backend.py

"""
ONNX Runtime inference backend for DeepNAPSI.

Uses the dynamically-quantised INT8 BEiT model for fast CPU inference.
The model is expected at model/model_int8.onnx (committed to the repo via
Git LFS). If that file is absent it is downloaded from the HF Hub.
"""

from __future__ import annotations

import concurrent.futures
import os
from pathlib import Path
from typing import List

import cv2
import numpy as np
import onnxruntime as ort
from PIL import Image

from nail_detection import get_nails_and_landmarks, draw_hand

# ---------------------------------------------------------------------------
# Constants
# ---------------------------------------------------------------------------

MODEL_LOCAL = Path(__file__).parent / "model" / "model_int8.onnx"
HF_REPO_ID  = os.environ.get("DEEPNAPSI_HF_REPO", "lfolle/DeepNAPSIModel")
HF_FILENAME = "model_int8.onnx"

# BEiT preprocessing parameters (from timm resolve_data_config)
INPUT_SIZE  = 384
MEAN        = np.array([0.5, 0.5, 0.5], dtype=np.float32)
STD         = np.array([0.5, 0.5, 0.5], dtype=np.float32)

FINGER_NAMES = ["Thumb", "Index", "Middle", "Ring", "Pinky"]
NUM_CLASSES  = 5
NUM_THREADS  = int(os.environ.get("ORT_NUM_THREADS", "16"))


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

def _get_model_path() -> Path:
    # Env-var override (useful for local dev pointing at hf_space/model/)
    env_path = os.environ.get("DEEPNAPSI_MODEL_PATH", "")
    if env_path and Path(env_path).exists():
        return Path(env_path)
    # Default local path (committed to Space via Git LFS, or pre-downloaded)
    if MODEL_LOCAL.exists():
        return MODEL_LOCAL
    # Fallback: download from private HF Hub model repo.
    # Requires DEEPNAPSI_HF_TOKEN env var (set as a Space secret).
    from huggingface_hub import hf_hub_download
    token = os.environ.get("DEEPNAPSI_HF_TOKEN") or os.environ.get("DeepNAPSIModel")
    if not token:
        raise FileNotFoundError(
            f"Model not found at {MODEL_LOCAL} and neither DEEPNAPSI_HF_TOKEN nor "
            "DeepNAPSIModel secret is set. Set one in the Space settings."
        )
    print(f"[backend] Downloading model from private repo {HF_REPO_ID} …")
    # Run inside a ThreadPoolExecutor so that huggingface_hub's internal asyncio
    # event loop is isolated; avoids the harmless-but-noisy
    # "Invalid file descriptor: -1" Python 3.12 GC warning on Space startup.
    with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
        future = pool.submit(hf_hub_download, HF_REPO_ID, HF_FILENAME, token=token)
        path = future.result()
    return Path(path)


def _build_session(model_path: Path) -> ort.InferenceSession:
    opts = ort.SessionOptions()
    opts.intra_op_num_threads = NUM_THREADS
    opts.inter_op_num_threads = NUM_THREADS
    opts.execution_mode = ort.ExecutionMode.ORT_PARALLEL
    opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
    return ort.InferenceSession(
        str(model_path),
        sess_options=opts,
        providers=["CPUExecutionProvider"],
    )


# ---------------------------------------------------------------------------
# Preprocessing (replaces timm transforms, no heavy ML dependency at serve time)
# ---------------------------------------------------------------------------

def _preprocess(nail_rgb: np.ndarray) -> np.ndarray:
    """
    Resize → CenterCrop → ToTensor → Normalize, matching BEiT training config.
    Returns float32 array [1, 3, 384, 384].
    """
    img = Image.fromarray(nail_rgb).convert("RGB")
    # Resize shortest side to INPUT_SIZE with bicubic
    w, h = img.size
    scale = INPUT_SIZE / min(w, h)
    new_w, new_h = max(INPUT_SIZE, round(w * scale)), max(INPUT_SIZE, round(h * scale))
    img = img.resize((new_w, new_h), Image.BICUBIC)
    # CenterCrop
    left = (new_w - INPUT_SIZE) // 2
    top  = (new_h - INPUT_SIZE) // 2
    img  = img.crop((left, top, left + INPUT_SIZE, top + INPUT_SIZE))
    # To float [0,1], normalise
    arr = np.asarray(img, dtype=np.float32) / 255.0
    arr = (arr - MEAN) / STD
    return arr.transpose(2, 0, 1)[None]  # [1, C, H, W]


# ---------------------------------------------------------------------------
# Inference with 3-view TTA
# ---------------------------------------------------------------------------

def _tta_logits(session: ort.InferenceSession, pixel_values: np.ndarray) -> np.ndarray:
    """Average logits over original + hflip + vflip views."""
    views = [
        pixel_values,
        pixel_values[:, :, :, ::-1].copy(),  # horizontal flip
        pixel_values[:, :, ::-1, :].copy(),  # vertical flip
    ]
    logits = np.stack(
        [session.run(None, {"pixel_values": v})[0] for v in views]
    ).mean(axis=0)
    return logits  # [B, 5]


# ---------------------------------------------------------------------------
# Top-level backend class
# ---------------------------------------------------------------------------

class Backend:
    def __init__(self) -> None:
        model_path = _get_model_path()
        self._session = _build_session(model_path)
        print(f"[backend] Loaded model from {model_path} with {NUM_THREADS} ORT threads.")

    def predict(self, image_rgb: np.ndarray) -> dict:
        """
        Run the full DeepNAPSI pipeline on a hand image.

        Args:
            image_rgb: HxWx3 uint8 RGB array (Gradio default).

        Returns:
            dict with keys:
              annotated_image  – RGB image with hand skeleton drawn
              nails            – list of 5 RGB nail crop arrays
              napsi_scores     – list of 5 int NAPSI predictions (0-4)
              napsi_sum        – int, sum of all 5 scores
              error            – str | None
        """
        image_bgr = cv2.cvtColor(image_rgb, cv2.COLOR_RGB2BGR)
        nails, landmarks = get_nails_and_landmarks(image_bgr)

        if nails is None or landmarks is None:
            return {
                "annotated_image": image_rgb,
                "nails":           [np.zeros((64, 64, 3), dtype=np.uint8)] * 5,
                "napsi_scores":    [-1] * 5,
                "napsi_sum":       -1,
                "error":           "No hand detected. Please upload a clear photo of one hand.",
            }

        # Draw skeleton on a copy
        annotated = image_bgr.copy()
        draw_hand(annotated, landmarks)
        annotated_rgb = cv2.cvtColor(annotated, cv2.COLOR_BGR2RGB)

        # Run classification on all 5 nails
        napsi_scores: List[int] = []
        nail_rgbs:    List[np.ndarray] = []

        for nail_bgr in nails:
            # Nail crops come out as RGB from extract_nails (already BGR→RGB swapped inside)
            nail_rgb = nail_bgr  # already RGB after the ::-1 flip in extract_nails
            nail_rgbs.append(nail_rgb)

            pixel_values = _preprocess(nail_rgb)
            logits = _tta_logits(self._session, pixel_values)  # [1, 5]
            pred   = int(np.argmax(logits, axis=-1)[0])
            napsi_scores.append(pred)

        return {
            "annotated_image": annotated_rgb,
            "nails":           nail_rgbs,
            "napsi_scores":    napsi_scores,
            "napsi_sum":       sum(napsi_scores),
            "error":           None,
        }