app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from PIL import Image
import re 
import requests
from io import BytesIO

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-VL-Chat-Int4", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-VL-Chat-Int4", device_map="auto", trust_remote_code=True).eval()

def generate_predictions(image_input, text_input, with_grounding):
    user_image_path = "/tmp/user_input_test_image.jpg"
    Image.fromarray((255 - (image_input * 255).astype('uint8'))).save(user_image_path)
    
    if with_grounding == "Yes":
        text_input += " with grounding"
    
    query = tokenizer.from_list_format([
        {'image': user_image_path},
        {'text': text_input},
    ])
    inputs = tokenizer(query, return_tensors='pt')
    inputs = inputs.to(model.device)
    
    pred = model.generate(**inputs)
    full_response = tokenizer.decode(pred.cpu()[0], skip_special_tokens=False)
    
    frontend_response = re.sub(r'Picture \d+:|<.*?>|\/tmp\/.*\.jpg', '', full_response).replace(text_input, '').strip()
    print("Generated Caption:", frontend_response)  # Debugging line
    
    image_with_boxes = tokenizer.draw_bbox_on_latest_picture(full_response)
    
    if image_with_boxes:
        temp_path = "/tmp/image_with_boxes.jpg"
        image_with_boxes.save(temp_path)
        image_with_boxes = Image.open(temp_path)
    
    return image_with_boxes, frontend_response

iface = gr.Interface(
    fn=generate_predictions, 
    inputs=[
        gr.inputs.Image(label="Image Input"), 
        gr.inputs.Textbox(default="Generate a caption for that image:", label="Prompt"),
        gr.inputs.Radio(["No", "Yes"], label="With Grounding", default="No")
    ], 
    outputs=[
        gr.outputs.Image(type='pil', label="Image"),
        gr.outputs.Textbox(label="Generated")
    ],
    title="Qwen-VL Demonstration",
    description = """
## Qwen-VL: A Multimodal Large Vision Language Model by Alibaba Cloud
**Space by [@Artificialguybr](https://twitter.com/artificialguybr)**
    
### Key Features:
- **Strong Performance**: Surpasses existing LVLMs on multiple English benchmarks including Zero-shot Captioning and VQA.
- **Multi-lingual Support**: Supports English, Chinese, and multi-lingual conversation.
- **High Resolution**: Utilizes 448*448 resolution for fine-grained recognition and understanding.
""",
)
iface.launch()