inference/e2e_model.py

# Purpose: ArtifactNet 7ch inference pipeline — HF Spaces (CPU, ONNX Runtime)
# Dependencies: onnxruntime, torch (HPSS/Mel only), huggingface_hub, scipy

"""ArtifactNet v9.4 inference — onnxruntime CPU.

UNet + CNN 은 .onnx (public-safe) 로 실행, HPSS + Mel + 7ch feature 는
pytorch CPU 로 처리 (가중치 없는 고정 연산이라 노출 위험 없음).
"""

import os
from pathlib import Path

import numpy as np
import onnxruntime as ort
import torch
from huggingface_hub import hf_hub_download
from scipy import stats as sp_stats

from config import (
    HF_MODEL_REPO, UNET_ONNX_FILENAME, CNN_ONNX_FILENAME,
    SR, N_FFT, HOP_LENGTH, CHUNK_SAMPLES, BATCH_SIZE,
)
from .audio_utils import sliding_chunks
from .model import (
    DifferentiableMel, hpss_gpu_pure, compute_forensic_features_7ch,
)

N_MELS = 128

FREQ_BANDS = [
    ("sub",    0,     250),
    ("low",    250,   2000),
    ("mid",    2000,  6000),
    ("hi_mid", 6000,  10000),
    ("hi",     10000, 16000),
    ("air",    16000, 22050),
]


# ============================================================
# Lazy singletons
# ============================================================

_unet_sess: ort.InferenceSession | None = None
_cnn_sess:  ort.InferenceSession | None = None
_mel: DifferentiableMel | None = None
_stft_window: torch.Tensor | None = None


def _ort_threads() -> int:
    """HF Spaces CPU basic = 2 vCPU. 환경변수로 override 가능."""
    try:
        return int(os.environ.get("ORT_THREADS", "2"))
    except ValueError:
        return 2


def _resolve_onnx(filename: str, env_var: str) -> str:
    """로컬 override (ARTIFACTNET_UNET_ONNX / _CNN_ONNX) 있으면 그걸 사용, 아니면 HF Hub."""
    local = os.environ.get(env_var)
    if local and Path(local).is_file():
        return local
    return hf_hub_download(HF_MODEL_REPO, filename)


def load_models():
    """ONNX 세션 + Mel/Window 초기화 (import 후 1회)."""
    global _unet_sess, _cnn_sess, _mel, _stft_window
    if _unet_sess is not None:
        return

    unet_path = _resolve_onnx(UNET_ONNX_FILENAME, "ARTIFACTNET_UNET_ONNX")
    cnn_path  = _resolve_onnx(CNN_ONNX_FILENAME,  "ARTIFACTNET_CNN_ONNX")

    opts = ort.SessionOptions()
    opts.intra_op_num_threads = _ort_threads()
    opts.inter_op_num_threads = 1
    opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL

    _unet_sess = ort.InferenceSession(unet_path, sess_options=opts,
                                      providers=["CPUExecutionProvider"])
    _cnn_sess  = ort.InferenceSession(cnn_path,  sess_options=opts,
                                      providers=["CPUExecutionProvider"])

    _mel = DifferentiableMel(sr=SR, n_fft=N_FFT, n_mels=N_MELS)
    _mel.eval()
    _stft_window = torch.hann_window(N_FFT)

    print(f"[hf-spaces] ONNX sessions ready (intra_threads={_ort_threads()})", flush=True)


# ============================================================
# Feature extraction helpers (75-dim Router + 28-dim Verdict)
# ============================================================

def _extract_router_verdict_features(
    all_mag, all_res, all_H, all_P, all_mask, all_mel_res, probs,
):
    """infer.py extract_features()와 동일한 로직 (device=CPU)."""
    freq_hz = torch.linspace(0, SR / 2, all_mag.shape[2])
    orig_total = all_mag.pow(2).mean().item() + 1e-8
    res_total = all_res.pow(2).mean().item() + 1e-8

    band_idx = []
    for _, flo, fhi in FREQ_BANDS:
        lo = (freq_hz >= flo).nonzero(as_tuple=True)[0]
        hi = (freq_hz >= fhi).nonzero(as_tuple=True)[0]
        band_idx.append((
            lo[0].item() if len(lo) else 0,
            hi[0].item() if len(hi) else all_mag.shape[2],
        ))

    rf = []
    for i0, i1 in band_idx:
        oe = all_mag[:, :, i0:i1, :].pow(2).mean().item() / orig_total
        re = all_res[:, :, i0:i1, :].pow(2).mean().item() / res_total
        rf.extend([oe, re, re / (oe + 1e-8)])

    mel_profile = all_mel_res.mean(dim=[0, 3]).squeeze().cpu().numpy()
    step = N_MELS // 32
    compressed = mel_profile[:32 * step].reshape(32, step).mean(axis=1)
    compressed = compressed - compressed.mean()
    norm = np.abs(compressed).max() + 1e-8
    rf.extend((compressed / norm).tolist())

    H_total = all_H.pow(2).mean().item() + 1e-8
    P_total = all_P.pow(2).mean().item() + 1e-8
    hp_ratio = H_total / (H_total + P_total)
    rf.append(hp_ratio)

    for i0, i1 in band_idx:
        rf.extend([
            all_H[:, :, i0:i1, :].pow(2).mean().item() / H_total,
            all_P[:, :, i0:i1, :].pow(2).mean().item() / P_total,
        ])

    mask_np = all_mask.cpu().numpy().flatten()
    rf.extend([
        float(mask_np.mean()), float(mask_np.std()),
        float(np.percentile(mask_np, 10)), float(np.percentile(mask_np, 25)),
        float(np.percentile(mask_np, 75)), float(np.percentile(mask_np, 90)),
        float(np.median(mask_np)),
    ])

    rf.extend([
        float(probs.mean()), float(probs.std()), float(np.median(probs)),
        float(np.percentile(probs, 10)), float(np.percentile(probs, 90)),
    ])

    router_feat = np.nan_to_num(np.array(rf, dtype=np.float32))

    arr = probs.astype(np.float64)
    n = len(arr)
    cnn_20 = np.array([
        n, arr.mean(), arr.std(), np.median(arr),
        arr.min(), arr.max(), arr.max() - arr.min(),
        np.percentile(arr, 10), np.percentile(arr, 25),
        np.percentile(arr, 75), np.percentile(arr, 90),
        (arr >= 0.3).mean(), (arr >= 0.5).mean(),
        (arr >= 0.7).mean(), (arr >= 0.8).mean(), (arr >= 0.9).mean(),
        float(sp_stats.skew(arr))          if n >= 3 else 0.0,
        float(sp_stats.kurtosis(arr))      if n >= 3 else 0.0,
        float(np.diff(arr).std())          if n >= 2 else 0.0,
        float(np.abs(np.diff(arr)).max())  if n >= 2 else 0.0,
    ], dtype=np.float32)

    hf8k_i = (freq_hz >= 8000).nonzero(as_tuple=True)[0]
    hf8k_i = hf8k_i[0].item() if len(hf8k_i) else all_mag.shape[2]
    ai0, ai1 = band_idx[5]

    res_8 = np.array([
        all_res[:, :, hf8k_i:, :].pow(2).mean().item() / res_total,
        all_res[:, :, ai0:ai1, :].pow(2).mean().item() / res_total,
        all_H[:, :, ai0:ai1, :].pow(2).mean().item() / H_total,
        all_P[:, :, ai0:ai1, :].pow(2).mean().item() / P_total,
        float(mel_profile[-1]),
        float(mel_profile[0]),
        float(mask_np.mean()),
        float(hp_ratio),
    ], dtype=np.float32)

    verdict_feat = np.nan_to_num(np.concatenate([cnn_20, res_8]))
    return router_feat, verdict_feat


# ============================================================
# Inference
# ============================================================

@torch.no_grad()
def run_e2e_inference(wav_mono_tensor: torch.Tensor):
    """mono waveform -> (probs, placeholder, metadata, forensic_stats, router_feat, verdict_feat).

    ONNX Runtime CPU + pytorch HPSS/Mel.
    """
    if _unet_sess is None:
        load_models()

    chunk_data = sliding_chunks(wav_mono_tensor, CHUNK_SAMPLES)
    if not chunk_data:
        return [], torch.zeros_like(wav_mono_tensor), [], {}, \
               np.zeros(75, dtype=np.float32), np.zeros(28, dtype=np.float32)
    chunks = [chunk for chunk, _ in chunk_data]
    metadata_list = [meta for _, meta in chunk_data]

    probs = []
    all_features = []
    all_mag_list, all_res_list, all_H_list, all_P_list = [], [], [], []
    all_mask_list, all_mel_res_list = [], []

    for i in range(0, len(chunks), BATCH_SIZE):
        batch = torch.stack(chunks[i:i + BATCH_SIZE])  # (B, CHUNK_SAMPLES)

        # STFT (torch, CPU)
        stft = torch.stft(
            batch, N_FFT, HOP_LENGTH,
            window=_stft_window, return_complex=True)
        stft_mag = stft.abs().unsqueeze(1)  # (B, 1, F, T)

        # UNet mask via ONNX
        mask_np = _unet_sess.run(
            ["mask"],
            {"stft_mag": stft_mag.numpy().astype(np.float32)},
        )[0]
        mask = torch.from_numpy(mask_np)
        res_mag = mask * stft_mag

        # HPSS — CPU median filter (unfold + median) 로 학습 분포 유지.
        # librosa.decompose.hpss 는 결과가 달라 v9.4 CNN 오판 (CLAUDE.md 경고 참조).
        H_mag, P_mag = hpss_gpu_pure(res_mag)

        # Mel 3-band
        mel_res = _mel(res_mag)
        mel_H   = _mel(H_mag)
        mel_P   = _mel(P_mag)

        features_7ch = compute_forensic_features_7ch(mel_res, mel_H, mel_P)
        all_features.append(features_7ch)

        # CNN logit via ONNX → sigmoid
        logits = _cnn_sess.run(
            ["logit"],
            {"features_7ch": features_7ch.numpy().astype(np.float32)},
        )[0]
        batch_probs = (1.0 / (1.0 + np.exp(-np.clip(logits, -30, 30)))).tolist()
        probs.extend(batch_probs)

        all_mag_list.append(stft_mag)
        all_res_list.append(res_mag)
        all_H_list.append(H_mag)
        all_P_list.append(P_mag)
        all_mask_list.append(mask)
        all_mel_res_list.append(mel_res)

    if all_features:
        all_feat_tensor = torch.cat(all_features, dim=0)
        channel_means = all_feat_tensor.mean(dim=[2, 3])
        feature_medians = channel_means.median(dim=0).values
        feat_min = channel_means.min(dim=0).values
        feat_max = channel_means.max(dim=0).values
        feat_range = feat_max - feat_min + 1e-8
        normalized = ((feature_medians - feat_min) / feat_range).clamp(0, 1)

        forensic_stats = {
            "residual_energy":     float(normalized[0]),
            "harmonic_strength":   float(normalized[1]),
            "percussive_strength": float(normalized[2]),
            "temporal_delta":      float(normalized[3]),
            "temporal_accel":      float(normalized[4]),
            "hp_ratio":            float(normalized[5]),
            "spectral_flux":       float(normalized[6]),
        }
    else:
        forensic_stats = {}

    probs_arr = np.array(probs, dtype=np.float32)
    if all_mag_list:
        all_mag = torch.cat(all_mag_list, dim=0)
        all_res = torch.cat(all_res_list, dim=0)
        all_H   = torch.cat(all_H_list,   dim=0)
        all_P   = torch.cat(all_P_list,   dim=0)
        all_mask = torch.cat(all_mask_list, dim=0)
        all_mel_res = torch.cat(all_mel_res_list, dim=0)

        router_feat, verdict_feat = _extract_router_verdict_features(
            all_mag, all_res, all_H, all_P, all_mask, all_mel_res, probs_arr,
        )
    else:
        router_feat = np.zeros(75, dtype=np.float32)
        verdict_feat = np.zeros(28, dtype=np.float32)

    residual_placeholder = torch.zeros_like(wav_mono_tensor)
    return probs, residual_placeholder, metadata_list, forensic_stats, router_feat, verdict_feat