File size: 9,474 Bytes
cd6775d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2020-present, Facebook, Inc.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
#
import torch
from torch import nn

from .utils import get_slices


def mlp(sizes, bias=True, batchnorm=True, groups=1):
    """
    Generate a feedforward neural network.
    """
    assert len(sizes) >= 2
    pairs = [(sizes[i], sizes[i + 1]) for i in range(len(sizes) - 1)]
    layers = []

    for i, (dim_in, dim_out) in enumerate(pairs):
        if groups == 1 or i == 0:
            layers.append(nn.Linear(dim_in, groups * dim_out, bias=bias))
        else:
            layers.append(GroupedLinear(groups * dim_in, groups * dim_out, bias=bias, groups=groups))
        if batchnorm:
            layers.append(nn.BatchNorm1d(groups * dim_out))
        if i < len(pairs) - 1:
            layers.append(nn.ReLU())

    return nn.Sequential(*layers)


def convs(channel_sizes, kernel_sizes, bias=True, batchnorm=True, residual=False, groups=1):
    """
    Generate a convolutional neural network.
    """
    assert len(channel_sizes) >= 2
    assert len(channel_sizes) == len(kernel_sizes) + 1
    pairs = [(channel_sizes[i], channel_sizes[i + 1]) for i in range(len(channel_sizes) - 1)]
    layers = []

    for i, (dim_in, dim_out) in enumerate(pairs):
        ks = (kernel_sizes[i], kernel_sizes[i])
        in_group = 1 if i == 0 else groups
        _dim_in = dim_in * in_group
        _dim_out = dim_out * groups
        if not residual:
            layers.append(nn.Conv2d(_dim_in, _dim_out, ks, padding=[k // 2 for k in ks], bias=bias, groups=in_group))
            if batchnorm:
                layers.append(nn.BatchNorm2d(_dim_out))
            if i < len(pairs) - 1:
                layers.append(nn.ReLU())
        else:
            layers.append(BottleneckResidualConv2d(
                _dim_in, _dim_out, ks, bias=bias,
                batchnorm=batchnorm, groups=in_group
            ))
            if i == len(pairs) - 1:
                layers.append(nn.Conv2d(_dim_out, _dim_out, (1, 1), bias=bias))

    return nn.Sequential(*layers)


class GroupedLinear(nn.Module):

    def __init__(self, in_features, out_features, bias=True, groups=1):

        super().__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.groups = groups
        self.bias = bias
        assert groups > 1

        self.layer = nn.Conv1d(in_features, out_features, bias=bias, kernel_size=1, groups=groups)

    def forward(self, input):
        assert input.dim() == 2 and input.size(1) == self.in_features
        return self.layer(input.unsqueeze(2)).squeeze(2)

    def extra_repr(self):
        return 'in_features={}, out_features={}, groups={}, bias={}'.format(
            self.in_features, self.out_features, self.groups, self.bias is not None
        )


class BottleneckResidualConv2d(nn.Module):

    def __init__(self, input_channels, output_channels, kernel_size, bias=True, batchnorm=True, groups=1):

        super().__init__()
        hidden_channels = min(input_channels, output_channels)
        assert all(k % 2 == 1 for k in kernel_size)

        self.conv1 = nn.Conv2d(input_channels, hidden_channels, kernel_size, padding=[k // 2 for k in kernel_size], bias=bias, groups=groups)
        self.conv2 = nn.Conv2d(hidden_channels, output_channels, kernel_size, padding=[k // 2 for k in kernel_size], bias=bias, groups=groups)
        self.act = nn.ReLU()

        self.batchnorm = batchnorm
        if self.batchnorm:
            self.bn1 = nn.BatchNorm2d(hidden_channels)
            self.bn2 = nn.BatchNorm2d(output_channels)

        if input_channels == output_channels:
            self.residual = nn.Sequential()
        else:
            self.residual = nn.Conv2d(input_channels, output_channels, (1, 1), bias=False, groups=groups)

    def forward(self, input):
        x = self.conv1(input)
        x = self.bn1(x) if self.batchnorm else x
        x = self.act(x)
        x = self.conv2(x)
        x = self.bn2(x) if self.batchnorm else x
        x = self.act(x + self.residual(input))
        return x


class QueryIdentity(nn.Module):

    def __init__(self, input_dim, heads, shuffle_hidden):
        super().__init__()
        self.input_dim = input_dim
        self.heads = heads
        self.shuffle_query = shuffle_hidden
        assert shuffle_hidden is False or heads > 1
        assert shuffle_hidden is False or self.input_dim % (2 ** self.heads) == 0
        if shuffle_hidden:
            self.slices = {head_id: get_slices(input_dim, head_id) for head_id in range(heads)}

    def forward(self, input):
        """
        Generate queries from hidden states by either
        repeating them or creating some shuffled version.
        """
        assert input.shape[-1] == self.input_dim
        input = input.contiguous().view(-1, self.input_dim) if input.dim() > 2 else input
        bs = len(input)

        if self.heads == 1:
            query = input

        elif not self.shuffle_query:
            query = input.unsqueeze(1).repeat(1, self.heads, 1)
            query = query.view(bs * self.heads, self.input_dim)

        else:
            query = torch.cat([
                input[:, a:b]
                for head_id in range(self.heads)
                for a, b in self.slices[head_id]
            ], 1).view(bs * self.heads, self.input_dim)

        assert query.shape == (bs * self.heads, self.input_dim)
        return query


class QueryMLP(nn.Module):

    def __init__(
        self, input_dim, heads, k_dim, product_quantization, multi_query_net,
        sizes, bias=True, batchnorm=True, grouped_conv=False
    ):
        super().__init__()
        self.input_dim = input_dim
        self.heads = heads
        self.k_dim = k_dim
        self.sizes = sizes
        self.grouped_conv = grouped_conv
        assert not multi_query_net or product_quantization or heads >= 2
        assert sizes[0] == input_dim
        assert sizes[-1] == (k_dim // 2) if multi_query_net else (heads * k_dim)
        assert self.grouped_conv is False or len(sizes) > 2

        # number of required MLPs
        self.groups = (2 * heads) if multi_query_net else 1

        # MLPs
        if self.grouped_conv:
            self.query_mlps = mlp(sizes, bias=bias, batchnorm=batchnorm, groups=self.groups)
        elif len(self.sizes) == 2:
            sizes_ = list(sizes)
            sizes_[-1] = sizes_[-1] * self.groups
            self.query_mlps = mlp(sizes_, bias=bias, batchnorm=batchnorm, groups=1)
        else:
            self.query_mlps = nn.ModuleList([
                mlp(sizes, bias=bias, batchnorm=batchnorm, groups=1)
                for _ in range(self.groups)
            ])

    def forward(self, input):
        """
        Compute queries using either grouped 1D convolutions or ModuleList + concat.
        """
        assert input.shape[-1] == self.input_dim
        input = input.contiguous().view(-1, self.input_dim) if input.dim() > 2 else input
        bs = len(input)

        if self.grouped_conv or len(self.sizes) == 2:
            query = self.query_mlps(input)
        else:
            outputs = [m(input) for m in self.query_mlps]
            query = torch.cat(outputs, 1) if len(outputs) > 1 else outputs[0]

        assert query.shape == (bs, self.heads * self.k_dim)
        return query.view(bs * self.heads, self.k_dim)


class QueryConv(nn.Module):

    def __init__(
        self, input_dim, heads, k_dim, product_quantization, multi_query_net,
        sizes, kernel_sizes, bias=True, batchnorm=True,
        residual=False, grouped_conv=False
    ):
        super().__init__()
        self.input_dim = input_dim
        self.heads = heads
        self.k_dim = k_dim
        self.sizes = sizes
        self.grouped_conv = grouped_conv
        assert not multi_query_net or product_quantization or heads >= 2
        assert sizes[0] == input_dim
        assert sizes[-1] == (k_dim // 2) if multi_query_net else (heads * k_dim)
        assert self.grouped_conv is False or len(sizes) > 2
        assert len(sizes) == len(kernel_sizes) + 1 >= 2 and all(ks % 2 == 1 for ks in kernel_sizes)

        # number of required CNNs
        self.groups = (2 * heads) if multi_query_net else 1

        # CNNs
        if self.grouped_conv:
            self.query_convs = convs(sizes, kernel_sizes, bias=bias, batchnorm=batchnorm, residual=residual, groups=self.groups)
        elif len(self.sizes) == 2:
            sizes_ = list(sizes)
            sizes_[-1] = sizes_[-1] * self.groups
            self.query_convs = convs(sizes_, kernel_sizes, bias=bias, batchnorm=batchnorm, residual=residual, groups=1)
        else:
            self.query_convs = nn.ModuleList([
                convs(sizes, kernel_sizes, bias=bias, batchnorm=batchnorm, residual=residual, groups=1)
                for _ in range(self.groups)
            ])

    def forward(self, input):

        bs, nf, h, w = input.shape
        assert nf == self.input_dim

        if self.grouped_conv or len(self.sizes) == 2:
            query = self.query_convs(input)
        else:
            outputs = [m(input) for m in self.query_convs]
            query = torch.cat(outputs, 1) if len(outputs) > 1 else outputs[0]

        assert query.shape == (bs, self.heads * self.k_dim, h, w)
        query = query.transpose(1, 3).contiguous().view(bs * w * h * self.heads, self.k_dim)
        return query