What is the format of structured data and what is the proposed way to read it?
#2
by
boydcheung - opened
Is there any nice python code to do convert the structured code to tabular like in panda?
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))
achang changed discussion status to
closed