File size: 7,289 Bytes
18793b8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
""" DropBlock, DropPath

PyTorch implementations of DropBlock and DropPath (Stochastic Depth) regularization layers.

Papers:
DropBlock: A regularization method for convolutional networks (https://arxiv.org/abs/1810.12890)

Deep Networks with Stochastic Depth (https://arxiv.org/abs/1603.09382)

Code:
DropBlock impl inspired by two Tensorflow impl that I liked:
 - https://github.com/tensorflow/tpu/blob/master/models/official/resnet/resnet_model.py#L74
 - https://github.com/clovaai/assembled-cnn/blob/master/nets/blocks.py

Hacked together by / Copyright 2020 Ross Wightman
"""
import torch
import torch.nn as nn
import torch.nn.functional as F


def drop_block_2d(
    x,
    drop_prob: float = 0.1,
    block_size: int = 7,
    gamma_scale: float = 1.0,
    with_noise: bool = False,
    inplace: bool = False,
    batchwise: bool = False,
):
    """DropBlock. See https://arxiv.org/pdf/1810.12890.pdf

    DropBlock with an experimental gaussian noise option. This layer has been tested on a few training
    runs with success, but needs further validation and possibly optimization for lower runtime impact.
    """
    _, C, H, W = x.shape
    total_size = W * H
    clipped_block_size = min(block_size, min(W, H))
    # seed_drop_rate, the gamma parameter
    gamma = (
        gamma_scale
        * drop_prob
        * total_size
        / clipped_block_size**2
        / ((W - block_size + 1) * (H - block_size + 1))
    )

    # Forces the block to be inside the feature map.
    w_i, h_i = torch.meshgrid(
        torch.arange(W).to(x.device), torch.arange(H).to(x.device)
    )
    valid_block = (
        (w_i >= clipped_block_size // 2) & (w_i < W - (clipped_block_size - 1) // 2)
    ) & ((h_i >= clipped_block_size // 2) & (h_i < H - (clipped_block_size - 1) // 2))
    valid_block = torch.reshape(valid_block, (1, 1, H, W)).to(dtype=x.dtype)

    if batchwise:
        # one mask for whole batch, quite a bit faster
        uniform_noise = torch.rand((1, C, H, W), dtype=x.dtype, device=x.device)
    else:
        uniform_noise = torch.rand_like(x)
    block_mask = ((2 - gamma - valid_block + uniform_noise) >= 1).to(dtype=x.dtype)
    block_mask = -F.max_pool2d(
        -block_mask,
        kernel_size=clipped_block_size,  # block_size,
        stride=1,
        padding=clipped_block_size // 2,
    )

    if with_noise:
        normal_noise = (
            torch.randn((1, C, H, W), dtype=x.dtype, device=x.device)
            if batchwise
            else torch.randn_like(x)
        )
        if inplace:
            x.mul_(block_mask).add_(normal_noise * (1 - block_mask))
        else:
            x = x * block_mask + normal_noise * (1 - block_mask)
    else:
        normalize_scale = (
            block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-7)
        ).to(x.dtype)
        if inplace:
            x.mul_(block_mask * normalize_scale)
        else:
            x = x * block_mask * normalize_scale
    return x


def drop_block_fast_2d(
    x: torch.Tensor,
    drop_prob: float = 0.1,
    block_size: int = 7,
    gamma_scale: float = 1.0,
    with_noise: bool = False,
    inplace: bool = False,
):
    """DropBlock. See https://arxiv.org/pdf/1810.12890.pdf

    DropBlock with an experimental gaussian noise option. Simplied from above without concern for valid
    block mask at edges.
    """
    _, _, H, W = x.shape
    total_size = W * H
    clipped_block_size = min(block_size, min(W, H))
    gamma = (
        gamma_scale
        * drop_prob
        * total_size
        / clipped_block_size**2
        / ((W - block_size + 1) * (H - block_size + 1))
    )

    block_mask = torch.empty_like(x).bernoulli_(gamma)
    block_mask = F.max_pool2d(
        block_mask.to(x.dtype),
        kernel_size=clipped_block_size,
        stride=1,
        padding=clipped_block_size // 2,
    )

    if with_noise:
        normal_noise = torch.empty_like(x).normal_()
        if inplace:
            x.mul_(1.0 - block_mask).add_(normal_noise * block_mask)
        else:
            x = x * (1.0 - block_mask) + normal_noise * block_mask
    else:
        block_mask = 1 - block_mask
        normalize_scale = (
            block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-6)
        ).to(dtype=x.dtype)
        if inplace:
            x.mul_(block_mask * normalize_scale)
        else:
            x = x * block_mask * normalize_scale
    return x


class DropBlock2d(nn.Module):
    """DropBlock. See https://arxiv.org/pdf/1810.12890.pdf"""

    def __init__(
        self,
        drop_prob: float = 0.1,
        block_size: int = 7,
        gamma_scale: float = 1.0,
        with_noise: bool = False,
        inplace: bool = False,
        batchwise: bool = False,
        fast: bool = True,
    ):
        super(DropBlock2d, self).__init__()
        self.drop_prob = drop_prob
        self.gamma_scale = gamma_scale
        self.block_size = block_size
        self.with_noise = with_noise
        self.inplace = inplace
        self.batchwise = batchwise
        self.fast = fast  # FIXME finish comparisons of fast vs not

    def forward(self, x):
        if not self.training or not self.drop_prob:
            return x
        if self.fast:
            return drop_block_fast_2d(
                x,
                self.drop_prob,
                self.block_size,
                self.gamma_scale,
                self.with_noise,
                self.inplace,
            )
        else:
            return drop_block_2d(
                x,
                self.drop_prob,
                self.block_size,
                self.gamma_scale,
                self.with_noise,
                self.inplace,
                self.batchwise,
            )


def drop_path(
    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True
):
    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
    'survival rate' as the argument.

    """
    if drop_prob == 0.0 or not training:
        return x
    keep_prob = 1 - drop_prob
    shape = (x.shape[0],) + (1,) * (
        x.ndim - 1
    )  # work with diff dim tensors, not just 2D ConvNets
    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
    if keep_prob > 0.0 and scale_by_keep:
        random_tensor.div_(keep_prob)
    return x * random_tensor


class DropPath(nn.Module):
    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks)."""

    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):
        super(DropPath, self).__init__()
        self.drop_prob = drop_prob
        self.scale_by_keep = scale_by_keep

    def forward(self, x):
        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)

    def extra_repr(self):
        return f"drop_prob={round(self.drop_prob,3):0.3f}"