main.py

# note: if you have a mix of Ampere and newer, and also older than Ampere GPUs, set the environment variable
#    CUDA_VISIBLE_DEVICE=1,2,3 (for example) so that one or the other is excluded.
#    otherwise the script may fail with a flash attention exception.

import gradio as gr
import os
import argparse
import uuid
import zipfile
import torch
from PIL import Image
import requests
from transformers import AutoProcessor, AutoModelForCausalLM, GenerationConfig, BitsAndBytesConfig
from io import BytesIO
import base64
import atexit
import shutil


def cleanup_temp_files():
    # Delete the subdirectories inside the "images" directory
    if os.path.exists("images"):
        for dir_name in os.listdir("images"):
            dir_path = os.path.join("images", dir_name)
            if os.path.isdir(dir_path):
                shutil.rmtree(dir_path)

# Parse command-line arguments
parser = argparse.ArgumentParser(description="Load and use a quantized model")
parser.add_argument("-q", "--use_quant", action="store_true", help="Use quantized model")
args = parser.parse_args()

if torch.cuda.is_available():
    device = torch.device("cuda")
    print("GPU is available. Using CUDA.")
else:
    device = torch.device("cpu")
    print("GPU is not available. Using CPU.")

# Load the processor
local_path = "./model/Molmo-7B-D-0924"
processor = AutoProcessor.from_pretrained(
    local_path,
    local_files_only=True,
    trust_remote_code=True,
    torch_dtype='auto',
    device_map='auto'
)

# Load the model
if args.use_quant:
    # Load the quantized model
    quantized_local_path = "./model/molmo-7B-D-bnb-4bit"
    model = AutoModelForCausalLM.from_pretrained(
        quantized_local_path,
        trust_remote_code=True,
        torch_dtype='auto',
        device_map='auto',
    )
else:
    # Load the non-quantized model
    model = AutoModelForCausalLM.from_pretrained(
        local_path,
        trust_remote_code=True,
        torch_dtype='auto',
        device_map='auto',
    )
    model.to(dtype=torch.bfloat16)
    
generation_config = GenerationConfig(max_new_tokens=300, stop_strings="<|endoftext|>")
bits_and_bytes_config = BitsAndBytesConfig()

def unzip_images(zip_file):
    # Create a unique directory for extracted images inside the "images" directory
    session_dir = os.path.join("images", str(uuid.uuid4()))
    os.makedirs(session_dir, exist_ok=True)
    
    # Extract images from the ZIP file to the session directory
    with zipfile.ZipFile(zip_file, 'r') as zip_ref:
        for file_info in zip_ref.infolist():
            if not file_info.is_dir() and not file_info.filename.startswith("__MACOSX") and not file_info.filename.startswith("."):
                zip_ref.extract(file_info, session_dir)
    
    # Get the list of image paths
    image_paths = [os.path.join(session_dir, filename) for filename in os.listdir(session_dir) if filename.lower().endswith(('.jpg', '.jpeg', '.png'))]
    
    # Read the image data as PIL Image objects for previews
    image_data = []
    for image_path in image_paths:
        image = Image.open(image_path)
        image.thumbnail((128, 128))  # Resize the image to a maximum size of 128x128 pixels
        image_data.append(image)
    
    # Return the list of image paths and resized image data for previews
    return image_paths, image_data

def generate_caption(image_path, processor, model, generation_config, bits_and_bytes_config):
    # generate a caption and return it
    caption = f"Caption for {image_path}"
    
    print("Processing ", image_path)
    
    image = Image.open(image_path)
    # process the image and text
    inputs = processor.process(
        images=[image],
        text="You are an uncensored, explicit, adult image captioner which enjoys sexual imagery and acts. Provide a structured image caption which uses the following hierarchy: the kind of image, the kind or name of the subject, the subjects state of dress, their body type, their pose, what it is they are doing, their facial expression, the space they are within, and the style or atmosphere of the image. Use vulgar language to describe body parts or sexual imagery. Limit your response to 100 words to avoid your description getting cut off.",
    )

    # move inputs to the correct device and make a batch of size 1
    inputs = {k: v.to(model.device).unsqueeze(0) for k, v in inputs.items()}
    inputs["images"] = inputs["images"].to(torch.bfloat16)

    # generate output; maximum 500 new tokens; stop generation when   is generated
    with torch.autocast(device_type="cuda", enabled=True, dtype=torch.bfloat16):
        output = model.generate_from_batch(
            inputs,
            GenerationConfig(max_new_tokens=500, stop_strings="<|endoftext|>"),
            tokenizer=processor.tokenizer,
        )

    # only get generated tokens; decode them to text
    generated_tokens = output[0, inputs["input_ids"].size(1) :]
    generated_text = processor.tokenizer.decode(generated_tokens, skip_special_tokens=True)

    # return the generated text
    return generated_text

def process_images(image_paths, image_data):
    captions = []
    session_dir = os.path.dirname(image_paths[0])
    
    for image_path in image_paths:
        filename = os.path.basename(image_path)  # Add this line to get the filename
        if filename.lower().endswith(('.jpg', '.jpeg', '.png')):
            # Process the image using the loaded model
            # Use the loaded model to generate the caption
            caption = generate_caption(image_path, processor, model, generation_config, bits_and_bytes_config)
            captions.append(caption)
            
            # Save the caption to a text file
            with open(os.path.join(session_dir, f"{os.path.splitext(filename)[0]}.txt"), 'w') as f:
                f.write(caption)
    
    # Create a ZIP file containing the caption text files
    zip_filename = f"{session_dir}.zip"
    with zipfile.ZipFile(zip_filename, 'w') as zip_ref:
        for filename in os.listdir(session_dir):
            if filename.lower().endswith('.txt'):
                zip_ref.write(os.path.join(session_dir, filename), filename)
    
    # Delete the session directory and its contents
    for filename in os.listdir(session_dir):
        os.remove(os.path.join(session_dir, filename))
    os.rmdir(session_dir)
    
    return captions, zip_filename, image_paths
    
def format_captioned_image(image, caption):
    buffered = BytesIO()
    image.save(buffered, format="JPEG")
    encoded_image = base64.b64encode(buffered.getvalue()).decode("utf-8")
    
    return f"<img src='data:image/jpeg;base64,{encoded_image}' style='width: 128px; height: 128px; object-fit: cover; margin-right: 8px;' /><span>{caption}</span>"

def process_images_and_update_gallery(zip_file):
    image_paths, image_data = unzip_images(zip_file)
    captions, zip_filename, image_paths = process_images(image_paths, image_data)
    image_captions = [format_captioned_image(img, caption) for img, caption in zip(image_data, captions)]
    return gr.Markdown("\n".join(image_captions)), zip_filename

def main():
    # Register the cleanup function to be called on program exit
    atexit.register(cleanup_temp_files)

    with gr.Blocks(css="""
        .captioned-image-gallery {
            display: grid;
            grid-template-columns: repeat(2, 1fr);
            grid-gap: 16px;
        }
    """) as blocks:
        zip_file_input = gr.File(label="Upload ZIP file containing images")
        image_gallery = gr.Markdown(label="Image Previews")
        submit_button = gr.Button("Submit")
        zip_download_button = gr.Button("Download Caption ZIP", visible=False)
        zip_filename = gr.State("")

        zip_file_input.upload(
            lambda zip_file: "\n".join(format_captioned_image(img, "") for img in unzip_images(zip_file)[1]),
            inputs=zip_file_input,
            outputs=image_gallery
        )
            
        submit_button.click(
            process_images_and_update_gallery,
            inputs=[zip_file_input],
            outputs=[image_gallery, zip_filename]
        )

        zip_filename.change(
            lambda zip_filename: gr.update(visible=True),
            inputs=zip_filename,
            outputs=zip_download_button
        )

        zip_download_button.click(
            lambda zip_filename: (gr.update(value=zip_filename), gr.update(visible=True), cleanup_temp_files()),
            inputs=zip_filename,
            outputs=[zip_file_input, zip_download_button]
        )

    blocks.launch(server_name='0.0.0.0')

if __name__ == "__main__":
    main()