkpoint = use_checkpoint
378
+
379
+ # build blocks
380
+ self.blocks = nn.ModuleList([
381
+ SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
382
+ num_heads=num_heads, window_size=window_size,
383
+ shift_size=0 if (i % 2 == 0) else window_size // 2,
384
+ mlp_ratio=mlp_ratio,
385
+ qkv_bias=qkv_bias, qk_scale=qk_scale,
386
+ drop=drop, attn_drop=attn_drop,
387
+ drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
388
+ norm_layer=norm_layer)
389
+ for i in range(depth)])
390
+
391
+ # patch merging layer
392
+ if downsample is not None:
393
+ self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
394
+ else:
395
+ self.downsample = None
396
+
397
+ def forward(self, x, x_size):
398
+ for blk in self.blocks:
399
+ if self.use_checkpoint:
400
+ x = checkpoint.checkpoint(blk, x, x_size)
401
+ else:
402
+ x = blk(x, x_size)
403
+ if self.downsample is not None:
404
+ x = self.downsample(x)
405
+ return x
406
+
407
+ def extra_repr(self) -> str:
408
+ return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
409
+
410
+ def flops(self):
411
+ flops = 0
412
+ for blk in self.blocks:
413
+ flops += blk.flops()
414
+ if self.downsample is not None:
415
+ flops += self.downsample.flops()
416
+ return flops
417
+
418
+
419
+ class RSTB(nn.Module):
420
+ """Residual Swin Transformer Block (RSTB).
421
+
422
+ Args:
423
+ dim (int): Number of input channels.
424
+ input_resolution (tuple[int]): Input resolution.
425
+ depth (int): Number of blocks.
426
+ num_heads (int): Number of attention heads.
427
+ window_size (int): Local window size.
428
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
429
+ qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
430
+ qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
431
+ drop (float, optional): Dropout rate. Default: 0.0
432
+ attn_drop (float, optional): Attention dropout rate. Default: 0.0
433
+ drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
434
+ norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
435
+ downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
436
+ use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
437
+ img_size: Input image size.
438
+ patch_size: Patch size.
439
+ resi_connection: The convolutional block before residual connection.
440
+ """
441
+
442
+ def __init__(self, dim, input_resolution, depth, num_heads, window_size,
443
+ mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
444
+ drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
445
+ img_size=224, patch_size=4, resi_connection='1conv'):
446
+ super(RSTB, self).__init__()
447
+
448
+ self.dim = dim
449
+ self.input_resolution = input_resolution
450
+
451
+ self.residual_group = BasicLayer(dim=dim,
452
+ input_resolution=input_resolution,
453
+ depth=depth,
454
+ num_heads=num_heads,
455
+ window_size=window_size,
456
+ mlp_ratio=mlp_ratio,
457
+ qkv_bias=qkv_bias, qk_scale=qk_scale,
458
+ drop=drop, attn_drop=attn_drop,
459
+ drop_path=drop_path,
460
+ norm_layer=norm_layer,
461
+ downsample=downsample,
462
+ use_checkpoint=use_checkpoint)
463
+
464
+ if resi_connection == '1conv':
465
+ self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
466
+ elif resi_connection == '3conv':
467
+ # to save parameters and memory
468
+ self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True),
469
+ nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
470
+ nn.LeakyReLU(negative_slope=0.2, inplace=True),
471
+ nn.Conv2d(dim // 4, dim, 3, 1, 1))
472
+
473
+ self.patch_embed = PatchEmbed(
474
+ img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
475
+ norm_layer=None)
476
+
477
+ self.patch_unembed = PatchUnEmbed(
478
+ img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
479
+ norm_layer=None)
480
+
481
+ def forward(self, x, x_size):
482
+ return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x
483
+
484
+ def flops(self):
485
+ flops = 0
486
+ flops += self.residual_group.flops()
487
+ H, W = self.input_resolution
488
+ flops += H * W * self.dim * self.dim * 9
489
+ flops += self.patch_embed.flops()
490
+ flops += self.patch_unembed.flops()
491
+
492
+ return flops
493
+
494
+
495
+ class PatchEmbed(nn.Module):
496
+ r""" Image to Patch Embedding
497
+
498
+ Args:
499
+ img_size (int): Image size. Default: 224.
500
+ patch_size (int): Patch token size. Default: 4.
501
+ in_chans (int): Number of input image channels. Default: 3.
502
+ embed_dim (int): Number of linear projection output channels. Default: 96.
503
+ norm_layer (nn.Module, optional): Normalization layer. Default: None
504
+ """
505
+
506
+ def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
507
+ super().__init__()
508
+ img_size = to_2tuple(img_size)
509
+ patch_size = to_2tuple(patch_size)
510
+ patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
511
+ self.img_size = img_size
512
+ self.patch_size = patch_size
513
+ self.patches_resolution = patches_resolution
514
+ self.num_patches = patches_resolution[0] * patches_resolution[1]
515
+
516
+ self.in_chans = in_chans
517
+ self.embed_dim = embed_dim
518
+
519
+ if norm_layer is not None:
520
+ self.norm = norm_layer(embed_dim)
521
+ else:
522
+ self.norm = None
523
+
524
+ def forward(self, x):
525
+ x = x.flatten(2).transpose(1, 2) # B Ph*Pw C
526
+ if self.norm is not None:
527
+ x = self.norm(x)
528
+ return x
529
+
530
+ def flops(self):
531
+ flops = 0
532
+ H, W = self.img_size
533
+ if self.norm is not None:
534
+ flops += H * W * self.embed_dim
535
+ return flops
536
+
537
+
538
+ class PatchUnEmbed(nn.Module):
539
+ r""" Image to Patch Unembedding
540
+
541
+ Args:
542
+ img_size (int): Image size. Default: 224.
543
+ patch_size (int): Patch token size. Default: 4.
544
+ in_chans (int): Number of input image channels. Default: 3.
545
+ embed_dim (int): Number of linear projection output channels. Default: 96.
546
+ norm_layer (nn.Module, optional): Normalization layer. Default: None
547
+ """
548
+
549
+ def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
550
+ super().__init__()
551
+ img_size = to_2tuple(img_size)
552
+ patch_size = to_2tuple(patch_size)
553
+ patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
554
+ self.img_size = img_size
555
+ self.patch_size = patch_size
556
+ self.patches_resolution = patches_resolution
557
+ self.num_patches = patches_resolution[0] * patches_resolution[1]
558
+
559
+ self.in_chans = in_chans
560
+ self.embed_dim = embed_dim
561
+
562
+ def forward(self, x, x_size):
563
+ B, HW, C = x.shape
564
+ x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C
565
+ return x
566
+
567
+ def flops(self):
568
+ flops = 0
569
+ return flops
570
+
571
+
572
+ class Upsample(nn.Sequential):
573
+ """Upsample module.
574
+
575
+ Args:
576
+ scale (int): Scale factor. Supported scales: 2^n and 3.
577
+ num_feat (int): Channel number of intermediate features.
578
+ """
579
+
580
+ def __init__(self, scale, num_feat):
581
+ m = []
582
+ if (scale & (scale - 1)) == 0: # scale = 2^n
583
+ for _ in range(int(math.log(scale, 2))):
584
+ m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
585
+ m.append(nn.PixelShuffle(2))
586
+ elif scale == 3:
587
+ m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
588
+ m.append(nn.PixelShuffle(3))
589
+ else:
590
+ raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
591
+ super(Upsample, self).__init__(*m)
592
+
593
+
594
+ class UpsampleOneStep(nn.Sequential):
595
+ """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
596
+ Used in lightweight SR to save parameters.
597
+
598
+ Args:
599
+ scale (int): Scale factor. Supported scales: 2^n and 3.
600
+ num_feat (int): Channel number of intermediate features.
601
+
602
+ """
603
+
604
+ def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
605
+ self.num_feat = num_feat
606
+ self.input_resolution = input_resolution
607
+ m = []
608
+ m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1))
609
+ m.append(nn.PixelShuffle(scale))
610
+ super(UpsampleOneStep, self).__init__(*m)
611
+
612
+ def flops(self):
613
+ H, W = self.input_resolution
614
+ flops = H * W * self.num_feat * 3 * 9
615
+ return flops
616
+
617
+
618
+ class SwinIR(nn.Module):
619
+ r""" SwinIR
620
+ A PyTorch impl of : `SwinIR: Image Restoration Using Swin Transformer`, based on Swin Transformer.
621
+
622
+ Args:
623
+ img_size (int | tuple(int)): Input image size. Default 64
624
+ patch_size (int | tuple(int)): Patch size. Default: 1
625
+ in_chans (int): Number of input image channels. Default: 3
626
+ embed_dim (int): Patch embedding dimension. Default: 96
627
+ depths (tuple(int)): Depth of each Swin Transformer layer.
628
+ num_heads (tuple(int)): Number of attention heads in different layers.
629
+ window_size (int): Window size. Default: 7
630
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
631
+ qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
632
+ qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
633
+ drop_rate (float): Dropout rate. Default: 0
634
+ attn_drop_rate (float): Attention dropout rate. Default: 0
635
+ drop_path_rate (float): Stochastic depth rate. Default: 0.1
636
+ norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
637
+ ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
638
+ patch_norm (bool): If True, add normalization after patch embedding. Default: True
639
+ use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
640
+ upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
641
+ img_range: Image range. 1. or 255.
642
+ upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
643
+ resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
644
+ """
645
+
646
+ def __init__(self, img_size=64, patch_size=1, in_chans=3,
647
+ embed_dim=96, depths=[6, 6, 6, 6], num_heads=[6, 6, 6, 6],
648
+ window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None,
649
+ drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
650
+ norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
651
+ use_checkpoint=False, upscale=2, img_range=1., upsampler='', resi_connection='1conv',
652
+ **kwargs):
653
+ super(SwinIR, self).__init__()
654
+ num_in_ch = in_chans
655
+ num_out_ch = in_chans
656
+ num_feat = 64
657
+ self.img_range = img_range
658
+ if in_chans == 3:
659
+ rgb_mean = (0.4488, 0.4371, 0.4040)
660
+ self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
661
+ else:
662
+ self.mean = torch.zeros(1, 1, 1, 1)
663
+ self.upscale = upscale
664
+ self.upsampler = upsampler
665
+ self.window_size = window_size
666
+
667
+ #####################################################################################################
668
+ ################################### 1, shallow feature extraction ###################################
669
+ self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
670
+
671
+ #####################################################################################################
672
+ ################################### 2, deep feature extraction ######################################
673
+ self.num_layers = len(depths)
674
+ self.embed_dim = embed_dim
675
+ self.ape = ape
676
+ self.patch_norm = patch_norm
677
+ self.num_features = embed_dim
678
+ self.mlp_ratio = mlp_ratio
679
+
680
+ # split image into non-overlapping patches
681
+ self.patch_embed = PatchEmbed(
682
+ img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
683
+ norm_layer=norm_layer if self.patch_norm else None)
684
+ num_patches = self.patch_embed.num_patches
685
+ patches_resolution = self.patch_embed.patches_resolution
686
+ self.patches_resolution = patches_resolution
687
+
688
+ # merge non-overlapping patches into image
689
+ self.patch_unembed = PatchUnEmbed(
690
+ img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
691
+ norm_layer=norm_layer if self.patch_norm else None)
692
+
693
+ # absolute position embedding
694
+ if self.ape:
695
+ self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
696
+ trunc_normal_(self.absolute_pos_embed, std=.02)
697
+
698
+ self.pos_drop = nn.Dropout(p=drop_rate)
699
+
700
+ # stochastic depth
701
+ dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
702
+
703
+ # build Residual Swin Transformer blocks (RSTB)
704
+ self.layers = nn.ModuleList()
705
+ for i_layer in range(self.num_layers):
706
+ layer = RSTB(dim=embed_dim,
707
+ input_resolution=(patches_resolution[0],
708
+ patches_resolution[1]),
709
+ depth=depths[i_layer],
710
+ num_heads=num_heads[i_layer],
711
+ window_size=window_size,
712
+ mlp_ratio=self.mlp_ratio,
713
+ qkv_bias=qkv_bias, qk_scale=qk_scale,
714
+ drop=drop_rate, attn_drop=attn_drop_rate,
715
+ drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], # no impact on SR results
716
+ norm_layer=norm_layer,
717
+ downsample=None,
718
+ use_checkpoint=use_checkpoint,
719
+ img_size=img_size,
720
+ patch_size=patch_size,
721
+ resi_connection=resi_connection
722
+
723
+ )
724
+ self.layers.append(layer)
725
+ self.norm = norm_layer(self.num_features)
726
+
727
+ # build the last conv layer in deep feature extraction
728
+ if resi_connection == '1conv':
729
+ self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
730
+ elif resi_connection == '3conv':
731
+ # to save parameters and memory
732
+ self.conv_after_body = nn.Sequential(nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
733
+ nn.LeakyReLU(negative_slope=0.2, inplace=True),
734
+ nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
735
+ nn.LeakyReLU(negative_slope=0.2, inplace=True),
736
+ nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1))
737
+
738
+ #####################################################################################################
739
+ ################################ 3, high quality image reconstruction ################################
740
+ if self.upsampler == 'pixelshuffle':
741
+ # for classical SR
742
+ self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
743
+ nn.LeakyReLU(inplace=True))
744
+ self.upsample = Upsample(upscale, num_feat)
745
+ self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
746
+ elif self.upsampler == 'pixelshuffledirect':
747
+ # for lightweight SR (to save parameters)
748
+ self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch,
749
+ (patches_resolution[0], patches_resolution[1]))
750
+ elif self.upsampler == 'nearest+conv':
751
+ # for real-world SR (less artifacts)
752
+ self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
753
+ nn.LeakyReLU(inplace=True))
754
+ self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
755
+ if self.upscale == 4:
756
+ self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
757
+ self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
758
+ self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
759
+ self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
760
+ else:
761
+ # for image denoising and JPEG compression artifact reduction
762
+ self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
763
+
764
+ self.apply(self._init_weights)
765
+
766
+ def _init_weights(self, m):
767
+ if isinstance(m, nn.Linear):
768
+ trunc_normal_(m.weight, std=.02)
769
+ if isinstance(m, nn.Linear) and m.bias is not None:
770
+ nn.init.constant_(m.bias, 0)
771
+ elif isinstance(m, nn.LayerNorm):
772
+ nn.init.constant_(m.bias, 0)
773
+ nn.init.constant_(m.weight, 1.0)
774
+
775
+ @torch.jit.ignore
776
+ def no_weight_decay(self):
777
+ return {'absolute_pos_embed'}
778
+
779
+ @torch.jit.ignore
780
+ def no_weight_decay_keywords(self):
781
+ return {'relative_position_bias_table'}
782
+
783
+ def check_image_size(self, x):
784
+ _, _, h, w = x.size()
785
+ mod_pad_h = (self.window_size - h % self.window_size) % self.window_size
786
+ mod_pad_w = (self.window_size - w % self.window_size) % self.window_size
787
+ x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect')
788
+ return x
789
+
790
+ def forward_features(self, x):
791
+ x_size = (x.shape[2], x.shape[3])
792
+ x = self.patch_embed(x)
793
+ if self.ape:
794
+ x = x + self.absolute_pos_embed
795
+ x = self.pos_drop(x)
796
+
797
+ for layer in self.layers:
798
+ x = layer(x, x_size)
799
+
800
+ x = self.norm(x) # B L C
801
+ x = self.patch_unembed(x, x_size)
802
+
803
+ return x
804
+
805
+ def forward(self, x):
806
+ H, W = x.shape[2:]
807
+ x = self.check_image_size(x)
808
+
809
+ self.mean = self.mean.type_as(x)
810
+ x = (x - self.mean) * self.img_range
811
+
812
+ if self.upsampler == 'pixelshuffle':
813
+ # for classical SR
814
+ x = self.conv_first(x)
815
+ x = self.conv_after_body(self.forward_features(x)) + x
816
+ x = self.conv_before_upsample(x)
817
+ x = self.conv_last(self.upsample(x))
818
+ elif self.upsampler == 'pixelshuffledirect':
819
+ # for lightweight SR
820
+ x = self.conv_first(x)
821
+ x = self.conv_after_body(self.forward_features(x)) + x
822
+ x = self.upsample(x)
823
+ elif self.upsampler == 'nearest+conv':
824
+ # for real-world SR
825
+ x = self.conv_first(x)
826
+ x = self.conv_after_body(self.forward_features(x)) + x
827
+ x = self.conv_before_upsample(x)
828
+ x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
829
+ if self.upscale == 4:
830
+ x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
831
+ x = self.conv_last(self.lrelu(self.conv_hr(x)))
832
+ else:
833
+ # for image denoising and JPEG compression artifact reduction
834
+ x_first = self.conv_first(x)
835
+ res = self.conv_after_body(self.forward_features(x_first)) + x_first
836
+ x = x + self.conv_last(res)
837
+
838
+ x = x / self.img_range + self.mean
839
+
840
+ return x[:, :, :H*self.upscale, :W*self.upscale]
841
+
842
+ def flops(self):
843
+ flops = 0
844
+ H, W = self.patches_resolution
845
+ flops += H * W * 3 * self.embed_dim * 9
846
+ flops += self.patch_embed.flops()
847
+ for i, layer in enumerate(self.layers):
848
+ flops += layer.flops()
849
+ flops += H * W * 3 * self.embed_dim * self.embed_dim
850
+ flops += self.upsample.flops()
851
+ return flops
852
+
853
+
854
+ if __name__ == '__main__':
855
+ upscale = 4
856
+ window_size = 8
857
+ height = (1024 // upscale // window_size + 1) * window_size
858
+ width = (720 // upscale // window_size + 1) * window_size
859
+ model = SwinIR(upscale=2, img_size=(height, width),
860
+ window_size=window_size, img_range=1., depths=[6, 6, 6, 6],
861
+ embed_dim=60, num_heads=[6, 6, 6, 6], mlp_ratio=2, upsampler='pixelshuffledirect')
862
+ print(model)
863
+ print(height, width, model.flops() / 1e9)
864
+
865
+ x = torch.randn((1, 3, height, width))
866
+ x = model(x)
867
+ print(x.shape)
packages.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ python3-opencv
render0001.png ADDED
render1546.png ADDED
render1682.png ADDED
requirements.txt ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ opencv-python
2
+ scikit-image
3
+ pillow
4
+ torchvision
5
+ hdf5storage
6
+ ninja
7
+ lmdb
8
+ requests
9
+ timm
10
+ einops
11
+ gradio