File size: 7,372 Bytes
231edce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import numpy as np
import torch
import torch.nn as nn

from transformers.models.distilbert.modeling_distilbert import Transformer as T
from .pooling import create_pool1d_layer


class Config:

    def __init__(self, **kwargs):
        for k, v in kwargs.items():
            setattr(self, k, v)


class Transformer(nn.Module):
    """
    If predict_sequence is True, then the model will predict an output
    for each element in the sequence. If False, then the model will
    predict a single output for the sequence. 
    
    e.g., classifying each image in a CT scan vs the entire CT scan
    """
    def __init__(self,
                 num_classes,
                 embedding_dim=512,
                 hidden_dim=1024,
                 n_layers=4,
                 n_heads=16,
                 dropout=0.2,
                 attention_dropout=0.1,
                 output_attentions=False,
                 activation='gelu',
                 output_hidden_states=False,
                 chunk_size_feed_forward=0,
                 predict_sequence=True,
                 pool=None
    ):
        super().__init__()
        config = Config(**{
                'dim': embedding_dim,
                'hidden_dim': hidden_dim,
                'n_layers': n_layers,
                'n_heads': n_heads,
                'dropout': dropout,
                'attention_dropout': attention_dropout,
                'output_attentions': output_attentions,
                'activation': activation,
                'output_hidden_states': output_hidden_states,
                'chunk_size_feed_forward': chunk_size_feed_forward
            })

        self.transformer = T(config)
        self.predict_sequence = predict_sequence

        if not predict_sequence:
            if isinstance(pool, str):
                self.pool_layer = create_pool1d_layer(pool)
                if pool == "catavgmax":
                    embedding_dim *= 2
            else:
                self.pool_layer = nn.Identity()

        self.classifier = nn.Linear(embedding_dim, num_classes)

    def extract_features(self, x):
        x, mask = x
        x = self.transformer(x, attn_mask=mask, head_mask=[None]*x.size(1))
        x = x[0]

        if not self.predict_sequence:
            if isinstance(self.pool_layer, nn.Identity):
                # Just take the last vector in the sequence
                x = x[:, 0]
            else:
                x = self.pool_layer(x.transpose(-1, -2))

        return x

    def classify(self, x):

        if not self.predict_sequence:
            if isinstance(self.pool_layer, nn.Identity):
                # Just take the last vector in the sequence
                x = x[:, 0]
            else:
                x = self.pool_layer(x.transpose(-1, -2))

        out = self.classifier(x)

        if self.classifier.out_features == 1:
            return out[..., 0]
        else:
            return out

    def forward_tr(self, x, mask):
        output = self.transformer(x, attn_mask=mask, head_mask=[None]*x.size(1))
        return self.classify(output[0])

    def forward(self, x):
        x, mask = x
        return self.forward_tr(x, mask)


class DualTransformer(nn.Module):
    """
    Essentially the same as above except predicts both sequence and study labels.
    """
    def __init__(self,
                 num_classes,
                 embedding_dim=512,
                 hidden_dim=1024,
                 n_layers=4,
                 n_heads=16,
                 dropout=0.2,
                 attention_dropout=0.1,
                 output_attentions=False,
                 activation='gelu',
                 output_hidden_states=False,
                 chunk_size_feed_forward=0,
                 pool=None
    ):
        super().__init__()
        config = Config(**{
                'dim': embedding_dim,
                'hidden_dim': hidden_dim,
                'n_layers': n_layers,
                'n_heads': n_heads,
                'dropout': dropout,
                'attention_dropout': attention_dropout,
                'output_attentions': output_attentions,
                'activation': activation,
                'output_hidden_states': output_hidden_states,
                'chunk_size_feed_forward': chunk_size_feed_forward
            })

        self.transformer = T(config)

        self.pool_layer = nn.Identity()

        self.classifier1 = nn.Linear(embedding_dim, num_classes)
        self.classifier2 = nn.Linear(embedding_dim, num_classes)

    def classify(self, x):

        if isinstance(self.pool_layer, nn.Identity):
            # Just take the last vector in the sequence
            x_summ = x[:, 0]
        else:
            x_summ = self.pool_layer(x.transpose(-1, -2))

        # Element-wise labels
        out1 = self.classifier1(x)[:, :, 0]
        # Single label for whole sequence
        out2 = self.classifier2(x_summ)
        out = torch.cat([out1, out2], dim=1)
        return out

    def forward_tr(self, x, mask):
        output = self.transformer(x, attn_mask=mask, head_mask=[None]*x.size(1))
        return self.classify(output[0])

    def forward(self, x):
        x, mask = x
        return self.forward_tr(x, mask)


class DualTransformerV2(nn.Module):
    """
    More complicated variant of DualTransformer.
    Returns tuple of: (element-wise prediction, sequence prediction)
    """
    def __init__(self,
                 num_seq_classes,
                 num_classes,
                 embedding_dim=512,
                 hidden_dim=1024,
                 n_layers=4,
                 n_heads=16,
                 dropout=0.2,
                 attention_dropout=0.1,
                 output_attentions=False,
                 activation='gelu',
                 output_hidden_states=False,
                 chunk_size_feed_forward=0,
                 pool=None
        ):
        super().__init__()
        config = Config(**{
                'dim': embedding_dim,
                'hidden_dim': hidden_dim,
                'n_layers': n_layers,
                'n_heads': n_heads,
                'dropout': dropout,
                'attention_dropout': attention_dropout,
                'output_attentions': output_attentions,
                'activation': activation,
                'output_hidden_states': output_hidden_states,
                'chunk_size_feed_forward': chunk_size_feed_forward
            })

        self.transformer = T(config)

        self.pool_layer = nn.Identity()

        self.classifier1 = nn.Linear(embedding_dim, num_seq_classes)
        self.classifier2 = nn.Linear(embedding_dim, num_classes)

    def classify(self, x):

        if isinstance(self.pool_layer, nn.Identity):
            # Just take the last vector in the sequence
            x_summ = x[:, 0]
        else:
            x_summ = self.pool_layer(x.transpose(-1, -2))

        # Element-wise labels
        out1 = self.classifier1(x)
        if self.classifier1.out_features == 1:
            out1 = out1[:, :, 0]
        # Single label for whole sequence
        out2 = self.classifier2(x_summ)
        return out1, out2

    def forward_tr(self, x, mask):
        output = self.transformer(x, attn_mask=mask, head_mask=[None]*x.size(1))
        return self.classify(output[0])

    def forward(self, x):
        x, mask = x
        return self.forward_tr(x, mask)