-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathmain_exp.py
More file actions
339 lines (283 loc) · 18.7 KB
/
main_exp.py
File metadata and controls
339 lines (283 loc) · 18.7 KB
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
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
import os
# os.environ['CUDA_VISIBLE_DEVICES']='0,1,3'
import argparse
from model_utils.exp_model_utils import add_optimizer_args, configure_optimizers, get_total_steps
from dataset_utils.clip_dataset_utils import add_datasets_args, load_data_raw
from dataset_utils.rec_universal_datamodule import UniversalDataModule
from pytorch_lightning import Trainer, LightningModule
from pytorch_lightning.callbacks.early_stopping import EarlyStopping
from ckpt_utils.exp_universal_checkpoint import UniversalCheckpoint
import torch.nn as nn
from torch.utils.data._utils.collate import default_collate
from pytorch_lightning.strategies.ddp import DDPStrategy
import torch
from transformers import T5Tokenizer, T5EncoderModel, T5ForConditionalGeneration
from models.t5_adapter import T5Attention_adapter, T5DenseActDense_adapter, T5DenseActDense_adapter_FFT
from models.t5_explain import Fuse_add_encoder_rep
import numpy as np
import copy
from models.evaluate_utils import rouge_score, bleu_score
def report_memory(name):
"""Simple GPU memory report."""
mega_bytes = 1024.0 * 1024.0
string = name + ' memory (MB)'
string += ' | allocated: {}'.format(torch.cuda.memory_allocated() / mega_bytes)
string += ' | max allocated: {}'.format(torch.cuda.max_memory_allocated() / mega_bytes)
string += ' | reserved: {}'.format(torch.cuda.memory_reserved() / mega_bytes)
string += ' | max reserved: {}'.format(torch.cuda.max_memory_reserved() / mega_bytes)
print(string)
class Collator_exp_gep():
def __init__(self, args):
self.args = args
self.seq_item = 10
self.exp_review_len = 128
self.tokenizer = T5Tokenizer.from_pretrained('./pretrain_model/t5/t5-small')
self.title_seq = args.title_seq
def __call__(self, inputs):
examples = []
for input_temp in inputs: ## title, review, summary, description
example = {}
title_list = input_temp[0]
title_list = title_list[-self.seq_item:]
title_token = {'input_ids': [], 'attention_mask': []}
for title_temp in title_list[:-1]:
title_temp = 'cls ' + title_temp
token_temp = self.tokenizer(title_temp, truncation=True, padding='max_length', max_length=self.title_seq)
cut = sum(token_temp['attention_mask'])
title_token['input_ids'] += token_temp['input_ids'][:cut]
title_token['attention_mask'] += token_temp['attention_mask'][:cut]
## batch size > 1
title_token['input_ids'] = torch.tensor(title_token['input_ids'] + ((self.seq_item-1)*self.title_seq - len(title_token['input_ids']))*[0])
title_token['attention_mask'] = torch.tensor(title_token['attention_mask'] + ((self.seq_item-1)*self.title_seq - len(title_token['attention_mask']))*[0])
### batch size = 1
# title_token['input_ids'] = torch.tensor(title_token['input_ids'])
# title_token['attention_mask'] = torch.tensor(title_token['attention_mask'])
example['title_seq_token'] = title_token
exp_input = {'input_ids': [], 'attention_mask': []}
title_last_tokenizer = self.tokenizer(title_list[-1], truncation=True, padding='max_length', max_length=self.title_seq)
try:
des_temp = input_temp[-1][-1][0]
except:
des_temp = ''
description_last_tokenizer = self.tokenizer(des_temp, truncation=True, padding='max_length', max_length=self.exp_review_len)
input_ids_exp_input = (title_last_tokenizer['input_ids'] + description_last_tokenizer['input_ids'])[:self.exp_review_len]
attention_mask_exp_input = (title_last_tokenizer['attention_mask'] + description_last_tokenizer['attention_mask'])[:self.exp_review_len]
exp_input['input_ids'] = torch.tensor(input_ids_exp_input + (self.exp_review_len-len(input_ids_exp_input))*[0])
exp_input['attention_mask'] = torch.tensor(attention_mask_exp_input + (self.exp_review_len-len(attention_mask_exp_input))*[0])
example['exp_description'] = exp_input
example['review'] = self.tokenizer(input_temp[-2][-1], truncation=True, padding='max_length', max_length=self.exp_review_len, return_tensors="pt")
examples.append(example)
return default_collate(examples)
class Rec_seq_exp(LightningModule):
def add_module_specific_args(parent_parser):
parser = parent_parser.add_argument_group('Title_seq=15, item_num=10, small lr=0.0005')
parser.add_argument('--dataset_name', default='Instruments')
parser.add_argument('--hidden_size', type=int, default=512) ## T5-small: 512, T5-base: 768, T5-large: 1024
parser.add_argument('--title_seq', type=int, default=15)
parser.add_argument('--encoder_ckpt_path', default=None)
parser.add_argument('--seq_items', type=int, default=10)
return parent_parser
def __init__(self, args):
super().__init__()
self.tokenizer = T5Tokenizer.from_pretrained('./pretrain_model/t5/t5-small')
self.encoder = T5EncoderModel.from_pretrained('./pretrain_model/t5/t5-small')
self.exp_gen = T5ForConditionalGeneration.from_pretrained('./pretrain_model/t5/t5-small')
encoder_config = copy.deepcopy(self.encoder.config)
encoder_config.use_cache = False
encoder_config.is_encoder_decoder = False
for i in self.encoder.encoder.block:
temp_selfatt = i.layer[0].SelfAttention
try:
i.layer[0].SelfAttention.relative_attention_bias
hashas_relative_attention_bias = True
except:
hashas_relative_attention_bias = False
i.layer[0].SelfAttention = T5Attention_adapter(config=encoder_config, has_relative_attention_bias=hashas_relative_attention_bias)
for name_temp in temp_selfatt.named_parameters():
exec('i.layer[0].SelfAttention.{} = temp_selfatt.{}'.format(name_temp[0], name_temp[0]))
temp_ffn = i.layer[1].DenseReluDense
i.layer[1].DenseReluDense = T5DenseActDense_adapter_FFT(config=encoder_config, len_seq=args.title_seq*(args.seq_items), len_title=args.title_seq)
for name_temp in temp_ffn.named_parameters():
exec('i.layer[1].DenseReluDense.{} = temp_ffn.{}'.format(name_temp[0], name_temp[0]))
encoder_weight = torch.load(args.encoder_ckpt_path)['state_dict']
for key_temp in list(encoder_weight.keys()):
encoder_weight[key_temp.replace('encoder.', '', 1)] = encoder_weight.pop(key_temp)
self.encoder.load_state_dict(encoder_weight)
self.exp_gen.encoder = self.encoder.encoder
for param in self.encoder.parameters():
param.requires_grad = False
for param in self.exp_gen.parameters():
param.requires_grad = False
exp_gen_config = copy.deepcopy(self.exp_gen.config)
exp_gen_config.use_cache = False
exp_gen_config.is_encoder_decoder = False
for i in self.exp_gen.decoder.block:
"""
temp_selfatt = i.layer[0].SelfAttention
try:
i.layer[0].SelfAttention.relative_attention_bias
hashas_relative_attention_bias = True
except:
hashas_relative_attention_bias = False
i.layer[0].SelfAttention = T5Attention_adapter(config=exp_gen_config, has_relative_attention_bias=hashas_relative_attention_bias)
for name_temp in temp_selfatt.named_parameters():
exec('i.layer[0].SelfAttention.{} = temp_selfatt.{}'.format(name_temp[0], name_temp[0]))
"""
temp_ffn = i.layer[2].DenseReluDense
i.layer[2].DenseReluDense = T5DenseActDense_adapter_FFT(config=exp_gen_config, len_seq=args.title_seq*(10-1), len_title=args.title_seq)
# i.layer[1].DenseReluDense = T5DenseActDense_adapter(config=encoder_config)
for name_temp in temp_ffn.named_parameters():
exec('i.layer[2].DenseReluDense.{} = temp_ffn.{}'.format(name_temp[0], name_temp[0]))
for param in self.exp_gen.decoder.block[-1].parameters(): ## small: 5, base:11
param.requires_grad = True
## decoder fine-tuning full parameter fine-tuning
# for param in self.exp_gen.decoder.parameters(): ## small: 5, base:11
# param.requires_grad = True
self.title_seq = args.title_seq
self.PPF = Fuse_add_encoder_rep(args.hidden_size)
self.loss_ce = nn.CrossEntropyLoss()
self.save_hyperparameters(args)
def setup(self, stage) -> None:
if stage == 'fit':
self.total_steps = get_total_steps(self.trainer, self.hparams)
print('Total steps: {}' .format(self.total_steps))
def configure_optimizers(self):
return configure_optimizers(self)
def training_step(self, batch, batch_idx):
self.exp_gen.train()
seq_titles = batch['title_seq_token']
last_hidden_state = self.encoder(**seq_titles).last_hidden_state
rep_seq = torch.gather(last_hidden_state, 1, (torch.sum(seq_titles['attention_mask'], dim=-1)-1).unsqueeze(1).repeat(1, last_hidden_state.shape[-1]).unsqueeze(1)).squeeze(1)
rep_description_out = self.exp_gen.encoder(**batch['exp_description'], return_dict=True)
ppf_rep = self.PPF(rep_seq, rep_description_out.last_hidden_state, mask=(batch['exp_description']['input_ids']>0).float().to(rep_seq.device))
rep_description_out.last_hidden_state = ppf_rep
labels = batch['review']['input_ids'].to(rep_seq.device).squeeze(1)
labels[labels == self.tokenizer.pad_token_id] = -100
out_exp = self.exp_gen(input_ids=batch['exp_description']['input_ids'].to(rep_seq.device), labels=labels, encoder_outputs=rep_description_out, attention_mask=batch['exp_description']['attention_mask'].to(rep_seq.device), decoder_attention_mask=batch['review']['attention_mask'].to(rep_seq.device).squeeze(1))
loss = out_exp.loss
self.log("train_loss", loss.item(), on_epoch=False, prog_bar=True, logger=True)
if self.trainer.global_rank == 0 and self.global_step == 100:
report_memory('Seq exp gen')
return {"loss": loss}
def validation_step(self, batch, batch_idx):
self.exp_gen.eval()
seq_titles = batch['title_seq_token']
last_hidden_state = self.encoder(**seq_titles).last_hidden_state
rep_seq = torch.gather(last_hidden_state, 1, (torch.sum(seq_titles['attention_mask'], dim=-1)-1).unsqueeze(1).repeat(1, last_hidden_state.shape[-1]).unsqueeze(1)).squeeze(1)
rep_description_out = self.exp_gen.encoder(**batch['exp_description'], return_dict=True)
# rep_description_out = self.encoder(**batch['exp_description'], return_dict=True)
ppf_rep = self.PPF(rep_seq, rep_description_out.last_hidden_state, mask=(batch['exp_description']['input_ids']>0).float().to(rep_seq.device))
rep_description_out.last_hidden_state = ppf_rep
out_exp = self.exp_gen.generate(input_ids=batch['exp_description']['input_ids'].to(rep_seq.device), attention_mask=batch['exp_description']['attention_mask'].to(rep_seq.device), encoder_outputs=rep_description_out)
# out_exp = self.exp_gen_r.generate(input_ids=batch['exp_description']['input_ids'].to(rep_seq.device), attention_mask=batch['exp_description']['attention_mask'].to(rep_seq.device))
generation_exp = self.tokenizer.batch_decode(out_exp, skip_special_tokens=True)
reviews = self.tokenizer.batch_decode(batch['review']['input_ids'].squeeze(1), skip_special_tokens=True)
new_tokens_generate = [l.split() for l in generation_exp]
new_tokens_reference = [ll.split() for ll in reviews]
BLEU1 = bleu_score(new_tokens_reference, new_tokens_generate, n_gram=1, smooth=False)
BLEU4 = bleu_score(new_tokens_reference, new_tokens_generate, n_gram=4, smooth=False)
ROUGE = rouge_score(reviews, generation_exp)
metrics = {}
metrics['BLEU-1'] = BLEU1
metrics['BLEU-4'] = BLEU4
metrics['ROUGE-1/f'] = ROUGE['rouge_1/f_score']
metrics['ROUGE-1/r'] = ROUGE['rouge_1/r_score']
metrics['ROUGE-1/p'] = ROUGE['rouge_1/p_score']
metrics['ROUGE-2/f'] = ROUGE['rouge_2/f_score']
metrics['ROUGE-2/r'] = ROUGE['rouge_2/r_score']
metrics['ROUGE-2/p'] = ROUGE['rouge_2/p_score']
metrics['ROUGE-l/f'] = ROUGE['rouge_l/f_score']
metrics['ROUGE-l/r'] = ROUGE['rouge_l/r_score']
metrics['ROUGE-l/p'] = ROUGE['rouge_l/p_score']
return {"metrics": metrics}
def validation_epoch_end(self, validation_step_outputs) -> None:
print('validation_epoch_end')
metrics_all = self.all_gather(validation_step_outputs)
val_metrics_dict = {'BLEU-1': [], 'BLEU-4': [], 'ROUGE-1/f': [], 'ROUGE-1/r': [], 'ROUGE-1/p': [], 'ROUGE-2/f': [], 'ROUGE-2/r': [], 'ROUGE-2/p': [], 'ROUGE-l/f': [], 'ROUGE-l/r': [], 'ROUGE-l/p': []}
val_metrics_dict_mean = {}
for temp in metrics_all:
for key_temp, val_temp in temp['metrics'].items():
val_metrics_dict[key_temp].append(torch.mean(val_temp).cpu().item())
for key_temp, values_temp in val_metrics_dict.items():
values_mean = round(np.mean(values_temp), 4)
val_metrics_dict_mean[key_temp] = values_mean
print(val_metrics_dict_mean)
self.log("Val_Metrics", val_metrics_dict_mean)
self.log("BLEU-1", val_metrics_dict_mean['BLEU-1'])
def test_step(self, batch, batch_idx):
self.exp_gen.eval()
seq_titles = batch['title_seq_token']
last_hidden_state = self.encoder(**seq_titles).last_hidden_state
rep_seq = torch.gather(last_hidden_state, 1, (torch.sum(seq_titles['attention_mask'], dim=-1)-1).unsqueeze(1).repeat(1, last_hidden_state.shape[-1]).unsqueeze(1)).squeeze(1)
rep_description_out = self.exp_gen.encoder(**batch['exp_description'], return_dict=True)
ppf_rep = self.PPF(rep_seq, rep_description_out.last_hidden_state, mask=(batch['exp_description']['input_ids']>0).float().to(rep_seq.device))
rep_description_out.last_hidden_state = ppf_rep
out_exp = self.exp_gen.generate(input_ids=batch['exp_description']['input_ids'].to(rep_seq.device), attention_mask=batch['exp_description']['attention_mask'].to(rep_seq.device), encoder_outputs=rep_description_out)
generation_exp = self.tokenizer.batch_decode(out_exp, skip_special_tokens=True)
reviews = self.tokenizer.batch_decode(batch['review']['input_ids'].squeeze(1), skip_special_tokens=True)
new_tokens_generate = [l.split() for l in generation_exp]
new_tokens_reference = [ll.split() for ll in reviews]
BLEU1 = bleu_score(new_tokens_reference, new_tokens_generate, n_gram=1, smooth=False)
BLEU4 = bleu_score(new_tokens_reference, new_tokens_generate, n_gram=4, smooth=False)
ROUGE = rouge_score(reviews, generation_exp)
metrics = {}
metrics['BLEU-1'] = BLEU1
metrics['BLEU-4'] = BLEU4
metrics['ROUGE-1/f'] = ROUGE['rouge_1/f_score']
metrics['ROUGE-1/r'] = ROUGE['rouge_1/r_score']
metrics['ROUGE-1/p'] = ROUGE['rouge_1/p_score']
metrics['ROUGE-2/f'] = ROUGE['rouge_2/f_score']
metrics['ROUGE-2/r'] = ROUGE['rouge_2/r_score']
metrics['ROUGE-2/p'] = ROUGE['rouge_2/p_score']
metrics['ROUGE-l/f'] = ROUGE['rouge_l/f_score']
metrics['ROUGE-l/r'] = ROUGE['rouge_l/r_score']
metrics['ROUGE-l/p'] = ROUGE['rouge_l/p_score']
return {"metrics": metrics}
def test_epoch_end(self, test_step_outputs) -> None:
print('test_epoch_end')
metrics_all = self.all_gather(test_step_outputs)
test_metrics_dict = {'BLEU-1': [], 'BLEU-4': [], 'ROUGE-1/f': [], 'ROUGE-1/r': [], 'ROUGE-1/p': [], 'ROUGE-2/f': [], 'ROUGE-2/r': [], 'ROUGE-2/p': [], 'ROUGE-l/f': [], 'ROUGE-l/r': [], 'ROUGE-l/p': []}
test_metrics_dict_mean = {}
for temp in metrics_all:
for key_temp, val_temp in temp['metrics'].items():
test_metrics_dict[key_temp].append(torch.mean(val_temp).cpu().item())
for key_temp, values_temp in test_metrics_dict.items():
values_mean = round(np.mean(values_temp), 4)
test_metrics_dict_mean[key_temp] = values_mean
print(test_metrics_dict_mean)
def main():
print("t5 base decoder parameter efficient fine-tuning")
args_parser = argparse.ArgumentParser()
args_parser = add_optimizer_args(args_parser)
args_parser = add_datasets_args(args_parser)
args_parser = UniversalDataModule.add_data_specific_args(args_parser)
args_parser = Trainer.add_argparse_args(args_parser)
args_parser = UniversalCheckpoint.add_argparse_args(args_parser)
args_parser = Rec_seq_exp.add_module_specific_args(args_parser)
custom_parser = [
'--datasets_path_train', 'data/Instruments/title_review_summary_description/train.pkl',
'--datasets_path_test', 'data/Instruments/title_review_summary_description/test.pkl',
'--datasets_path_val', 'data/Instruments/title_review_summary_description/val.pkl',
'--train_batchsize', '256',
'--val_batchsize', '256',
'--test_batchsize', '256',
'--title_seq', '15',
'--max_epochs', '250',
'--random_seed', '72',
'--dataset_name', 'Instruments',
'--encoder_ckpt_path', 'ckpt/rec/instruments_title_rep_rec_inner_prod_negpredict_unisrec_data_t5_small_seqlen_10_patinence_5_check_epoch_1_title_len_15_ffn_filter_last_onlyseq_weight_qk_adapterdot_real_last_t5block_train_cls_cut_50epoch_bestsave/last.ckpt',
]
args = args_parser.parse_args(args=custom_parser)
print(args)
datasets = load_data_raw(args)
checkpoint_callback = UniversalCheckpoint(args)
collate_fn = Collator_exp_gep(args)
datamodule = UniversalDataModule(collate_fn=collate_fn, args=args, datasets=datasets)
early_stop_callback_step = EarlyStopping(monitor='BLEU-1', min_delta=0.00, patience=2, verbose=False, mode='max')
# trainer = Trainer(devices=1, accelerator="gpu", strategy = DDPStrategy(find_unused_parameters=False), callbacks=[checkpoint_callback, early_stop_callback], max_epochs=args.max_epochs, check_val_every_n_epoch=10)
trainer = Trainer(devices=4, accelerator="gpu", strategy = DDPStrategy (find_unused_parameters=True), callbacks=[checkpoint_callback, early_stop_callback_step], max_epochs=args.max_epochs, check_val_every_n_epoch=1)
model = Rec_seq_exp(args)
trainer.fit(model, datamodule)
trainer.test(model, datamodule)
if __name__ =="__main__":
main()