app.py

import gradio as gr

import argparse
from functools import partial
import cv2
import requests
import os
from io import BytesIO
from PIL import Image
import numpy as np
from pathlib import Path


import warnings

import torch
warnings.filterwarnings("ignore")

from groundingdino.models import build_model
from groundingdino.util.slconfig import SLConfig
from groundingdino.util.utils import clean_state_dict
from groundingdino.util.inference import annotate, load_image, predict
import groundingdino.datasets.transforms as T

from huggingface_hub import hf_hub_download


# Use this command for evaluate the GLIP-T model
config_file = "groundingdino/config/GroundingDINO_SwinB_cfg.py"
ckpt_repo_id = "ShilongLiu/GroundingDINO"
ckpt_filenmae = "groundingdino_swinb_cogcoor.pth"


def load_model_hf(model_config_path, repo_id, filename, device='cpu'):
    args = SLConfig.fromfile(model_config_path) 
    model = build_model(args)
    args.device = device

    cache_file = hf_hub_download(repo_id=repo_id, filename=filename)
    checkpoint = torch.load(cache_file, map_location=device)
    log = model.load_state_dict(clean_state_dict(checkpoint['model']), strict=False)
    print("Model loaded from {} \n => {}".format(cache_file, log))
    _ = model.eval()
    return model    

def image_transform_grounding(init_image):
    transform = T.Compose([
        T.RandomResize([800], max_size=1333),
        T.ToTensor(),
        T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ])
    image, _ = transform(init_image, None) # 3, h, w
    return init_image, image

def image_transform_grounding_for_vis(init_image):
    transform = T.Compose([
        T.RandomResize([800], max_size=1333),
    ])
    image, _ = transform(init_image, None) # 3, h, w
    return image

model = load_model_hf(config_file, ckpt_repo_id, ckpt_filenmae)

def run_grounding(input_image, grounding_caption, box_threshold, text_threshold):
    # Convert numpy array to PIL Image if needed
    if isinstance(input_image, np.ndarray):
        if input_image.ndim == 3:
            input_image = cv2.cvtColor(input_image, cv2.COLOR_BGR2RGB)
        input_image = Image.fromarray(input_image)
    
    init_image = input_image.convert("RGB")
    original_size = init_image.size

    _, image_tensor = image_transform_grounding(init_image)
    image_pil: Image = image_transform_grounding_for_vis(init_image)

    # run grounidng
    boxes, logits, phrases = predict(model, image_tensor, grounding_caption, box_threshold, text_threshold, device='cpu')
    annotated_frame = annotate(image_source=np.asarray(image_pil), boxes=boxes, logits=logits, phrases=phrases)
    image_with_box = Image.fromarray(cv2.cvtColor(annotated_frame, cv2.COLOR_BGR2RGB))

    return image_with_box

if __name__ == "__main__":
    
    parser = argparse.ArgumentParser("Grounding DINO demo", add_help=True)
    parser.add_argument("--debug", action="store_true", help="using debug mode")
    parser.add_argument("--share", action="store_true", help="share the app")
    args = parser.parse_args()
    
    css = """
  #mkd {
    height: 500px; 
    overflow: auto; 
    border: 1px solid #ccc; 
  }
"""
    with gr.Blocks(css=css) as demo:
        gr.Markdown("<h1><center>Grounding DINO<h1><center>")
        gr.Markdown("<h3><center>Open-World Detection with <a href='https://github.com/IDEA-Research/GroundingDINO'>Grounding DINO</a><h3><center>")
        gr.Markdown("<h3><center>Running on CPU, so it may take a while to run the model.<h3><center>")

        with gr.Row():
            with gr.Column():
                input_image = gr.Image(label="Input Image", type="pil")
                grounding_caption = gr.Textbox(label="Detection Prompt")
                run_button = gr.Button("Run")
                
                with gr.Accordion("Advanced options", open=False):
                    box_threshold = gr.Slider(
                        minimum=0.0, maximum=1.0, value=0.25, step=0.001, 
                        label="Box Threshold"
                    )
                    text_threshold = gr.Slider(
                        minimum=0.0, maximum=1.0, value=0.25, step=0.001, 
                        label="Text Threshold"
                    )

            with gr.Column():
                gallery = gr.Image(
                    label="Detection Result",
                    type="pil"
                )

        run_button.click(
            fn=run_grounding, 
            inputs=[input_image, grounding_caption, box_threshold, text_threshold], 
            outputs=[gallery]
        )
        
        gr.Examples(
            examples=[["this_is_fine.png", "coffee cup", 0.25, 0.25]],
            inputs=[input_image, grounding_caption, box_threshold, text_threshold],
            outputs=[gallery],
            fn=run_grounding,
            cache_examples=True,
        )
    
    demo.launch(share=args.share, debug=args.debug, show_error=True)