Add Got-OCR 2 Fast image processor and refactor slow one

[yonigozlan]

What does this PR do?

Refactor slow image processor of Got-OCR 2 in order to make only one call to preprocess, and not a separate call to crop_to_patches
Also add a fast image processor, with some nice speedups :)

def benchmark_image_processor(image_processor, images,benchmark_it=10, warmup_it=10):
    # warm up
    for _ in range(warmup_it):
        _ = image_processor(images=images, return_tensors="pt", device=device)
    # benchmark
    start_time = time.time()
    for _ in range(benchmark_it):
        _ = image_processor(images=images, return_tensors="pt", device=device)
    end_time = time.time()

    return (end_time - start_time) / benchmark_it

image = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
checkpoint = "stepfun-ai/GOT-OCR-2.0-hf"
image_processor_fast = AutoImageProcessor.from_pretrained(checkpoint, use_fast=True)
image_processor_slow = AutoImageProcessor.from_pretrained(checkpoint)
device = "cuda"
batch_size = 4

slow_time_one = benchmark_image_processor(image_processor_slow, image, benchmark_it=10)
fast_time_one = benchmark_image_processor(image_processor_fast, image, benchmark_it=10)
slow_time_batch = benchmark_image_processor(image_processor_slow, [image]*batch_size, benchmark_it=10)
fast_time_batch = benchmark_image_processor(image_processor_fast, [image]*batch_size, benchmark_it=10)

print(f"slow_time_one: {slow_time_one}, fast_time_one: {fast_time_one}, speedup: {slow_time_one/fast_time_one}")
print(f"slow_time_batch: {slow_time_batch}, fast_time_batch: {fast_time_batch}, speedup: {slow_time_batch/fast_time_batch}")

CPU:

slow_time_one: 0.03414120674133301, fast_time_one: 0.007321953773498535, speedup: 4.662854723954348
slow_time_batch: 0.1300074815750122, fast_time_batch: 0.030838775634765624, speedup: 4.215714758417654

CUDA:

slow_time_one: 0.03233206272125244, fast_time_one: 0.0006550788879394531, speedup: 49.35598340369778
slow_time_batch: 0.1255408763885498, fast_time_batch: 0.0017477989196777344, speedup: 71.82798603153816

Would be great to merge soon as this image processor will also be used for InternVL!

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

Just 1 comment sorry that it took so long