-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathfinetune.py
202 lines (179 loc) · 9.42 KB
/
finetune.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
193
194
195
196
197
198
199
200
201
import os
import json
import copy
import torch
import numpy as np
import argparse
from Distiller.glue_preprocess import load_and_cache_examples, glue_compute_metrics
from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoConfig, get_linear_schedule_with_warmup
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm
class STLR(torch.optim.lr_scheduler._LRScheduler):
def __init__(self, optimizer, max_mul, ratio, steps_per_cycle, decay=1, last_epoch=-1):
self.max_mul = max_mul - 1
self.turning_point = steps_per_cycle // (ratio + 1)
self.steps_per_cycle = steps_per_cycle
self.decay = decay
super().__init__(optimizer, last_epoch)
def get_lr(self):
residual = self.last_epoch % self.steps_per_cycle
multiplier = self.decay ** (self.last_epoch // self.steps_per_cycle)
if residual <= self.turning_point:
multiplier *= self.max_mul * (residual / self.turning_point)
else:
multiplier *= self.max_mul * (
(self.steps_per_cycle - residual) /
(self.steps_per_cycle - self.turning_point))
return [lr * (1 + multiplier) for lr in self.base_lrs]
def eval(args, model, tokenizer):
dataset, s_dataset, features, s_features, examples = load_and_cache_examples(args, tokenizer, mode="dev",
return_examples=True)
eval_sampler = SequentialSampler(dataset)
eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
preds = []
label_list = []
model.eval()
for batch in tqdm(eval_dataloader):
# labels = batch['labels']
# batch = tuple(t.to(args.device) for t in batch)
batch = {key: value.to(args.device) for key, value in batch.items()}
with torch.no_grad():
if model.config.model_type in ['distilbert']:
batch.pop('token_type_ids')
outputs = model(**batch)
# outputs = model(**batch)
predictions = outputs.logits.detach().cpu()
if args.task_name not in ["stsb","cloth"]:
predictions = predictions.argmax(dim=-1)
else:
predictions = predictions[:, 0]
label_list.extend(batch['labels'].cpu().tolist())
preds.extend(predictions.tolist())
# eval_metric_compute = metric.compute()
eval_metric = glue_compute_metrics(args.task_name, np.array(preds), np.array(label_list))
if args.task_name == 'mnli':
mnli_args = copy.deepcopy(args)
mnli_args.task_name = 'mnli-mm'
dataset, s_dataset, features, s_features, examples = load_and_cache_examples(mnli_args, tokenizer, mode="dev",
return_examples=True)
eval_sampler = SequentialSampler(dataset)
eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
preds = []
label_list = []
model.eval()
for batch in tqdm(eval_dataloader):
batch = {key: value.to(args.device) for key, value in batch.items()}
with torch.no_grad():
if model.config.model_type in ['distilbert']:
batch.pop('token_type_ids')
outputs = model(**batch)
# outputs = model(**batch)
predictions = outputs.logits.detach().cpu()
predictions = predictions.argmax(dim=-1)
label_list.extend(batch['labels'].cpu().tolist())
preds.extend(predictions.tolist())
mm_eval_metric = glue_compute_metrics('mnli-mm', np.array(preds), np.array(label_list))
eval_metric["mnli-mm/acc"] = mm_eval_metric["mnli-mm/acc"]
eval_metric['acc'] = (eval_metric['acc']+eval_metric["mnli-mm/acc"]) / 2
model.train()
print(f"Eval result: {eval_metric}")
return eval_metric
def main(args):
best_result = 0.0
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.output_dir)
if not os.path.exists(args.data_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.data_dir)
# Setup CUDA, GPU & distributed training
if args.local_rank == -1:
device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend="nccl")
args.n_gpu = 1
args.device = device
config = AutoConfig.from_pretrained(args.model_path)
config.num_labels = args.num_labels
model = AutoModelForSequenceClassification.from_pretrained(args.model_path, config=config)
tokenizer = AutoTokenizer.from_pretrained(args.model_path,use_fast=False,
config=config)
dataset = load_and_cache_examples(args, tokenizer, "train", False)
train_sampler = RandomSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
train_dataloader = DataLoader(dataset, sampler=train_sampler, batch_size=args.train_batch_size)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0}
]
optimizer = torch.optim.AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=0.1*len(train_dataloader)*args.epoch, num_training_steps=len(train_dataloader)*args.epoch)
model.to(args.device)
model.train()
if args.train:
for i in range(args.epoch):
print(f"Epoch {i+1}")
for step, batch in tqdm(enumerate(train_dataloader), total=len(train_dataloader)):
batch = {key: value.to(args.device) for key, value in batch.items()}
optimizer.zero_grad()
outputs = model(**batch)
loss = outputs.loss
loss.backward()
optimizer.step()
scheduler.step()
eval_result = eval(args, model, tokenizer)
model.train()
if eval_result['acc'] > best_result:
best_result = eval_result['acc']
model_to_save = model.module if hasattr(model,
"module") else model # Take care of distributed/parallel training
model_to_save.save_pretrained(args.output_dir)
if tokenizer:
tokenizer.save_pretrained(args.output_dir)
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
output_eval_file = os.path.join(args.output_dir, f"eval_results.txt")
with open(output_eval_file, "a") as writer:
writer.write(f"Output: {json.dumps(eval_result, indent=2)}\n")
with open(os.path.join(args.output_dir, "training_args.json"), 'w') as f:
arg_dict = vars(args)
arg_dict['device'] = str(arg_dict['device'])
json.dump(arg_dict, f)
if args.eval:
eval_result = eval(args, model, tokenizer)
print(eval_result)
# model_to_save = model.module if hasattr(model,
# "module") else model # Take care of distributed/parallel training
# model_to_save.save_pretrained(args.output_dir)
# if tokenizer:
# tokenizer.save_pretrained(args.output_dir)
# torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# with open(os.path.join(args.output_dir, "training_args.json"), 'w') as f:
# arg_dict = vars(args)
# arg_dict['device'] = str(arg_dict['device'])
# json.dump(arg_dict, f)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--model_path", required=True)
parser.add_argument("--output_dir", default="finetuned/")
parser.add_argument("--data_dir", required=True)
parser.add_argument("--max_seq_length", default=512, type=int)
parser.add_argument("--epoch", default=5, type=int)
parser.add_argument("--local_rank", default=-1)
parser.add_argument("--task_name", default="kaggle")
parser.add_argument("--overwrite_cache", default=False)
parser.add_argument("--learning_rate", default=5e-5,type=float)
parser.add_argument("--adam_epsilon", default=1e-8)
parser.add_argument("--train_batch_size", default=16, type=int)
parser.add_argument("--eval_batch_size", default=64, type=int)
parser.add_argument("--weight_decay", default=1e-4, type=float,
help="Weight decay if we apply some.")
parser.add_argument("--num_labels", default=2, type=int, required=True)
parser.add_argument("--cuda", action="store_true")
parser.add_argument("--train", action="store_true")
parser.add_argument("--eval", action="store_true")
args = parser.parse_args()
args.output_dir = args.output_dir + args.task_name + '/' + args.model_path
main(args)