-
Notifications
You must be signed in to change notification settings - Fork 15
/
modeling.py
247 lines (201 loc) · 13.1 KB
/
modeling.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
import torch
from torch.nn import Module, Linear, LayerNorm, Dropout
from transformers import BertPreTrainedModel, LongformerModel
from transformers.modeling_bert import ACT2FN
from utils import extract_clusters, extract_mentions_to_predicted_clusters_from_clusters, mask_tensor
class FullyConnectedLayer(Module):
def __init__(self, config, input_dim, output_dim, dropout_prob):
super(FullyConnectedLayer, self).__init__()
self.input_dim = input_dim
self.output_dim = output_dim
self.dropout_prob = dropout_prob
self.dense = Linear(self.input_dim, self.output_dim)
self.layer_norm = LayerNorm(self.output_dim, eps=config.layer_norm_eps)
self.activation_func = ACT2FN[config.hidden_act]
self.dropout = Dropout(self.dropout_prob)
def forward(self, inputs):
temp = inputs
temp = self.dense(temp)
temp = self.activation_func(temp)
temp = self.layer_norm(temp)
temp = self.dropout(temp)
return temp
class S2E(BertPreTrainedModel):
def __init__(self, config, args):
super().__init__(config)
self.max_span_length = args.max_span_length
self.top_lambda = args.top_lambda
self.ffnn_size = args.ffnn_size
self.do_mlps = self.ffnn_size > 0
self.ffnn_size = self.ffnn_size if self.do_mlps else config.hidden_size
self.normalise_loss = args.normalise_loss
self.longformer = LongformerModel(config)
self.start_mention_mlp = FullyConnectedLayer(config, config.hidden_size, self.ffnn_size, args.dropout_prob) if self.do_mlps else None
self.end_mention_mlp = FullyConnectedLayer(config, config.hidden_size, self.ffnn_size, args.dropout_prob) if self.do_mlps else None
self.start_coref_mlp = FullyConnectedLayer(config, config.hidden_size, self.ffnn_size, args.dropout_prob) if self.do_mlps else None
self.end_coref_mlp = FullyConnectedLayer(config, config.hidden_size, self.ffnn_size, args.dropout_prob) if self.do_mlps else None
self.start_coref_mlp = FullyConnectedLayer(config, config.hidden_size, self.ffnn_size, args.dropout_prob) if self.do_mlps else None
self.end_coref_mlp = FullyConnectedLayer(config, config.hidden_size, self.ffnn_size, args.dropout_prob) if self.do_mlps else None
self.mention_start_classifier = Linear(self.ffnn_size, 1)
self.mention_end_classifier = Linear(self.ffnn_size, 1)
self.mention_s2e_classifier = Linear(self.ffnn_size, self.ffnn_size)
self.antecedent_s2s_classifier = Linear(self.ffnn_size, self.ffnn_size)
self.antecedent_e2e_classifier = Linear(self.ffnn_size, self.ffnn_size)
self.antecedent_s2e_classifier = Linear(self.ffnn_size, self.ffnn_size)
self.antecedent_e2s_classifier = Linear(self.ffnn_size, self.ffnn_size)
self.init_weights()
def _get_span_mask(self, batch_size, k, max_k):
"""
:param batch_size: int
:param k: tensor of size [batch_size], with the required k for each example
:param max_k: int
:return: [batch_size, max_k] of zero-ones, where 1 stands for a valid span and 0 for a padded span
"""
size = (batch_size, max_k)
idx = torch.arange(max_k, device=self.device).unsqueeze(0).expand(size)
len_expanded = k.unsqueeze(1).expand(size)
return (idx < len_expanded).int()
def _prune_topk_mentions(self, mention_logits, attention_mask):
"""
:param mention_logits: Shape [batch_size, seq_length, seq_length]
:param attention_mask: [batch_size, seq_length]
:param top_lambda:
:return:
"""
batch_size, seq_length, _ = mention_logits.size()
actual_seq_lengths = torch.sum(attention_mask, dim=-1) # [batch_size]
k = (actual_seq_lengths * self.top_lambda).int() # [batch_size]
max_k = int(torch.max(k)) # This is the k for the largest input in the batch, we will need to pad
_, topk_1d_indices = torch.topk(mention_logits.view(batch_size, -1), dim=-1, k=max_k) # [batch_size, max_k]
span_mask = self._get_span_mask(batch_size, k, max_k) # [batch_size, max_k]
topk_1d_indices = (topk_1d_indices * span_mask) + (1 - span_mask) * ((seq_length ** 2) - 1) # We take different k for each example
sorted_topk_1d_indices, _ = torch.sort(topk_1d_indices, dim=-1) # [batch_size, max_k]
topk_mention_start_ids = sorted_topk_1d_indices // seq_length # [batch_size, max_k]
topk_mention_end_ids = sorted_topk_1d_indices % seq_length # [batch_size, max_k]
topk_mention_logits = mention_logits[torch.arange(batch_size).unsqueeze(-1).expand(batch_size, max_k),
topk_mention_start_ids, topk_mention_end_ids] # [batch_size, max_k]
topk_mention_logits = topk_mention_logits.unsqueeze(-1) + topk_mention_logits.unsqueeze(-2) # [batch_size, max_k, max_k]
return topk_mention_start_ids, topk_mention_end_ids, span_mask, topk_mention_logits
def _mask_antecedent_logits(self, antecedent_logits, span_mask):
# We now build the matrix for each pair of spans (i,j) - whether j is a candidate for being antecedent of i?
antecedents_mask = torch.ones_like(antecedent_logits, dtype=self.dtype).tril(diagonal=-1) # [batch_size, k, k]
antecedents_mask = antecedents_mask * span_mask.unsqueeze(-1) # [batch_size, k, k]
antecedent_logits = mask_tensor(antecedent_logits, antecedents_mask)
return antecedent_logits
def _get_cluster_labels_after_pruning(self, span_starts, span_ends, all_clusters):
"""
:param span_starts: [batch_size, max_k]
:param span_ends: [batch_size, max_k]
:param all_clusters: [batch_size, max_cluster_size, max_clusters_num, 2]
:return: [batch_size, max_k, max_k + 1] - [b, i, j] == 1 if i is antecedent of j
"""
batch_size, max_k = span_starts.size()
new_cluster_labels = torch.zeros((batch_size, max_k, max_k + 1), device='cpu')
all_clusters_cpu = all_clusters.cpu().numpy()
for b, (starts, ends, gold_clusters) in enumerate(zip(span_starts.cpu().tolist(), span_ends.cpu().tolist(), all_clusters_cpu)):
gold_clusters = extract_clusters(gold_clusters)
mention_to_gold_clusters = extract_mentions_to_predicted_clusters_from_clusters(gold_clusters)
gold_mentions = set(mention_to_gold_clusters.keys())
for i, (start, end) in enumerate(zip(starts, ends)):
if (start, end) not in gold_mentions:
continue
for j, (a_start, a_end) in enumerate(list(zip(starts, ends))[:i]):
if (a_start, a_end) in mention_to_gold_clusters[(start, end)]:
new_cluster_labels[b, i, j] = 1
new_cluster_labels = new_cluster_labels.to(self.device)
no_antecedents = 1 - torch.sum(new_cluster_labels, dim=-1).bool().float()
new_cluster_labels[:, :, -1] = no_antecedents
return new_cluster_labels
def _get_marginal_log_likelihood_loss(self, coref_logits, cluster_labels_after_pruning, span_mask):
"""
:param coref_logits: [batch_size, max_k, max_k]
:param cluster_labels_after_pruning: [batch_size, max_k, max_k]
:param span_mask: [batch_size, max_k]
:return:
"""
gold_coref_logits = mask_tensor(coref_logits, cluster_labels_after_pruning)
gold_log_sum_exp = torch.logsumexp(gold_coref_logits, dim=-1) # [batch_size, max_k]
all_log_sum_exp = torch.logsumexp(coref_logits, dim=-1) # [batch_size, max_k]
gold_log_probs = gold_log_sum_exp - all_log_sum_exp
losses = - gold_log_probs
losses = losses * span_mask
per_example_loss = torch.sum(losses, dim=-1) # [batch_size]
if self.normalise_loss:
per_example_loss = per_example_loss / losses.size(-1)
loss = per_example_loss.mean()
return loss
def _get_mention_mask(self, mention_logits_or_weights):
"""
Returns a tensor of size [batch_size, seq_length, seq_length] where valid spans
(start <= end < start + max_span_length) are 1 and the rest are 0
:param mention_logits_or_weights: Either the span mention logits or weights, size [batch_size, seq_length, seq_length]
"""
mention_mask = torch.ones_like(mention_logits_or_weights, dtype=self.dtype)
mention_mask = mention_mask.triu(diagonal=0)
mention_mask = mention_mask.tril(diagonal=self.max_span_length - 1)
return mention_mask
def _calc_mention_logits(self, start_mention_reps, end_mention_reps):
start_mention_logits = self.mention_start_classifier(start_mention_reps).squeeze(-1) # [batch_size, seq_length]
end_mention_logits = self.mention_end_classifier(end_mention_reps).squeeze(-1) # [batch_size, seq_length]
temp = self.mention_s2e_classifier(start_mention_reps) # [batch_size, seq_length]
joint_mention_logits = torch.matmul(temp,
end_mention_reps.permute([0, 2, 1])) # [batch_size, seq_length, seq_length]
mention_logits = joint_mention_logits + start_mention_logits.unsqueeze(-1) + end_mention_logits.unsqueeze(-2)
mention_mask = self._get_mention_mask(mention_logits) # [batch_size, seq_length, seq_length]
mention_logits = mask_tensor(mention_logits, mention_mask) # [batch_size, seq_length, seq_length]
return mention_logits
def _calc_coref_logits(self, top_k_start_coref_reps, top_k_end_coref_reps):
# s2s
temp = self.antecedent_s2s_classifier(top_k_start_coref_reps) # [batch_size, max_k, dim]
top_k_s2s_coref_logits = torch.matmul(temp,
top_k_start_coref_reps.permute([0, 2, 1])) # [batch_size, max_k, max_k]
# e2e
temp = self.antecedent_e2e_classifier(top_k_end_coref_reps) # [batch_size, max_k, dim]
top_k_e2e_coref_logits = torch.matmul(temp,
top_k_end_coref_reps.permute([0, 2, 1])) # [batch_size, max_k, max_k]
# s2e
temp = self.antecedent_s2e_classifier(top_k_start_coref_reps) # [batch_size, max_k, dim]
top_k_s2e_coref_logits = torch.matmul(temp,
top_k_end_coref_reps.permute([0, 2, 1])) # [batch_size, max_k, max_k]
# e2s
temp = self.antecedent_e2s_classifier(top_k_end_coref_reps) # [batch_size, max_k, dim]
top_k_e2s_coref_logits = torch.matmul(temp,
top_k_start_coref_reps.permute([0, 2, 1])) # [batch_size, max_k, max_k]
# sum all terms
coref_logits = top_k_s2e_coref_logits + top_k_e2s_coref_logits + top_k_s2s_coref_logits + top_k_e2e_coref_logits # [batch_size, max_k, max_k]
return coref_logits
def forward(self, input_ids, attention_mask=None, gold_clusters=None, return_all_outputs=False):
outputs = self.longformer(input_ids, attention_mask=attention_mask)
sequence_output = outputs[0] # [batch_size, seq_len, dim]
# Compute representations
start_mention_reps = self.start_mention_mlp(sequence_output) if self.do_mlps else sequence_output
end_mention_reps = self.end_mention_mlp(sequence_output) if self.do_mlps else sequence_output
start_coref_reps = self.start_coref_mlp(sequence_output) if self.do_mlps else sequence_output
end_coref_reps = self.end_coref_mlp(sequence_output) if self.do_mlps else sequence_output
# mention scores
mention_logits = self._calc_mention_logits(start_mention_reps, end_mention_reps)
# prune mentions
mention_start_ids, mention_end_ids, span_mask, topk_mention_logits = self._prune_topk_mentions(mention_logits, attention_mask)
batch_size, _, dim = start_coref_reps.size()
max_k = mention_start_ids.size(-1)
size = (batch_size, max_k, dim)
# Antecedent scores
# gather reps
topk_start_coref_reps = torch.gather(start_coref_reps, dim=1, index=mention_start_ids.unsqueeze(-1).expand(size))
topk_end_coref_reps = torch.gather(end_coref_reps, dim=1, index=mention_end_ids.unsqueeze(-1).expand(size))
coref_logits = self._calc_coref_logits(topk_start_coref_reps, topk_end_coref_reps)
final_logits = topk_mention_logits + coref_logits
final_logits = self._mask_antecedent_logits(final_logits, span_mask)
# adding zero logits for null span
final_logits = torch.cat((final_logits, torch.zeros((batch_size, max_k, 1), device=self.device)), dim=-1) # [batch_size, max_k, max_k + 1]
if return_all_outputs:
outputs = (mention_start_ids, mention_end_ids, final_logits, mention_logits)
else:
outputs = tuple()
if gold_clusters is not None:
losses = {}
labels_after_pruning = self._get_cluster_labels_after_pruning(mention_start_ids, mention_end_ids, gold_clusters)
loss = self._get_marginal_log_likelihood_loss(final_logits, labels_after_pruning, span_mask)
losses.update({"loss": loss})
outputs = (loss,) + outputs + (losses,)
return outputs