example.py

"""Minimal inference example for google/vit-base-patch16-224 (ViT-Base) using ExecuTorch.

Loads a quantized INT8 .pte model and runs image classification on a single
image, printing the top-5 ImageNet-1k predictions.

Preprocessing: resize the shorter edge to 232px (bilinear), center-crop to
224x224, convert to RGB, scale pixel values to [0, 1], then normalize with
mean=[0.5, 0.5, 0.5] and std=[0.5, 0.5, 0.5] (the stats reported by the
checkpoint's image processor). This matches the resize-232/center-crop-224
recipe used to calibrate and evaluate this model in the Arm optimization
pipeline.

Requires: executorch, torch, torchvision, pillow, json
"""

import argparse
import json
from pathlib import Path

import torch
from executorch.runtime import Runtime
from PIL import Image, ImageDraw, ImageFont
from torchvision import transforms

# ── Configuration ────────────────────────────────────────────────────────────
MODEL_PATH = "google__vit-base-patch16-224_raspberry_executorch_optimized.pte"
IMAGE_PATH = "sample_input.jpg"
LABELS_PATH = "imagenet_classes.json"
RESIZE_SIZE = 232  # shorter edge, before center crop
CROP_SIZE = 224
MEAN = (0.5, 0.5, 0.5)
STD = (0.5, 0.5, 0.5)
TOP_K = 5

_TRANSFORM = transforms.Compose(
    [
        transforms.Resize(RESIZE_SIZE),
        transforms.CenterCrop(CROP_SIZE),
        transforms.ToTensor(),
    ]
)


def load_model(model_path: str):
    """Load an ExecuTorch .pte model and return its forward method.

    The method is loaded once and reused across calls.
    """
    runtime = Runtime.get()
    program = runtime.load_program(model_path)
    return program.load_method("forward")


def load_labels(labels_path: str) -> dict[str, str]:
    """Load the ImageNet-1k index-to-label mapping."""
    with open(labels_path) as f:
        return json.load(f)


def preprocess(image_path: str) -> torch.Tensor:
    """Load an image and prepare it as a normalized [1, 3, 224, 224] tensor."""
    image = Image.open(image_path).convert("RGB")
    tensor = _TRANSFORM(image)  # [3, 224, 224] in [0, 1]

    mean = torch.tensor(MEAN, dtype=torch.float32).view(3, 1, 1)
    std = torch.tensor(STD, dtype=torch.float32).view(3, 1, 1)
    tensor = (tensor - mean) / std

    return tensor.unsqueeze(0).contiguous()  # [1, 3, H, W]


def run_inference(method, input_tensor: torch.Tensor) -> torch.Tensor:
    """Run a forward pass and return the raw logits tensor [1, num_classes]."""
    outputs = method.execute([input_tensor])
    return outputs[0]


def postprocess(logits: torch.Tensor, labels: dict[str, str]) -> list[dict]:
    """Apply softmax and return the top-k class predictions."""
    if logits.dim() == 1:
        logits = logits.unsqueeze(0)

    probabilities = torch.softmax(logits, dim=-1)
    k = min(TOP_K, probabilities.shape[-1])
    top_scores, top_indices = probabilities.topk(k, dim=-1)

    results = []
    for score, idx in zip(
        top_scores[0].tolist(), top_indices[0].tolist(), strict=False
    ):
        results.append(
            {
                "class_index": idx,
                "label": labels.get(str(idx), f"class_{idx}"),
                "probability": round(score, 6),
            }
        )
    return results


def save_results(results: list[dict], output_path: Path) -> None:
    """Save the top-k predictions as JSON next to the script."""
    with open(output_path, "w") as f:
        json.dump(results, f, indent=2)
    print(f"Predictions saved to: {output_path}")


def _load_fonts() -> tuple:
    """Load DejaVuSans fonts, falling back to the built-in bitmap font."""
    font_dir = Path("/usr/share/fonts/truetype/dejavu")
    try:
        regular = ImageFont.truetype(str(font_dir / "DejaVuSans.ttf"), 14)
        bold = ImageFont.truetype(str(font_dir / "DejaVuSans-Bold.ttf"), 15)
    except OSError:
        regular = ImageFont.load_default()
        bold = regular
    return regular, bold


def save_output_image(image_path: str, results: list[dict], output_path: Path) -> None:
    """Render the input image alongside its top-k predictions and save as sample_output.jpg.

    Layout (700 x 320 px):
      left panel: input image (280x280)
      right panel: top-5 prediction bars with class names and probabilities
    """
    MARGIN = 20
    IMG_SZ = 280
    PANEL_W = 360
    CANVAS_W = MARGIN + IMG_SZ + MARGIN + PANEL_W + MARGIN  # 700
    CANVAS_H = MARGIN + IMG_SZ + MARGIN  # 320
    TITLE_H = 28
    ROW_H = (IMG_SZ - TITLE_H - 8) // TOP_K

    BG_COLOR = "#f5f5f5"
    PANEL_BG = "#ffffff"
    TOP1_BG = "#e3f2fd"
    BAR_TOP1 = "#1565c0"
    BAR_REST = "#90caf9"
    TEXT_DARK = "#212121"
    TEXT_GRAY = "#616161"
    BORDER = "#bdbdbd"

    font, font_bold = _load_fonts()

    canvas = Image.new("RGB", (CANVAS_W, CANVAS_H), BG_COLOR)
    draw = ImageDraw.Draw(canvas)

    # Left panel: input image
    orig = Image.open(image_path).convert("RGB").resize((IMG_SZ, IMG_SZ), Image.LANCZOS)
    canvas.paste(orig, (MARGIN, MARGIN))
    draw.rectangle(
        [MARGIN, MARGIN, MARGIN + IMG_SZ - 1, MARGIN + IMG_SZ - 1],
        outline=BORDER,
        width=1,
    )

    # Right panel: predictions
    px = MARGIN + IMG_SZ + MARGIN
    py = MARGIN
    draw.rectangle(
        [px, py, px + PANEL_W, py + IMG_SZ], fill=PANEL_BG, outline=BORDER, width=1
    )

    draw.text((px + 10, py + 6), "Top-5 Predictions", font=font_bold, fill=TEXT_DARK)
    draw.line(
        [px + 1, py + TITLE_H, px + PANEL_W - 1, py + TITLE_H], fill=BORDER, width=1
    )

    max_prob = results[0]["probability"] if results else 1.0
    BAR_MAX_W = PANEL_W - 24

    for rank, pred in enumerate(results):
        ry = py + TITLE_H + 8 + rank * ROW_H

        if rank == 0:
            draw.rectangle([px + 2, ry, px + PANEL_W - 2, ry + ROW_H - 2], fill=TOP1_BG)

        badge_text = f"#{rank + 1}"
        draw.text(
            (px + 8, ry + 4),
            badge_text,
            font=font_bold,
            fill=BAR_TOP1 if rank == 0 else TEXT_GRAY,
        )

        label = pred["label"]
        max_chars = 32
        if len(label) > max_chars:
            label = label[: max_chars - 1] + "…"
        draw.text((px + 36, ry + 4), label, font=font, fill=TEXT_DARK)

        bar_w = max(4, int((pred["probability"] / max_prob) * BAR_MAX_W))
        bar_y = ry + ROW_H - 14
        bar_color = BAR_TOP1 if rank == 0 else BAR_REST
        draw.rectangle([px + 8, bar_y, px + 8 + bar_w, bar_y + 8], fill=bar_color)

        pct_text = f"{pred['probability'] * 100:.2f}%"
        draw.text((px + PANEL_W - 56, ry + 4), pct_text, font=font, fill=TEXT_GRAY)

    canvas.save(str(output_path), quality=95)
    print(f"Output image saved to: {output_path}")


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--output", type=Path, help="Write predictions as JSON")
    parser.add_argument("--output-image", type=Path, help="Write the rendered result image")
    args = parser.parse_args()

    script_dir = Path(__file__).resolve().parent
    model_path = str(script_dir / MODEL_PATH)
    image_path = str(script_dir / IMAGE_PATH)
    labels_path = str(script_dir / LABELS_PATH)

    method = load_model(model_path)
    labels = load_labels(labels_path)

    input_tensor = preprocess(image_path)
    logits = run_inference(method, input_tensor)
    results = postprocess(logits, labels)

    print(f"Top-{TOP_K} predictions for {IMAGE_PATH}:")
    for rank, pred in enumerate(results, start=1):
        print(f"  {rank}. {pred['label']} ({pred['probability']:.4f})")

    if args.output:
        save_results(results, args.output)
    if args.output_image:
        save_output_image(image_path, results, args.output_image)


if __name__ == "__main__":
    main()