from argparse import ArgumentParser
from itertools import chain

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments, default_data_collator


parser = ArgumentParser()
parser.add_argument('-i', '--input', default='data',
                    help='training data input file')
parser.add_argument('-o', '--output', default='model',
                    help='output directory for trained model')
args = parser.parse_args()


# Load and tokenize dataset
raw_dataset = load_dataset('text', data_files={'train': args.input}, keep_linebreaks=True)
tokenizer = AutoTokenizer.from_pretrained('gpt2-large', use_fast=True)
tokenized_dataset = raw_dataset.map(lambda examples: tokenizer(examples['text']),
                                    batched=True, remove_columns='text')


# Generate chunks of block_size
block_size = tokenizer.model_max_length

# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
    # Concatenate all texts.
    concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
    total_length = len(concatenated_examples[list(examples.keys())[0]])
    # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
    # customize this part to your needs.
    if total_length >= block_size:
        total_length = (total_length // block_size) * block_size
    # Split by chunks of max_len.
    result = {
        k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
        for k, t in concatenated_examples.items()
    }
    result['labels'] = result['input_ids'].copy()
    return result

lm_dataset = tokenized_dataset.map(group_texts, batched=True)


# Create and train the model
model = AutoModelForCausalLM.from_pretrained('gpt2-large', low_cpu_mem_usage=True).to('cuda')
trainer = Trainer(model, TrainingArguments(output_dir=args.output, save_strategy='no',
    per_device_train_batch_size=1, gradient_checkpointing=True, optim='adafactor'),
    default_data_collator, lm_dataset['train'])
trainer.train()
trainer.save_model()