Datasets:
Tasks:
Text Classification
Modalities:
Text
Formats:
parquet
Sub-tasks:
sentiment-classification
Languages:
English
Size:
10K - 100K
ArXiv:
License:
Delete loading script auxiliary file
Browse files- lm_finetuning.py +0 -250
lm_finetuning.py
DELETED
|
@@ -1,250 +0,0 @@
|
|
| 1 |
-
'''
|
| 2 |
-
wandb offline
|
| 3 |
-
export WANDB_DISABLED='true'
|
| 4 |
-
export RAY_RESULTS='ray_results'
|
| 5 |
-
|
| 6 |
-
python lm_finetuning.py -m "roberta-large" -o "ckpt/2021/roberta-large" --push-to-hub --hf-organization "cardiffnlp" -a "roberta-large-tweet-topic-multi-all" --split-train "train_all" --split-valid "validation_2021" --split-test "test_2021"
|
| 7 |
-
python lm_finetuning.py -m "roberta-large" -o "ckpt/2020/roberta-large" --push-to-hub --hf-organization "cardiffnlp" -a "roberta-large-tweet-topic-multi-2020" --split-train "train_2020" --split-valid "validation_2020" --split-test "test_2021"
|
| 8 |
-
|
| 9 |
-
python lm_finetuning.py -m "roberta-base" -o "ckpt/2021/roberta_base" --push-to-hub --hf-organization "cardiffnlp" -a "roberta-base-tweet-topic-multi-all" --split-train "train_all" --split-valid "validation_2021" --split-test "test_2021"
|
| 10 |
-
python lm_finetuning.py -m "roberta-base" -o "ckpt/2020/roberta_base" --push-to-hub --hf-organization "cardiffnlp" -a "roberta-base-tweet-topic-multi-2020" --split-train "train_2020" --split-valid "validation_2020" --split-test "test_2021"
|
| 11 |
-
|
| 12 |
-
python lm_finetuning.py -m "cardiffnlp/twitter-roberta-base-2019-90m" -o "ckpt/2021/twitter-roberta-base-2019-90m" --push-to-hub --hf-organization "cardiffnlp" -a "twitter-roberta-base-2019-90m-tweet-topic-multi-all" --split-train "train_all" --split-valid "validation_2021" --split-test "test_2021"
|
| 13 |
-
python lm_finetuning.py -m "cardiffnlp/twitter-roberta-base-2019-90m" -o "ckpt/2020/twitter-roberta-base-2019-90m" --push-to-hub --hf-organization "cardiffnlp" -a "twitter-roberta-base-2019-90m-tweet-topic-multi-2020" --split-train "train_2020" --split-valid "validation_2020" --split-test "test_2021"
|
| 14 |
-
|
| 15 |
-
python lm_finetuning.py -m "cardiffnlp/twitter-roberta-base-dec2020" -o "ckpt/2021/twitter-roberta-base-dec2020" --push-to-hub --hf-organization "cardiffnlp" -a "twitter-roberta-base-dec2020-tweet-topic-multi-all" --split-train "train_all" --split-valid "validation_2021" --split-test "test_2021"
|
| 16 |
-
python lm_finetuning.py -m "cardiffnlp/twitter-roberta-base-dec2020" -o "ckpt/2020/twitter-roberta-base-dec2020" --push-to-hub --hf-organization "cardiffnlp" -a "twitter-roberta-base-dec2020-tweet-topic-multi-2020" --split-train "train_2020" --split-valid "validation_2020" --split-test "test_2021"
|
| 17 |
-
|
| 18 |
-
python lm_finetuning.py -m "cardiffnlp/twitter-roberta-base-dec2021" -o "ckpt/2021/twitter-roberta-base-dec2021" --push-to-hub --hf-organization "cardiffnlp" -a "twitter-roberta-base-dec2021-tweet-topic-multi-all" --split-train "train_all" --split-valid "validation_2021" --split-test "test_2021"
|
| 19 |
-
python lm_finetuning.py -m "cardiffnlp/twitter-roberta-base-dec2021" -o "ckpt/2020/twitter-roberta-base-dec2021" --push-to-hub --hf-organization "cardiffnlp" -a "twitter-roberta-base-dec2021-tweet-topic-multi-2020" --split-train "train_2020" --split-valid "validation_2020" --split-test "test_2021"
|
| 20 |
-
'''
|
| 21 |
-
|
| 22 |
-
import argparse
|
| 23 |
-
import json
|
| 24 |
-
import logging
|
| 25 |
-
import os
|
| 26 |
-
import math
|
| 27 |
-
import shutil
|
| 28 |
-
import urllib.request
|
| 29 |
-
import multiprocessing
|
| 30 |
-
from os.path import join as pj
|
| 31 |
-
|
| 32 |
-
import torch
|
| 33 |
-
import numpy as np
|
| 34 |
-
from huggingface_hub import create_repo
|
| 35 |
-
from datasets import load_dataset, load_metric
|
| 36 |
-
from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
|
| 37 |
-
from ray import tune
|
| 38 |
-
|
| 39 |
-
from readme import get_readme
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
logging.basicConfig(format='%(asctime)s %(levelname)-8s %(message)s', level=logging.INFO, datefmt='%Y-%m-%d %H:%M:%S')
|
| 43 |
-
|
| 44 |
-
PARALLEL = bool(int(os.getenv("PARALLEL", 1)))
|
| 45 |
-
RAY_RESULTS = os.getenv("RAY_RESULTS", "ray_results")
|
| 46 |
-
LABEL2ID = {
|
| 47 |
-
"arts_&_culture": 0,
|
| 48 |
-
"business_&_entrepreneurs": 1,
|
| 49 |
-
"celebrity_&_pop_culture": 2,
|
| 50 |
-
"diaries_&_daily_life": 3,
|
| 51 |
-
"family": 4,
|
| 52 |
-
"fashion_&_style": 5,
|
| 53 |
-
"film_tv_&_video": 6,
|
| 54 |
-
"fitness_&_health": 7,
|
| 55 |
-
"food_&_dining": 8,
|
| 56 |
-
"gaming": 9,
|
| 57 |
-
"learning_&_educational": 10,
|
| 58 |
-
"music": 11,
|
| 59 |
-
"news_&_social_concern": 12,
|
| 60 |
-
"other_hobbies": 13,
|
| 61 |
-
"relationships": 14,
|
| 62 |
-
"science_&_technology": 15,
|
| 63 |
-
"sports": 16,
|
| 64 |
-
"travel_&_adventure": 17,
|
| 65 |
-
"youth_&_student_life": 18
|
| 66 |
-
}
|
| 67 |
-
ID2LABEL = {v: k for k, v in LABEL2ID.items()}
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
def internet_connection(host='http://google.com'):
|
| 71 |
-
try:
|
| 72 |
-
urllib.request.urlopen(host)
|
| 73 |
-
return True
|
| 74 |
-
except:
|
| 75 |
-
return False
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
def sigmoid(x):
|
| 79 |
-
return 1 / (1 + math.exp(-x))
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
def get_metrics():
|
| 83 |
-
metric_accuracy = load_metric("accuracy", "multilabel")
|
| 84 |
-
metric_f1 = load_metric("f1", "multilabel")
|
| 85 |
-
|
| 86 |
-
# metric_f1.compute(predictions=[[0, 1, 1], [1, 1, 0]], references=[[0, 1, 1], [0, 1, 0]], average='micro')
|
| 87 |
-
# metric_accuracy.compute(predictions=[[0, 1, 1], [1, 1, 0]], references=[[0, 1, 1], [0, 1, 0]])
|
| 88 |
-
|
| 89 |
-
def compute_metric_search(eval_pred):
|
| 90 |
-
logits, labels = eval_pred
|
| 91 |
-
predictions = np.array([[int(sigmoid(j) > 0.5) for j in i] for i in logits])
|
| 92 |
-
return metric_f1.compute(predictions=predictions, references=labels, average='micro')
|
| 93 |
-
|
| 94 |
-
def compute_metric_all(eval_pred):
|
| 95 |
-
logits, labels = eval_pred
|
| 96 |
-
predictions = np.array([[int(sigmoid(j) > 0.5) for j in i] for i in logits])
|
| 97 |
-
return {
|
| 98 |
-
'f1': metric_f1.compute(predictions=predictions, references=labels, average='micro')['f1'],
|
| 99 |
-
'f1_macro': metric_f1.compute(predictions=predictions, references=labels, average='macro')['f1'],
|
| 100 |
-
'accuracy': metric_accuracy.compute(predictions=predictions, references=labels)['accuracy']
|
| 101 |
-
}
|
| 102 |
-
return compute_metric_search, compute_metric_all
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
def main():
|
| 106 |
-
parser = argparse.ArgumentParser(description='Fine-tuning language model.')
|
| 107 |
-
parser.add_argument('-m', '--model', help='transformer LM', default='roberta-base', type=str)
|
| 108 |
-
parser.add_argument('-d', '--dataset', help='', default='cardiffnlp/tweet_topic_multi', type=str)
|
| 109 |
-
parser.add_argument('--split-train', help='', required=True, type=str)
|
| 110 |
-
parser.add_argument('--split-validation', help='', required=True, type=str)
|
| 111 |
-
parser.add_argument('--split-test', help='', required=True, type=str)
|
| 112 |
-
parser.add_argument('-l', '--seq-length', help='', default=128, type=int)
|
| 113 |
-
parser.add_argument('--random-seed', help='', default=42, type=int)
|
| 114 |
-
parser.add_argument('--eval-step', help='', default=50, type=int)
|
| 115 |
-
parser.add_argument('-o', '--output-dir', help='Directory to output', default='ckpt_tmp', type=str)
|
| 116 |
-
parser.add_argument('-t', '--n-trials', default=10, type=int)
|
| 117 |
-
parser.add_argument('--push-to-hub', action='store_true')
|
| 118 |
-
parser.add_argument('--use-auth-token', action='store_true')
|
| 119 |
-
parser.add_argument('--hf-organization', default=None, type=str)
|
| 120 |
-
parser.add_argument('-a', '--model-alias', help='', default=None, type=str)
|
| 121 |
-
parser.add_argument('--summary-file', default='metric_summary.json', type=str)
|
| 122 |
-
parser.add_argument('--skip-train', action='store_true')
|
| 123 |
-
parser.add_argument('--skip-eval', action='store_true')
|
| 124 |
-
opt = parser.parse_args()
|
| 125 |
-
assert opt.summary_file.endswith('.json'), f'`--summary-file` should be a json file {opt.summary_file}'
|
| 126 |
-
# setup data
|
| 127 |
-
dataset = load_dataset(opt.dataset)
|
| 128 |
-
network = internet_connection()
|
| 129 |
-
# setup model
|
| 130 |
-
tokenizer = AutoTokenizer.from_pretrained(opt.model, local_files_only=not network)
|
| 131 |
-
model = AutoModelForSequenceClassification.from_pretrained(
|
| 132 |
-
opt.model,
|
| 133 |
-
id2label=ID2LABEL,
|
| 134 |
-
label2id=LABEL2ID,
|
| 135 |
-
num_labels=len(dataset[opt.split_train]['label'][0]),
|
| 136 |
-
local_files_only=not network,
|
| 137 |
-
problem_type="multi_label_classification"
|
| 138 |
-
)
|
| 139 |
-
tokenized_datasets = dataset.map(
|
| 140 |
-
lambda x: tokenizer(x["text"], padding="max_length", truncation=True, max_length=opt.seq_length),
|
| 141 |
-
batched=True)
|
| 142 |
-
# setup metrics
|
| 143 |
-
compute_metric_search, compute_metric_all = get_metrics()
|
| 144 |
-
|
| 145 |
-
if not opt.skip_train:
|
| 146 |
-
# setup trainer
|
| 147 |
-
trainer = Trainer(
|
| 148 |
-
model=model,
|
| 149 |
-
args=TrainingArguments(
|
| 150 |
-
output_dir=opt.output_dir,
|
| 151 |
-
evaluation_strategy="steps",
|
| 152 |
-
eval_steps=opt.eval_step,
|
| 153 |
-
seed=opt.random_seed
|
| 154 |
-
),
|
| 155 |
-
train_dataset=tokenized_datasets[opt.split_train],
|
| 156 |
-
eval_dataset=tokenized_datasets[opt.split_validation],
|
| 157 |
-
compute_metrics=compute_metric_search,
|
| 158 |
-
model_init=lambda x: AutoModelForSequenceClassification.from_pretrained(
|
| 159 |
-
opt.model,
|
| 160 |
-
return_dict=True,
|
| 161 |
-
num_labels=len(dataset[opt.split_train]['label'][0]),
|
| 162 |
-
id2label=ID2LABEL,
|
| 163 |
-
label2id=LABEL2ID
|
| 164 |
-
)
|
| 165 |
-
)
|
| 166 |
-
# parameter search
|
| 167 |
-
if PARALLEL:
|
| 168 |
-
best_run = trainer.hyperparameter_search(
|
| 169 |
-
hp_space=lambda x: {
|
| 170 |
-
"learning_rate": tune.loguniform(1e-6, 1e-4),
|
| 171 |
-
"num_train_epochs": tune.choice(list(range(1, 6))),
|
| 172 |
-
"per_device_train_batch_size": tune.choice([4, 8, 16, 32, 64]),
|
| 173 |
-
},
|
| 174 |
-
local_dir=RAY_RESULTS, direction="maximize", backend="ray", n_trials=opt.n_trials,
|
| 175 |
-
resources_per_trial={'cpu': multiprocessing.cpu_count(), "gpu": torch.cuda.device_count()},
|
| 176 |
-
|
| 177 |
-
)
|
| 178 |
-
else:
|
| 179 |
-
best_run = trainer.hyperparameter_search(
|
| 180 |
-
hp_space=lambda x: {
|
| 181 |
-
"learning_rate": tune.loguniform(1e-6, 1e-4),
|
| 182 |
-
"num_train_epochs": tune.choice(list(range(1, 6))),
|
| 183 |
-
"per_device_train_batch_size": tune.choice([4, 8, 16, 32, 64]),
|
| 184 |
-
},
|
| 185 |
-
local_dir=RAY_RESULTS, direction="maximize", backend="ray", n_trials=opt.n_trials
|
| 186 |
-
)
|
| 187 |
-
# finetuning
|
| 188 |
-
for n, v in best_run.hyperparameters.items():
|
| 189 |
-
setattr(trainer.args, n, v)
|
| 190 |
-
trainer.train()
|
| 191 |
-
trainer.save_model(pj(opt.output_dir, 'best_model'))
|
| 192 |
-
best_model_path = pj(opt.output_dir, 'best_model')
|
| 193 |
-
else:
|
| 194 |
-
best_model_path = pj(opt.output_dir, 'best_model')
|
| 195 |
-
|
| 196 |
-
# evaluation
|
| 197 |
-
model = AutoModelForSequenceClassification.from_pretrained(
|
| 198 |
-
best_model_path,
|
| 199 |
-
num_labels=len(dataset[opt.split_train]['label'][0]),
|
| 200 |
-
local_files_only=not network,
|
| 201 |
-
problem_type="multi_label_classification",
|
| 202 |
-
id2label=ID2LABEL,
|
| 203 |
-
label2id=LABEL2ID
|
| 204 |
-
)
|
| 205 |
-
trainer = Trainer(
|
| 206 |
-
model=model,
|
| 207 |
-
args=TrainingArguments(
|
| 208 |
-
output_dir=opt.output_dir,
|
| 209 |
-
evaluation_strategy="no",
|
| 210 |
-
seed=opt.random_seed
|
| 211 |
-
),
|
| 212 |
-
train_dataset=tokenized_datasets[opt.split_train],
|
| 213 |
-
eval_dataset=tokenized_datasets[opt.split_test],
|
| 214 |
-
compute_metrics=compute_metric_all
|
| 215 |
-
)
|
| 216 |
-
summary_file = pj(opt.output_dir, opt.summary_file)
|
| 217 |
-
if not opt.skip_eval:
|
| 218 |
-
result = {f'test/{k}': v for k, v in trainer.evaluate().items()}
|
| 219 |
-
logging.info(json.dumps(result, indent=4))
|
| 220 |
-
with open(summary_file, 'w') as f:
|
| 221 |
-
json.dump(result, f)
|
| 222 |
-
|
| 223 |
-
if opt.push_to_hub:
|
| 224 |
-
assert opt.hf_organization is not None, f'specify hf organization `--hf-organization`'
|
| 225 |
-
assert opt.model_alias is not None, f'specify hf organization `--model-alias`'
|
| 226 |
-
url = create_repo(opt.model_alias, organization=opt.hf_organization, exist_ok=True)
|
| 227 |
-
# if not opt.skip_train:
|
| 228 |
-
args = {"use_auth_token": opt.use_auth_token, "repo_url": url, "organization": opt.hf_organization}
|
| 229 |
-
trainer.model.push_to_hub(opt.model_alias, **args)
|
| 230 |
-
tokenizer.push_to_hub(opt.model_alias, **args)
|
| 231 |
-
if os.path.exists(summary_file):
|
| 232 |
-
shutil.copy2(summary_file, opt.model_alias)
|
| 233 |
-
extra_desc = f"This model is fine-tuned on `{opt.split_train}` split and validated on `{opt.split_test}` split of tweet_topic."
|
| 234 |
-
readme = get_readme(
|
| 235 |
-
model_name=f"{opt.hf_organization}/{opt.model_alias}",
|
| 236 |
-
metric=summary_file,
|
| 237 |
-
language_model=opt.model,
|
| 238 |
-
extra_desc= extra_desc
|
| 239 |
-
)
|
| 240 |
-
with open(f"{opt.model_alias}/README.md", "w") as f:
|
| 241 |
-
f.write(readme)
|
| 242 |
-
os.system(
|
| 243 |
-
f"cd {opt.model_alias} && git lfs install && git add . && git commit -m 'model update' && git push && cd ../")
|
| 244 |
-
shutil.rmtree(f"{opt.model_alias}") # clean up the cloned repo
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
if __name__ == '__main__':
|
| 248 |
-
main()
|
| 249 |
-
|
| 250 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|