app.py

import glob
import gradio as gr
import torch
from transformers import ViTForImageClassification, AutoImageProcessor
from PIL import Image, ImageDraw, ImageFont
import random
import os

# ✅ Load model and processor from Hugging Face
model_name = "akridge/noaa-esd-coral-bleaching-vit-classifier-v1"
model = ViTForImageClassification.from_pretrained(model_name)
processor = AutoImageProcessor.from_pretrained(model_name, use_fast=True)

# 🗂️ Ensure id2label keys are integers and labels are uppercase
id2label = {int(k): v.upper() for k, v in model.config.id2label.items()}

# 🎨 Label colors (RGBA)
LABEL_COLORS = {
    "CORAL": ((0, 0, 255, 80), (0, 0, 200)),        # Fill: Blue transparent, Border: Dark blue
    "CORAL_BL": ((255, 255, 255, 150), (150, 150, 150)),  # Fill: White transparent, Border: Gray
}

def predict_and_overlay(image, rows=2, cols=5, patch_size=224):
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(device).eval()

    # ✅ Load image
    image = image.convert('RGB')
    width, height = image.size
    scale_factor = max(width, height) / 800
    font_size = max(12, int(scale_factor * 8))
    border_width = max(3, int(scale_factor * 3))

    # 🖌️ Create overlay
    overlay = Image.new('RGBA', image.size, (255, 255, 255, 0))
    overlay_draw = ImageDraw.Draw(overlay)

    # ✅ Generate sampled points
    cell_width, cell_height = width / cols, height / rows
    sampled_points = [
        (random.randint(int(col * cell_width), int((col + 1) * cell_width - 1)),
         random.randint(int(row * cell_height), int((row + 1) * cell_height - 1)))
        for row in range(rows) for col in range(cols)
    ]

    predictions = []

    # ✅ Load font
    try:
        font = ImageFont.truetype("arial.ttf", size=font_size)
    except IOError:
        font = ImageFont.load_default()

    # 🔍 Predict and draw patches
    for x, y in sampled_points:
        left, upper = max(0, x - patch_size // 2), max(0, y - patch_size // 2)
        right, lower = min(width, left + patch_size), min(height, upper + patch_size)

        # 🧠 Predict label
        patch = image.crop((left, upper, right, lower))
        inputs = processor(images=patch, return_tensors="pt").to(device)
        with torch.no_grad():
            pred_id = model(**inputs).logits.argmax(-1).item()
            pred_label = id2label.get(pred_id, "UNKNOWN")
            predictions.append(pred_label)

        # 🎨 Fill and border colors
        fill_color, border_color = LABEL_COLORS.get(pred_label, ((200, 200, 200, 100), (100, 100, 100)))

        # 🟦 Draw filled rectangle and label
        overlay_draw.rectangle([(left, upper), (right, lower)], fill=fill_color, outline=border_color, width=border_width)

        label_text = pred_label
        bbox = overlay_draw.textbbox((0, 0), label_text, font=font)
        text_width, text_height = bbox[2] - bbox[0], bbox[3] - bbox[1]
        text_bg_coords = [(left, upper - text_height - 6), (left + text_width + 6, upper)]

        overlay_draw.rectangle(text_bg_coords, fill=(0, 0, 0, 200))
        overlay_draw.text((left + 3, upper - text_height - 4), label_text, fill="white", font=font)

    # 🖼️ Merge overlay with original
    final_image = Image.alpha_composite(image.convert('RGBA'), overlay).convert('RGB')

    return final_image, ", ".join(predictions)

# Load example images from coral_images folder
example_images = glob.glob("coral_images/*.[jp][pn]g")

# 🚀 Gradio interface
def gradio_interface(image, rows=2, cols=5):
    if image is None:
        return None, "No image uploaded. Please upload an image or select from examples."
    final_image, predictions = predict_and_overlay(image, rows, cols)
    return final_image, predictions

app_title = "🌊 NOAA ESD Coral Bleaching ViT Classifier"
app_description = """
Upload a coral image or select from example images to sample points and predict coral bleaching using the ViT classifier model hosted on Hugging Face.

**Model:** [akridge/noaa-esd-coral-bleaching-vit-classifier-v1](https://huggingface.co/akridge/noaa-esd-coral-bleaching-vit-classifier-v1)
"""

# Custom CSS for improved styling
custom_css = """
    .gradio-container h1 {
        font-size: 2.2em;
        text-align: center;
    }
    .gradio-container p {
        font-size: 1.2em;
    }
    .gradio-container .gr-button {
        font-size: 1.2em;
    }
"""

with gr.Blocks(theme=gr.themes.Ocean(), css=custom_css, title=app_title) as interface:
    gr.Markdown(f"<h1>{app_title}</h1>")
    gr.Markdown(app_description)

    with gr.Row():
        image_input = gr.Image(type="pil", label="Upload Coral Image")
        result_output = gr.Image(type="pil", label="Predicted Results")

    with gr.Row():
        rows_slider = gr.Slider(1, 10, value=2, label="Rows of Sample Points")
        cols_slider = gr.Slider(1, 10, value=5, label="Columns of Sample Points")

    with gr.Row():
        run_button = gr.Button("Run Prediction", variant="primary")
        clear_button = gr.Button("Clear")

    # Add example images section
    gr.Examples(
        examples=[[img] for img in example_images],
        inputs=[image_input],
        outputs=[result_output],
        examples_per_page=6,
        label="Example Coral Images"
    )

    # Define button actions
    run_button.click(
        fn=gradio_interface,
        inputs=[image_input, rows_slider, cols_slider],
        outputs=[result_output, gr.Textbox(label="Predictions")]
    )

    clear_button.click(lambda: (None, ""), outputs=[image_input, result_output])

interface.launch()