T5 Model For Summarization On GPU

Beginners
1

Hi Folks,

I have started learning and implementing models on huggingface, I was trying to fine tune T5 model on custom dataset but I am getting this error

—> 58 return data.pin_memory(device)

  • 59     elif isinstance(data, (str, bytes)):*

  • 60         return data*

RuntimeError: cannot pin ‘torch.cuda.LongTensor’ only dense CPU tensors can be pinned


tokenizer = AutoTokenizer.from_pretrained('t5-small')

def create_dataset_item(text, labels, model_name='t5-small'):

    input_text = "generate text: " + text
    target_text = labels

    inputs = tokenizer(
        input_text,
        return_tensors="pt",
        padding="max_length",
        truncation=True,
        max_length=1024,
    )
    lb = tokenizer(
        target_text,
        return_tensors="pt",
        padding="max_length",
        truncation=True,
        max_length=128,
    )
    inputs['labels'] = lb['input_ids']

    return {
        'input_ids': inputs['input_ids'].squeeze().to('cuda'),
        'attention_mask': inputs['attention_mask'].squeeze().to('cuda'),
        'labels': inputs['labels'].squeeze().to('cuda')
    }

def create_custom_dataset(data, model_name='t5-small'):
    dataset = []

    for _, item in data.iterrows():
        dataset.append(create_dataset_item(item["text"], item["labels"], model_name))

    return dataset


custom_dataset = create_custom_dataset(parsing_data)

model = AutoModelForSeq2SeqLM.from_pretrained('t5-small').to('cuda')

more …

training_args = Seq2SeqTrainingArguments(
    output_dir="pdf-modeling-v1",
    num_train_epochs=1,
    per_device_train_batch_size=8,
    save_steps=10_000,
    save_total_limit=2,
    learning_rate=5e-5,
    use_cpu=False,
)

trainer = Seq2SeqTrainer(
    model=model,
    args=training_args,
    data_collator=None,  # Use default collate function
    train_dataset=custom_dataset,
)

# Train the model
trainer.train()

Please let me know if i am making any mistakes