Logic123456789
commited on
Commit
•
e16b4bd
1
Parent(s):
915e12e
add the models.py
Browse files
models.py
ADDED
@@ -0,0 +1,524 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import torch.distributed as dist
|
5 |
+
|
6 |
+
from simcse.modeling_glm import GLMModel, GLMPreTrainedModel
|
7 |
+
import simcse.mse_loss
|
8 |
+
|
9 |
+
import transformers
|
10 |
+
from transformers import RobertaTokenizer, AutoModel, PreTrainedModel
|
11 |
+
from transformers.models.roberta.modeling_roberta import RobertaPreTrainedModel, RobertaModel, RobertaLMHead
|
12 |
+
from transformers.models.bert.modeling_bert import BertPreTrainedModel, BertModel, BertLMPredictionHead
|
13 |
+
from transformers.activations import gelu
|
14 |
+
from transformers.file_utils import (
|
15 |
+
add_code_sample_docstrings,
|
16 |
+
add_start_docstrings,
|
17 |
+
add_start_docstrings_to_model_forward,
|
18 |
+
replace_return_docstrings,
|
19 |
+
)
|
20 |
+
from transformers.modeling_outputs import SequenceClassifierOutput, BaseModelOutputWithPoolingAndCrossAttentions
|
21 |
+
|
22 |
+
glm_model = None
|
23 |
+
|
24 |
+
def init_glm(path):
|
25 |
+
global glm_model
|
26 |
+
glm_model = GLMModel.from_pretrained(path, trust_remote_code=True).to("cuda:0")
|
27 |
+
for param in glm_model.parameters():
|
28 |
+
param.requires_grad = False
|
29 |
+
|
30 |
+
|
31 |
+
|
32 |
+
class MLPLayer(nn.Module):
|
33 |
+
"""
|
34 |
+
Head for getting sentence representations over RoBERTa/BERT's CLS representation.
|
35 |
+
"""
|
36 |
+
|
37 |
+
def __init__(self, config):
|
38 |
+
super().__init__()
|
39 |
+
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
40 |
+
# 1536
|
41 |
+
self.fc = nn.Linear(config.hidden_size, 1536)
|
42 |
+
self.activation = nn.Tanh()
|
43 |
+
|
44 |
+
def forward(self, features, **kwargs):
|
45 |
+
x = self.dense(features)
|
46 |
+
x = self.fc(x)
|
47 |
+
x = self.activation(x)
|
48 |
+
|
49 |
+
return x
|
50 |
+
|
51 |
+
class Similarity(nn.Module):
|
52 |
+
"""
|
53 |
+
Dot product or cosine similarity
|
54 |
+
"""
|
55 |
+
|
56 |
+
def __init__(self, temp):
|
57 |
+
super().__init__()
|
58 |
+
self.temp = temp
|
59 |
+
self.cos = nn.CosineSimilarity(dim=-1)
|
60 |
+
|
61 |
+
def forward(self, x, y):
|
62 |
+
return self.cos(x, y) / self.temp
|
63 |
+
|
64 |
+
|
65 |
+
class Pooler(nn.Module):
|
66 |
+
"""
|
67 |
+
Parameter-free poolers to get the sentence embedding
|
68 |
+
'cls': [CLS] representation with BERT/RoBERTa's MLP pooler.
|
69 |
+
'cls_before_pooler': [CLS] representation without the original MLP pooler.
|
70 |
+
'avg': average of the last layers' hidden states at each token.
|
71 |
+
'avg_top2': average of the last two layers.
|
72 |
+
'avg_first_last': average of the first and the last layers.
|
73 |
+
"""
|
74 |
+
|
75 |
+
def __init__(self, pooler_type):
|
76 |
+
super().__init__()
|
77 |
+
self.pooler_type = pooler_type
|
78 |
+
assert self.pooler_type in ["cls", "cls_before_pooler", "avg", "avg_top2",
|
79 |
+
"avg_first_last"], "unrecognized pooling type %s" % self.pooler_type
|
80 |
+
|
81 |
+
def forward(self, attention_mask, outputs):
|
82 |
+
last_hidden = outputs.last_hidden_state
|
83 |
+
# pooler_output = outputs.pooler_output
|
84 |
+
hidden_states = outputs.hidden_states
|
85 |
+
|
86 |
+
if self.pooler_type in ['cls_before_pooler', 'cls']:
|
87 |
+
return last_hidden[:, 0]
|
88 |
+
elif self.pooler_type == "avg":
|
89 |
+
return ((last_hidden * attention_mask.unsqueeze(-1)).sum(1) / attention_mask.sum(-1).unsqueeze(-1))
|
90 |
+
elif self.pooler_type == "avg_first_last":
|
91 |
+
first_hidden = hidden_states[1]
|
92 |
+
last_hidden = hidden_states[-1]
|
93 |
+
pooled_result = ((first_hidden + last_hidden) / 2.0 * attention_mask.unsqueeze(-1)).sum(
|
94 |
+
1) / attention_mask.sum(-1).unsqueeze(-1)
|
95 |
+
return pooled_result
|
96 |
+
elif self.pooler_type == "avg_top2":
|
97 |
+
second_last_hidden = hidden_states[-2]
|
98 |
+
last_hidden = hidden_states[-1]
|
99 |
+
pooled_result = ((last_hidden + second_last_hidden) / 2.0 * attention_mask.unsqueeze(-1)).sum(
|
100 |
+
1) / attention_mask.sum(-1).unsqueeze(-1)
|
101 |
+
return pooled_result
|
102 |
+
else:
|
103 |
+
raise NotImplementedError
|
104 |
+
|
105 |
+
|
106 |
+
def cl_init(cls, config):
|
107 |
+
"""
|
108 |
+
Contrastive learning class init function.
|
109 |
+
"""
|
110 |
+
cls.pooler_type = cls.model_args.pooler_type
|
111 |
+
cls.pooler = Pooler(cls.model_args.pooler_type)
|
112 |
+
if cls.model_args.pooler_type == "cls":
|
113 |
+
cls.mlp = MLPLayer(config)
|
114 |
+
cls.sim = Similarity(temp=cls.model_args.temp)
|
115 |
+
cls.init_weights()
|
116 |
+
|
117 |
+
|
118 |
+
def cl_forward(cls,
|
119 |
+
encoder,
|
120 |
+
input_ids=None,
|
121 |
+
attention_mask=None,
|
122 |
+
token_type_ids=None,
|
123 |
+
position_ids=None,
|
124 |
+
head_mask=None,
|
125 |
+
inputs_embeds=None,
|
126 |
+
labels=None,
|
127 |
+
output_attentions=None,
|
128 |
+
output_hidden_states=None,
|
129 |
+
return_dict=None,
|
130 |
+
mlm_input_ids=None,
|
131 |
+
mlm_labels=None,
|
132 |
+
left_emb=None,
|
133 |
+
right_emb=None,
|
134 |
+
kl_loss=False
|
135 |
+
):
|
136 |
+
return_dict = return_dict if return_dict is not None else cls.config.use_return_dict
|
137 |
+
ori_input_ids = input_ids
|
138 |
+
batch_size = input_ids.size(0)
|
139 |
+
# Number of sentences in one instance
|
140 |
+
# 2: pair instance; 3: pair instance with a hard negative
|
141 |
+
num_sent = input_ids.size(1)
|
142 |
+
|
143 |
+
mlm_outputs = None
|
144 |
+
# Flatten input for encoding
|
145 |
+
input_ids = input_ids.view((-1, input_ids.size(-1))) # (bs * num_sent, len)
|
146 |
+
attention_mask = attention_mask.view((-1, attention_mask.size(-1))) # (bs * num_sent len)
|
147 |
+
if token_type_ids is not None:
|
148 |
+
token_type_ids = token_type_ids.view((-1, token_type_ids.size(-1))) # (bs * num_sent, len)
|
149 |
+
|
150 |
+
if inputs_embeds is not None:
|
151 |
+
input_ids = None
|
152 |
+
|
153 |
+
# Get raw embeddings
|
154 |
+
outputs = encoder(
|
155 |
+
input_ids,
|
156 |
+
attention_mask=attention_mask,
|
157 |
+
token_type_ids=token_type_ids,
|
158 |
+
position_ids=position_ids,
|
159 |
+
head_mask=head_mask,
|
160 |
+
inputs_embeds=inputs_embeds,
|
161 |
+
output_attentions=output_attentions,
|
162 |
+
output_hidden_states=True if cls.model_args.pooler_type in ['avg_top2', 'avg_first_last'] else False,
|
163 |
+
return_dict=True,
|
164 |
+
)
|
165 |
+
|
166 |
+
# MLM auxiliary objective
|
167 |
+
if mlm_input_ids is not None:
|
168 |
+
mlm_input_ids = mlm_input_ids.view((-1, mlm_input_ids.size(-1)))
|
169 |
+
mlm_outputs = encoder(
|
170 |
+
mlm_input_ids,
|
171 |
+
attention_mask=attention_mask,
|
172 |
+
token_type_ids=token_type_ids,
|
173 |
+
position_ids=position_ids,
|
174 |
+
head_mask=head_mask,
|
175 |
+
inputs_embeds=inputs_embeds,
|
176 |
+
output_attentions=output_attentions,
|
177 |
+
output_hidden_states=True if cls.model_args.pooler_type in ['avg_top2', 'avg_first_last'] else False,
|
178 |
+
return_dict=True,
|
179 |
+
)
|
180 |
+
|
181 |
+
# Pooling
|
182 |
+
pooler_output = cls.pooler(attention_mask, outputs)
|
183 |
+
pooler_output = pooler_output.view((batch_size, num_sent, pooler_output.size(-1))) # (bs, num_sent, hidden)
|
184 |
+
# If using "cls", we add an extra MLP layer
|
185 |
+
# (same as BERT's original implementation) over the representation.
|
186 |
+
if cls.pooler_type == "cls":
|
187 |
+
pooler_output = cls.mlp(pooler_output)
|
188 |
+
|
189 |
+
# Separate representation
|
190 |
+
z1, z2 = pooler_output[:, 0], pooler_output[:, 1]
|
191 |
+
|
192 |
+
tensor_left = left_emb
|
193 |
+
tensor_right = right_emb
|
194 |
+
|
195 |
+
# Hard negative
|
196 |
+
if num_sent == 3:
|
197 |
+
z3 = pooler_output[:, 2]
|
198 |
+
|
199 |
+
# Gather all embeddings if using distributed training
|
200 |
+
if dist.is_initialized() and cls.training:
|
201 |
+
# Gather hard negative
|
202 |
+
if num_sent >= 3:
|
203 |
+
z3_list = [torch.zeros_like(z3) for _ in range(dist.get_world_size())]
|
204 |
+
dist.all_gather(tensor_list=z3_list, tensor=z3.contiguous())
|
205 |
+
z3_list[dist.get_rank()] = z3
|
206 |
+
z3 = torch.cat(z3_list, 0)
|
207 |
+
|
208 |
+
# Dummy vectors for allgather
|
209 |
+
z1_list = [torch.zeros_like(z1) for _ in range(dist.get_world_size())]
|
210 |
+
z2_list = [torch.zeros_like(z2) for _ in range(dist.get_world_size())]
|
211 |
+
# Allgather
|
212 |
+
dist.all_gather(tensor_list=z1_list, tensor=z1.contiguous())
|
213 |
+
dist.all_gather(tensor_list=z2_list, tensor=z2.contiguous())
|
214 |
+
|
215 |
+
# Since allgather results do not have gradients, we replace the
|
216 |
+
# current process's corresponding embeddings with original tensors
|
217 |
+
z1_list[dist.get_rank()] = z1
|
218 |
+
z2_list[dist.get_rank()] = z2
|
219 |
+
# Get full batch embeddings: (bs x N, hidden)
|
220 |
+
z1 = torch.cat(z1_list, 0)
|
221 |
+
z2 = torch.cat(z2_list, 0)
|
222 |
+
|
223 |
+
mse_loss = F.mse_loss(z1, tensor_left) + F.mse_loss(z2, tensor_right)
|
224 |
+
|
225 |
+
# softmax_row, softmax_col = simcse.mse_loss.giveMeMatrix(tensor_left, tensor_right)
|
226 |
+
# softmax_row_model, softmax_col_model = simcse.mse_loss.giveMeMatrix(z1,z2)
|
227 |
+
# ziang_labels = torch.tensor([i for i in range(8)], device='cuda:0')
|
228 |
+
|
229 |
+
"""
|
230 |
+
this is KL div loss
|
231 |
+
"""
|
232 |
+
|
233 |
+
KL_loss = nn.KLDivLoss(reduction="batchmean")
|
234 |
+
beta = 5
|
235 |
+
|
236 |
+
# openai的embed,giveMeMatrix返回一个normalized过前后向量,相乘后的矩阵
|
237 |
+
cos_sim_matrix_openai = simcse.mse_loss.giveMeMatrix(tensor_left, tensor_right)
|
238 |
+
beta_scaled_cos_sim_matrix_openai = beta * cos_sim_matrix_openai
|
239 |
+
|
240 |
+
# 我们的embed,giveMeMatrix返回一个normalized过前后向量,相乘后的矩阵
|
241 |
+
cos_sim_matrix_data = simcse.mse_loss.giveMeMatrix(z1, z2)
|
242 |
+
beta_scaled_cos_sim_matrix_data = beta * cos_sim_matrix_data
|
243 |
+
|
244 |
+
beta_scaled_cos_sim_matrix_openai_vertical = beta_scaled_cos_sim_matrix_openai.softmax(dim=1)
|
245 |
+
beta_scaled_cos_sim_matrix_openai_horizontal = beta_scaled_cos_sim_matrix_openai.softmax(dim=0)
|
246 |
+
|
247 |
+
beta_scaled_cos_sim_matrix_data_vertical = beta_scaled_cos_sim_matrix_data.softmax(dim=1)
|
248 |
+
beta_scaled_cos_sim_matrix_data_horizontal = beta_scaled_cos_sim_matrix_data.softmax(dim=0)
|
249 |
+
|
250 |
+
# remove reduction="batchmean"
|
251 |
+
KL_vertical_loss = KL_loss(beta_scaled_cos_sim_matrix_data_vertical.log(), beta_scaled_cos_sim_matrix_openai_vertical)
|
252 |
+
KL_horizontal_loss = KL_loss(beta_scaled_cos_sim_matrix_data_horizontal.log(), beta_scaled_cos_sim_matrix_openai_horizontal)
|
253 |
+
|
254 |
+
KL_loss = (KL_vertical_loss + KL_horizontal_loss) / 2
|
255 |
+
|
256 |
+
# KL_row_loss = F.kl_div(softmax_row_model.log(), softmax_row, reduction='batchmean')
|
257 |
+
# KL_col_loss = F.kl_div(softmax_col_model.log(), softmax_col, reduction='batchmean')
|
258 |
+
# KL_loss = (KL_row_loss + KL_col_loss) / 2
|
259 |
+
|
260 |
+
ziang_loss = KL_loss + mse_loss
|
261 |
+
|
262 |
+
cos_sim = cls.sim(z1.unsqueeze(1), z2.unsqueeze(0))
|
263 |
+
|
264 |
+
# Hard negative
|
265 |
+
if num_sent >= 3:
|
266 |
+
z1_z3_cos = cls.sim(z1.unsqueeze(1), z3.unsqueeze(0))
|
267 |
+
cos_sim = torch.cat([cos_sim, z1_z3_cos], 1)
|
268 |
+
|
269 |
+
labels = torch.arange(cos_sim.size(0)).long().to(cls.device)
|
270 |
+
loss_fct = nn.CrossEntropyLoss()
|
271 |
+
|
272 |
+
# Calculate loss with hard negatives
|
273 |
+
if num_sent == 3:
|
274 |
+
# Note that weights are actually logits of weights
|
275 |
+
z3_weight = cls.model_args.hard_negative_weight
|
276 |
+
weights = torch.tensor(
|
277 |
+
[[0.0] * (cos_sim.size(-1) - z1_z3_cos.size(-1)) + [0.0] * i + [z3_weight] + [0.0] * (
|
278 |
+
z1_z3_cos.size(-1) - i - 1) for i in range(z1_z3_cos.size(-1))]
|
279 |
+
).to(cls.device)
|
280 |
+
cos_sim = cos_sim + weights
|
281 |
+
|
282 |
+
loss = loss_fct(cos_sim, labels)
|
283 |
+
|
284 |
+
# Calculate loss for MLM
|
285 |
+
if mlm_outputs is not None and mlm_labels is not None:
|
286 |
+
mlm_labels = mlm_labels.view(-1, mlm_labels.size(-1))
|
287 |
+
prediction_scores = cls.lm_head(mlm_outputs.last_hidden_state)
|
288 |
+
masked_lm_loss = loss_fct(prediction_scores.view(-1, cls.config.vocab_size), mlm_labels.view(-1))
|
289 |
+
loss = loss + cls.model_args.mlm_weight * masked_lm_loss
|
290 |
+
|
291 |
+
if not return_dict:
|
292 |
+
output = (cos_sim,) + outputs[2:]
|
293 |
+
return ((loss,) + output) if loss is not None else output
|
294 |
+
|
295 |
+
return SequenceClassifierOutput(
|
296 |
+
loss=ziang_loss,
|
297 |
+
logits=cos_sim,
|
298 |
+
hidden_states=outputs.hidden_states,
|
299 |
+
)
|
300 |
+
|
301 |
+
|
302 |
+
def sentemb_forward(
|
303 |
+
cls,
|
304 |
+
encoder,
|
305 |
+
input_ids=None,
|
306 |
+
attention_mask=None,
|
307 |
+
token_type_ids=None,
|
308 |
+
position_ids=None,
|
309 |
+
head_mask=None,
|
310 |
+
inputs_embeds=None,
|
311 |
+
labels=None,
|
312 |
+
output_attentions=None,
|
313 |
+
output_hidden_states=None,
|
314 |
+
return_dict=None,
|
315 |
+
):
|
316 |
+
return_dict = return_dict if return_dict is not None else cls.config.use_return_dict
|
317 |
+
|
318 |
+
if inputs_embeds is not None:
|
319 |
+
input_ids = None
|
320 |
+
|
321 |
+
outputs = encoder(
|
322 |
+
input_ids,
|
323 |
+
attention_mask=attention_mask,
|
324 |
+
token_type_ids=token_type_ids,
|
325 |
+
position_ids=position_ids,
|
326 |
+
head_mask=head_mask,
|
327 |
+
inputs_embeds=inputs_embeds,
|
328 |
+
output_attentions=output_attentions,
|
329 |
+
output_hidden_states=True if cls.pooler_type in ['avg_top2', 'avg_first_last'] else False,
|
330 |
+
return_dict=True,
|
331 |
+
)
|
332 |
+
|
333 |
+
pooler_output = cls.pooler(attention_mask, outputs)
|
334 |
+
if cls.pooler_type == "cls" and not cls.model_args.mlp_only_train:
|
335 |
+
pooler_output = cls.mlp(pooler_output)
|
336 |
+
|
337 |
+
if not return_dict:
|
338 |
+
return (outputs[0], pooler_output) + outputs[2:]
|
339 |
+
|
340 |
+
return BaseModelOutputWithPoolingAndCrossAttentions(
|
341 |
+
pooler_output=pooler_output,
|
342 |
+
last_hidden_state=outputs.last_hidden_state,
|
343 |
+
hidden_states=outputs.hidden_states,
|
344 |
+
)
|
345 |
+
|
346 |
+
|
347 |
+
class BertForCL(BertPreTrainedModel):
|
348 |
+
_keys_to_ignore_on_load_missing = [r"position_ids"]
|
349 |
+
|
350 |
+
def __init__(self, config, *model_args, **model_kargs):
|
351 |
+
super().__init__(config)
|
352 |
+
self.model_args = model_kargs["model_args"]
|
353 |
+
self.bert = BertModel(config, add_pooling_layer=False)
|
354 |
+
|
355 |
+
if self.model_args.do_mlm:
|
356 |
+
self.lm_head = BertLMPredictionHead(config)
|
357 |
+
|
358 |
+
if self.model_args.init_embeddings_model:
|
359 |
+
if "glm" in self.model_args.init_embeddings_model:
|
360 |
+
init_glm(self.model_args.init_embeddings_model)
|
361 |
+
self.fc = nn.Linear(glm_model.config.hidden_size, config.hidden_size)
|
362 |
+
else:
|
363 |
+
raise NotImplementedError
|
364 |
+
|
365 |
+
cl_init(self, config)
|
366 |
+
|
367 |
+
def forward(self,
|
368 |
+
input_ids=None,
|
369 |
+
attention_mask=None,
|
370 |
+
token_type_ids=None,
|
371 |
+
position_ids=None,
|
372 |
+
head_mask=None,
|
373 |
+
inputs_embeds=None,
|
374 |
+
labels=None,
|
375 |
+
output_attentions=None,
|
376 |
+
output_hidden_states=None,
|
377 |
+
return_dict=None,
|
378 |
+
sent_emb=False,
|
379 |
+
mlm_input_ids=None,
|
380 |
+
mlm_labels=None,
|
381 |
+
left_emb=None,
|
382 |
+
right_emb=None,
|
383 |
+
):
|
384 |
+
if self.model_args.init_embeddings_model:
|
385 |
+
input_ids_for_glm = input_ids.view((-1, input_ids.size(-1))) # (bs * num_sent, len)
|
386 |
+
attention_mask_for_glm = attention_mask.view((-1, attention_mask.size(-1))) # (bs * num_sent len)
|
387 |
+
if token_type_ids is not None:
|
388 |
+
token_type_ids_for_glm = token_type_ids.view((-1, token_type_ids.size(-1))) # (bs * num_sent, len)
|
389 |
+
|
390 |
+
outputs_from_glm = glm_model(input_ids_for_glm,
|
391 |
+
attention_mask=attention_mask_for_glm,
|
392 |
+
token_type_ids=token_type_ids_for_glm,
|
393 |
+
position_ids=position_ids,
|
394 |
+
head_mask=head_mask,
|
395 |
+
inputs_embeds=inputs_embeds,
|
396 |
+
labels=labels,
|
397 |
+
output_attentions=output_attentions,
|
398 |
+
output_hidden_states=output_hidden_states,
|
399 |
+
return_dict=return_dict,
|
400 |
+
)
|
401 |
+
|
402 |
+
inputs_embeds = self.fc(outputs_from_glm.last_hidden_state)
|
403 |
+
|
404 |
+
if sent_emb:
|
405 |
+
return sentemb_forward(self, self.bert,
|
406 |
+
input_ids=input_ids,
|
407 |
+
attention_mask=attention_mask,
|
408 |
+
token_type_ids=token_type_ids,
|
409 |
+
position_ids=position_ids,
|
410 |
+
head_mask=head_mask,
|
411 |
+
inputs_embeds=inputs_embeds,
|
412 |
+
labels=labels,
|
413 |
+
output_attentions=output_attentions,
|
414 |
+
output_hidden_states=output_hidden_states,
|
415 |
+
return_dict=return_dict,
|
416 |
+
)
|
417 |
+
else:
|
418 |
+
return cl_forward(self, self.bert,
|
419 |
+
input_ids=input_ids,
|
420 |
+
attention_mask=attention_mask,
|
421 |
+
token_type_ids=token_type_ids,
|
422 |
+
position_ids=position_ids,
|
423 |
+
head_mask=head_mask,
|
424 |
+
inputs_embeds=inputs_embeds,
|
425 |
+
labels=labels,
|
426 |
+
output_attentions=output_attentions,
|
427 |
+
output_hidden_states=output_hidden_states,
|
428 |
+
return_dict=return_dict,
|
429 |
+
mlm_input_ids=mlm_input_ids,
|
430 |
+
mlm_labels=mlm_labels,
|
431 |
+
left_emb=left_emb,
|
432 |
+
right_emb=right_emb,
|
433 |
+
)
|
434 |
+
|
435 |
+
|
436 |
+
class RobertaForCL(RobertaPreTrainedModel):
|
437 |
+
_keys_to_ignore_on_load_missing = [r"position_ids"]
|
438 |
+
|
439 |
+
def __init__(self, config, *model_args, **model_kargs):
|
440 |
+
super().__init__(config)
|
441 |
+
self.model_args = model_kargs["model_args"]
|
442 |
+
self.roberta = RobertaModel(config, add_pooling_layer=False)
|
443 |
+
|
444 |
+
if self.model_args.do_mlm:
|
445 |
+
self.lm_head = RobertaLMHead(config)
|
446 |
+
|
447 |
+
if self.model_args.init_embeddings_model:
|
448 |
+
if "glm" in self.model_args.init_embeddings_model:
|
449 |
+
init_glm(self.model_args.init_embeddings_model)
|
450 |
+
self.fc = nn.Linear(glm_model.config.hidden_size, config.hidden_size)
|
451 |
+
else:
|
452 |
+
raise NotImplementedError
|
453 |
+
|
454 |
+
cl_init(self, config)
|
455 |
+
|
456 |
+
def forward(self,
|
457 |
+
input_ids=None,
|
458 |
+
attention_mask=None,
|
459 |
+
token_type_ids=None,
|
460 |
+
position_ids=None,
|
461 |
+
head_mask=None,
|
462 |
+
inputs_embeds=None,
|
463 |
+
labels=None,
|
464 |
+
output_attentions=None,
|
465 |
+
output_hidden_states=None,
|
466 |
+
return_dict=None,
|
467 |
+
sent_emb=False,
|
468 |
+
mlm_input_ids=None,
|
469 |
+
mlm_labels=None,
|
470 |
+
left_emb=None,
|
471 |
+
right_emb=None,
|
472 |
+
):
|
473 |
+
|
474 |
+
if self.model_args.init_embeddings_model and not sent_emb:
|
475 |
+
input_ids_for_glm = input_ids.view((-1, input_ids.size(-1))) # (bs * num_sent, len)
|
476 |
+
attention_mask_for_glm = attention_mask.view((-1, attention_mask.size(-1))) # (bs * num_sent len)
|
477 |
+
if token_type_ids is not None:
|
478 |
+
token_type_ids_for_glm = token_type_ids.view((-1, token_type_ids.size(-1))) # (bs * num_sent, len)
|
479 |
+
|
480 |
+
outputs_from_glm = glm_model(input_ids_for_glm,
|
481 |
+
attention_mask=attention_mask_for_glm,
|
482 |
+
token_type_ids=token_type_ids_for_glm,
|
483 |
+
position_ids=position_ids,
|
484 |
+
head_mask=head_mask,
|
485 |
+
inputs_embeds=inputs_embeds,
|
486 |
+
labels=labels,
|
487 |
+
output_attentions=output_attentions,
|
488 |
+
output_hidden_states=output_hidden_states,
|
489 |
+
return_dict=return_dict,
|
490 |
+
)
|
491 |
+
|
492 |
+
inputs_embeds = self.fc(outputs_from_glm.last_hidden_state)
|
493 |
+
|
494 |
+
if sent_emb:
|
495 |
+
return sentemb_forward(self, self.roberta,
|
496 |
+
input_ids=input_ids,
|
497 |
+
attention_mask=attention_mask,
|
498 |
+
token_type_ids=token_type_ids,
|
499 |
+
position_ids=position_ids,
|
500 |
+
head_mask=head_mask,
|
501 |
+
inputs_embeds=inputs_embeds,
|
502 |
+
labels=labels,
|
503 |
+
output_attentions=output_attentions,
|
504 |
+
output_hidden_states=output_hidden_states,
|
505 |
+
return_dict=return_dict,
|
506 |
+
)
|
507 |
+
else:
|
508 |
+
return cl_forward(self, self.roberta,
|
509 |
+
input_ids=input_ids,
|
510 |
+
attention_mask=attention_mask,
|
511 |
+
token_type_ids=token_type_ids,
|
512 |
+
position_ids=position_ids,
|
513 |
+
head_mask=head_mask,
|
514 |
+
inputs_embeds=inputs_embeds,
|
515 |
+
labels=labels,
|
516 |
+
output_attentions=output_attentions,
|
517 |
+
output_hidden_states=output_hidden_states,
|
518 |
+
return_dict=return_dict,
|
519 |
+
mlm_input_ids=mlm_input_ids,
|
520 |
+
mlm_labels=mlm_labels,
|
521 |
+
left_emb=left_emb,
|
522 |
+
right_emb=right_emb,
|
523 |
+
)
|
524 |
+
|