-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
191 lines (161 loc) · 7.33 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
import logging
import os
import warnings
import evaluate
import numpy as np
import torch
from datasets import concatenate_datasets, load_dataset
from peft import LoraConfig, TaskType, get_peft_model
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
DataCollatorForSeq2Seq,
HfArgumentParser,
Trainer,
)
from transformers.trainer_utils import is_main_process
from arguments import DatasetsArguments, ModelArguments, MyTrainingArguments
from prompt import INPUT_PROMPT, TOTAL_PROMPT
from utils import category_map, output_map, seed_everything
logger = logging.getLogger(__name__)
os.environ["TOKENIZERS_PARALLELISM"] = "false"
warnings.filterwarnings("ignore")
def main(model_args: ModelArguments, data_args: DatasetsArguments, training_args: MyTrainingArguments):
seed_everything(training_args.seed)
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, padding_size="right")
if tokenizer.pad_token == None:
tokenizer.pad_token = tokenizer.eos_token
train_dataset = load_dataset("json", data_files=data_args.train_data_path, split="train")
if training_args.do_aug:
train_aug_dataset = load_dataset("json", data_files=data_args.aug_data_path, split="train")
train_dataset = concatenate_datasets([train_dataset, train_aug_dataset])
valid_dataset = load_dataset("json", data_files=data_args.valid_data_path, split="train")
def preprocess_data(raw):
category = raw["input"]["category"]
inference_1 = raw["input"]["inference_1"]
inference_2 = raw["input"]["inference_2"]
inference_3 = raw["input"]["inference_3"]
dialog_info_list = raw["input"]["conversation"]
dialog_list = []
for dialog_info in dialog_info_list:
ref = dialog_info["utterance_id"].split(".")[-1]
speaker = dialog_info["speaker"]
utterance = dialog_info["utterance"].replace("\n", " ")
dialog_list.append(f"화자{speaker}: {utterance}")
ref_ids = []
for i, ref_id in enumerate(raw["input"]["reference_id"]):
tmp = int(ref_id.split(".")[-1])
ref_ids.append(f"{dialog_list[tmp - 1]}")
raw["reference_id"] = "\n".join(ref_ids)
raw["dialog"] = "\n".join(dialog_list)
raw["inference_1"] = inference_1
raw["inference_2"] = inference_2
raw["inference_3"] = inference_3
raw["category"] = category
return raw
train_dataset = train_dataset.map(preprocess_data, keep_in_memory=True)
valid_dataset = valid_dataset.map(preprocess_data, keep_in_memory=True)
def preprocess_prompt(raw):
dialog = raw["dialog"]
category = raw["category"]
inference_1 = raw["inference_1"]
inference_2 = raw["inference_2"]
inference_3 = raw["inference_3"]
ref = raw["reference_id"]
definition = category_map[raw["category"]]
full_text = TOTAL_PROMPT.format_map(
{
"category": category,
"dialog": dialog,
"def": definition,
"ref": ref,
"inference_1": inference_1,
"inference_2": inference_2,
"inference_3": inference_3,
"answer": output_map[raw["output"]],
}
)
tokenized_full_prompt = tokenizer(full_text)
user_prompt = INPUT_PROMPT.format_map(
{
"category": category,
"dialog": dialog,
"def": definition,
"ref": ref,
"inference_1": inference_1,
"inference_2": inference_2,
"inference_3": inference_3,
}
)
tokenized_user_prompt = tokenizer(user_prompt)
user_prompt_len = len(tokenized_user_prompt["input_ids"])
tokenized_full_prompt["labels"] = tokenized_full_prompt["input_ids"].copy()
tokenized_full_prompt["labels"] = [-100] * user_prompt_len + tokenized_full_prompt["labels"][user_prompt_len:]
return tokenized_full_prompt
train_dataset = train_dataset.map(preprocess_prompt, remove_columns=train_dataset.column_names)
valid_dataset = valid_dataset.map(preprocess_prompt, remove_columns=valid_dataset.column_names)
logger.info(tokenizer.decode(valid_dataset[3]["input_ids"][:]))
logger.info(valid_dataset[3]["input_ids"][-20:])
logger.info(valid_dataset[3]["labels"][-20:])
logger.info(
f"train size: {len(train_dataset)}\nvalid size: {len(valid_dataset)}\nratio: {len(valid_dataset)/(len(train_dataset) + len(valid_dataset))}"
)
model = AutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path,
torch_dtype=torch.bfloat16 if training_args.bf16 else torch.float16,
low_cpu_mem_usage=True,
use_cache=False,
attn_implementation="eager" if "gemma" in model_args.model_name_or_path.lower() else "sdpa",
trust_remote_code=True if "EXA" in model_args.model_name_or_path.lower() else "sdpa",
)
peft_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
target_modules=model_args.target_modules,
inference_mode=False,
r=model_args.r,
lora_alpha=model_args.lora_alpha,
use_rslora=model_args.use_rslora,
lora_dropout=model_args.lora_dropout,
)
model = get_peft_model(model, peft_config)
model.print_trainable_parameters()
model.enable_input_require_grads()
def preprocess_logits_for_metrics(logits, labels):
logits = logits if not isinstance(logits, tuple) else logits[0]
logit_idx = [tokenizer.vocab["A"], tokenizer.vocab["B"], tokenizer.vocab["C"]]
logits = logits[:, -2, logit_idx]
return logits
acc_metric = evaluate.load("accuracy")
int_output_map = {"A": 0, "B": 1, "C": 2}
def compute_metrics(evaluation_result):
logits, labels = evaluation_result
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
labels = list(map(lambda x: int_output_map[x], labels))
probs = torch.nn.functional.softmax(torch.tensor(logits), dim=-1)
predictions = np.argmax(probs, axis=-1)
acc = acc_metric.compute(predictions=predictions, references=labels)
return acc
trainer = Trainer(
model=model,
data_collator=DataCollatorForSeq2Seq(tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True),
args=training_args,
train_dataset=train_dataset,
eval_dataset=valid_dataset,
preprocess_logits_for_metrics=preprocess_logits_for_metrics,
compute_metrics=compute_metrics,
)
trainer.train()
if training_args.local_rank == 0:
model = model.merge_and_unload()
model.save_pretrained(training_args.output_dir)
tokenizer.save_pretrained(training_args.output_dir)
if __name__ == "__main__":
parser = HfArgumentParser((ModelArguments, DatasetsArguments, MyTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
)
main(model_args=model_args, data_args=data_args, training_args=training_args)