-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtrain_script.py
More file actions
180 lines (132 loc) · 5.84 KB
/
Copy pathtrain_script.py
File metadata and controls
180 lines (132 loc) · 5.84 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
from transformers import TrainingArguments, Trainer
import yaml
import argparse
from easydict import EasyDict as edict
import os
from utils import *
# set to suppress the following warning:
# huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
# probably fine since we tokenize all data first before passing to trainer
os.environ['TOKENIZERS_PARALLELISM'] = 'true'
# hacky way to prevent loss/metrics being printed to stdout
# while still enabling logging to wandb
# https://github.com/huggingface/transformers/issues/18093
from transformers.trainer_callback import ProgressCallback
def on_log(self, args, state, control, logs=None, **kwargs):
if state.is_local_process_zero and self.training_bar is not None:
_ = logs.pop("total_flos", None)
ProgressCallback.on_log = on_log
def main(configs):
if configs.type == 'llama':
# set wandb project in which to store logs
if 'wandb_project' in configs:
os.environ['WANDB_PROJECT'] = configs.wandb_project
### Prepare Model and Tokenizer ###
print('Preparing Model and Tokenizer')
print(os.getenv("HF_HOME"))
model = get_model(configs)
tokenizer = get_tokenizer(configs.model_id)
### Prepare data ###
print('Preparing and tokenizing data')
t_dataset, e_dataset = get_datasets_llama(configs, tokenizer, model_type="llama")
collate_fn = get_collate_func(tokenizer)
### Get custom loss objective and metrics ###
prm_compute_loss_func = get_compute_loss_func()
prm_compute_metrics = get_compute_metrics_llama()
### training loop ###
training_args = TrainingArguments(label_names=["labels", "accuracy"], **configs.training_args)
print('Training loop started')
trainer = Trainer(
model,
training_args,
train_dataset=t_dataset,
eval_dataset=e_dataset,
data_collator=collate_fn,
processing_class=tokenizer,
compute_loss_func=prm_compute_loss_func,
compute_metrics=prm_compute_metrics,
)
# train
checkpoint = None
if 'resume_from_checkpoint' in configs:
checkpoint = configs.resume_from_checkpoint
trainer.train(resume_from_checkpoint=checkpoint)
if configs.type == 'deberta':
# set wandb project in which to store logs
if 'wandb_project' in configs:
os.environ['WANDB_PROJECT'] = configs.wandb_project
### Prepare Model and Tokenizer ###
print('Preparing Model and Tokenizer')
print(os.getenv("HF_HOME"))
tokenizer = get_tokenizer_bert(configs.model_id)
model = get_model_bert(configs, tokenizer)
### Prepare data ###
print('Preparing and tokenizing data')
t_dataset, e_dataset = get_datasets(configs, tokenizer, model_type="deberta")
collate_fn = get_collate_func(tokenizer)
### Get custom loss objective and metrics ###
prm_compute_loss_func = get_compute_loss_func_bert()
prm_compute_metrics = get_compute_metrics_bert()
### training loop ###
training_args = TrainingArguments(**configs.training_args)
print('Training loop started')
trainer = Trainer(
model,
training_args,
train_dataset=t_dataset,
eval_dataset=e_dataset,
data_collator=collate_fn,
processing_class=tokenizer,
compute_loss_func=prm_compute_loss_func,
compute_metrics=prm_compute_metrics
)
# train
checkpoint = None
if 'resume_from_checkpoint' in configs:
checkpoint = configs.resume_from_checkpoint
trainer.train(resume_from_checkpoint=checkpoint)
if configs.type == 'default':
# set wandb project in which to store logs
if 'wandb_project' in configs:
os.environ['WANDB_PROJECT'] = configs.wandb_project
### Prepare Model and Tokenizer ###
print('Preparing Model and Tokenizer')
print(os.getenv("HF_HOME"))
model = get_model(configs)
tokenizer = get_tokenizer(configs.model_id)
### Prepare data ###
print('Preparing and tokenizing data')
t_dataset, e_dataset = get_datasets(configs, tokenizer, model_type="llama")
collate_fn = get_collate_func(tokenizer)
### Get custom loss objective and metrics ###
prm_compute_loss_func = get_compute_loss_func()
prm_compute_metrics = get_compute_metrics()
### training loop ###
training_args = TrainingArguments(**configs.training_args)
print('Training loop started')
trainer = Trainer(
model,
training_args,
train_dataset=t_dataset,
eval_dataset=e_dataset,
data_collator=collate_fn,
processing_class=tokenizer,
compute_loss_func=prm_compute_loss_func,
compute_metrics=prm_compute_metrics
)
# train
checkpoint = None
if 'resume_from_checkpoint' in configs:
checkpoint = configs.resume_from_checkpoint
trainer.train(resume_from_checkpoint=checkpoint)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Training script for training PRM models')
parser.add_argument('-c','--config', type=str, help='Path to config json', default='./train_configs/llama_prm800k.yml')
parser.add_argument('--local_rank', type=int, default=-1, help='Used by torch.distributed.launch/torchrun')
args = parser.parse_args()
with open(args.config) as stream:
try:
configs = edict(yaml.safe_load(stream))
except yaml.YAMLError as exc:
print(exc)
main(configs)