diff options
Diffstat (limited to 'train.py')
-rw-r--r-- | train.py | 12 |
1 files changed, 7 insertions, 5 deletions
@@ -1,6 +1,7 @@ from argparse import ArgumentParser from itertools import chain +from torch import float16 from datasets import load_dataset from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments, default_data_collator @@ -15,13 +16,13 @@ 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('distilgpt2', use_fast=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 +block_size = 256 # 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): @@ -44,8 +45,9 @@ lm_dataset = tokenized_dataset.map(group_texts, batched=True) # Create and train the model -model = AutoModelForCausalLM.from_pretrained('distilgpt2') -trainer = Trainer(model, TrainingArguments(output_dir=args.output), - default_data_collator, lm_dataset['train']) +model = AutoModelForCausalLM.from_pretrained('gpt2-large', + torch_dtype=float16, low_cpu_mem_usage=True).to('cuda') +trainer = Trainer(model, TrainingArguments(output_dir=args.output, per_device_train_batch_size=1, + gradient_accumulation_steps=8), default_data_collator, lm_dataset['train']) trainer.train() trainer.save_model() |