diff options
Diffstat (limited to 'train.py')
-rw-r--r-- | train.py | 13 |
1 files changed, 6 insertions, 7 deletions
@@ -1,7 +1,6 @@ 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 @@ -22,7 +21,7 @@ tokenized_dataset = raw_dataset.map(lambda examples: tokenizer(examples['text']) # Generate chunks of block_size -block_size = 256 # tokenizer.model_max_length +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): @@ -38,16 +37,16 @@ def group_texts(examples): 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() + 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', - torch_dtype=float16, low_cpu_mem_usage=True).to('cuda') -trainer = Trainer(model, TrainingArguments(output_dir=args.output, per_device_train_batch_size=1), - default_data_collator, lm_dataset['train']) +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() |