-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain.py
192 lines (157 loc) · 7.3 KB
/
train.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
import os
from transformers import set_seed, TrainingArguments, HfArgumentParser, PretrainedConfig
from transformers import AutoTokenizer
from datasets import load_dataset
import torch
import wandb
from args import ModelArguments, DatasetArguments
from model import MSSEModel
from trainer import MyTrainer
from mteb import MTEB
from prettytable import PrettyTable
from config import MSSEConfig
import logging
import pandas as pd
import random
import torch.optim as optim
import pandas as pd
from datasets import Dataset
logger = logging.getLogger(__name__)
def preprocess_logits_for_metrics(logits, labels):
pred_ids = torch.argmax(logits[0], dim=-1)
return pred_ids
def eval_mteb(model, batch_size):
tasks = [
"STS12",
"STS13",
"STS14",
"STS15",
"STS16",
"STSBenchmark",
"SICK-R",
]
evaluation = MTEB(tasks=tasks, task_langs=["en"], task_categories=['S2S'])
results = evaluation.run(model, overwrite_results=True, batch_size=batch_size, eval_splits=['test'], output_folder='mteb_results/'+wandb.run.name)
sts12 = results['STS12']['test']['cos_sim']['spearman']
sts13 = results['STS13']['test']['cos_sim']['spearman']
sts14 = results['STS14']['test']['cos_sim']['spearman']
sts15 = results['STS15']['test']['cos_sim']['spearman']
sts16 = results['STS16']['test']['cos_sim']['spearman']
sickr = results['SICK-R']['test']['cos_sim']['spearman']
stsb = results['STSBenchmark']['test']['cos_sim']['spearman']
avg_sts = (sts12 + sts13 + sts14 + sts15 + sts16 + sickr + stsb) / 7
wandb.summary['STS12'] = sts12
wandb.summary['STS13'] = sts13
wandb.summary['STS14'] = sts14
wandb.summary['STS15'] = sts15
wandb.summary['STS16'] = sts16
wandb.summary['SICK-R'] = sickr
wandb.summary['STSBenchmark'] = stsb
wandb.summary['mteb_avg_sts'] = avg_sts
return results
# For few-shot sentence embeddings to obtain fixed ratio of the dataset.
def split_train_dataset(datasets, split_ratio=0.1, random_seed=None):
random.seed(random_seed)
train_dataset_length = len(datasets["train"])
shuffled_indices = list(range(train_dataset_length))
random.shuffle(shuffled_indices)
num_samples = int(split_ratio * train_dataset_length)
random_indices = shuffled_indices[:num_samples]
train_dataset_split = datasets["train"].select(random_indices)
return train_dataset_split
if __name__ == "__main__":
parser = HfArgumentParser((ModelArguments, TrainingArguments, DatasetArguments))
model_args, training_args, dataset_args = parser.parse_args_into_dataclasses()
wandb.init(project='MSSE-49', name=model_args.output_dir_name, dir=model_args.output_dir_name)
set_seed(training_args.seed)
wandb.config.update(model_args)
wandb.config.update(training_args)
wandb.config.update(dataset_args)
training_args.output_dir = 'results/' + wandb.run.name
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)
config = MSSEConfig(
encoder_name_or_path=model_args.model_name_or_path,
max_length=model_args.max_length,
decoder_num_heads=model_args.decoder_num_heads,
decoder_num_layers=model_args.decoder_num_layers,
decoder_noise_dropout=model_args.decoder_target_dropout,
pooler=model_args.pooler,
do_contrastive=model_args.do_contrastive,
do_generative=model_args.do_generative,
prompt_format=model_args.prompt_format,
contrastive_weight=model_args.contrastive_weight,
generative_weight=model_args.generative_weight,
contrastive_temp=model_args.contrastive_temp,
)
print(config)
model = MSSEModel(config)
def map_fn(example):
max_length = model_args.max_length
if config.pooler == 'mask':
prompt_len = len(tokenizer(config.prompt_format, add_special_tokens=False)['input_ids'])
def preprocess_sentence(sentence):
tokenized_sentence = tokenizer.decode(
tokenizer(sentence, padding=True, truncation=True, max_length=config.max_length)['input_ids'],
skip_special_tokens=True
)
return config.prompt_format.replace('[X]', tokenized_sentence).replace('[MASK]', tokenizer.mask_token)
example['sent0'] = preprocess_sentence(example['sent0'])
example['sent1'] = preprocess_sentence(example['sent1'])
if 'hard_neg' in example:
example['hard_neg'] = preprocess_sentence(example['hard_neg'])
max_length = max_length + prompt_len
original_inputs = tokenizer(example['sent0'], padding='max_length', truncation=True, max_length=max_length)
example['input_ids'] = original_inputs['input_ids']
example['attention_mask'] = original_inputs['attention_mask']
positive_inputs = tokenizer(example['sent1'], padding='max_length', truncation=True, max_length=max_length)
example['positive_input_ids'] = positive_inputs['input_ids']
example['positive_attention_mask'] = positive_inputs['attention_mask']
if 'hard_neg' in example:
negative_inputs = tokenizer(example['hard_neg'], padding='max_length', truncation=True, max_length=max_length)
example['negative_input_ids'] = negative_inputs['input_ids']
example['negative_attention_mask'] = negative_inputs['attention_mask']
return example
def load_and_clean_csv(file_path):
try:
df = pd.read_csv(file_path, delimiter="\t" if "tsv" in file_path else ",", encoding='utf-8')
except UnicodeDecodeError:
print("UTF-8 decoding failed. Trying latin1...")
df = pd.read_csv(file_path, delimiter="\t" if "tsv" in file_path else ",", encoding='latin1')
df = df.applymap(lambda x: x if isinstance(x, str) else str(x))
return Dataset.from_pandas(df)
if dataset_args.train_dataset == "data/output.csv":
dataset = load_and_clean_csv(dataset_args.train_dataset)
else:
raise NotImplementedError()
dataset = dataset.map(
map_fn,
batched=False,
num_proc=12,
# remove_columns=column_names,
load_from_cache_file=True,
).train_test_split(0.1, seed=training_args.seed, shuffle=True)
test_valid = dataset['test'].train_test_split(0.01)
trainer = MyTrainer(
model=model,
args=training_args,
tokenizer=tokenizer,
# train_dataset=dataset,
train_dataset=dataset['train'],
eval_dataset=test_valid['test'],
preprocess_logits_for_metrics=preprocess_logits_for_metrics,
)
trainer.train()
mteb_results = eval_mteb(model, batch_size=training_args.eval_batch_size)
table = PrettyTable(["Name", "Value"])
# Add rows
table.add_row(["STS12", wandb.summary['STS12']])
table.add_row(["STS13", wandb.summary['STS13']])
table.add_row(["STS14", wandb.summary['STS14']])
table.add_row(["STS15", wandb.summary['STS15']])
table.add_row(["STS16", wandb.summary['STS16']])
table.add_row(["SICK-R", wandb.summary['SICK-R']])
table.add_row(["STSBenchmark", wandb.summary['STSBenchmark']])
table.add_row(["Avg.", wandb.summary['mteb_avg_sts']])
# Print the table
print(table)
wandb.finish()