Datasets:

achang

/

plot_qa

like 7

I made the processed dataset here jonathanjordan21/plot_qa_processed

Here is example how to use

def pre_process(sample, processor, max_length=512, ignore_id=-100):
    # create tensor from image
    try:
        image = sample["image"].convert('RGB')
        pixel_values = processor(
            image, return_tensors="pt").pixel_values.squeeze()
    except Exception as e:
        print(sample)
        print(f"Error: {e}")
        return {}

    # tokenize document
    input_ids = processor.tokenizer(
        sample["text"],
        add_special_tokens=False,
        max_length=max_length,
        padding="max_length",
        truncation=True,
        return_tensors="pt",
    )["input_ids"].squeeze(0)

    labels = input_ids.clone()
    # model doesn't need to predict pad token
    labels[labels == processor.tokenizer.pad_token_id] = ignore_id
    return {"pixel_values": pixel_values, "labels": labels, "target_sequence": sample["text"]}

dataset = load_dataset("achang/plot_qa", streaming=True)
processor = DonutProcessor.from_pretrained("naver-clova-ix/donut-base")
processed_dataset = dataset.map(partial(pre_process, processor=processor, max_length=512, ignore_id=-100))