/
__init__.py
327 lines (258 loc) · 13.2 KB
/
__init__.py
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
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
"""
---
title: Deep Residual Learning for Image Recognition (ResNet)
summary: >
A PyTorch implementation/tutorial of Deep Residual Learning for Image Recognition (ResNet).
---
# Deep Residual Learning for Image Recognition (ResNet)
This is a [PyTorch](https://pytorch.org) implementation of the paper
[Deep Residual Learning for Image Recognition](https://papers.labml.ai/paper/1512.03385).
ResNets train layers as residual functions to overcome the
*degradation problem*.
The degradation problem is the accuracy of deep neural networks degrading when
the number of layers becomes very high.
The accuracy increases as the number of layers increase, then saturates,
and then starts to degrade.
The paper argues that deeper models should perform at least as well as shallower
models because the extra layers can just learn to perform an identity mapping.
## Residual Learning
If $\mathcal{H}(x)$ is the mapping that needs to be learned by a few layers,
they train the residual function
$$\mathcal{F}(x) = \mathcal{H}(x) - x$$
instead. And the original function becomes $\mathcal{F}(x) + x$.
In this case, learning identity mapping for $\mathcal{H}(x)$ is
equivalent to learning $\mathcal{F}(x)$ to be $0$, which is easier to
learn.
In the parameterized form this can be written as,
$$\mathcal{F}(x, \{W_i\}) + x$$
and when the feature map sizes of $\mathcal{F}(x, {W_i})$ and $x$ are different
the paper suggests doing a linear projection, with learned weights $W_s$.
$$\mathcal{F}(x, \{W_i\}) + W_s x$$
Paper experimented with zero padding instead of linear projections and found linear projections
to work better. Also when the feature map sizes match they found identity mapping
to be better than linear projections.
$\mathcal{F}$ should have more than one layer, otherwise the sum $\mathcal{F}(x, \{W_i\}) + W_s x$
also won't have non-linearities and will be like a linear layer.
Here is [the training code](experiment.html) for training a ResNet on CIFAR-10.
[](https://app.labml.ai/run/fc5ad600e4af11ebbafd23b8665193c1)
"""
from typing import List, Optional
import torch
from torch import nn
from labml_helpers.module import Module
class ShortcutProjection(Module):
"""
## Linear projections for shortcut connection
This does the $W_s x$ projection described above.
"""
def __init__(self, in_channels: int, out_channels: int, stride: int):
"""
* `in_channels` is the number of channels in $x$
* `out_channels` is the number of channels in $\mathcal{F}(x, \{W_i\})$
* `stride` is the stride length in the convolution operation for $F$.
We do the same stride on the shortcut connection, to match the feature-map size.
"""
super().__init__()
# Convolution layer for linear projection $W_s x$
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride)
# Paper suggests adding batch normalization after each convolution operation
self.bn = nn.BatchNorm2d(out_channels)
def forward(self, x: torch.Tensor):
# Convolution and batch normalization
return self.bn(self.conv(x))
class ResidualBlock(Module):
"""
<a id="residual_block"></a>
## Residual Block
This implements the residual block described in the paper.
It has two $3 \times 3$ convolution layers.
The first convolution layer maps from `in_channels` to `out_channels`,
where the `out_channels` is higher than `in_channels` when we reduce the
feature map size with a stride length greater than $1$.
The second convolution layer maps from `out_channels` to `out_channels` and
always has a stride length of 1.
Both convolution layers are followed by batch normalization.
"""
def __init__(self, in_channels: int, out_channels: int, stride: int):
"""
* `in_channels` is the number of channels in $x$
* `out_channels` is the number of output channels
* `stride` is the stride length in the convolution operation.
"""
super().__init__()
# First $3 \times 3$ convolution layer, this maps to `out_channels`
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1)
# Batch normalization after the first convolution
self.bn1 = nn.BatchNorm2d(out_channels)
# First activation function (ReLU)
self.act1 = nn.ReLU()
# Second $3 \times 3$ convolution layer
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
# Batch normalization after the second convolution
self.bn2 = nn.BatchNorm2d(out_channels)
# Shortcut connection should be a projection if the stride length is not $1$
# of if the number of channels change
if stride != 1 or in_channels != out_channels:
# Projection $W_s x$
self.shortcut = ShortcutProjection(in_channels, out_channels, stride)
else:
# Identity $x$
self.shortcut = nn.Identity()
# Second activation function (ReLU) (after adding the shortcut)
self.act2 = nn.ReLU()
def forward(self, x: torch.Tensor):
"""
* `x` is the input of shape `[batch_size, in_channels, height, width]`
"""
# Get the shortcut connection
shortcut = self.shortcut(x)
# First convolution and activation
x = self.act1(self.bn1(self.conv1(x)))
# Second convolution
x = self.bn2(self.conv2(x))
# Activation function after adding the shortcut
return self.act2(x + shortcut)
class BottleneckResidualBlock(Module):
"""
<a id="bottleneck_residual_block"></a>
## Bottleneck Residual Block
This implements the bottleneck block described in the paper.
It has $1 \times 1$, $3 \times 3$, and $1 \times 1$ convolution layers.
The first convolution layer maps from `in_channels` to `bottleneck_channels` with a $1x1$
convolution,
where the `bottleneck_channels` is lower than `in_channels`.
The second $3x3$ convolution layer maps from `bottleneck_channels` to `bottleneck_channels`.
This can have a stride length greater than $1$ when we want to compress the
feature map size.
The third, final $1x1$ convolution layer maps to `out_channels`.
`out_channels` is higher than `in_channels` if the stride length is greater than $1$;
otherwise, $out_channels$ is equal to `in_channels`.
`bottleneck_channels` is less than `in_channels` and the $3x3$ convolution is performed
on this shrunk space (hence the bottleneck). The two $1x1$ convolution decreases and increases
the number of channels.
"""
def __init__(self, in_channels: int, bottleneck_channels: int, out_channels: int, stride: int):
"""
* `in_channels` is the number of channels in $x$
* `bottleneck_channels` is the number of channels for the $3x3$ convlution
* `out_channels` is the number of output channels
* `stride` is the stride length in the $3x3$ convolution operation.
"""
super().__init__()
# First $1 \times 1$ convolution layer, this maps to `bottleneck_channels`
self.conv1 = nn.Conv2d(in_channels, bottleneck_channels, kernel_size=1, stride=1)
# Batch normalization after the first convolution
self.bn1 = nn.BatchNorm2d(bottleneck_channels)
# First activation function (ReLU)
self.act1 = nn.ReLU()
# Second $3 \times 3$ convolution layer
self.conv2 = nn.Conv2d(bottleneck_channels, bottleneck_channels, kernel_size=3, stride=stride, padding=1)
# Batch normalization after the second convolution
self.bn2 = nn.BatchNorm2d(bottleneck_channels)
# Second activation function (ReLU)
self.act2 = nn.ReLU()
# Third $1 \times 1$ convolution layer, this maps to `out_channels`.
self.conv3 = nn.Conv2d(bottleneck_channels, out_channels, kernel_size=1, stride=1)
# Batch normalization after the second convolution
self.bn3 = nn.BatchNorm2d(out_channels)
# Shortcut connection should be a projection if the stride length is not $1$
# of if the number of channels change
if stride != 1 or in_channels != out_channels:
# Projection $W_s x$
self.shortcut = ShortcutProjection(in_channels, out_channels, stride)
else:
# Identity $x$
self.shortcut = nn.Identity()
# Second activation function (ReLU) (after adding the shortcut)
self.act3 = nn.ReLU()
def forward(self, x: torch.Tensor):
"""
* `x` is the input of shape `[batch_size, in_channels, height, width]`
"""
# Get the shortcut connection
shortcut = self.shortcut(x)
# First convolution and activation
x = self.act1(self.bn1(self.conv1(x)))
# Second convolution and activation
x = self.act2(self.bn2(self.conv2(x)))
# Third convolution
x = self.bn3(self.conv3(x))
# Activation function after adding the shortcut
return self.act3(x + shortcut)
class ResNetBase(Module):
"""
## ResNet Model
This is a the base of the resnet model without
the final linear layer and softmax for classification.
The resnet is made of stacked [residual blocks](#residual_block) or
[bottleneck residual blocks](#bottleneck_residual_block).
The feature map size is halved after a few blocks with a block of stride length $2$.
The number of channels is increased when the feature map size is reduced.
Finally the feature map is average pooled to get a vector representation.
"""
def __init__(self, n_blocks: List[int], n_channels: List[int],
bottlenecks: Optional[List[int]] = None,
img_channels: int = 3, first_kernel_size: int = 7):
"""
* `n_blocks` is a list of of number of blocks for each feature map size.
* `n_channels` is the number of channels for each feature map size.
* `bottlenecks` is the number of channels the bottlenecks.
If this is `None`, [residual blocks](#residual_block) are used.
* `img_channels` is the number of channels in the input.
* `first_kernel_size` is the kernel size of the initial convolution layer
"""
super().__init__()
# Number of blocks and number of channels for each feature map size
assert len(n_blocks) == len(n_channels)
# If [bottleneck residual blocks](#bottleneck_residual_block) are used,
# the number of channels in bottlenecks should be provided for each feature map size
assert bottlenecks is None or len(bottlenecks) == len(n_channels)
# Initial convolution layer maps from `img_channels` to number of channels in the first
# residual block (`n_channels[0]`)
self.conv = nn.Conv2d(img_channels, n_channels[0],
kernel_size=first_kernel_size, stride=2, padding=first_kernel_size // 2)
# Batch norm after initial convolution
self.bn = nn.BatchNorm2d(n_channels[0])
# List of blocks
blocks = []
# Number of channels from previous layer (or block)
prev_channels = n_channels[0]
# Loop through each feature map size
for i, channels in enumerate(n_channels):
# The first block for the new feature map size, will have a stride length of $2$
# except fro the very first block
stride = 2 if len(blocks) == 0 else 1
if bottlenecks is None:
# [residual blocks](#residual_block) that maps from `prev_channels` to `channels`
blocks.append(ResidualBlock(prev_channels, channels, stride=stride))
else:
# [bottleneck residual blocks](#bottleneck_residual_block)
# that maps from `prev_channels` to `channels`
blocks.append(BottleneckResidualBlock(prev_channels, bottlenecks[i], channels,
stride=stride))
# Change the number of channels
prev_channels = channels
# Add rest of the blocks - no change in feature map size or channels
for _ in range(n_blocks[i] - 1):
if bottlenecks is None:
# [residual blocks](#residual_block)
blocks.append(ResidualBlock(channels, channels, stride=1))
else:
# [bottleneck residual blocks](#bottleneck_residual_block)
blocks.append(BottleneckResidualBlock(channels, bottlenecks[i], channels, stride=1))
# Stack the blocks
self.blocks = nn.Sequential(*blocks)
def forward(self, x: torch.Tensor):
"""
* `x` has shape `[batch_size, img_channels, height, width]`
"""
# Initial convolution and batch normalization
x = self.bn(self.conv(x))
# Residual (or bottleneck) blocks
x = self.blocks(x)
# Change `x` from shape `[batch_size, channels, h, w]` to `[batch_size, channels, h * w]`
x = x.view(x.shape[0], x.shape[1], -1)
# Global average pooling
return x.mean(dim=-1)