app.py

import gradio as gr
from transformers import AutoProcessor, AutoTokenizer, AutoImageProcessor, AutoModelForCausalLM, BlipForConditionalGeneration, Blip2ForConditionalGeneration, VisionEncoderDecoderModel
import torch
import open_clip

from huggingface_hub import hf_hub_download


git_processor_large_coco = AutoProcessor.from_pretrained("microsoft/git-large-coco")
git_model_large_coco = AutoModelForCausalLM.from_pretrained("microsoft/git-large-coco")

git_processor_large_textcaps = AutoProcessor.from_pretrained("microsoft/git-large-r-textcaps")
git_model_large_textcaps = AutoModelForCausalLM.from_pretrained("microsoft/git-large-r-textcaps")



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

git_model_large_coco.to(device)
git_model_large_textcaps.to(device)


def generate_caption(processor, model, image, tokenizer=None, use_float_16=False):
    inputs = processor(images=image, return_tensors="pt").to(device)

    if use_float_16:
        inputs = inputs.to(torch.float16)
    
    generated_ids = model.generate(pixel_values=inputs.pixel_values, max_length=50)

    if tokenizer is not None:
        generated_caption = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    else:
        generated_caption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
   
    return generated_caption


def generate_caption_coca(model, transform, image):
    im = transform(image).unsqueeze(0).to(device)
    with torch.no_grad(), torch.cuda.amp.autocast():
        generated = model.generate(im, seq_len=20)
    return open_clip.decode(generated[0].detach()).split("<end_of_text>")[0].replace("<start_of_text>", "")


def generate_captions(image):


    caption_git_large_coco = generate_caption(git_processor_large_coco, git_model_large_coco, image)

    caption_git_large_textcaps = generate_caption(git_processor_large_textcaps, git_model_large_textcaps, image)

    return caption_git_large_coco, caption_git_large_textcaps

   
outputs = [gr.outputs.Textbox(label="Caption generated by GIT-large fine-tuned on COCO"), gr.outputs.Textbox(label="Caption generated by GIT-large fine-tuned on TextCaps")] 

title = "Interactive demo: comparing image captioning models"
description = "Gradio Demo to compare GIT 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 = ""

interface = gr.Interface(fn=generate_captions, 
                         inputs=gr.inputs.Image(type="pil"),
                         outputs=outputs, 
                         title=title,
                         description=description,
                         article=article, 
                         enable_queue=True)
interface.launch(debug=True)