1
from transformers import DonutProcessor, VisionEncoderDecoderModel
2
import torch
3

4
device = "cuda" if torch.cuda.is_available() else "cpu"
5

6
def model_fn(model_dir):
7
    # Load our model from Hugging Face
8
    processor = DonutProcessor.from_pretrained(model_dir)
9
    model = VisionEncoderDecoderModel.from_pretrained(model_dir)
10

11
    # Move model to GPU
12
    model.to(device)
13

14
    return model, processor
15

16

17
def predict_fn(data, model_and_processor):
18
    # unpack model and tokenizer
19
    model, processor = model_and_processor
20
    
21
    image = data.get("inputs")
22
    pixel_values = processor.feature_extractor(image, return_tensors="pt").pixel_values
23
    task_prompt = "<s>" # start of sequence token for decoder since we are not having a user prompt
24
    decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids
25

26
    # run inference
27
    outputs = model.generate(
28
        pixel_values.to(device),
29
        decoder_input_ids=decoder_input_ids.to(device),
30
        max_length=model.decoder.config.max_position_embeddings,
31
        early_stopping=True,
32
        pad_token_id=processor.tokenizer.pad_token_id,
33
        eos_token_id=processor.tokenizer.eos_token_id,
34
        use_cache=True,
35
        num_beams=1,
36
        bad_words_ids=[[processor.tokenizer.unk_token_id]],
37
        return_dict_in_generate=True,
38
    )
39

40
    # process output
41
    prediction = processor.batch_decode(outputs.sequences)[0]
42
    prediction = processor.token2json(prediction)
43

44
    return prediction
45

46

47