-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
146 lines (125 loc) · 7.26 KB
/
models.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
import torch
import torch.nn as nn
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
class BERTRNN(nn.Module):
def __init__(self, model_name, n_rnn_layers, dropout=0.5, n_finetune_layers=0):
super(BERTRNN, self).__init__()
if model_name in ['bert-base-uncased', 'gpt2']:
from transformers import AutoModel
self.bert = AutoModel.from_pretrained(model_name)
else:
from transformers import T5EncoderModel
self.bert = T5EncoderModel.from_pretrained(model_name)
nhid = self.bert.config.hidden_size
# only finetune the top several layers of BERT
n_layers = 12
if n_finetune_layers > 0:
for param in self.bert.parameters():
param.requires_grad = False
if model_name == 'bert-base-uncased':
for param in self.bert.pooler.parameters():
param.requires_grad = True
for i in range(n_layers - 1, n_layers - 1 - n_finetune_layers, -1):
for param in self.bert.encoder.layer[i].parameters():
param.requires_grad = True
elif model_name == 'gpt2':
for param in self.bert.ln_f.parameters():
param.requires_grad = True
for i in range(n_layers - 1, n_layers - 1 - n_finetune_layers, -1):
for param in self.bert.h[i].parameters():
param.requires_grad = True
elif model_name == 't5-base':
for param in self.bert.encoder.final_layer_norm.parameters():
param.requires_grad = True
for i in range(n_layers - 1, n_layers - 1 - n_finetune_layers, -1):
for param in self.bert.encoder.block[i].parameters():
param.requires_grad = True
self.rnn = nn.GRU(nhid, nhid // 2, num_layers=n_rnn_layers, dropout=dropout, bidirectional=True)
self.fc = nn.Linear(nhid, 2)
self.dropout = nn.Dropout(p=dropout)
self.model_name = model_name
def forward(self, input_ids, attention_mask, conv_lens):
conv_lens = conv_lens.to('cpu')
batch_size, max_conv_len, seq_len = input_ids.shape
# import ipdb; ipdb.set_trace()
input_ids = input_ids.reshape(-1, seq_len) # (batch_size*max_conv_len, seq_len)
attention_mask = attention_mask.reshape(-1, seq_len) # (batch_size*max_conv_len, seq_len)
outputs = self.bert(input_ids, attention_mask=attention_mask)
if self.model_name == 'bert-base-uncased':
last_hidden_states = outputs.pooler_output # (batch_size*max_conv_len, hidden_size)
else:
last_hidden_states = outputs.last_hidden_state # (batch_size*max_conv_len, seq_len, hidden_size)
attention_mask = attention_mask.to(torch.float)
attention_mask = attention_mask / attention_mask.sum(dim=1, keepdims=True)
last_hidden_states = torch.einsum('abc,ab->ac', last_hidden_states, attention_mask) # (batch_size*max_conv_len, hidden_size)
last_hidden_states = last_hidden_states.reshape(batch_size, max_conv_len, -1) # (batch_size, max_conv_len, hidden_size)
hidden_size = last_hidden_states.shape[-1]
# reshape BERT embeddings to fit into RNN
last_hidden_states = last_hidden_states.permute(1, 0, 2) # (max_conv_len, batch_size, hidden_size)
# sequence modeling of act tags using RNN
last_hidden_states = pack_padded_sequence(last_hidden_states, conv_lens, enforce_sorted=False)
self.rnn.flatten_parameters()
outputs, _ = self.rnn(last_hidden_states)
outputs, _ = pad_packed_sequence(outputs) # (batch_max_conv_len, batch_size, hidden_size)
if outputs.shape[0] < max_conv_len:
outputs_padding = torch.zeros(max_conv_len - outputs.shape[0], batch_size, hidden_size, device=outputs.device)
outputs = torch.cat([outputs, outputs_padding], dim=0) # (max_conv_len, batch_size, hidden_size)
outputs = self.dropout(outputs)
outputs = self.fc(outputs) # (max_conv_len, batch_size, 2)
outputs = outputs.permute(1, 0, 2) # (batch_size, max_conv_len, 2)
print('**********', conv_lens, '**********')
last_cell_indices = [list(range(batch_size)), (conv_lens-1).numpy().tolist()]
outputs = outputs[last_cell_indices] # (batch_size, 2)
return outputs
class BERT(nn.Module):
def __init__(self, model_name, dropout=0.5, n_finetune_layers=0):
super(BERT, self).__init__()
if model_name in ['bert-base-uncased', 'gpt2']:
from transformers import AutoModel
self.bert = AutoModel.from_pretrained(model_name)
else:
from transformers import T5EncoderModel
self.bert = T5EncoderModel.from_pretrained(model_name)
nhid = self.bert.config.hidden_size
# only finetune the top several layers of BERT
n_layers = 12
if n_finetune_layers > 0:
print('Freezing some top layers....')
for param in self.bert.parameters():
param.requires_grad = False
if model_name == 'bert-base-uncased':
for param in self.bert.pooler.parameters():
param.requires_grad = True
for i in range(n_layers - 1, n_layers - 1 - n_finetune_layers, -1):
for param in self.bert.encoder.layer[i].parameters():
param.requires_grad = True
elif model_name == 'gpt2':
for param in self.bert.ln_f.parameters():
param.requires_grad = True
for i in range(n_layers - 1, n_layers - 1 - n_finetune_layers, -1):
for param in self.bert.h[i].parameters():
param.requires_grad = True
elif model_name == 't5-base':
for param in self.bert.encoder.final_layer_norm.parameters():
param.requires_grad = True
for i in range(n_layers - 1, n_layers - 1 - n_finetune_layers, -1):
for param in self.bert.encoder.block[i].parameters():
param.requires_grad = True
self.fc = nn.Linear(nhid, 2)
self.dropout = nn.Dropout(p=dropout)
self.model_name = model_name
def forward(self, input_ids, attention_mask):
batch_size, max_conv_len, seq_len = input_ids.shape
input_ids = input_ids.reshape(-1, seq_len) # (batch_size, seq_len)
attention_mask = attention_mask.reshape(-1, seq_len) # (batch_size, seq_len)
outputs = self.bert(input_ids, attention_mask=attention_mask)
if self.model_name == 'bert-base-uncased':
last_hidden_states = outputs.pooler_output # (batch_size, hidden_size)
else:
last_hidden_states = outputs.last_hidden_state # (batch_size, seq_len, hidden_size)
attention_mask = attention_mask.to(torch.float)
attention_mask = attention_mask / attention_mask.sum(dim=1, keepdims=True)
last_hidden_states = torch.einsum('abc,ab->ac', last_hidden_states, attention_mask) # (batch_size, hidden_size)
outputs = self.dropout(last_hidden_states)
outputs = self.fc(outputs) # (batch_size, 2)
return outputs