app.py

import gradio as gr
from transformers import AutoProcessor, AutoModelForCausalLM, BlipForQuestionAnswering, ViltForQuestionAnswering
import torch
import math

torch.hub.download_url_to_file('http://images.cocodataset.org/val2017/000000039769.jpg', 'cats.jpg')
torch.hub.download_url_to_file('https://huggingface.co/datasets/nielsr/textcaps-sample/resolve/main/stop_sign.png', 'stop_sign.png')
torch.hub.download_url_to_file('https://cdn.openai.com/dall-e-2/demos/text2im/astronaut/horse/photo/0.jpg', 'astronaut.jpg')

git_processor_base = AutoProcessor.from_pretrained("microsoft/git-base-vqav2")
git_model_base = AutoModelForCausalLM.from_pretrained("microsoft/git-base-vqav2")

# git_processor_large = AutoProcessor.from_pretrained("microsoft/git-large-vqav2")
# git_model_large = AutoModelForCausalLM.from_pretrained("microsoft/git-large-vqav2")

blip_processor_base = AutoProcessor.from_pretrained("Salesforce/blip-vqa-base")
blip_model_base = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base")

# blip_processor_large = AutoProcessor.from_pretrained("Salesforce/blip-vqa-capfilt-large")
# blip_model_large = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-capfilt-large")

vilt_processor = AutoProcessor.from_pretrained("dandelin/vilt-b32-finetuned-vqa")
vilt_model = ViltForQuestionAnswering.from_pretrained("dandelin/vilt-b32-finetuned-vqa")

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

git_model_base.to(device)
blip_model_base.to(device)
#git_model_large.to(device)
#blip_model_large.to(device)
vilt_model.to(device)

def generate_answer_git(processor, model, image, question):
    # prepare image
    pixel_values = processor(images=image, return_tensors="pt").pixel_values

    # prepare question
    input_ids = processor(text=question, add_special_tokens=False).input_ids
    input_ids = [processor.tokenizer.cls_token_id] + input_ids
    input_ids = torch.tensor(input_ids).unsqueeze(0)
    
    generated_ids = model.generate(pixel_values=pixel_values, input_ids=input_ids, max_length=50, return_dict_in_generate=True, output_scores=True)
    print(generated_ids.scores)
    print(generated_scores)
    # generated_answer = processor.batch_decode(generated_ids, skip_special_tokens=True)
   
    return 'haha'


def generate_answer_blip(processor, model, image, question):
    # prepare image + question
    inputs = processor(images=image, text=question, return_tensors="pt")
    
    generated_ids = model.generate(**inputs, max_length=50, output_scores=True)
    print(generated_ids)
    generated_answer = processor.batch_decode(generated_ids, skip_special_tokens=True)
   
    return generated_answer


def generate_answer_vilt(processor, model, image, question):
    # prepare image + question
    encoding = processor(images=image, text=question, return_tensors="pt")

    with torch.no_grad():
        outputs = model(**encoding)
    print(outputs.logits)
    print(torch.softmax(outputs.logits, dim=1))
    predicted_class_idx = outputs.logits.argmax(-1).item()
    print(f"prdicted_class_idx: {predicted_class_idx}")
    logitsList = torch.softmax(outputs.logits, dim=1).flatten().tolist()
    print(f"predicted_class_idx_in_list = {logitsList.index(max(logitsList))}")
    m = max(logitsList)
    s = -math.inf 
    for logit in logitsList:
        if s <= logit < m:
            s = logit
    t = sum(logitsList)
    pm, ps = m/t, s/t
    print(f"{pm}, {ps}")
    print(f"scaled: {pm/(pm + ps)}, {ps/(pm + ps)}")
    print(f"runnerup_idx_in_list = {logitsList.index(s)}")
    print(f"runnerup val: {model.config.id2label[logitsList.index(s)]}")
    return model.config.id2label[predicted_class_idx]


def generate_answers(image, question):
    answer_git_base = generate_answer_git(git_processor_base, git_model_base, image, question)

    # answer_git_large = generate_answer_git(git_processor_large, git_model_large, image, question)

    answer_blip_base = generate_answer_blip(blip_processor_base, blip_model_base, image, question)

    # answer_blip_large = generate_answer_blip(blip_processor_large, blip_model_large, image, question)

    answer_vilt = generate_answer_vilt(vilt_processor, vilt_model, image, question)

    return answer_git_base, answer_blip_base, answer_vilt

   
examples = [["cats.jpg", "How many cats are there?"], ["stop_sign.png", "What's behind the stop sign?"], ["astronaut.jpg", "What's the astronaut riding on?"]]
outputs = [gr.outputs.Textbox(label="Answer generated by GIT-base"), gr.outputs.Textbox(label="Answer generated by BLIP-base"), gr.outputs.Textbox(label="Answer generated by ViLT")] 

title = "Interactive demo: comparing visual question answering (VQA) models"
description = "Gradio Demo to compare GIT, BLIP and ViLT, 3 state-of-the-art vision+language models. To use it, simply upload your image and click 'submit', or click one of the examples to load them. Read more at the links below."
article = "<p style='text-align: center'><a href='https://huggingface.co/docs/transformers/main/model_doc/blip' target='_blank'>BLIP docs</a> | <a href='https://huggingface.co/docs/transformers/main/model_doc/git' target='_blank'>GIT docs</a></p>"

interface = gr.Interface(fn=generate_answers, 
                         inputs=[gr.inputs.Image(type="pil"), gr.inputs.Textbox(label="Question")],
                         outputs=outputs,
                         examples=examples, 
                         title=title,
                         description=description,
                         article=article, 
                         enable_queue=True)
interface.launch()