from transformers import VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
from PIL import Image

model = VisionEncoderDecoderModel.from_pretrained("vit-gpt2-image-captioning")
feature_extractor = ViTImageProcessor.from_pretrained("vit-gpt2-image-captioning")
tokenizer = AutoTokenizer.from_pretrained("vit-gpt2-image-captioning")

max_length = 16
num_beams = 4
gen_kwargs = {"max_length": max_length, "num_beams": num_beams}


def predict_step(img_array):
    i_image = Image.fromarray(img_array)

    if i_image.mode != "RGB":
        i_image = i_image.convert(mode="RGB")

    pixel_values = feature_extractor(images=i_image, return_tensors="pt", do_normalize=True).pixel_values

    output_ids = model.generate(pixel_values, **gen_kwargs)

    pred = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
    pred = [p.strip() for p in pred]
    return pred