Spaces:
Running
Running
Commit
·
0d88c28
1
Parent(s):
b0ebb46
Upload 2 files
Browse files- models/code/code_classification.py +544 -0
- models/code/code_generation.py +201 -0
models/code/code_classification.py
ADDED
@@ -0,0 +1,544 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
# @Time : 2023/3/11 8:02 上午
|
3 |
+
# @Author : NuoChen
|
4 |
+
# @File : code_classification.py
|
5 |
+
|
6 |
+
## ======== Roberta ========
|
7 |
+
import torch
|
8 |
+
from torch import nn
|
9 |
+
from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
|
10 |
+
from transformers import RobertaModel
|
11 |
+
from transformers.activations import ACT2FN
|
12 |
+
from transformers.models.electra import ElectraModel
|
13 |
+
from transformers.models.roformer import RoFormerModel
|
14 |
+
from transformers.models.albert import AlbertModel
|
15 |
+
from transformers.models.bert import BertModel, BertPreTrainedModel
|
16 |
+
from transformers.models.deberta_v2 import DebertaV2Model, DebertaV2PreTrainedModel
|
17 |
+
from transformers.modeling_outputs import SequenceClassifierOutput
|
18 |
+
from transformers.models.roberta import RobertaPreTrainedModel
|
19 |
+
from transformers.models.bert.modeling_bert import BertForSequenceClassification
|
20 |
+
from transformers.models.megatron_bert import MegatronBertPreTrainedModel, MegatronBertModel
|
21 |
+
import logging
|
22 |
+
from typing import Optional, List, Union, Tuple
|
23 |
+
import torch
|
24 |
+
from torch._C import NoopLogger
|
25 |
+
from torch.autograd import Variable
|
26 |
+
import copy
|
27 |
+
import torch.nn
|
28 |
+
import torch.nn.functional as F
|
29 |
+
from torch import Tensor
|
30 |
+
from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
|
31 |
+
|
32 |
+
from transformers import RobertaModel, RobertaPreTrainedModel
|
33 |
+
from transformers.models.plbart.modeling_plbart import PLBartPreTrainedModel, PLBartClassificationHead, PLBartModel
|
34 |
+
from transformers.models.plbart.configuration_plbart import PLBartConfig
|
35 |
+
from transformers.models.t5.modeling_t5 import T5PreTrainedModel#, T5ClassificationHead, T5Model
|
36 |
+
from transformers.models.t5.modeling_t5 import T5Config,T5Stack
|
37 |
+
from transformers.modeling_outputs import SequenceClassifierOutput, Seq2SeqSequenceClassifierOutput, SequenceClassifierOutputWithPast
|
38 |
+
from models.basic_modules.prefix_encoder import PrefixEncoder
|
39 |
+
|
40 |
+
from models.basic_modules.adapter import BertAdaModel, RobertaAdaModel, init_adapter
|
41 |
+
from tools.model_utils.parameter_freeze import ParameterFreeze
|
42 |
+
|
43 |
+
freezer = ParameterFreeze()
|
44 |
+
|
45 |
+
## ======== Roberta ========
|
46 |
+
# Vanilla Fine-tuning For Roberta
|
47 |
+
class RobertaForCodeClassification(RobertaPreTrainedModel):
|
48 |
+
def __init__(self, config):
|
49 |
+
super().__init__(config)
|
50 |
+
self.num_labels = config.num_labels
|
51 |
+
self.config = config
|
52 |
+
self.roberta = RobertaModel(config)
|
53 |
+
if self.config.use_freezing:
|
54 |
+
self.roberta = freezer.freeze_lm(self.roberta)
|
55 |
+
self.dropout = torch.nn.Dropout(config.hidden_dropout_prob)
|
56 |
+
self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels)
|
57 |
+
self.init_weights()
|
58 |
+
|
59 |
+
def freeze_backbone(self, use_freezing: bool=True):
|
60 |
+
if use_freezing:
|
61 |
+
self.roberta = freezer.freeze_lm(self.roberta)
|
62 |
+
else:
|
63 |
+
self.roberta = freezer.unfreeze_lm(self.roberta)
|
64 |
+
|
65 |
+
def forward(
|
66 |
+
self,
|
67 |
+
input_ids=None,
|
68 |
+
attention_mask=None,
|
69 |
+
token_type_ids=None,
|
70 |
+
position_ids=None,
|
71 |
+
head_mask=None,
|
72 |
+
inputs_embeds=None,
|
73 |
+
labels=None,
|
74 |
+
output_attentions=None,
|
75 |
+
output_hidden_states=None,
|
76 |
+
return_dict=None,
|
77 |
+
):
|
78 |
+
r"""
|
79 |
+
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
|
80 |
+
Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
|
81 |
+
config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
|
82 |
+
If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
83 |
+
"""
|
84 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
85 |
+
|
86 |
+
outputs = self.roberta(
|
87 |
+
input_ids,
|
88 |
+
attention_mask=attention_mask,
|
89 |
+
token_type_ids=token_type_ids,
|
90 |
+
position_ids=position_ids,
|
91 |
+
head_mask=head_mask,
|
92 |
+
inputs_embeds=inputs_embeds,
|
93 |
+
output_attentions=output_attentions,
|
94 |
+
output_hidden_states=output_hidden_states,
|
95 |
+
return_dict=return_dict,
|
96 |
+
)
|
97 |
+
|
98 |
+
pooled_output = outputs[1]
|
99 |
+
|
100 |
+
pooled_output = self.dropout(pooled_output)
|
101 |
+
logits = self.classifier(pooled_output)
|
102 |
+
|
103 |
+
loss = None
|
104 |
+
if labels is not None:
|
105 |
+
loss_fct = CrossEntropyLoss()
|
106 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
107 |
+
|
108 |
+
if not return_dict:
|
109 |
+
output = (logits,) + outputs[2:]
|
110 |
+
return ((loss,) + output) if loss is not None else output
|
111 |
+
|
112 |
+
return SequenceClassifierOutput(
|
113 |
+
loss=loss,
|
114 |
+
logits=logits,
|
115 |
+
hidden_states=outputs.hidden_states,
|
116 |
+
attentions=outputs.attentions,
|
117 |
+
)
|
118 |
+
|
119 |
+
## ======== CodeBERT ========
|
120 |
+
# Vanilla Fine-tuning For CodeBERT
|
121 |
+
class CodeBERTForCodeClassification(RobertaPreTrainedModel):
|
122 |
+
def __init__(self, config):
|
123 |
+
super().__init__(config)
|
124 |
+
self.num_labels = config.num_labels
|
125 |
+
self.config = config
|
126 |
+
self.roberta = RobertaModel(config)
|
127 |
+
if self.config.use_freezing:
|
128 |
+
self.roberta = freezer.freeze_lm(self.roberta)
|
129 |
+
self.dropout = torch.nn.Dropout(config.hidden_dropout_prob)
|
130 |
+
self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels)
|
131 |
+
self.init_weights()
|
132 |
+
|
133 |
+
def freeze_backbone(self, use_freezing: bool=True):
|
134 |
+
if use_freezing:
|
135 |
+
self.roberta = freezer.freeze_lm(self.roberta)
|
136 |
+
else:
|
137 |
+
self.roberta = freezer.unfreeze_lm(self.roberta)
|
138 |
+
|
139 |
+
def forward(
|
140 |
+
self,
|
141 |
+
input_ids=None,
|
142 |
+
attention_mask=None,
|
143 |
+
token_type_ids=None,
|
144 |
+
position_ids=None,
|
145 |
+
head_mask=None,
|
146 |
+
inputs_embeds=None,
|
147 |
+
labels=None,
|
148 |
+
output_attentions=None,
|
149 |
+
output_hidden_states=None,
|
150 |
+
return_dict=None,
|
151 |
+
):
|
152 |
+
r"""
|
153 |
+
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
|
154 |
+
Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
|
155 |
+
config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
|
156 |
+
If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
157 |
+
"""
|
158 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
159 |
+
|
160 |
+
outputs = self.roberta(
|
161 |
+
input_ids,
|
162 |
+
attention_mask=attention_mask,
|
163 |
+
token_type_ids=token_type_ids,
|
164 |
+
position_ids=position_ids,
|
165 |
+
head_mask=head_mask,
|
166 |
+
inputs_embeds=inputs_embeds,
|
167 |
+
output_attentions=output_attentions,
|
168 |
+
output_hidden_states=output_hidden_states,
|
169 |
+
return_dict=return_dict,
|
170 |
+
)
|
171 |
+
|
172 |
+
pooled_output = outputs[1]
|
173 |
+
|
174 |
+
pooled_output = self.dropout(pooled_output)
|
175 |
+
logits = self.classifier(pooled_output)
|
176 |
+
|
177 |
+
loss = None
|
178 |
+
if labels is not None:
|
179 |
+
if self.config.problem_type is None:
|
180 |
+
if self.num_labels == 1:
|
181 |
+
self.config.problem_type = "regression"
|
182 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
183 |
+
self.config.problem_type = "single_label_classification"
|
184 |
+
else:
|
185 |
+
self.config.problem_type = "multi_label_classification"
|
186 |
+
|
187 |
+
if self.config.problem_type == "regression":
|
188 |
+
loss_fct = MSELoss()
|
189 |
+
if self.num_labels == 1:
|
190 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
191 |
+
else:
|
192 |
+
loss = loss_fct(logits, labels)
|
193 |
+
elif self.config.problem_type == "single_label_classification":
|
194 |
+
loss_fct = CrossEntropyLoss()
|
195 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
196 |
+
elif self.config.problem_type == "multi_label_classification":
|
197 |
+
loss_fct = BCEWithLogitsLoss()
|
198 |
+
loss = loss_fct(logits, labels)
|
199 |
+
if not return_dict:
|
200 |
+
output = (logits,) + outputs[2:]
|
201 |
+
return ((loss,) + output) if loss is not None else output
|
202 |
+
|
203 |
+
return SequenceClassifierOutput(
|
204 |
+
loss=loss,
|
205 |
+
logits=logits,
|
206 |
+
hidden_states=outputs.hidden_states,
|
207 |
+
attentions=outputs.attentions,
|
208 |
+
)
|
209 |
+
|
210 |
+
## ======== GraphCodeBERT ========
|
211 |
+
|
212 |
+
# Vanilla Fine-tuning For GraphCodeBERT
|
213 |
+
class GraphCodeBERTForCodeClassification(RobertaPreTrainedModel):
|
214 |
+
def __init__(self, config):
|
215 |
+
super().__init__(config)
|
216 |
+
self.num_labels = config.num_labels
|
217 |
+
self.config = config
|
218 |
+
self.roberta = RobertaModel(config)
|
219 |
+
if self.config.use_freezing:
|
220 |
+
self.roberta = freezer.freeze_lm(self.roberta)
|
221 |
+
self.dropout = torch.nn.Dropout(config.hidden_dropout_prob)
|
222 |
+
self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels)
|
223 |
+
self.init_weights()
|
224 |
+
|
225 |
+
def freeze_backbone(self, use_freezing: bool=True):
|
226 |
+
if use_freezing:
|
227 |
+
self.roberta = freezer.freeze_lm(self.roberta)
|
228 |
+
else:
|
229 |
+
self.roberta = freezer.unfreeze_lm(self.roberta)
|
230 |
+
|
231 |
+
def forward(
|
232 |
+
self,
|
233 |
+
input_ids=None,
|
234 |
+
attention_mask=None,
|
235 |
+
token_type_ids=None,
|
236 |
+
position_ids=None,
|
237 |
+
head_mask=None,
|
238 |
+
inputs_embeds=None,
|
239 |
+
labels=None,
|
240 |
+
output_attentions=None,
|
241 |
+
output_hidden_states=None,
|
242 |
+
return_dict=None,
|
243 |
+
):
|
244 |
+
r"""
|
245 |
+
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
|
246 |
+
Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
|
247 |
+
config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
|
248 |
+
If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
249 |
+
"""
|
250 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
251 |
+
|
252 |
+
outputs = self.roberta(
|
253 |
+
input_ids,
|
254 |
+
attention_mask=attention_mask,
|
255 |
+
token_type_ids=token_type_ids,
|
256 |
+
position_ids=position_ids,
|
257 |
+
head_mask=head_mask,
|
258 |
+
inputs_embeds=inputs_embeds,
|
259 |
+
output_attentions=output_attentions,
|
260 |
+
output_hidden_states=output_hidden_states,
|
261 |
+
return_dict=return_dict,
|
262 |
+
)
|
263 |
+
|
264 |
+
pooled_output = outputs[1]
|
265 |
+
|
266 |
+
pooled_output = self.dropout(pooled_output)
|
267 |
+
logits = self.classifier(pooled_output)
|
268 |
+
|
269 |
+
loss = None
|
270 |
+
if labels is not None:
|
271 |
+
loss_fct = CrossEntropyLoss()
|
272 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
273 |
+
|
274 |
+
if not return_dict:
|
275 |
+
output = (logits,) + outputs[2:]
|
276 |
+
return ((loss,) + output) if loss is not None else output
|
277 |
+
|
278 |
+
return SequenceClassifierOutput(
|
279 |
+
loss=loss,
|
280 |
+
logits=logits,
|
281 |
+
hidden_states=outputs.hidden_states,
|
282 |
+
attentions=outputs.attentions,
|
283 |
+
)
|
284 |
+
|
285 |
+
## ======== PLBART ========
|
286 |
+
|
287 |
+
# Vanilla Fine-tuning For PLBART
|
288 |
+
class PLBARTForCodeClassification(PLBartPreTrainedModel):
|
289 |
+
|
290 |
+
_keys_to_ignore_on_load_missing = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]
|
291 |
+
|
292 |
+
def __init__(self, config: PLBartConfig, **kwargs):
|
293 |
+
super().__init__(config, **kwargs)
|
294 |
+
self.model = PLBartModel(config)
|
295 |
+
self.classification_head = PLBartClassificationHead(
|
296 |
+
config.d_model,
|
297 |
+
config.d_model,
|
298 |
+
config.num_labels,
|
299 |
+
config.classifier_dropout,
|
300 |
+
)
|
301 |
+
self.model._init_weights(self.classification_head.dense)
|
302 |
+
self.model._init_weights(self.classification_head.out_proj)
|
303 |
+
|
304 |
+
|
305 |
+
# Copied from transformers.models.bart.modeling_bart.BartForSequenceClassification.forward
|
306 |
+
def forward(
|
307 |
+
self,
|
308 |
+
input_ids: torch.LongTensor = None,
|
309 |
+
attention_mask: Optional[torch.Tensor] = None,
|
310 |
+
decoder_input_ids: Optional[torch.LongTensor] = None,
|
311 |
+
decoder_attention_mask: Optional[torch.LongTensor] = None,
|
312 |
+
head_mask: Optional[torch.Tensor] = None,
|
313 |
+
decoder_head_mask: Optional[torch.Tensor] = None,
|
314 |
+
cross_attn_head_mask: Optional[torch.Tensor] = None,
|
315 |
+
encoder_outputs: Optional[List[torch.FloatTensor]] = None,
|
316 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
317 |
+
decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
|
318 |
+
labels: Optional[torch.LongTensor] = None,
|
319 |
+
use_cache: Optional[bool] = None,
|
320 |
+
output_attentions: Optional[bool] = None,
|
321 |
+
output_hidden_states: Optional[bool] = None,
|
322 |
+
return_dict: Optional[bool] = None,
|
323 |
+
) -> Union[Tuple, Seq2SeqSequenceClassifierOutput]:
|
324 |
+
r"""
|
325 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
326 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
327 |
+
config.num_labels - 1]`. If `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
328 |
+
"""
|
329 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
330 |
+
if labels is not None:
|
331 |
+
use_cache = False
|
332 |
+
|
333 |
+
if input_ids is None and inputs_embeds is not None:
|
334 |
+
raise NotImplementedError(
|
335 |
+
f"Passing input embeddings is currently not supported for {self.__class__.__name__}"
|
336 |
+
)
|
337 |
+
|
338 |
+
outputs = self.model(
|
339 |
+
input_ids,
|
340 |
+
attention_mask=attention_mask,
|
341 |
+
decoder_input_ids=decoder_input_ids,
|
342 |
+
decoder_attention_mask=decoder_attention_mask,
|
343 |
+
head_mask=head_mask,
|
344 |
+
decoder_head_mask=decoder_head_mask,
|
345 |
+
cross_attn_head_mask=cross_attn_head_mask,
|
346 |
+
encoder_outputs=encoder_outputs,
|
347 |
+
inputs_embeds=inputs_embeds,
|
348 |
+
decoder_inputs_embeds=decoder_inputs_embeds,
|
349 |
+
use_cache=use_cache,
|
350 |
+
output_attentions=output_attentions,
|
351 |
+
output_hidden_states=output_hidden_states,
|
352 |
+
return_dict=return_dict,
|
353 |
+
)
|
354 |
+
hidden_states = outputs[0] # last hidden state
|
355 |
+
|
356 |
+
eos_mask = input_ids.eq(self.config.eos_token_id).to(hidden_states.device)
|
357 |
+
|
358 |
+
if len(torch.unique_consecutive(eos_mask.sum(1))) > 1:
|
359 |
+
raise ValueError("All examples must have the same number of <eos> tokens.")
|
360 |
+
sentence_representation = hidden_states[eos_mask, :].view(hidden_states.size(0), -1, hidden_states.size(-1))[
|
361 |
+
:, -1, :
|
362 |
+
]
|
363 |
+
logits = self.classification_head(sentence_representation)
|
364 |
+
|
365 |
+
loss = None
|
366 |
+
if labels is not None:
|
367 |
+
if self.config.problem_type is None:
|
368 |
+
if self.config.num_labels == 1:
|
369 |
+
self.config.problem_type = "regression"
|
370 |
+
elif self.config.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
371 |
+
self.config.problem_type = "single_label_classification"
|
372 |
+
else:
|
373 |
+
self.config.problem_type = "multi_label_classification"
|
374 |
+
|
375 |
+
if self.config.problem_type == "regression":
|
376 |
+
loss_fct = MSELoss()
|
377 |
+
if self.config.num_labels == 1:
|
378 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
379 |
+
else:
|
380 |
+
loss = loss_fct(logits, labels)
|
381 |
+
elif self.config.problem_type == "single_label_classification":
|
382 |
+
loss_fct = CrossEntropyLoss()
|
383 |
+
loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
|
384 |
+
elif self.config.problem_type == "multi_label_classification":
|
385 |
+
loss_fct = BCEWithLogitsLoss()
|
386 |
+
loss = loss_fct(logits, labels)
|
387 |
+
if not return_dict:
|
388 |
+
output = (logits,) + outputs[1:]
|
389 |
+
return ((loss,) + output) if loss is not None else output
|
390 |
+
|
391 |
+
return Seq2SeqSequenceClassifierOutput(
|
392 |
+
loss=loss,
|
393 |
+
logits=logits,
|
394 |
+
past_key_values=outputs.past_key_values,
|
395 |
+
decoder_hidden_states=outputs.decoder_hidden_states,
|
396 |
+
decoder_attentions=outputs.decoder_attentions,
|
397 |
+
cross_attentions=outputs.cross_attentions,
|
398 |
+
encoder_last_hidden_state=outputs.encoder_last_hidden_state,
|
399 |
+
encoder_hidden_states=outputs.encoder_hidden_states,
|
400 |
+
encoder_attentions=outputs.encoder_attentions,
|
401 |
+
)
|
402 |
+
|
403 |
+
|
404 |
+
## ======== CodeT5 ========
|
405 |
+
|
406 |
+
# Vanilla Fine-tuning For CodeT5
|
407 |
+
class CodeT5ForCodeClassification(T5PreTrainedModel):
|
408 |
+
_keys_to_ignore_on_load_missing = [r"encoder.embed_tokens.weight"]
|
409 |
+
|
410 |
+
def __init__(self, config: T5Config):
|
411 |
+
super().__init__(config)
|
412 |
+
self.model_dim = config.d_model
|
413 |
+
self.config.problem_type = None
|
414 |
+
self.config.is_encoder_decoder = False
|
415 |
+
|
416 |
+
self.shared = nn.Embedding(config.vocab_size, config.d_model)
|
417 |
+
|
418 |
+
encoder_config = copy.deepcopy(config)
|
419 |
+
encoder_config.is_decoder = False
|
420 |
+
encoder_config.is_encoder_decoder = False
|
421 |
+
encoder_config.use_cache = False
|
422 |
+
self.encoder = T5Stack(encoder_config, self.shared)
|
423 |
+
|
424 |
+
classifier_dropout = (
|
425 |
+
config.classifier_dropout if hasattr(config, 'classifier_dropout') else config.dropout_rate
|
426 |
+
)
|
427 |
+
self.dropout = nn.Dropout(classifier_dropout)
|
428 |
+
self.classifier = nn.Linear(config.d_model, config.num_labels)
|
429 |
+
|
430 |
+
# Initialize weights and apply final processing
|
431 |
+
self.post_init()
|
432 |
+
|
433 |
+
# Model parallel
|
434 |
+
self.model_parallel = False
|
435 |
+
self.device_map = None
|
436 |
+
|
437 |
+
def parallelize(self, device_map=None):
|
438 |
+
self.device_map = (
|
439 |
+
get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))
|
440 |
+
if device_map is None
|
441 |
+
else device_map
|
442 |
+
)
|
443 |
+
assert_device_map(self.device_map, len(self.encoder.block))
|
444 |
+
self.encoder.parallelize(self.device_map)
|
445 |
+
self.classifier.to(self.encoder.first_device)
|
446 |
+
self.model_parallel = True
|
447 |
+
|
448 |
+
def deparallelize(self):
|
449 |
+
self.encoder.deparallelize()
|
450 |
+
self.encoder = self.encoder.to("cpu")
|
451 |
+
self.classifier = self.classifier.to("cpu")
|
452 |
+
self.model_parallel = False
|
453 |
+
self.device_map = None
|
454 |
+
torch.cuda.empty_cache()
|
455 |
+
|
456 |
+
def get_input_embeddings(self):
|
457 |
+
return self.shared
|
458 |
+
|
459 |
+
def set_input_embeddings(self, new_embeddings):
|
460 |
+
self.shared = new_embeddings
|
461 |
+
self.encoder.set_input_embeddings(new_embeddings)
|
462 |
+
|
463 |
+
def get_encoder(self):
|
464 |
+
return self.encoder
|
465 |
+
|
466 |
+
def _prune_heads(self, heads_to_prune):
|
467 |
+
"""
|
468 |
+
Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
|
469 |
+
class PreTrainedModel
|
470 |
+
"""
|
471 |
+
for layer, heads in heads_to_prune.items():
|
472 |
+
self.encoder.block[layer].layer[0].SelfAttention.prune_heads(heads)
|
473 |
+
|
474 |
+
def forward(
|
475 |
+
self,
|
476 |
+
input_ids: Optional[torch.LongTensor] = None,
|
477 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
478 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
479 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
480 |
+
labels: Optional[torch.LongTensor] = None,
|
481 |
+
output_attentions: Optional[bool] = None,
|
482 |
+
output_hidden_states: Optional[bool] = None,
|
483 |
+
return_dict: Optional[bool] = None,
|
484 |
+
) -> Union[Tuple[torch.FloatTensor], SequenceClassifierOutput]:
|
485 |
+
r"""
|
486 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
487 |
+
Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
|
488 |
+
Returns:
|
489 |
+
"""
|
490 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
491 |
+
|
492 |
+
outputs = self.encoder(
|
493 |
+
input_ids=input_ids,
|
494 |
+
attention_mask=attention_mask,
|
495 |
+
inputs_embeds=inputs_embeds,
|
496 |
+
head_mask=head_mask,
|
497 |
+
output_attentions=output_attentions,
|
498 |
+
output_hidden_states=output_hidden_states,
|
499 |
+
return_dict=return_dict,
|
500 |
+
)
|
501 |
+
|
502 |
+
# Get last hidden indices
|
503 |
+
# (batch_size) -> (batch_size, 1) -> (batch_size, hidden_size) -> (batch_size, 1, hidden_size)
|
504 |
+
last_hidden_indices = (
|
505 |
+
(input_ids != self.config.pad_token_id).sum(dim=-1) - 1
|
506 |
+
).unsqueeze(dim=-1).repeat(1, outputs[0].size(-1)).unsqueeze(1)
|
507 |
+
sequence_output = outputs[0].gather(dim=1, index=last_hidden_indices).squeeze(1)
|
508 |
+
|
509 |
+
sequence_output = self.dropout(sequence_output)
|
510 |
+
logits = self.classifier(sequence_output)
|
511 |
+
|
512 |
+
loss = None
|
513 |
+
if labels is not None:
|
514 |
+
if self.config.problem_type is None:
|
515 |
+
if self.config.num_labels == 1:
|
516 |
+
self.config.problem_type = "regression"
|
517 |
+
elif self.config.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
518 |
+
self.config.problem_type = "single_label_classification"
|
519 |
+
else:
|
520 |
+
self.config.problem_type = "multi_label_classification"
|
521 |
+
|
522 |
+
if self.config.problem_type == "regression":
|
523 |
+
loss_fct = nn.MSELoss()
|
524 |
+
if self.config.num_labels == 1:
|
525 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
526 |
+
else:
|
527 |
+
loss = loss_fct(logits, labels)
|
528 |
+
elif self.config.problem_type == "single_label_classification":
|
529 |
+
loss_fct = nn.CrossEntropyLoss()
|
530 |
+
loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
|
531 |
+
elif self.config.problem_type == "multi_label_classification":
|
532 |
+
loss_fct = nn.BCEWithLogitsLoss()
|
533 |
+
loss = loss_fct(logits, labels)
|
534 |
+
|
535 |
+
if not return_dict:
|
536 |
+
output = (logits,) + outputs[2:]
|
537 |
+
return ((loss,) + output) if loss is not None else output
|
538 |
+
|
539 |
+
return SequenceClassifierOutput(
|
540 |
+
loss=loss,
|
541 |
+
logits=logits,
|
542 |
+
hidden_states=outputs.hidden_states,
|
543 |
+
attentions=outputs.attentions
|
544 |
+
)
|
models/code/code_generation.py
ADDED
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
# @Time : 2023/4/05 18:02 下午
|
3 |
+
# @Author : NuoChen
|
4 |
+
# @File : code_generation.py
|
5 |
+
|
6 |
+
from transformers import PLBartTokenizer, PLBartForSequenceClassification, PLBartConfig, PLBartForConditionalGeneration
|
7 |
+
from typing import Any, Dict, List, Optional, Tuple, Union
|
8 |
+
from transformers.modeling_outputs import (
|
9 |
+
BaseModelOutput,
|
10 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
11 |
+
CausalLMOutputWithCrossAttentions,
|
12 |
+
Seq2SeqLMOutput,
|
13 |
+
Seq2SeqModelOutput,
|
14 |
+
Seq2SeqSequenceClassifierOutput,
|
15 |
+
)
|
16 |
+
import torch
|
17 |
+
from torch import nn
|
18 |
+
from typing import Optional, List, Union, Tuple
|
19 |
+
from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
|
20 |
+
|
21 |
+
from transformers import RobertaModel, RobertaPreTrainedModel
|
22 |
+
from transformers.models.plbart.modeling_plbart import PLBartPreTrainedModel, PLBartModel
|
23 |
+
from transformers.models.plbart.configuration_plbart import PLBartConfig
|
24 |
+
|
25 |
+
|
26 |
+
def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int):
|
27 |
+
"""
|
28 |
+
Shift input ids one token to the right, and wrap the last non pad token (the <LID> token) Note that MBart does not
|
29 |
+
have a single `decoder_start_token_id` in contrast to other Bart-like models.
|
30 |
+
"""
|
31 |
+
prev_output_tokens = input_ids.clone()
|
32 |
+
|
33 |
+
if pad_token_id is None:
|
34 |
+
raise ValueError("self.model.config.pad_token_id has to be defined.")
|
35 |
+
# replace possible -100 values in labels by `pad_token_id`
|
36 |
+
prev_output_tokens.masked_fill_(prev_output_tokens == -100, pad_token_id)
|
37 |
+
|
38 |
+
index_of_eos = (prev_output_tokens.ne(pad_token_id).sum(dim=1) - 1).unsqueeze(-1)
|
39 |
+
decoder_start_tokens = prev_output_tokens.gather(1, index_of_eos).squeeze()
|
40 |
+
prev_output_tokens[:, 1:] = prev_output_tokens[:, :-1].clone()
|
41 |
+
prev_output_tokens[:, 0] = decoder_start_tokens
|
42 |
+
|
43 |
+
return prev_output_tokens
|
44 |
+
|
45 |
+
class PLBARTForCodeGeneration(PLBartPreTrainedModel):
|
46 |
+
base_model_prefix = "model"
|
47 |
+
_keys_to_ignore_on_load_missing = [
|
48 |
+
r"final_logits_bias",
|
49 |
+
r"encoder.version",
|
50 |
+
r"decoder.version",
|
51 |
+
r"lm_head.weight",
|
52 |
+
]
|
53 |
+
|
54 |
+
def __init__(self, config: PLBartConfig):
|
55 |
+
super().__init__(config)
|
56 |
+
self.model = PLBartModel(config)
|
57 |
+
self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
|
58 |
+
self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)
|
59 |
+
|
60 |
+
self.init_weights()
|
61 |
+
|
62 |
+
def get_encoder(self):
|
63 |
+
return self.model.get_encoder()
|
64 |
+
|
65 |
+
def get_decoder(self):
|
66 |
+
return self.model.get_decoder()
|
67 |
+
|
68 |
+
def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
|
69 |
+
new_embeddings = super().resize_token_embeddings(new_num_tokens)
|
70 |
+
self._resize_final_logits_bias(new_num_tokens)
|
71 |
+
return new_embeddings
|
72 |
+
|
73 |
+
def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
|
74 |
+
old_num_tokens = self.final_logits_bias.shape[-1]
|
75 |
+
if new_num_tokens <= old_num_tokens:
|
76 |
+
new_bias = self.final_logits_bias[:, :new_num_tokens]
|
77 |
+
else:
|
78 |
+
extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
|
79 |
+
new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
|
80 |
+
self.register_buffer("final_logits_bias", new_bias)
|
81 |
+
|
82 |
+
def get_output_embeddings(self):
|
83 |
+
return self.lm_head
|
84 |
+
|
85 |
+
def set_output_embeddings(self, new_embeddings):
|
86 |
+
self.lm_head = new_embeddings
|
87 |
+
|
88 |
+
def forward(
|
89 |
+
self,
|
90 |
+
input_ids: Optional[torch.LongTensor] = None,
|
91 |
+
attention_mask: Optional[torch.LongTensor] = None,
|
92 |
+
decoder_input_ids: Optional[torch.LongTensor] = None,
|
93 |
+
decoder_attention_mask: Optional[torch.Tensor] = None,
|
94 |
+
head_mask: Optional[torch.Tensor] = None,
|
95 |
+
decoder_head_mask: Optional[torch.LongTensor] = None,
|
96 |
+
cross_attn_head_mask: Optional[torch.Tensor] = None,
|
97 |
+
encoder_outputs: Optional[List[torch.FloatTensor]] = None,
|
98 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
99 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
100 |
+
decoder_inputs_embeds=None,
|
101 |
+
labels: Optional[torch.Tensor] = None,
|
102 |
+
use_cache: Optional[bool] = None,
|
103 |
+
output_attentions: Optional[bool] = None,
|
104 |
+
output_hidden_states: Optional[bool] = None,
|
105 |
+
return_dict: Optional[bool] = None,
|
106 |
+
) -> Union[Tuple[torch.Tensor], Seq2SeqLMOutput]:
|
107 |
+
r"""
|
108 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
109 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
110 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
111 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
112 |
+
|
113 |
+
Returns:
|
114 |
+
|
115 |
+
"""
|
116 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
117 |
+
|
118 |
+
if labels is not None:
|
119 |
+
if decoder_input_ids is None:
|
120 |
+
decoder_input_ids = shift_tokens_right(labels, self.config.pad_token_id)
|
121 |
+
|
122 |
+
outputs = self.model(
|
123 |
+
input_ids,
|
124 |
+
attention_mask=attention_mask,
|
125 |
+
decoder_input_ids=decoder_input_ids,
|
126 |
+
encoder_outputs=encoder_outputs,
|
127 |
+
decoder_attention_mask=decoder_attention_mask,
|
128 |
+
head_mask=head_mask,
|
129 |
+
decoder_head_mask=decoder_head_mask,
|
130 |
+
cross_attn_head_mask=cross_attn_head_mask,
|
131 |
+
past_key_values=past_key_values,
|
132 |
+
inputs_embeds=inputs_embeds,
|
133 |
+
decoder_inputs_embeds=decoder_inputs_embeds,
|
134 |
+
use_cache=use_cache,
|
135 |
+
output_attentions=output_attentions,
|
136 |
+
output_hidden_states=output_hidden_states,
|
137 |
+
return_dict=return_dict,
|
138 |
+
)
|
139 |
+
lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias
|
140 |
+
|
141 |
+
masked_lm_loss = None
|
142 |
+
if labels is not None:
|
143 |
+
loss_fct = CrossEntropyLoss()
|
144 |
+
masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
|
145 |
+
|
146 |
+
if not return_dict:
|
147 |
+
output = (lm_logits,) + outputs[1:]
|
148 |
+
return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
|
149 |
+
|
150 |
+
return Seq2SeqLMOutput(
|
151 |
+
loss=masked_lm_loss,
|
152 |
+
logits=lm_logits,
|
153 |
+
past_key_values=outputs.past_key_values,
|
154 |
+
decoder_hidden_states=outputs.decoder_hidden_states,
|
155 |
+
decoder_attentions=outputs.decoder_attentions,
|
156 |
+
cross_attentions=outputs.cross_attentions,
|
157 |
+
encoder_last_hidden_state=outputs.encoder_last_hidden_state,
|
158 |
+
encoder_hidden_states=outputs.encoder_hidden_states,
|
159 |
+
encoder_attentions=outputs.encoder_attentions,
|
160 |
+
)
|
161 |
+
|
162 |
+
def prepare_inputs_for_generation(
|
163 |
+
self,
|
164 |
+
decoder_input_ids: torch.LongTensor,
|
165 |
+
past: Optional[List[torch.FloatTensor]] = None,
|
166 |
+
attention_mask: Optional[torch.LongTensor] = None,
|
167 |
+
head_mask: Optional[torch.Tensor] = None,
|
168 |
+
decoder_head_mask: Optional[torch.Tensor] = None,
|
169 |
+
cross_attn_head_mask: Optional[torch.Tensor] = None,
|
170 |
+
use_cache: Optional[bool] = None,
|
171 |
+
encoder_outputs: Optional[List[torch.FloatTensor]] = None,
|
172 |
+
**kwargs # TODO: Check if this is needed. It is unused?
|
173 |
+
) -> Dict[str, Any]:
|
174 |
+
# cut decoder_input_ids if past is used
|
175 |
+
if past is not None:
|
176 |
+
decoder_input_ids = decoder_input_ids[:, -1:]
|
177 |
+
|
178 |
+
return {
|
179 |
+
"input_ids": None, # encoder_outputs is defined. input_ids not needed
|
180 |
+
"encoder_outputs": encoder_outputs,
|
181 |
+
"past_key_values": past,
|
182 |
+
"decoder_input_ids": decoder_input_ids,
|
183 |
+
"attention_mask": attention_mask,
|
184 |
+
"head_mask": head_mask,
|
185 |
+
"decoder_head_mask": decoder_head_mask,
|
186 |
+
"cross_attn_head_mask": cross_attn_head_mask,
|
187 |
+
"use_cache": use_cache, # change this to avoid caching (presumably for debugging)
|
188 |
+
}
|
189 |
+
|
190 |
+
def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
|
191 |
+
return shift_tokens_right(labels, self.config.pad_token_id)
|
192 |
+
|
193 |
+
@staticmethod
|
194 |
+
def _reorder_cache(past, beam_idx):
|
195 |
+
reordered_past = ()
|
196 |
+
for layer_past in past:
|
197 |
+
# cached cross_attention states don't have to be reordered -> they are always the same
|
198 |
+
reordered_past += (
|
199 |
+
tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
|
200 |
+
)
|
201 |
+
return reordered_past
|