app.py

import gradio as gr
import subprocess
import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM

subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)

device = "cuda" if torch.cuda.is_available() else "cpu"
florence_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval()
florence_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True)

def generate_captions(image):
    if not isinstance(image, Image.Image):
        image = Image.fromarray(image)
    
    inputs = florence_processor(text="<MORE_DETAILED_CAPTION>", images=image, return_tensors="pt").to(device)
    
    captions = []
    for i in range(3):
        generated_ids = florence_model.generate(
            input_ids=inputs["input_ids"],
            pixel_values=inputs["pixel_values"],
            max_new_tokens=1024,
            early_stopping=False,
            do_sample=True,  
            temperature=0.7 + i * 0.1,
            num_beams=3
        )
        generated_text = florence_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
        parsed_answer = florence_processor.post_process_generation(
            generated_text,
            task="<MORE_DETAILED_CAPTION>",
            image_size=(image.width, image.height)
        )
        prompt = parsed_answer["<MORE_DETAILED_CAPTION>"]
        captions.append(prompt)
        print(f"\n\nGeneration {i+1} completed!:" + prompt)
    
    return "\n\n".join([f"Caption {i+1}: {caption}" for i, caption in enumerate(captions)])

io = gr.Interface(
    generate_captions,
    inputs=[gr.Image(label="Input Image")],
    outputs=[gr.Textbox(label="Output Captions", lines=10, show_copy_button=True)]
)

io.launch(debug=True)