ExamSicurezza/Florence-2-Image-Caption

1import gradio as gr
2import subprocess
3import torch
4from PIL import Image
5from transformers import AutoProcessor, AutoModelForCausalLM
6
7try:
8    subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, check=True, shell=True)
9except subprocess.CalledProcessError as e:
10    print(f"Error installing flash-attn: {e}")
11    print("Continuing without flash-attn.")
12
13device = "cuda" if torch.cuda.is_available() else "cpu"
14vision_language_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval()
15vision_language_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True)
16
17def describe_image(uploaded_image):
18    """
19    Generates a detailed description of the input image.
20
21    Args:
22        uploaded_image (PIL.Image.Image or numpy.ndarray): The image to describe.
23
24    Returns:
25        str: A detailed textual description of the image.
26    """
27    if not isinstance(uploaded_image, Image.Image):
28        uploaded_image = Image.fromarray(uploaded_image)
29
30    inputs = vision_language_processor(text="<MORE_DETAILED_CAPTION>", images=uploaded_image, return_tensors="pt").to(device)
31    with torch.no_grad():
32        generated_ids = vision_language_model.generate(
33            input_ids=inputs["input_ids"],
34            pixel_values=inputs["pixel_values"],
35            max_new_tokens=1024,
36            early_stopping=False,
37            do_sample=False,
38            num_beams=3,
39        )
40    generated_text = vision_language_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
41    processed_description = vision_language_processor.post_process_generation(
42        generated_text,
43        task="<MORE_DETAILED_CAPTION>",
44        image_size=(uploaded_image.width, uploaded_image.height)
45    )
46    image_description = processed_description["<MORE_DETAILED_CAPTION>"]
47    print("\nImage description generated!:", image_description)
48    return image_description
49
50image_description_interface = gr.Interface(
51    fn=describe_image,
52    inputs=gr.Image(label="Upload Image"),
53    outputs=gr.Textbox(label="Generated Caption", lines=4, show_copy_button=True),
54    live=False,
55)
56
57image_description_interface.launch(debug=True, ssr_mode=False)