example.py

"""Inference example for DeiT-Tiny INT8 (LiteRT .tflite).

Requirements:
    pip install ai-edge-litert pillow torch numpy

Usage:
    python example.py                          # uses sample_input.jpg
    python example.py --image path/to/img.jpg
"""

from __future__ import annotations

import argparse
import json
import math
from pathlib import Path
from torchvision.models import GoogLeNet_Weights

import numpy as np
from PIL import Image

# ---------------------------------------------------------------------------
# Paths
# ---------------------------------------------------------------------------
SCRIPT_DIR = Path(__file__).parent.resolve()
_model_in_parent = SCRIPT_DIR.parent / "facebook__deit-tiny-patch16-224_litert_optimized.tflite"
MODEL_PATH = _model_in_parent if _model_in_parent.exists() else SCRIPT_DIR / "facebook__deit-tiny-patch16-224_litert_optimized.tflite"
IMAGE_PATH = SCRIPT_DIR / "sample_input.jpg"

# ---------------------------------------------------------------------------
# Constants
# ---------------------------------------------------------------------------
INPUT_SIZE = 224
RESIZE_SIZE = 256
IMAGENET_MEAN = [0.485, 0.456, 0.406]
IMAGENET_STD = [0.229, 0.224, 0.225]
TOP_K = 5


# ---------------------------------------------------------------------------
# Model loading
# ---------------------------------------------------------------------------

def load_model(model_path: Path):
    """Load LiteRT .tflite model and return allocated interpreter."""
    from ai_edge_litert.interpreter import Interpreter  # noqa: PLC0415
    interpreter = Interpreter(model_path=str(model_path))
    interpreter.allocate_tensors()
    return interpreter


# ---------------------------------------------------------------------------
# Preprocessing
# ---------------------------------------------------------------------------

def preprocess(image_path: Path) -> np.ndarray:
    """Load and preprocess an image for DeiT-Tiny input.

    Steps:
      1. Resize the shorter edge to 256 using bicubic interpolation.
      2. Center crop to 224x224.
      3. Normalize to [0, 1] and apply ImageNet mean/std.

    Returns:
        Array of shape (1, 3, 224, 224), dtype float32.
    """
    img = Image.open(image_path).convert("RGB")
    w, h = img.size
    scale = RESIZE_SIZE / min(w, h)
    new_w, new_h = math.ceil(w * scale), math.ceil(h * scale)
    img = img.resize((new_w, new_h), Image.BICUBIC)
    left = (new_w - INPUT_SIZE) // 2
    top = (new_h - INPUT_SIZE) // 2
    img = img.crop((left, top, left + INPUT_SIZE, top + INPUT_SIZE))
    arr = np.array(img, dtype=np.float32) / 255.0
    mean = np.array(IMAGENET_MEAN, dtype=np.float32).reshape(1, 1, 3)
    std = np.array(IMAGENET_STD, dtype=np.float32).reshape(1, 1, 3)
    arr = (arr - mean) / std
    return arr.transpose(2, 0, 1)[np.newaxis, ...]  # (1, 3, H, W)


# ---------------------------------------------------------------------------
# Postprocessing
# ---------------------------------------------------------------------------

def postprocess(logits: np.ndarray, top_k: int = TOP_K) -> list[dict]:
    """Apply softmax and return top-k predictions.

    Args:
        logits: Array of shape (1, 1000) with raw class scores.
        top_k:  Number of top predictions to return.

    Returns:
        List of dicts: {"rank": int, "class": str, "probability": float}.
    """
    logits = logits[0]
    exp = np.exp(logits - logits.max())
    probs = exp / exp.sum()
    top_indices = np.argsort(probs)[::-1][:top_k]
    # ImageNet class labels (1000 classes)
    WEIGHTS = GoogLeNet_Weights.IMAGENET1K_V1
    classes = WEIGHTS.meta["categories"]
    return [
        {"rank": i + 1, "class": classes[idx], "probability": float(probs[idx])}
        for i, idx in enumerate(top_indices)
    ]


# ---------------------------------------------------------------------------
# Inference
# ---------------------------------------------------------------------------

def run_inference(interpreter, input_array: np.ndarray) -> np.ndarray:
    """Run forward pass and return dequantized logits.

    If the output tensor is INT8 (as produced by PT2E static quantization),
    it is dequantized to float32 using the tensor's scale and zero_point
    before being returned. Float32 outputs are returned as-is.
    """
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()
    interpreter.set_tensor(input_details[0]["index"], input_array)
    interpreter.invoke()
    raw = interpreter.get_tensor(output_details[0]["index"])
    if raw.dtype == np.int8:
        quant_params = output_details[0].get("quantization_parameters", {})
        scales = quant_params.get("scales", None)
        zero_points = quant_params.get("zero_points", None)
        if scales is not None and len(scales) > 0:
            scale = float(scales[0])
            zero_point = int(zero_points[0])
        else:
            # Fall back to the legacy (scale, zero_point) tuple
            scale, zero_point = output_details[0]["quantization"]
        return (raw.astype(np.float32) - zero_point) * scale
    return raw.astype(np.float32)


# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------

def main(image_path: Path) -> None:
    print(f"Loading model from {MODEL_PATH} ...")
    interpreter = load_model(MODEL_PATH)

    print(f"Preprocessing image: {image_path}")
    input_array = preprocess(image_path)

    print("Running inference ...")
    logits = run_inference(interpreter, input_array)

    predictions = postprocess(logits)

    print("\nTop-5 predictions:")
    for pred in predictions:
        print(f"  Top-{pred['rank']}: {pred['class']} ({pred['probability']:.4f})")

    out_path = SCRIPT_DIR / "predictions.json"
    with open(out_path, "w") as f:
        json.dump({"predictions": predictions}, f, indent=2)
    print(f"\nPredictions saved to {out_path}")


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="DeiT-Tiny INT8 LiteRT inference")
    parser.add_argument(
        "--image",
        type=Path,
        default=IMAGE_PATH,
        help="Path to input image (default: sample_input.jpg next to this script)",
    )
    args = parser.parse_args()
    main(args.image)