Repository navigation
Expand file tree
/
Copy pathtrain_bart.py
More file actions
229 lines (195 loc) · 8.47 KB
/
Copy pathtrain_bart.py
File metadata and controls
229 lines (195 loc) · 8.47 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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
"""
train_bart.py
-------------
Fine-tunes facebook/bart-base or facebook/bart-large for AI-to-Human style transfer.
Mirrors the T5 approach from the paper (Section 4.3–4.4) but uses BART seq2seq.
Usage:
python train_bart.py --config bart_base.yaml
python train_bart.py --config bart_base.yaml --smoke
python train_bart.py --config bart_large.yaml --output_dir results/bart_large_run2
"""
import argparse
import json
import os
import sys
import yaml
def load_config(path: str, overrides: dict) -> dict:
with open(path) as f:
cfg = yaml.safe_load(f)
cfg.update({k: v for k, v in overrides.items() if v is not None})
return cfg
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--config", required=True)
parser.add_argument("--smoke", action="store_true",
help="Smoke test: 128 samples, 10 steps only")
parser.add_argument("--output_dir", default=None)
parser.add_argument("--max_train_samples", type=int, default=None)
args = parser.parse_args()
cfg = load_config(args.config, {"output_dir": args.output_dir})
if args.smoke:
cfg["max_train_samples"] = 128
cfg["max_eval_samples"] = 32
cfg["max_steps"] = 10
cfg["num_train_epochs"] = 1
cfg["logging_steps"] = 2
cfg["save_strategy"] = "no"
cfg["evaluation_strategy"] = "no"
print("=== SMOKE TEST MODE ===")
if args.max_train_samples:
cfg["max_train_samples"] = args.max_train_samples
# Lazy imports so --smoke on CPU doesn't pull in CUDA unnecessarily
import torch
from datasets import Dataset
from transformers import (
AutoTokenizer,
BartForConditionalGeneration,
DataCollatorForSeq2Seq,
Seq2SeqTrainer,
Seq2SeqTrainingArguments,
)
# ------------------------------------------------------------------ #
# Optional LoRA for BART
# ------------------------------------------------------------------ #
use_lora = cfg.get("use_lora", False)
if use_lora:
try:
from peft import LoraConfig, TaskType, get_peft_model
except ImportError:
print("ERROR: peft not installed. Run: pip install peft")
sys.exit(1)
# ------------------------------------------------------------------ #
# Load data
# ------------------------------------------------------------------ #
def read_jsonl(path, max_rows=None):
rows = []
with open(path) as f:
for i, line in enumerate(f):
if max_rows and i >= max_rows:
break
line = line.strip()
if line:
rows.append(json.loads(line))
return rows
print(f"Loading training data from {cfg['train_file']} ...")
train_rows = read_jsonl(cfg["train_file"], cfg.get("max_train_samples"))
val_rows = read_jsonl(cfg["val_file"], cfg.get("max_eval_samples"))
print(f" Train: {len(train_rows):,} Val: {len(val_rows):,}")
prefix = cfg.get("prefix", "humanize: ")
# ------------------------------------------------------------------ #
# Tokenizer
# ------------------------------------------------------------------ #
print(f"Loading tokenizer: {cfg['model_name']}")
tokenizer = AutoTokenizer.from_pretrained(cfg["model_name"])
max_in = cfg.get("max_input_length", 256)
max_tgt = cfg.get("max_target_length", 256)
def tokenize(batch):
inputs = [prefix + t for t in batch["ai"]]
model_inputs = tokenizer(
inputs,
max_length=max_in,
truncation=True,
padding=False,
)
with tokenizer.as_target_tokenizer():
labels = tokenizer(
batch["human"],
max_length=max_tgt,
truncation=True,
padding=False,
)
model_inputs["labels"] = labels["input_ids"]
return model_inputs
train_ds = Dataset.from_list(train_rows).map(
tokenize, batched=True, remove_columns=Dataset.from_list(train_rows).column_names
)
val_ds = Dataset.from_list(val_rows).map(
tokenize, batched=True, remove_columns=Dataset.from_list(val_rows).column_names
)
# ------------------------------------------------------------------ #
# Model
# ------------------------------------------------------------------ #
print(f"Loading model: {cfg['model_name']}")
model = BartForConditionalGeneration.from_pretrained(cfg["model_name"])
if use_lora:
lora_cfg = LoraConfig(
task_type=TaskType.SEQ_2_SEQ_LM,
r=cfg.get("lora_r", 16),
lora_alpha=cfg.get("lora_alpha", 32),
lora_dropout=0.05,
target_modules=cfg.get("lora_target_modules",
["q_proj", "v_proj", "encoder_attn.q_proj", "encoder_attn.v_proj"]),
)
model = get_peft_model(model, lora_cfg)
model.print_trainable_parameters()
print("LoRA enabled for BART.")
# ------------------------------------------------------------------ #
# Training arguments
# ------------------------------------------------------------------ #
output_dir = cfg.get("output_dir", "results/bart")
os.makedirs(output_dir, exist_ok=True)
# Build kwargs dynamically so max_steps overrides epochs when set in smoke
training_kwargs = dict(
output_dir=output_dir,
num_train_epochs=cfg.get("num_train_epochs", 5),
per_device_train_batch_size=cfg.get("per_device_train_batch_size", 4),
per_device_eval_batch_size=cfg.get("per_device_eval_batch_size", 4),
gradient_accumulation_steps=cfg.get("gradient_accumulation_steps", 4),
learning_rate=cfg.get("learning_rate", 5e-5),
lr_scheduler_type=cfg.get("lr_scheduler_type", "linear"),
warmup_ratio=cfg.get("warmup_ratio", 0.05),
fp16=cfg.get("fp16", True),
bf16=cfg.get("bf16", False),
weight_decay=cfg.get("weight_decay", 0.01),
max_grad_norm=cfg.get("max_grad_norm", 1.0),
seed=cfg.get("seed", 42),
evaluation_strategy=cfg.get("evaluation_strategy", "epoch"),
save_strategy=cfg.get("save_strategy", "epoch"),
load_best_model_at_end=cfg.get("load_best_model_at_end", True),
metric_for_best_model=cfg.get("metric_for_best_model", "eval_loss"),
predict_with_generate=cfg.get("predict_with_generate", True),
generation_num_beams=cfg.get("num_beams", 4),
generation_max_new_tokens=cfg.get("max_new_tokens", 128),
logging_steps=cfg.get("logging_steps", 50),
report_to=cfg.get("report_to", "none"),
)
if "max_steps" in cfg and cfg["max_steps"] > 0:
training_kwargs["max_steps"] = cfg["max_steps"]
training_args = Seq2SeqTrainingArguments(**training_kwargs)
data_collator = DataCollatorForSeq2Seq(tokenizer, model=model, padding=True)
# ------------------------------------------------------------------ #
# Trainer
# ------------------------------------------------------------------ #
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_ds,
eval_dataset=val_ds,
tokenizer=tokenizer,
data_collator=data_collator,
)
print(f"\nStarting training - model: {cfg['model_name']} output: {output_dir}")
trainer.train()
print("\nSaving final model ...")
trainer.save_model(os.path.join(output_dir, "checkpoint-best"))
tokenizer.save_pretrained(os.path.join(output_dir, "checkpoint-best"))
# Save config alongside checkpoint for reproducibility
with open(os.path.join(output_dir, "train_config.yaml"), "w") as f:
yaml.dump(cfg, f)
print(f"\nDone. Checkpoint saved to {output_dir}/checkpoint-best")
# Quick smoke generation
if args.smoke and val_rows:
print("\n--- Smoke generation sample ---")
model.eval()
sample = val_rows[0]
inputs = tokenizer(prefix + sample["ai"], return_tensors="pt",
max_length=max_in, truncation=True)
inputs = {k: v.to(model.device) for k, v in inputs.items()}
with torch.no_grad():
out = model.generate(**inputs, num_beams=4, max_new_tokens=64)
generated = tokenizer.decode(out[0], skip_special_tokens=True)
print(f" AI input: {sample['ai'][:120]}...")
print(f" Generated: {generated[:120]}...")
print(f" Human ref: {sample['human'][:120]}...")
if __name__ == "__main__":
main()