init
Browse files- lm_finetuning.py +179 -0
lm_finetuning.py
ADDED
@@ -0,0 +1,179 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import json
|
3 |
+
import logging
|
4 |
+
import os
|
5 |
+
import shutil
|
6 |
+
import urllib.request
|
7 |
+
import multiprocessing
|
8 |
+
from os.path import join as pj
|
9 |
+
|
10 |
+
import torch
|
11 |
+
import numpy as np
|
12 |
+
from huggingface_hub import create_repo
|
13 |
+
from datasets import load_dataset, load_metric
|
14 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
|
15 |
+
from ray import tune
|
16 |
+
|
17 |
+
logging.basicConfig(format='%(asctime)s %(levelname)-8s %(message)s', level=logging.INFO, datefmt='%Y-%m-%d %H:%M:%S')
|
18 |
+
|
19 |
+
PARALLEL = bool(int(os.getenv("PARALLEL", 1)))
|
20 |
+
RAY_RESULTS = os.getenv("RAY_RESULTS", "ray_results")
|
21 |
+
|
22 |
+
|
23 |
+
def internet_connection(host='http://google.com'):
|
24 |
+
try:
|
25 |
+
urllib.request.urlopen(host)
|
26 |
+
return True
|
27 |
+
except:
|
28 |
+
return False
|
29 |
+
|
30 |
+
|
31 |
+
def get_metrics():
|
32 |
+
metric_accuracy = load_metric("accuracy", "multilabel")
|
33 |
+
metric_f1 = load_metric("f1", "multilabel")
|
34 |
+
|
35 |
+
# metric_f1.compute(predictions=[[0, 1, 1], [1, 1, 0]], references=[[0, 1, 1], [0, 1, 0]], average='micro')
|
36 |
+
# metric_accuracy.compute(predictions=[[0, 1, 1], [1, 1, 0]], references=[[0, 1, 1], [0, 1, 0]])
|
37 |
+
|
38 |
+
def compute_metric_search(eval_pred):
|
39 |
+
logits, labels = eval_pred
|
40 |
+
predictions = np.argmax(logits, axis=-1)
|
41 |
+
return metric_f1.compute(predictions=predictions, references=labels, average='micro')
|
42 |
+
|
43 |
+
def compute_metric_all(eval_pred):
|
44 |
+
logits, labels = eval_pred
|
45 |
+
predictions = np.argmax(logits, axis=-1)
|
46 |
+
return {
|
47 |
+
'f1': metric_f1.compute(predictions=predictions, references=labels, average='micro')['f1'],
|
48 |
+
'f1_macro': metric_f1.compute(predictions=predictions, references=labels, average='macro')['f1'],
|
49 |
+
'accuracy': metric_accuracy.compute(predictions=predictions, references=labels)['accuracy']
|
50 |
+
}
|
51 |
+
return compute_metric_search, compute_metric_all
|
52 |
+
|
53 |
+
|
54 |
+
def main():
|
55 |
+
parser = argparse.ArgumentParser(description='Fine-tuning language model.')
|
56 |
+
parser.add_argument('-m', '--model', help='transformer LM', default='roberta-base', type=str)
|
57 |
+
parser.add_argument('-d', '--dataset', help='', default='cardiffnlp/tweet_topic_multi', type=str)
|
58 |
+
parser.add_argument('--dataset-name', help='huggingface dataset name', default='citation_intent', type=str)
|
59 |
+
parser.add_argument('-l', '--seq-length', help='', default=128, type=int)
|
60 |
+
parser.add_argument('--random-seed', help='', default=42, type=int)
|
61 |
+
parser.add_argument('--eval-step', help='', default=50, type=int)
|
62 |
+
parser.add_argument('-o', '--output-dir', help='Directory to output', default='ckpt_tmp', type=str)
|
63 |
+
parser.add_argument('-t', '--n-trials', default=10, type=int)
|
64 |
+
parser.add_argument('--push-to-hub', action='store_true')
|
65 |
+
parser.add_argument('--use-auth-token', action='store_true')
|
66 |
+
parser.add_argument('--hf-organization', default=None, type=str)
|
67 |
+
parser.add_argument('-a', '--model-alias', help='', default=None, type=str)
|
68 |
+
parser.add_argument('--summary-file', default='metric_summary.json', type=str)
|
69 |
+
parser.add_argument('--skip-train', action='store_true')
|
70 |
+
parser.add_argument('--skip-eval', action='store_true')
|
71 |
+
opt = parser.parse_args()
|
72 |
+
assert opt.summary_file.endswith('.json'), f'`--summary-file` should be a json file {opt.summary_file}'
|
73 |
+
# setup data
|
74 |
+
dataset = load_dataset(opt.dataset, opt.dataset_name)
|
75 |
+
network = internet_connection()
|
76 |
+
# setup model
|
77 |
+
tokenizer = AutoTokenizer.from_pretrained(opt.model, local_files_only=not network)
|
78 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
79 |
+
opt.model,
|
80 |
+
num_labels=len(dataset['train'][0]['label']),
|
81 |
+
local_files_only=not network,
|
82 |
+
problem_type="multi_label_classification"
|
83 |
+
)
|
84 |
+
tokenized_datasets = dataset.map(
|
85 |
+
lambda x: tokenizer(x["text"], padding="max_length", truncation=True, max_length=opt.seq_length),
|
86 |
+
batched=True)
|
87 |
+
# setup metrics
|
88 |
+
compute_metric_search, compute_metric_all = get_metrics()
|
89 |
+
|
90 |
+
if not opt.skip_train:
|
91 |
+
# setup trainer
|
92 |
+
trainer = Trainer(
|
93 |
+
model=model,
|
94 |
+
args=TrainingArguments(
|
95 |
+
output_dir=opt.output_dir,
|
96 |
+
evaluation_strategy="steps",
|
97 |
+
eval_steps=opt.eval_step,
|
98 |
+
seed=opt.random_seed
|
99 |
+
),
|
100 |
+
train_dataset=tokenized_datasets["train"],
|
101 |
+
eval_dataset=tokenized_datasets["validation"],
|
102 |
+
compute_metrics=compute_metric_search,
|
103 |
+
model_init=lambda x: AutoModelForSequenceClassification.from_pretrained(
|
104 |
+
opt.model, return_dict=True, num_labels=dataset['train'].features['label'].num_classes)
|
105 |
+
)
|
106 |
+
# parameter search
|
107 |
+
if PARALLEL:
|
108 |
+
best_run = trainer.hyperparameter_search(
|
109 |
+
hp_space=lambda x: {
|
110 |
+
"learning_rate": tune.loguniform(1e-6, 1e-4),
|
111 |
+
"num_train_epochs": tune.choice(list(range(1, 6))),
|
112 |
+
"per_device_train_batch_size": tune.choice([4, 8, 16, 32, 64]),
|
113 |
+
},
|
114 |
+
local_dir=RAY_RESULTS, direction="maximize", backend="ray", n_trials=opt.n_trials,
|
115 |
+
resources_per_trial={'cpu': multiprocessing.cpu_count(), "gpu": torch.cuda.device_count()},
|
116 |
+
|
117 |
+
)
|
118 |
+
else:
|
119 |
+
best_run = trainer.hyperparameter_search(
|
120 |
+
hp_space=lambda x: {
|
121 |
+
"learning_rate": tune.loguniform(1e-6, 1e-4),
|
122 |
+
"num_train_epochs": tune.choice(list(range(1, 6))),
|
123 |
+
"per_device_train_batch_size": tune.choice([4, 8, 16, 32, 64]),
|
124 |
+
},
|
125 |
+
local_dir=RAY_RESULTS, direction="maximize", backend="ray", n_trials=opt.n_trials
|
126 |
+
)
|
127 |
+
# finetuning
|
128 |
+
for n, v in best_run.hyperparameters.items():
|
129 |
+
setattr(trainer.args, n, v)
|
130 |
+
trainer.train()
|
131 |
+
trainer.save_model(pj(opt.output_dir, 'best_model'))
|
132 |
+
best_model_path = pj(opt.output_dir, 'best_model')
|
133 |
+
else:
|
134 |
+
best_model_path = opt.output_dir
|
135 |
+
|
136 |
+
# evaluation
|
137 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
138 |
+
best_model_path,
|
139 |
+
num_labels=dataset['train'].features['label'].num_classes,
|
140 |
+
local_files_only=not network)
|
141 |
+
trainer = Trainer(
|
142 |
+
model=model,
|
143 |
+
args=TrainingArguments(
|
144 |
+
output_dir=opt.output_dir,
|
145 |
+
evaluation_strategy="no",
|
146 |
+
seed=opt.random_seed
|
147 |
+
),
|
148 |
+
train_dataset=tokenized_datasets["train"],
|
149 |
+
eval_dataset=tokenized_datasets["test"],
|
150 |
+
compute_metrics=compute_metric_all,
|
151 |
+
model_init=lambda x: AutoModelForSequenceClassification.from_pretrained(
|
152 |
+
opt.model, return_dict=True, num_labels=dataset['train'].features['label'].num_classes)
|
153 |
+
)
|
154 |
+
summary_file = pj(opt.output_dir, opt.summary_file)
|
155 |
+
if not opt.skip_eval:
|
156 |
+
result = {f'test/{k}': v for k, v in trainer.evaluate().items()}
|
157 |
+
logging.info(json.dumps(result, indent=4))
|
158 |
+
with open(summary_file, 'w') as f:
|
159 |
+
json.dump(result, f)
|
160 |
+
|
161 |
+
if opt.push_to_hub:
|
162 |
+
assert opt.hf_organization is not None, f'specify hf organization `--hf-organization`'
|
163 |
+
assert opt.model_alias is not None, f'specify hf organization `--model-alias`'
|
164 |
+
url = create_repo(opt.model_alias, organization=opt.hf_organization, exist_ok=True)
|
165 |
+
# if not opt.skip_train:
|
166 |
+
args = {"use_auth_token": opt.use_auth_token, "repo_url": url, "organization": opt.hf_organization}
|
167 |
+
trainer.model.push_to_hub(opt.model_alias, **args)
|
168 |
+
tokenizer.push_to_hub(opt.model_alias, **args)
|
169 |
+
if os.path.exists(summary_file):
|
170 |
+
shutil.copy2(summary_file, opt.model_alias)
|
171 |
+
os.system(
|
172 |
+
f"cd {opt.model_alias} && git lfs install && git add . && git commit -m 'model update' && git push && cd ../")
|
173 |
+
shutil.rmtree(f"{opt.model_alias}") # clean up the cloned repo
|
174 |
+
|
175 |
+
|
176 |
+
if __name__ == '__main__':
|
177 |
+
main()
|
178 |
+
|
179 |
+
|