/
model.py
305 lines (249 loc) · 12.9 KB
/
model.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
from transformers import BertTokenizer, BertTokenizerFast, BertPreTrainedModel, BertModel, BertConfig, AdamW, get_linear_schedule_with_warmup
from transformers import AutoTokenizer, AutoModel, AutoConfig
from torch import nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
import torch
class F1_Loss(nn.Module):
'''Calculate F1 score. Can work with gpu tensors
The original implmentation is written by Michal Haltuf on Kaggle.
Returns
-------
torch.Tensor
`ndim` == 1. epsilon <= val <= 1
Reference
---------
- https://www.kaggle.com/rejpalcz/best-loss-function-for-f1-score-metric
- https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score
- https://discuss.pytorch.org/t/calculating-precision-recall-and-f1-score-in-case-of-multi-label-classification/28265/6
- http://www.ryanzhang.info/python/writing-your-own-loss-function-module-for-pytorch/
'''
def __init__(self, epsilon=1e-7):
super().__init__()
self.epsilon = epsilon
def forward(self, y_pred, y_true, ):
assert y_pred.ndim == 1
assert y_true.ndim == 1
y_true = y_true.type(torch.int64).cuda()
y_true = F.one_hot(y_true, 2).to(torch.float32).cuda()
y_pred = F.sigmoid(y_pred).cuda()
y_pred_New = torch.zeros(y_true.shape,dtype=torch.float32).cuda()
y_pred_New[:, 1] = y_pred
y_pred_New[:, 0] = 1 - y_pred
y_pred = y_pred_New.cuda()
tp = (y_true * y_pred).sum(dim=0).to(torch.float32)
tn = ((1 - y_true) * (1 - y_pred)).sum(dim=0).to(torch.float32)
fp = ((1 - y_true) * y_pred).sum(dim=0).to(torch.float32)
fn = (y_true * (1 - y_pred)).sum(dim=0).to(torch.float32)
precision = tp / (tp + fp + self.epsilon)
recall = tp / (tp + fn + self.epsilon)
f1 = 2 * (precision * recall) / (precision + recall + self.epsilon)
f1 = f1.clamp(min=self.epsilon, max=1 - self.epsilon)
return 1 - f1.mean()
class MultiTaskBertForCovidEntityClassification(BertPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.bert = BertModel(config)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.subtasks = config.subtasks
# We will create a dictionary of classifiers based on the number of subtasks
self.classifiers = {subtask: nn.Linear(config.hidden_size, config.num_labels) for subtask in self.subtasks}
# self.classifier = nn.Linear(config.hidden_size, config.num_labels)
self.init_weights()
def forward(
self,
input_ids,
entity_start_positions, ## TODO check what is entity_start_positions
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
):
outputs = self.bert(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
)
# DEBUG:
# print("BERT model outputs shape", outputs[0].shape, outputs[1].shape)
# print(entity_start_positions[:, 0], entity_start_positions[:, 1])
# OLD CODE:
# pooled_output = outputs[1]
# input [8,68]
# NOTE: outputs[0] has all the hidden dimensions for the entire sequence output[0] [8,68,768]
# We will extract the embeddings indexed with entity_start_positions # TODO why start position embedding output[1] [8,768]
pooled_output = outputs[0][entity_start_positions[:, 0], entity_start_positions[:, 1], :]
pooled_output = self.dropout(pooled_output) ## [batch_size, 768]
# Get logits for each subtaskx
# logits = self.classifier(pooled_output) 10 (#subtask batch_size 8 2 (0,1)]
logits = {subtask: self.classifiers[subtask](pooled_output) for subtask in self.subtasks}
outputs = outputs[2:] + (logits,) # add hidden states and attention if they are here
if labels is not None:
loss_fct = nn.CrossEntropyLoss()
# DEBUG:
# print(f"Logits:{logits.view(-1, self.num_labels)}, \t, Labels:{labels.view(-1)}")
for i, subtask in enumerate(self.subtasks):
# print(labels[subtask].is_cuda)
if i == 0:
loss = loss_fct(logits[subtask].view(-1, self.num_labels), labels[subtask].view(-1))
else:
loss += loss_fct(logits[subtask].view(-1, self.num_labels), labels[subtask].view(-1))
outputs = outputs + (loss,)
return outputs # (loss), logits, (hidden_states), (attentions)
class MultiTaskBertForCovidEntityClassificationNew(nn.Module):
def __init__(self, auto_model_version, config):
super(MultiTaskBertForCovidEntityClassificationNew, self).__init__()
self.num_labels = config.num_labels
self.f1_loss = config.f1_loss
self.device = config.device
self.f1loss = F1_Loss().to(self.device)
self.event = config.event
self.embedding_type = config.embedding_type
self.subtasks = config.subtasks
config.num_labels = len(config.subtasks) ##TODO
self.weighting = config.weighting
self.encoder = AutoModel.from_pretrained(auto_model_version, config=config)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
if self.embedding_type == 2: # Concat
self.fc = nn.Linear(config.hidden_size * 4, config.hidden_size)
if self.embedding_type == 3: # multi-head-concat
self.fc1 = nn.Linear(config.hidden_size, config.hidden_size // 4)
self.fc2 = nn.Linear(config.hidden_size, config.hidden_size // 4)
self.fc3 = nn.Linear(config.hidden_size, config.hidden_size // 4)
self.fc4 = nn.Linear(config.hidden_size, config.hidden_size // 4)
self.fc_final = nn.Linear(config.hidden_size, config.hidden_size)
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
## input [batch size, hidden size]
## output [batch size, #subtask]
def resize_token_embeddings(self, length):
self.encoder.resize_token_embeddings(length)
@staticmethod
def build_loss_weight(y, factor=10):
weight = (y*(factor-1) + 1) / factor
return weight
def forward(self,
input_ids,
entity_start_positions,
labels=None):
outputs = self.encoder(input_ids)
if self.embedding_type == 0: # last layer
pooled_output = outputs[0][entity_start_positions[:, 0], entity_start_positions[:, 1], :]
elif self.embedding_type == 1: # sum
hidden_states = torch.stack(
[x[entity_start_positions[:, 0], entity_start_positions[:, 1], :]
for x in outputs[2][-4:]], dim=0)
pooled_output = torch.sum(hidden_states, dim=0)
elif self.embedding_type == 2: # Concat
hidden_states = tuple([x[entity_start_positions[:, 0], entity_start_positions[:, 1], :]
for x in outputs[2][-4:]])
pooled_output = self.fc(torch.cat(hidden_states, dim=1))
elif self.embedding_type == 3:
hidden_states = []
hidden_states.append(
self.fc1(outputs[2][-4][entity_start_positions[:, 0], entity_start_positions[:, 1], :]))
hidden_states.append(
self.fc2(outputs[2][-4][entity_start_positions[:, 0], entity_start_positions[:, 1], :]))
hidden_states.append(
self.fc3(outputs[2][-4][entity_start_positions[:, 0], entity_start_positions[:, 1], :]))
hidden_states.append(
self.fc4(outputs[2][-4][entity_start_positions[:, 0], entity_start_positions[:, 1], :]))
pooled_output = self.fc_final(torch.cat(hidden_states, dim=1))
pooled_output = self.dropout(pooled_output) ## [batch_size, 768]
# Get logits for each subtask
all_logits = self.classifier(pooled_output) #[batch size, # subtask]
if labels is not None:
y = torch.stack([labels[subtask] for subtask in labels.keys()], dim =1).type(torch.float)
if self.weighting:
weight = self.build_loss_weight(y)
else:
weight = None
# TODO: currently, weight is only applicable to BCE loss
if self.f1_loss:
loss = 0
for i in range(len(self.subtasks)):
loss += self.f1loss(all_logits[:,i],y[:,i])
else:
loss = F.binary_cross_entropy(torch.sigmoid(all_logits), y, weight=weight)
output = (all_logits, loss)
else:
output = (all_logits, )
return output # logits, (loss)
class MultiTaskBertForCovidEntityClassificationShare(nn.Module):
def __init__(self, auto_model_version, config):
super(MultiTaskBertForCovidEntityClassificationShare, self).__init__()
self.num_labels = config.num_labels
self.f1_loss = config.f1_loss
self.device = config.device
self.f1loss = F1_Loss().to(self.device)
self.embedding_type = config.embedding_type
self.subtasks = config.subtasks
config.num_labels = len(config.subtasks) ##TODO
self.weighting = config.weighting
self.encoder = AutoModel.from_pretrained(auto_model_version, config=config)
self.dropout = nn.Dropout(config.__dict__.get("hidden_dropout_prob", 0.1))
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
if self.embedding_type == 2: # Concat
self.fc = nn.Linear(config.hidden_size * 4, config.hidden_size)
if self.embedding_type == 3: # multi-head-concat
self.fc1 = nn.Linear(config.hidden_size, config.hidden_size // 4)
self.fc2 = nn.Linear(config.hidden_size, config.hidden_size // 4)
self.fc3 = nn.Linear(config.hidden_size, config.hidden_size // 4)
self.fc4 = nn.Linear(config.hidden_size, config.hidden_size // 4)
self.fc_final = nn.Linear(config.hidden_size, config.hidden_size)
def resize_token_embeddings(self, length):
self.encoder.resize_token_embeddings(length)
def build_loss_weight(self, y, factor=10):
weight = (y*(factor-1) + 1) / factor
return weight
def forward(self,
input_ids,
entity_start_positions,
y=None):
outputs = self.encoder(input_ids)
if self.embedding_type == 0: # last layer
pooled_output = outputs[0][entity_start_positions[:, 0], entity_start_positions[:, 1], :]
elif self.embedding_type == 1: # sum
hidden_states = torch.stack(
[x[entity_start_positions[:, 0], entity_start_positions[:, 1], :]
for x in outputs[2][-4:]], dim=0)
pooled_output = torch.sum(hidden_states, dim=0)
elif self.embedding_type == 2: # Concat
hidden_states = tuple([x[entity_start_positions[:, 0], entity_start_positions[:, 1], :]
for x in outputs[2][-4:]])
pooled_output = self.fc(torch.cat(hidden_states, dim=1))
elif self.embedding_type == 3:
hidden_states = []
hidden_states.append(
self.fc1(outputs[2][-4][entity_start_positions[:, 0], entity_start_positions[:, 1], :]))
hidden_states.append(
self.fc2(outputs[2][-4][entity_start_positions[:, 0], entity_start_positions[:, 1], :]))
hidden_states.append(
self.fc3(outputs[2][-4][entity_start_positions[:, 0], entity_start_positions[:, 1], :]))
hidden_states.append(
self.fc4(outputs[2][-4][entity_start_positions[:, 0], entity_start_positions[:, 1], :]))
pooled_output = self.fc_final(torch.cat(hidden_states, dim=1))
pooled_output = self.dropout(pooled_output) ## [batch_size, 768]
# Get logits for each subtask
all_logits = self.classifier(pooled_output) #[batch size, # subtask]
if y is not None:
if self.weighting:
weight = self.build_loss_weight(y)
else:
weight = None
# TODO: currently, weight is only applicable to BCE loss
if self.f1_loss:
loss = 0
for i in range(len(self.subtasks)):
loss += self.f1loss(all_logits[:,i],y[:,i])
else:
loss = F.binary_cross_entropy(torch.sigmoid(all_logits), y, weight=weight)
output = (all_logits, loss)
else:
output = (all_logits, )
return output # logits, (loss)