app _bk.py

import gradio as gr
import numpy as np
from PIL import Image, ImageDraw
import torch
import torchvision.transforms as transforms
import timm

# URL for the Hugging Face checkpoint
CHECKPOINT_URL = "https://huggingface.co/ReefNet/beit_global/resolve/main/checkpoint-60.pth"

# Class labels
all_classes = [
    'Acanthastrea', 'Acropora', 'Agaricia', 'Alveopora', 'Astrea', 'Astreopora',
    'Caulastraea', 'Coeloseris', 'Colpophyllia', 'Coscinaraea', 'Ctenactis',
    'Cycloseris', 'Cyphastrea', 'Dendrogyra', 'Dichocoenia', 'Diploastrea',
    'Diploria', 'Dipsastraea', 'Echinophyllia', 'Echinopora', 'Euphyllia',
    'Eusmilia', 'Favia', 'Favites', 'Fungia', 'Galaxea', 'Gardineroseris',
    'Goniastrea', 'Goniopora', 'Halomitra', 'Herpolitha', 'Hydnophora',
    'Isophyllia', 'Isopora', 'Leptastrea', 'Leptoria', 'Leptoseris',
    'Lithophyllon', 'Lobactis', 'Lobophyllia', 'Madracis', 'Meandrina', 'Merulina',
    'Montastraea', 'Montipora', 'Mussa', 'Mussismilia', 'Mycedium', 'Orbicella',
    'Oulastrea', 'Oulophyllia', 'Oxypora', 'Pachyseris', 'Pavona', 'Pectinia',
    'Physogyra', 'Platygyra', 'Plerogyra', 'Plesiastrea', 'Pocillopora',
    'Podabacia', 'Porites', 'Psammocora', 'Pseudodiploria', 'Sandalolitha',
    'Scolymia', 'Seriatopora', 'Siderastrea', 'Stephanocoenia', 'Stylocoeniella',
    'Stylophora', 'Tubastraea', 'Turbinaria'
]

# Function to load the BeIT model
def load_model(model_name):
    print(f"Loading {model_name} model...")
    if model_name == 'beit':
        args = type('', (), {})()
        args.model = 'beitv2_large_patch16_224.in1k_ft_in22k_in1k'
        args.nb_classes = len(all_classes)
        args.drop_path = 0.1

        # Create model
        model = timm.create_model(
            args.model,
            pretrained=False,
            num_classes=args.nb_classes,
            drop_path_rate=args.drop_path,
            use_rel_pos_bias=True,
            use_abs_pos_emb=True,
        )

        # Load checkpoint from Hugging Face
        checkpoint = torch.hub.load_state_dict_from_url(CHECKPOINT_URL, map_location="cpu")
        state_dict = checkpoint.get('model', checkpoint)

        # Filter state dict
        filtered_state_dict = {k: v for k, v in state_dict.items() if "relative_position_index" not in k}
        model.load_state_dict(filtered_state_dict, strict=False)
    else:
        raise ValueError(f"Model {model_name} not implemented!")

    # Move model to CUDA if available
    model.eval()
    if torch.cuda.is_available():
        model.cuda()
    return model

# Preprocessing transforms
preprocess = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

# Initialize selected model
selected_model_name = 'beit'
model = load_model(selected_model_name)

def predict_label(image):
    """Predict the label for the given image."""
    # Ensure the image is a PIL Image
    if isinstance(image, np.ndarray):
        image = Image.fromarray(image)
    elif not isinstance(image, Image.Image):
        raise TypeError(f"Unexpected type {type(image)}, expected PIL.Image or numpy.ndarray.")

    input_tensor = preprocess(image).unsqueeze(0)
    if torch.cuda.is_available():
        input_tensor = input_tensor.cuda()

    with torch.no_grad():
        outputs = model(input_tensor)
        predicted_class = torch.argmax(outputs, dim=1).item()

    return all_classes[predicted_class]


# Function to draw a rectangle on the image
def draw_rectangle(image, x, y, size=224):
    image_pil = image.copy()
    draw = ImageDraw.Draw(image_pil)
    draw.rectangle([x, y, x + size, y + size], outline="red", width=3)
    return image_pil

# Crop a region of interest
def crop_image(image, x, y, size=224):
    image_np = np.array(image)
    h, w, _ = image_np.shape
    x = min(max(x, 0), w - size)
    y = min(max(y, 0), h - size)
    cropped = image_np[y:y+size, x:x+size]
    return Image.fromarray(cropped)

# Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("## Coral Classification with BeIT Model")
    with gr.Row():
        with gr.Column():
            image_input = gr.Image(type="pil", label="Upload Image", interactive=True)
            x_slider = gr.Slider(0, 1000, step=1, value=0, label="X Coordinate")
            y_slider = gr.Slider(0, 1000, step=1, value=0, label="Y Coordinate")
        with gr.Column():
            interactive_image = gr.Image(label="Interactive Image")
            cropped_image = gr.Image(label="Cropped Patch")
            label_output = gr.Textbox(label="Predicted Label")
    
    # Interactions
    def update_selection(image, x, y):
        overlay_image = draw_rectangle(image, x, y)
        cropped = crop_image(image, x, y)
        return overlay_image, cropped

    def predict_from_cropped(cropped):
        return predict_label(cropped)

    crop_button = gr.Button("Crop")
    crop_button.click(fn=update_selection, inputs=[image_input, x_slider, y_slider], outputs=[interactive_image, cropped_image])

    predict_button = gr.Button("Predict")
    predict_button.click(fn=predict_from_cropped, inputs=cropped_image, outputs=label_output)

    def update_sliders(image):
        if image:
            width, height = image.size
            return gr.update(maximum=width - 224), gr.update(maximum=height - 224)
        return gr.update(), gr.update()

    image_input.change(fn=update_sliders, inputs=image_input, outputs=[x_slider, y_slider])

demo.launch(server_name="0.0.0.0", server_port=7860)