training_scripts/finetune_t5.py

""" Fine-tune T5 on topic classification (multi-label multi-class classification)
```
python finetune_t5.py --dataset-name ja --model-alias mt5-small-tweet-topic-ja --model-organization cardiffnlp --low-cpu-mem-usage
```
"""
import json
import logging
import os
import argparse
import gc
from glob import glob
from typing import List, Set
from shutil import copyfile
from statistics import mean
from distutils.dir_util import copy_tree

import torch
import transformers
from datasets import load_dataset
from transformers import Seq2SeqTrainer, Seq2SeqTrainingArguments, pipeline
from huggingface_hub import Repository


os.environ['TOKENIZERS_PARALLELISM'] = 'false'  # turn-off the warning message
os.environ['WANDB_DISABLED'] = 'true'  # disable wandb
_LR = [1e-6, 1e-5, 1e-4]
_BATCH = 32
_EPOCH = 5
_CLASS_MAP = {
    'Arts & Culture': ['Τέχνες & Πολιτισμός', 'Arte y cultura', 'アート&カルチャー'],
    'Business & Entrepreneurs': ['Επιχειρήσεις & Επιχειρηματίες', 'Negocios y emprendedores', 'ビジネス'],
    'Celebrity & Pop Culture': ['Διασημότητες & Ποπ κουλτούρα', 'Celebridades y cultura pop', '芸能'],
    'Diaries & Daily Life': ['Ημερολόγια & Καθημερινή ζωή', 'Diarios y vida diaria', '日常'],
    'Family': ['Οικογένεια', 'Familia', '家族'],
    'Fashion & Style': ['Μόδα & Στυλ', 'Moda y estilo', 'ファッション'],
    'Film, TV & Video': ['Ταινίες, τηλεόραση & βίντεο', 'Cine, televisión y video', '映画&ラジオ'],
    'Fitness & Health': ['Γυμναστική & Υεία', 'Estado físico y salud', 'フィットネス&健康'],
    'Food & Dining': ['Φαγητό & Δείπνο', 'Comida y comedor', '料理'],
    'Learning & Educational': ['Μάθηση & Εκπαίδευση', 'Aprendizaje y educación', '教育関連'],
    'News & Social Concern': ['Ειδήσεις & Κοινωνία', 'Noticias e interés social', '社会'],
    'Relationships': ['Σχέσεις', 'Relaciones', '人間関係'],
    'Science & Technology': ['Επιστήμη & Τεχνολογία', 'Ciencia y Tecnología', 'サイエンス'],
    'Youth & Student Life': ['Νεανική & Φοιτητική ζωή', 'Juventud y Vida Estudiantil', '学校'],
    'Music': ['Μουσική', 'Música', '音楽'],
    'Gaming': ['Παιχνίδια', 'Juegos', 'ゲーム'],
    'Sports': ['Αθλητισμός', 'Deportes', 'スポーツ'],
    'Travel & Adventure': ['Ταξίδια & Περιπέτεια', 'Viajes y aventuras', '旅行'],
    'Other Hobbies': ['Άλλα χόμπι', 'Otros pasatiempos', 'その他']
}


def load_model(
        model_name: str,
        use_auth_token: bool = False,
        low_cpu_mem_usage: bool = False) -> transformers.PreTrainedModel:
    """Load language model from huggingface model hub."""
    # config & tokenizer
    config = transformers.AutoConfig.from_pretrained(model_name, use_auth_token=use_auth_token)
    if config.model_type == 't5':  # T5 model requires T5ForConditionalGeneration class
        model_class = transformers.T5ForConditionalGeneration.from_pretrained
    elif config.model_type == 'mt5':
        model_class = transformers.MT5ForConditionalGeneration.from_pretrained
    elif config.model_type == 'bart':
        model_class = transformers.BartForConditionalGeneration.from_pretrained
    elif config.model_type == 'mbart':
        model_class = transformers.MBartForConditionalGeneration.from_pretrained
    else:
        raise ValueError(f'unsupported model type: {config.model_type}')
    param = {'config': config, 'use_auth_token': use_auth_token, 'low_cpu_mem_usage': low_cpu_mem_usage}
    return model_class(model_name, **param)


def train(
        model_name: str,
        model_low_cpu_mem_usage: bool,
        dataset: str,
        dataset_name: str,
        dataset_column_label: str,
        dataset_column_text: str,
        random_seed: int,
        use_auth_token: bool):
    """Fine-tune seq2seq model."""
    logging.info(f'[TRAIN]\n\t *LM: {model_name}, \n\t *Data: {dataset} ({dataset_name})')
    output_dir = f'ckpt/{os.path.basename(model_name)}.{os.path.basename(dataset)}.{dataset_name}'

    tokenizer = transformers.AutoTokenizer.from_pretrained(model_name, use_auth_token=use_auth_token)
    dataset_instance = load_dataset(dataset, dataset_name, split="train", use_auth_token=use_auth_token)
    tokenized_dataset = []
    for d in dataset_instance:
        model_inputs = tokenizer(d[dataset_column_text], truncation=True)
        model_inputs['labels'] = tokenizer(text_target=d[dataset_column_label], truncation=True)['input_ids']
        tokenized_dataset.append(model_inputs)

    for n, lr_tmp in enumerate(_LR):
        logging.info(f"[TRAIN {n}/{len(_LR)}] lr: {lr_tmp}")
        output_dir_tmp = f"{output_dir}/model_lr_{lr_tmp}"
        if os.path.exists(f"{output_dir_tmp}/pytorch_model.bin"):
            continue
        model = load_model(
            model_name=model_name, use_auth_token=use_auth_token, low_cpu_mem_usage=model_low_cpu_mem_usage
        )
        trainer = Seq2SeqTrainer(
            model=model,
            args=Seq2SeqTrainingArguments(
                num_train_epochs=_EPOCH,
                learning_rate=lr_tmp,
                output_dir=output_dir_tmp,
                save_strategy="epoch",
                evaluation_strategy="no",
                seed=random_seed,
                per_device_train_batch_size=_BATCH,
            ),
            data_collator=transformers.DataCollatorForSeq2Seq(tokenizer, model=model),
            train_dataset=tokenized_dataset.copy(),
        )
        # train
        trainer.train()
        del trainer
        del model
        gc.collect()
        torch.cuda.empty_cache()

    for model_path in glob(f"{output_dir}/*/*"):
        tokenizer.save_pretrained(model_path)


def get_f1_score(references: List[Set[str]], predictions: List[Set[str]]) -> float:
    scores = []
    for g, r in zip(references, predictions):
        tp = len(set(g).intersection(set(r)))
        fp = len([_g for _g in g if _g not in r])
        fn = len([_r for _r in r if _r not in g])
        f1 = 0 if tp == 0 else 2 * tp / (2 * tp + fp + fn)
        scores.append(f1)
    return mean(scores)


def unify_label(label: Set[str]):
    new_label = []
    for label_tmp in label:
        label_en = [k for k, v in _CLASS_MAP.items() if label_tmp in v]
        if label_en:
            new_label.append(label_en[0])
    return set(new_label)


def get_metric(
        prediction_file: str,
        metric_file: str,
        model_path: str,
        data: List[str],
        label: List[str]) -> float:
    if os.path.exists(metric_file):
        with open(metric_file) as f:
            eval_metric = json.load(f)
        return eval_metric['f1']
    if not os.path.exists(prediction_file):
        pipe = pipeline(
            'text2text-generation',
            model=model_path,
            device='cuda:0' if torch.cuda.is_available() else 'cpu',
        )
        output = pipe(data, batch_size=_BATCH)
        output = [i['generated_text'] for i in output]
        with open(prediction_file, 'w') as f:
            f.write('\n'.join(output))
    with open(prediction_file) as f:
        output = [unify_label(set(i.split(','))) for i in f.read().split('\n')]
    label = [unify_label(set(i.split(','))) for i in label]
    eval_metric = {'f1': get_f1_score(label, output)}
    logging.info(json.dumps(eval_metric, indent=4))
    with open(metric_file, 'w') as f:
        json.dump(eval_metric, f)
    return eval_metric['f1']


def validate(
        model_name: str,
        dataset: str,
        dataset_name: str,
        dataset_column_text: str,
        use_auth_token: bool,
        dataset_column_label: str):
    logging.info(f'[VALIDATE]\n\t *LM: {model_name}, \n\t *Data: {dataset} ({dataset_name})')
    output_dir = f'ckpt/{os.path.basename(model_name)}.{os.path.basename(dataset)}.{dataset_name}'
    dataset_instance = load_dataset(dataset, dataset_name, split='validation', use_auth_token=use_auth_token)
    label = [i[dataset_column_label] for i in dataset_instance]
    data = [i[dataset_column_text] for i in dataset_instance]
    model_score = []
    for model_path in glob(f"{output_dir}/*/*/pytorch_model.bin"):
        model_path = os.path.dirname(model_path)
        prediction_file = f"{model_path}/prediction.validate.{os.path.basename(dataset)}.{dataset_name}.txt"
        metric_file = f"{model_path}/metric.validate.{os.path.basename(dataset)}.{dataset_name}.json"
        metric = get_metric(
            prediction_file=prediction_file,
            metric_file=metric_file,
            model_path=model_path,
            label=label,
            data=data
        )
        model_score.append([model_path, metric])
    model_score = sorted(model_score, key=lambda x: x[1])
    logging.info('Validation Result')
    for k, v in model_score:
        logging.info(f'{k}: {v}')
    best_model = model_score[-1][0]
    best_model_path = f'{output_dir}/best_model'
    copy_tree(best_model, best_model_path)


def test(
        model_name: str,
        dataset: str,
        dataset_name: str,
        dataset_column_text: str,
        use_auth_token: bool,
        dataset_column_label: str):
    logging.info(f'[TEST]\n\t *LM: {model_name}, \n\t *Data: {dataset} ({dataset_name})')
    output_dir = f'ckpt/{os.path.basename(model_name)}.{os.path.basename(dataset)}.{dataset_name}'
    dataset_instance = load_dataset(dataset, dataset_name, split='test', use_auth_token=use_auth_token)
    label = [i[dataset_column_label] for i in dataset_instance]
    data = [i[dataset_column_text] for i in dataset_instance]
    model_path = f'{output_dir}/best_model'
    if not os.path.exists(model_path):
        model_path = os.path.basename(model_name)

    prediction_file = f"{model_path}/prediction.{os.path.basename(dataset)}.{dataset_name}.txt"
    metric_file = f"{model_path}/metric.{os.path.basename(dataset)}.{dataset_name}.json"
    metric = get_metric(
        prediction_file=prediction_file,
        metric_file=metric_file,
        model_path=model_path,
        label=label,
        data=data
    )
    logging.info(f'Test Result: {metric}')


def upload(
        model_name: str,
        dataset: str,
        dataset_name: str,
        dataset_column_text: str,
        use_auth_token: bool,
        model_alias: str,
        model_organization: str):
    assert model_alias is not None and model_organization is not None,\
        'model_organization must be specified when model_alias is specified'
    logging.info('uploading to huggingface')
    output_dir = f'ckpt/{os.path.basename(model_name)}.{os.path.basename(dataset)}.{dataset_name}'
    args = {'use_auth_token': use_auth_token, 'organization': model_organization}
    model_path = f'{output_dir}/best_model'
    if not os.path.exists(model_path):
        model_path = os.path.basename(model_name)
    model = load_model(model_name=model_path)
    tokenizer = transformers.AutoTokenizer.from_pretrained(model_name, use_auth_token=use_auth_token)
    model.push_to_hub(model_alias, **args)
    tokenizer.push_to_hub(model_alias, **args)
    repo = Repository(model_alias, f'{model_organization}/{model_alias}')
    for i in glob(f'{model_path}/*'):
        if not os.path.exists(f'{model_alias}/{os.path.basename(i)}'):
            copyfile(i, f'{model_alias}/{os.path.basename(i)}')
    dataset_instance = load_dataset(dataset, dataset_name, split='validation', use_auth_token=use_auth_token)
    sample = [i[dataset_column_text] for i in dataset_instance]
    sample = [i for i in sample if "'" not in i and '"' not in i][:3]
    widget = '\n'.join([f"- text: '{t}'\n  example_title: example {_n + 1}" for _n, t in enumerate(sample)])
    with open(f'{model_alias}/README.md', 'w') as f:
        f.write(f"""
---
widget:
{widget}
---

# {model_organization}/{model_alias}

This is [{model_name}](https://huggingface.co/{model_name}) fine-tuned on [{dataset} ({dataset_name})](https://huggingface.co/datasets/{dataset}).

### Usage

```python
from transformers import pipeline

pipe = pipeline('text2text-generation', model='{model_organization}/{model_alias}')
output = pipe('{sample[0]}')
```
        """)
    repo.push_to_hub()


if __name__ == '__main__':
    # arguments
    logging.basicConfig(format='%(asctime)s %(levelname)-8s %(message)s', level=logging.INFO, datefmt='%Y-%m-%d %H:%M:%S')
    parser = argparse.ArgumentParser(description='Seq2Seq LM Fine-tuning on topic classification.')
    parser.add_argument('-m', '--model-name', default='google/mt5-small', type=str)
    parser.add_argument('--low-cpu-mem-usage', action='store_true')
    parser.add_argument('-d', '--dataset', default='cardiffnlp/tweet_topic_multilingual', type=str)
    parser.add_argument('--dataset-name', default='ja', type=str)
    parser.add_argument('--dataset-column-label', default='label_name_flatten', type=str)
    parser.add_argument('--dataset-column-text', default='text', type=str)
    parser.add_argument('--random-seed', default=42, type=int)
    parser.add_argument('--use-auth-token', action='store_true')
    parser.add_argument('--model-alias', default=None, type=str)
    parser.add_argument('--model-organization', default=None, type=str)
    parser.add_argument('--skip-train', action='store_true')
    parser.add_argument('--skip-validate', action='store_true')
    parser.add_argument('--skip-test', action='store_true')
    parser.add_argument('--skip-upload', action='store_true')
    opt = parser.parse_args()

    if not opt.skip_train:
        train(
            model_name=opt.model_name,
            model_low_cpu_mem_usage=opt.low_cpu_mem_usage,
            dataset=opt.dataset,
            dataset_name=opt.dataset_name,
            dataset_column_label=opt.dataset_column_label,
            dataset_column_text=opt.dataset_column_text,
            random_seed=opt.random_seed,
            use_auth_token=opt.use_auth_token,
        )
    if not opt.skip_validate:
        validate(
            model_name=opt.model_name,
            dataset=opt.dataset,
            dataset_name=opt.dataset_name,
            dataset_column_label=opt.dataset_column_label,
            dataset_column_text=opt.dataset_column_text,
            use_auth_token=opt.use_auth_token
        )
    if not opt.skip_test:
        test(
            model_name=opt.model_name,
            dataset=opt.dataset,
            dataset_name=opt.dataset_name,
            dataset_column_label=opt.dataset_column_label,
            dataset_column_text=opt.dataset_column_text,
            use_auth_token=opt.use_auth_token
        )
    if not opt.skip_upload:
        upload(
            model_name=opt.model_name,
            dataset=opt.dataset,
            dataset_name=opt.dataset_name,
            dataset_column_text=opt.dataset_column_text,
            use_auth_token=opt.use_auth_token,
            model_alias=opt.model_alias,
            model_organization=opt.model_organization
        )