app.py

import os
import time
import numpy as np
from PIL import Image
import tensorflow as tf
import gradio as gr

# ─── Constants ────────────────────────────────────────────────────────────────
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
MODELS_DIR = os.path.join(BASE_DIR, "models")
IMG_SIZE = (224, 224)

MODEL_CONFIGS = [
    ("MobileNetV3 Small",  "mobilenetv3_small.tflite"),
    ("MobileNetV3 Large",  "mobilenetv3_large.tflite"),
    ("EfficientNetV2-B0",  "efficientnetv2b0.tflite"),
]

MEDAL = ["🥇", "🥈", "🥉"]

# ─── Load labels ──────────────────────────────────────────────────────────────
with open(os.path.join(MODELS_DIR, "labels.txt")) as f:
    CLASS_NAMES = [line.strip() for line in f if line.strip()]

# ─── Load TFLite interpreters at startup ──────────────────────────────────────
INTERPRETERS: list[tuple[str, tf.lite.Interpreter]] = []
for display_name, filename in MODEL_CONFIGS:
    path = os.path.join(MODELS_DIR, filename)
    interp = tf.lite.Interpreter(model_path=path)
    interp.allocate_tensors()
    INTERPRETERS.append((display_name, interp))

print(f"Loaded {len(INTERPRETERS)} TFLite models.")


# ─── Inference helpers ────────────────────────────────────────────────────────
def preprocess(pil_image: Image.Image) -> np.ndarray:
    """Resize, convert to float32 array, add batch dim. Values in [0, 255]."""
    img = pil_image.convert("RGB").resize(IMG_SIZE, Image.BILINEAR)
    arr = np.array(img, dtype=np.float32)          # (224, 224, 3)
    return np.expand_dims(arr, axis=0)              # (1, 224, 224, 3)


def run_model(interp: tf.lite.Interpreter, img_array: np.ndarray) -> tuple[np.ndarray, float]:
    """Run a single TFLite inference. Returns (probs, latency_ms)."""
    input_idx  = interp.get_input_details()[0]["index"]
    output_idx = interp.get_output_details()[0]["index"]
    interp.set_tensor(input_idx, img_array)
    t0 = time.perf_counter()
    interp.invoke()
    latency_ms = (time.perf_counter() - t0) * 1000
    probs = interp.get_tensor(output_idx)[0]        # shape (num_classes,)
    return probs, latency_ms


def fmt_model_card(display_name: str, probs: np.ndarray, latency_ms: float) -> str:
    """Build a Markdown string for one model's result card."""
    top3_idx = np.argsort(probs)[::-1][:3]
    lines = [f"### {display_name}", f"⏱ `{latency_ms:.1f} ms`", ""]
    for rank, idx in enumerate(top3_idx):
        pct = probs[idx] * 100
        bar = "█" * int(pct / 5)            # simple bar, max 20 chars
        lines.append(f"{MEDAL[rank]} **{CLASS_NAMES[idx]}** — {pct:.1f}%")
        lines.append(f"`{bar:<20}` ")
        lines.append("")
    return "\n".join(lines)


# ─── Main prediction function ─────────────────────────────────────────────────
def predict(pil_image):
    if pil_image is None:
        empty = "_(upload an image and press Run Inference)_"
        return empty, empty, empty

    img_array = preprocess(pil_image)
    cards: list[str] = []

    for display_name, interp in INTERPRETERS:
        probs, latency_ms = run_model(interp, img_array)
        cards.append(fmt_model_card(display_name, probs, latency_ms))

    return cards[0], cards[1], cards[2]


# ─── Gradio UI ────────────────────────────────────────────────────────────────
CSS = """
.model-card { background: var(--block-background-fill); border-radius: 12px; padding: 16px; }
footer { display: none !important; }
"""

with gr.Blocks(title="Animal Toy Classifier") as demo:
    gr.Markdown("# 🦁 Animal Toy Classifier\nUpload a photo of an animal toy — three models will predict the class and vote on the best answer.")

    with gr.Row():
        # ── Left: upload + button ──────────────────────────────────────────────
        with gr.Column(scale=1, min_width=260):
            img_in = gr.Image(
                type="pil",
                sources=["upload"],
                label="Input Image",
                height=280,
            )
            with gr.Row():
                run_btn   = gr.Button("▶  Run Inference", variant="primary", scale=3)
                clear_btn = gr.ClearButton(
                    components=[img_in],
                    value="✕ Clear",
                    scale=1,
                )

        # ── Right: per-model result cards ──────────────────────────────────────
        with gr.Column(scale=2):
            with gr.Row():
                with gr.Column(elem_classes="model-card"):
                    out_small = gr.Markdown("### MobileNetV3 Small\n\n_(waiting)_")
                with gr.Column(elem_classes="model-card"):
                    out_large = gr.Markdown("### MobileNetV3 Large\n\n_(waiting)_")
                with gr.Column(elem_classes="model-card"):
                    out_eff   = gr.Markdown("### EfficientNetV2-B0\n\n_(waiting)_")

    # ── Wire up ────────────────────────────────────────────────────────────────
    run_btn.click(
        fn=predict,
        inputs=[img_in],
        outputs=[out_small, out_large, out_eff],
    )

    # Reset outputs when Clear is pressed
    clear_btn.click(
        fn=lambda: (
            "### MobileNetV3 Small\n\n_(waiting)_",
            "### MobileNetV3 Large\n\n_(waiting)_",
            "### EfficientNetV2-B0\n\n_(waiting)_",
        ),
        inputs=[],
        outputs=[out_small, out_large, out_eff],
    )


if __name__ == "__main__":
    demo.launch(theme=gr.themes.Soft(), css=CSS)