LDN/layers.py at master · ivankreso/LDN · GitHub

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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as cp

checkpoint = lambda func, *inputs: cp.checkpoint(func, *inputs, preserve_rng_state=False)


class SpatialPyramidPooling(nn.Module):
    def __init__(self, conv_class, upsample_func, num_maps_in, bt_size=512, level_size=128,
                 out_size=256, grids=[6,3,2,1], square_grid=False):
        super(SpatialPyramidPooling, self).__init__()
        self.upsample = upsample_func
        self.grids = grids
        self.num_levels = len(grids)
        self.square_grid = square_grid
        self.spp = nn.Sequential()
        self.spp.add_module('spp_bn', conv_class(num_maps_in, bt_size, k=1))
        num_features = bt_size
        final_size = num_features
        for i in range(self.num_levels):
            final_size += level_size
            self.spp.add_module('spp'+str(i), conv_class(num_features, level_size, k=1))
        self.spp.add_module('spp_fuse', conv_class(final_size, out_size, k=1))


    def forward(self, x):
        levels = []
        target_size = x.size()[2:4]

        ar = target_size[1] / target_size[0]

        x = self.spp[0].forward(x)
        levels.append(x)
        num = len(self.spp) - 1

        for i in range(1, num):
            if not self.square_grid:
                grid_size = (self.grids[i-1], max(1, round(ar*self.grids[i-1])))
                x_pooled = F.adaptive_avg_pool2d(x, grid_size)
            else:
                x_pooled = F.adaptive_avg_pool2d(x, self.grids[i-1])
            level = self.spp[i].forward(x_pooled)
            level = self.upsample(level, target_size)
            levels.append(level)

        x = torch.cat(levels, 1)
        return self.spp[-1].forward(x)


class Upsample(nn.Module):
    def __init__(self, conv_class, upsample_func, num_maps_in, skip_maps_in, num_maps_out, k,
                 produce_aux=False, num_classes=0, dws_conv=False, checkpointing=False):
        super(Upsample, self).__init__()
        print('Upsample layer: in =', num_maps_in, ', skip =', skip_maps_in, ' out =', num_maps_out)
        self.upsample_func = upsample_func
        self.bottleneck = conv_class(skip_maps_in, num_maps_in, k=1)
        self.produce_aux = produce_aux
        self.has_blend_conv = num_maps_out > 0
        self.num_maps_out = num_maps_in
        self.checkpointing = checkpointing
        if produce_aux:
            self.aux_logits = conv_class(num_maps_in, num_classes, k=1, output_conv=True)
        if self.has_blend_conv:
            self.num_maps_out = num_maps_out
            bt_maps = 128
            self.blend_bt = None
            if not dws_conv and k >=3 and num_maps_in > bt_maps:
                print(f'Bottleneck before 3x3: {num_maps_in} -> {bt_maps}')
                self.blend_bt = conv_class(num_maps_in, bt_maps, k=1)
                num_maps_in = bt_maps
            self.blend_conv = conv_class(num_maps_in, num_maps_out, k=k)
        self.forward_func = self._get_forward_func()

    def _get_forward_func(self):
        def func(*inputs):
            x, skip = inputs
            skip = self.bottleneck(skip)
            skip_size = skip.size()[2:4]
            if self.produce_aux:
                aux = self.aux_logits(x)
            x = self.upsample_func(x, skip_size)
            x += skip
            if self.has_blend_conv:
                if self.blend_bt is not None:
                    x = self.blend_bt(x)
                x = self.blend_conv(x)
            if self.produce_aux:
                return x, aux
            return x
        return func

    def forward(self, bottom, skip):
        if self.checkpointing and self.training:
            return checkpoint(self.forward_func, *[bottom, skip])
        else:
            return self.forward_func(*[bottom, skip])


class UpsampleResidual(nn.Module):
    def __init__(self, conv_class, upsample_func, num_maps_in, skip_maps_in, num_maps_out, k,
                 produce_aux=False, num_classes=0, dws_conv=False):
        super(UpsampleResidual, self).__init__()
        print('Upsample layer: in =', num_maps_in, ', skip =', skip_maps_in, ' out =', num_maps_out)
        self.upsample_func = upsample_func
        self.bottleneck = conv_class(skip_maps_in, num_maps_in, k=1)
        self.produce_aux = produce_aux
        self.has_blend_conv = num_maps_out > 0
        self.num_maps_out = num_maps_in
        if num_maps_out != num_maps_in:
            self.skip_bt = conv_class(num_maps_in, num_maps_out, k=1)
            print(f'Bottleneck on residual: {num_maps_in} -> {num_maps_out}')
        else:
            self.skip_bt = None

        if produce_aux:
            self.aux_logits = conv_class(num_maps_in, num_classes, k=1, output_conv=True)

        if self.has_blend_conv:
            self.num_maps_out = num_maps_out
            bt_maps = 128
            self.blend_bt = None
            if not dws_conv and k >=3 and num_maps_in > bt_maps:
                print(f'Bottleneck before 3x3: {num_maps_in} -> {bt_maps}')
                self.blend_bt = conv_class(num_maps_in, bt_maps, k=1)
                num_maps_in = bt_maps
            self.blend_conv = conv_class(num_maps_in, num_maps_out, k=k)


    def forward(self, bottom, skip):
        skip = self.bottleneck(skip)
        skip_size = skip.size()[2:4]
        if self.produce_aux:
            aux = self.aux_logits(bottom)

        bottom = self.upsample_func(bottom, skip_size)
        x = skip
        x += bottom

        if self.has_blend_conv:
            if self.blend_bt is not None:
                x = self.blend_bt(x)
            x = self.blend_conv(x)
        if self.skip_bt is not None:
            bottom = self.skip_bt(bottom)
        x += bottom
        if self.produce_aux:
            return x, aux
        return x