app.py

import os

# Reduce TensorFlow log verbosity
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"

import numpy as np
import torch
import gradio as gr
import spaces
from huggingface_hub import hf_hub_download
from tensorflow.keras.models import load_model
from tensorflow.keras.preprocessing import image as keras_image
from ultralytics import YOLO


# =========================================================
# Your model repo (where you uploaded the trained models)
# =========================================================

MODEL_REPO = "SabaTariq510/waste-classification-models"


# =========================================================
# Download + load models (runs once when the Space starts)
# =========================================================

print("Downloading MobileNetV2...")
mobilenet_path = hf_hub_download(repo_id=MODEL_REPO, filename="best_mobilenet.keras")

print("Downloading YOLOv8...")
yolo_path = hf_hub_download(repo_id=MODEL_REPO, filename="bestyolomodel.pt")

print("Loading MobileNetV2...")
mobilenet_model = load_model(mobilenet_path)

print("Loading YOLOv8...")
yolo_model = YOLO(yolo_path, task="detect")

print("Both models loaded successfully!")


# =========================================================
# MobileNet Classes
# =========================================================

mobilenet_classes = [
    "cardboard",
    "glass",
    "metal",
    "paper",
    "plastic",
    "trash"
]


# =========================================================
# Recyclable Information
# =========================================================

recyclable = {
    "cardboard": "Recyclable",
    "glass": "Recyclable",
    "metal": "Recyclable",
    "paper": "Recyclable",
    "plastic": "Recyclable",
    "trash": "Non-Recyclable"
}


# =========================================================
# Prediction Function
# ZeroGPU: this function must be decorated with @spaces.GPU
# for it to actually get a GPU allocated to it, otherwise
# no GPU will be assigned at runtime.
# =========================================================

@spaces.GPU
def predict(img, model_choice):

    if img is None:
        return "⚠️ Please upload an image first."

    # Save a temporary copy (YOLO needs a file path as input)
    temp_path = "/tmp/uploaded_image.jpg"
    img.save(temp_path)

    # =====================================================
    # MobileNetV2
    # =====================================================
    if model_choice == "MobileNetV2":

        resized = img.convert("RGB").resize((224, 224))

        arr = keras_image.img_to_array(resized)
        arr = arr / 255.0
        arr = np.expand_dims(arr, axis=0)

        prediction = mobilenet_model.predict(arr, verbose=0)

        index = int(np.argmax(prediction))
        confidence = float(np.max(prediction)) * 100
        waste = mobilenet_classes[index]
        model_used = "MobileNetV2"

    # =====================================================
    # YOLOv8
    # =====================================================
    else:

        # Only use "cuda" if a GPU is actually available at runtime
        # (e.g. ZeroGPU Space). Otherwise fall back to CPU so the
        # prediction doesn't silently fail with CUDA_ERROR_NO_DEVICE.
        device = "cuda" if torch.cuda.is_available() else "cpu"

        try:
            results = yolo_model.predict(
                source=temp_path,
                conf=0.25,
                device=device
            )
        except Exception as e:
            return f"❌ YOLOv8 prediction failed: {e}"

        result = results[0]

        if len(result.boxes) > 0:

            # Confidence values for all detected boxes
            confidences = result.boxes.conf

            # Pick the detection with the highest confidence
            best_index = int(confidences.argmax())
            box = result.boxes[best_index]

            class_index = int(box.cls)
            confidence = float(box.conf) * 100
            waste = result.names[class_index]

        else:
            waste = "No Waste Detected"
            confidence = 0

        model_used = f"YOLOv8 ({device})"

    recycle = recyclable.get(waste, "Unknown")

    result_text = (
        f"**Model Used:** {model_used}\n\n"
        f"**Prediction:** {waste}\n\n"
        f"**Confidence:** {confidence:.2f}%\n\n"
        f"**Recyclable:** {recycle}"
    )

    return result_text


# =========================================================
# Gradio Interface
# =========================================================

with gr.Blocks(title="Smart Waste Classification") as demo:

    gr.Markdown("# ♻️ Smart Waste Classification")
    gr.Markdown("Upload an image and choose a model — get an instant prediction.")

    with gr.Row():

        with gr.Column():
            img_input = gr.Image(type="pil", label="Upload Waste Image")

            model_choice = gr.Radio(
                choices=["MobileNetV2", "YOLOv8"],
                value="MobileNetV2",
                label="Select Model"
            )

            predict_btn = gr.Button("Predict", variant="primary")

        with gr.Column():
            output = gr.Markdown(label="Result")

    predict_btn.click(
        fn=predict,
        inputs=[img_input, model_choice],
        outputs=output
    )


if __name__ == "__main__":
    demo.launch()