app.py

#!/usr/bin/env python

from __future__ import annotations

import gradio as gr
import PIL.Image
import spaces
import torch
from transformers import AutoProcessor, BlipForConditionalGeneration
from typing import Union
import os

DESCRIPTION = "# Image Captioning with LongCap"

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Using device: ", device)

model_id = "unography/blip-long-cap"
processor = AutoProcessor.from_pretrained(model_id)
model = BlipForConditionalGeneration.from_pretrained(model_id).to(device)

torch.hub.download_url_to_file("https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg", "demo.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"
)

@spaces.GPU()
def run(image: Union[str, PIL.Image.Image]) -> str:
    if isinstance(image, str):
        image = Image.open(image)
    inputs = processor(images=image, return_tensors="pt").to(device)
    out = model.generate(pixel_values=inputs.pixel_values, num_beams=3, repetition_penalty=2.5, max_length=300)
    generated_caption = processor.decode(out[0], skip_special_tokens=True)
    return generated_caption


with gr.Blocks(css="style.css") as demo:
    gr.Markdown(DESCRIPTION)
    input_image = gr.Image(type="pil")
    run_button = gr.Button("Caption")
    output = gr.Textbox(label="Result")
    gr.Examples(
        examples=[
            "demo.jpg",
            "stop_sign.png",
            "astronaut.jpg",
        ],
        inputs=input_image,
        outputs=output,
        fn=run,
        cache_examples=os.getenv("CACHE_EXAMPLES") == "1",
    )

    run_button.click(
        fn=run,
        inputs=input_image,
        outputs=output,
        api_name="caption",
    )

if __name__ == "__main__":
    demo.queue(max_size=20).launch()