-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathfinetune.py
executable file
·147 lines (119 loc) · 6.26 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
import os
import transformers
import re
from spy import Transformer
from datasets import load_dataset
from utils import get_args
import random
import huggingface_hub
from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments, DataCollatorForLanguageModeling
class PythonDataset:
def __init__(self, tokenizer, dataset, transformer, seq_length=1024, language='python', ratio=1.0):
self.tokenizer = tokenizer
self.concat_token_id = tokenizer.eos_token_id if tokenizer.eos_token_id is not None else args.eos_token_id
# self.dataset = dataset.shard(index=0, num_shards=10000)
self.dataset = dataset
self.seq_length = seq_length
self.transformer = transformer
self.language = language
self.re_first_line = re.compile(r'^.*\n')
if language.startswith('spython'):
self.dataset = self.dataset.filter(lambda x: random.random() < ratio)
self.dataset = self.dataset.map(self.convert_to_spy, batched=True, load_from_cache_file=False, remove_columns=['max_stars_repo_path', 'max_stars_repo_name', 'max_stars_count', 'id'])
self.dataset = self.dataset.map(self.tokenize_and_concate, batched=True, remove_columns=['content'], load_from_cache_file=False)
def convert_to_spy(self, examples):
spy_examples = []
for sample in examples['content']:
try:
first_line = self.re_first_line.match(sample).group()
except:
continue
if first_line.startswith('<'):
sample = self.re_first_line.sub('', sample)
if sample.startswith('#!/'):
continue
try:
spy_code = self.transformer.parse(sample)
except (ValueError, RecursionError):
continue
spy_examples.append(spy_code if self.language.startswith('spython') else sample)
return {'content': spy_examples}
def tokenize_and_concate(self, examples):
tokenized_example = self.tokenizer(examples['content'])
concatenated_examples = {}
for k in tokenized_example.keys():
all_token_ids = []
for tokenized_input in tokenized_example[k]:
all_token_ids.extend(tokenized_input + [self.concat_token_id])
concatenated_examples[k] = all_token_ids
total_length = len(concatenated_examples[list(concatenated_examples.keys())[0]])
result = {k:[] for k in concatenated_examples.keys()}
for k,t in concatenated_examples.items():
for i in range(0, total_length, self.seq_length):
if i+self.seq_length < total_length:
result[k].append(t[i:i+self.seq_length])
return {'input_ids': result['input_ids'], 'labels': result["input_ids"].copy()}
if __name__ == '__main__':
args = get_args()
transformer = Transformer()
tokenizer = AutoTokenizer.from_pretrained(args.model_path, cache_dir="./cached")
tokenizer.pad_token = tokenizer.eos_token
transformer.special_tokens = sorted(list(set(transformer.special_tokens)))
if args.language.startswith('spython'):
tokenizer.add_tokens(transformer.special_tokens)
model_config = transformers.AutoConfig.from_pretrained(args.model_path, cache_dir="./cached")
if args.language.startswith('spython'):
model_vocab_size = model_config.vocab_size + len(transformer.special_tokens)
if model_config.vocab_size > model_vocab_size:
model_vocab_size = model_config.vocab_size
elif model_config.vocab_size > tokenizer.vocab_size:
model_vocab_size = model_config.vocab_size
else:
model_vocab_size = tokenizer.vocab_size
if not args.further_train:
model = AutoModelForCausalLM.from_pretrained(args.model_path, cache_dir="./cached", load_in_8bit=False, vocab_size=model_vocab_size, ignore_mismatched_sizes=True)
elif args.from_scratch:
model = AutoModelForCausalLM.from_config(model_config, vocab_size=model_vocab_size, ignore_mismatched_sizes=True)
else:
model = AutoModelForCausalLM.from_pretrained(args.further_train, cache_dir="./cached", load_in_8bit=False, vocab_size=model_vocab_size, ignore_mismatched_sizes=True)
dataset = load_dataset(args.dataset_name, split="train", cache_dir="./cached")
dataset = dataset.train_test_split(test_size=0.05, shuffle=True)
train_dataset = dataset['train']
val_dataset = dataset['test']
train_data = PythonDataset(tokenizer, train_dataset, transformer, seq_length=args.seq_length, language=args.language, ratio=args.ratio).dataset
val_data = PythonDataset(tokenizer, val_dataset, transformer, seq_length=args.seq_length, language=args.language).dataset
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer,mlm=False)
if args.further_train:
run_name = f"{args.model_path.split('/')[-1]}-{args.language}-{args.ratio}"
else:
run_name = f"{args.model_path.split('/')[-1]}-{args.language}"
if args.from_scratch:
run_name += '-from_scratch'
training_args = TrainingArguments(
output_dir=os.path.join(args.output_dir, run_name),
dataloader_drop_last=True,
evaluation_strategy="epoch",
save_strategy="epoch",
load_best_model_at_end=True,
num_train_epochs=args.epoch,
eval_steps=args.eval_freq,
save_steps=args.save_freq,
per_device_train_batch_size=args.batch_size,
per_device_eval_batch_size=args.batch_size,
learning_rate=args.learning_rate,
torch_compile=True,
lr_scheduler_type=args.lr_scheduler_type,
warmup_steps=args.num_warmup_steps,
gradient_accumulation_steps=args.gradient_accumulation_steps,
fp16=True,
weight_decay=args.weight_decay,
run_name=run_name,
logging_steps=args.log_freq,
log_level='debug',
ddp_find_unused_parameters=False,
)
trainer = Trainer(model=model, args=training_args, train_dataset=train_data, eval_dataset=val_data, data_collator=data_collator)
print("Training...")
trainer.train(resume_from_checkpoint=args.resume_from_checkpoint)
print("Saving last checkpoint of the model")
model.save_pretrained(os.path.join(args.output_dir, run_name, 'best_model'))