File size: 5,118 Bytes
159cb3e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
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
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
import torch
import torch.nn as nn
import torch.nn.functional as F
try:
    from arch_model import EBlock, DBlock
    from arch_util import CustomSequential
except:
    from archs.arch_model import EBlock, DBlock
    from .arch_util import CustomSequential

class DarkIR(nn.Module):
    
    def __init__(self, img_channel=3, 
                 width=32, 
                 middle_blk_num_enc=2,
                 middle_blk_num_dec=2, 
                 enc_blk_nums=[1, 2, 3], 
                 dec_blk_nums=[3, 1, 1],  
                 dilations = [1, 4, 9], 
                 extra_depth_wise = True):
        super(DarkIR, self).__init__()
        
        self.intro = nn.Conv2d(in_channels=img_channel, out_channels=width, kernel_size=3, padding=1, stride=1, groups=1,
                                bias=True)
        self.ending = nn.Conv2d(in_channels=width, out_channels=img_channel, kernel_size=3, padding=1, stride=1, groups=1,
                              bias=True)

        self.encoders = nn.ModuleList()
        self.decoders = nn.ModuleList()
        self.middle_blks = nn.ModuleList()
        self.ups = nn.ModuleList()
        self.downs = nn.ModuleList()
        
        chan = width
        for num in enc_blk_nums:
            self.encoders.append(
                CustomSequential(
                    *[EBlock(chan, extra_depth_wise=extra_depth_wise) for _ in range(num)]
                )
            )
            self.downs.append(
                nn.Conv2d(chan, 2*chan, 2, 2)
            )
            chan = chan * 2

        self.middle_blks_enc = \
            CustomSequential(
                *[EBlock(chan, extra_depth_wise=extra_depth_wise) for _ in range(middle_blk_num_enc)]
            )
        self.middle_blks_dec = \
            CustomSequential(
                *[DBlock(chan, dilations=dilations, extra_depth_wise=extra_depth_wise) for _ in range(middle_blk_num_dec)]
            )

        for num in dec_blk_nums:
            self.ups.append(
                nn.Sequential(
                    nn.Conv2d(chan, chan * 2, 1, bias=False),
                    nn.PixelShuffle(2)
                )
            )
            chan = chan // 2
            self.decoders.append(
                CustomSequential(
                    *[DBlock(chan, dilations=dilations, extra_depth_wise=extra_depth_wise) for _ in range(num)]
                )
            )
        self.padder_size = 2 ** len(self.encoders)        
        
        # this layer is needed for the computing of the middle loss. It isn't necessary for anything else
        self.side_out = nn.Conv2d(in_channels = width * 2**len(self.encoders), out_channels = img_channel, 
                                kernel_size = 3, stride=1, padding=1)
        
    def forward(self, input, side_loss = False, use_adapter = None):

        _, _, H, W = input.shape

        input = self.check_image_size(input)
        x = self.intro(input)
        
        skips = []
        for encoder, down in zip(self.encoders, self.downs):
            x = encoder(x)
            skips.append(x)
            x = down(x)

        # we apply the encoder transforms
        x_light = self.middle_blks_enc(x)
        
        if side_loss:
            out_side = self.side_out(x_light)
        # apply the decoder transforms
        x = self.middle_blks_dec(x_light)
        x = x + x_light

        for decoder, up, skip in zip(self.decoders, self.ups, skips[::-1]):
            x = up(x)
            x = x + skip
            x = decoder(x)

        x = self.ending(x)
        x = x + input
        out = x[:, :, :H, :W] # we recover the original size of the image
        if side_loss:
            return out_side, out
        else:        
            return out

    def check_image_size(self, x):
        _, _, h, w = x.size()
        mod_pad_h = (self.padder_size - h % self.padder_size) % self.padder_size
        mod_pad_w = (self.padder_size - w % self.padder_size) % self.padder_size
        x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), value = 0)
        return x      

if __name__ == '__main__':
    
    img_channel = 3
    width = 64
    
    enc_blks = [1, 2, 3]
    middle_blk_num_enc = 2
    middle_blk_num_dec = 2
    dec_blks = [3, 1, 1]
    residual_layers = None
    dilations = [1, 4, 9]
    extra_depth_wise = True
    
    net = DarkIR(img_channel=img_channel, 
                  width=width, 
                  middle_blk_num_enc=middle_blk_num_enc,
                  middle_blk_num_dec= middle_blk_num_dec,
                  enc_blk_nums=enc_blks, 
                  dec_blk_nums=dec_blks,
                  dilations = dilations,
                  extra_depth_wise = extra_depth_wise)
    
    new_state_dict = net.state_dict()

    inp_shape = (3, 256, 256)

    net.load_state_dict(new_state_dict)

    from ptflops import get_model_complexity_info

    macs, params = get_model_complexity_info(net, inp_shape, verbose=False, print_per_layer_stat=False)

    print(macs, params)    
    
    weights = net.state_dict()
    adapter_weights = {k: v for k, v in weights.items() if 'adapter' not in k}