code/infer_bg2fg_variable.py

"""Variable-length bg→fg inference for SA + Frieren.

Short bg (< window): repeat-pad to window, infer, truncate to original length.
Long bg (> window):  50% overlap windows + linear-crossfade overlap-add.
Same wrapper for both models; per-model imports are lazy so a missing env on
one side doesn't break the other.

Mixture audio is generated alongside fg_pred:
  bg_norm  = LUFS-normalize(bg, -30)
  mix_raw  = bg_norm + fg_pred
  mix_norm = LUFS-normalize(mix_raw, -23)

Usage:
  conda activate stable-audio
  python infer_bg2fg_variable.py --model sa --ckpt <last.ckpt> \
      --bg-dir /home/dingqy/inference_demo/youtube \
      --out /home/dingqy/inference_demo/youtube_out/sa

  conda activate V2A
  python infer_bg2fg_variable.py --model frieren --ckpt <last.ckpt> \
      --bg-dir /home/dingqy/inference_demo/youtube \
      --out /home/dingqy/inference_demo/youtube_out/frieren
"""
import argparse, os, sys, json
from pathlib import Path
import numpy as np
import torch
import torchaudio

# ------------------------- variable-length glue -------------------------

def repeat_pad(bg_1d_np: np.ndarray, target_len: int) -> np.ndarray:
    """Tile bg until length >= target_len, then truncate."""
    L = len(bg_1d_np)
    if L >= target_len:
        return bg_1d_np[:target_len]
    n = (target_len + L - 1) // L
    return np.tile(bg_1d_np, n)[:target_len]


def overlap_stitch(bg_1d_np: np.ndarray, infer_one,
                   window_len: int, hop: int) -> np.ndarray:
    """Sliding 50%-overlap windows, infer each, weighted overlap-add.

    Triangular window for crossfade; weight normalization at the end so edges
    (covered by fewer windows) don't get attenuated.

    `infer_one` takes a 1-D float32 ndarray of length `window_len`,
    returns a 1-D float32 ndarray of the same length.
    """
    L = len(bg_1d_np)
    assert L > window_len, "use repeat_pad for short bg"

    # triangular window peaking at window_len // 2
    half = window_len // 2
    ramp_up = np.linspace(0.0, 1.0, half, endpoint=False, dtype=np.float32)
    ramp_dn = np.linspace(1.0, 0.0, window_len - half, endpoint=False, dtype=np.float32)
    win = np.concatenate([ramp_up, ramp_dn])

    starts = list(range(0, L - window_len + 1, hop))
    if starts[-1] + window_len < L:
        starts.append(L - window_len)        # last window flush to end

    fg_acc = np.zeros(L, dtype=np.float32)
    w_acc  = np.zeros(L, dtype=np.float32)
    for k, s in enumerate(starts):
        bg_chunk = bg_1d_np[s:s + window_len].astype(np.float32)
        fg_chunk = infer_one(bg_chunk).astype(np.float32)
        fg_acc[s:s + window_len] += fg_chunk * win
        w_acc [s:s + window_len] += win
        print(f"    window {k+1}/{len(starts)}  start={s/16000:.2f}s", flush=True)
    w_acc = np.maximum(w_acc, 1e-6)
    return fg_acc / w_acc


def render_3panel_spec(bg_1d, fg_1d, mix_1d, sr, title, out_path,
                       n_fft=2048, hop_length=512, n_mels=128, fmax=None):
    """3-panel mel spec (bg / fg_pred / mixture) using librosa.display.specshow."""
    import matplotlib
    matplotlib.use("Agg")
    import matplotlib.pyplot as plt
    import librosa.display as ld
    from librosa.filters import mel as librosa_mel_fn
    if fmax is None:
        fmax = sr // 2
    mb = torch.from_numpy(librosa_mel_fn(sr=sr, n_fft=n_fft, n_mels=n_mels,
                                         fmin=0, fmax=fmax)).float()
    win = torch.hann_window(n_fft)
    def to_mel(x):
        x_t = torch.from_numpy(x.astype(np.float32)).unsqueeze(0)
        spec = torch.stft(x_t, n_fft=n_fft, hop_length=hop_length, win_length=n_fft,
                          window=win, center=True, return_complex=True).abs()
        return ((mb @ spec[0]).clamp(min=1e-5).log10().numpy() * 20.0)
    mels = [to_mel(bg_1d), to_mel(fg_1d), to_mel(mix_1d)]
    labels = ["bg", "fg_pred", "mixture"]
    vmin = min(m.min() for m in mels); vmax = max(m.max() for m in mels)
    fig, axes = plt.subplots(3, 1, figsize=(12, 7.5), dpi=110, sharex=True)
    last_img = None
    for ax, m, lab in zip(axes, mels, labels):
        last_img = ld.specshow(m, x_axis="time", y_axis="mel",
                               sr=sr, hop_length=hop_length, fmax=fmax,
                               cmap="magma", ax=ax, vmin=vmin, vmax=vmax)
        ax.set_ylabel(f"{lab}\nmel (Hz)", fontsize=9)
    axes[0].set_title(title, fontsize=11)
    axes[-1].set_xlabel("time (s)")
    fig.colorbar(last_img, ax=axes, format="%+.1f", label="log-mel (dB)",
                 shrink=0.9, pad=0.02)
    fig.savefig(out_path, bbox_inches="tight")
    plt.close(fig)


def lufs_normalize(wav_1d: np.ndarray, target_lufs: float, sr: int) -> np.ndarray:
    import pyloudnorm as pyln
    meter = pyln.Meter(sr)
    cur = meter.integrated_loudness(wav_1d)
    if not np.isfinite(cur):
        return wav_1d
    return pyln.normalize.loudness(wav_1d, cur, target_lufs).astype(np.float32)


def find_lufs_mixture_gains(bg_mono, fg_mono, sr,
                             bg_lufs=-30.0, mix_lufs=-23.0,
                             tol=0.05, max_iter=25):
    """Hidingsound LUFS protocol — returns (bg_gain, fg_gain) linear scalars.
      - bg final LUFS in mixture = bg_lufs (FIXED, no further scaling).
      - LUFS(bg_gain*bg_mono + fg_gain*fg_mono) = mix_lufs.
    Caller multiplies these gains into the actual (possibly stereo) signals.

    Binary search on fg gain in dB-space. LUFS is monotone in s_db, so binary
    search converges robustly in ~log2(range/tol) ≈ 12-15 iterations regardless
    of bg/fg power balance.
    """
    import pyloudnorm as pyln
    meter = pyln.Meter(sr)
    L_bg = meter.integrated_loudness(bg_mono)
    if not np.isfinite(L_bg):
        return 1.0, 0.0
    bg_gain = 10 ** ((bg_lufs - L_bg) / 20)
    bg_norm_mono = bg_mono * bg_gain
    L_fg = meter.integrated_loudness(fg_mono)
    if not np.isfinite(L_fg):
        return float(bg_gain), 0.0
    e_target = 10 ** (mix_lufs / 10)
    e_bg     = 10 ** (bg_lufs / 10)
    if e_target <= e_bg:
        return float(bg_gain), 0.0                          # mix target ≤ bg → no fg

    def loudness_at(s_db):
        mix = bg_norm_mono + (10 ** (s_db / 20)) * fg_mono
        return meter.integrated_loudness(mix)

    # Initial bracket centered on the analytic energy-sum estimate ± 30 dB.
    s0_db = 10 * np.log10(max(e_target - e_bg, 0)) - L_fg
    lo, hi = s0_db - 30.0, s0_db + 30.0
    L_lo, L_hi = loudness_at(lo), loudness_at(hi)
    # Expand if bracket doesn't span the target.
    while np.isfinite(L_lo) and L_lo > mix_lufs and lo > -120.0:
        lo -= 20.0; L_lo = loudness_at(lo)
    while np.isfinite(L_hi) and L_hi < mix_lufs and hi < 60.0:
        hi += 20.0; L_hi = loudness_at(hi)
    if not np.isfinite(L_hi) or L_hi < mix_lufs:
        return float(bg_gain), float(10 ** (hi / 20))       # un-reachable; clamp at hi

    mid = 0.5 * (lo + hi)
    for _ in range(max_iter):
        mid = 0.5 * (lo + hi)
        L_mid = loudness_at(mid)
        if not np.isfinite(L_mid): break
        if abs(L_mid - mix_lufs) < tol: break
        if L_mid < mix_lufs:
            lo = mid
        else:
            hi = mid
    return float(bg_gain), float(10 ** (mid / 20))


def make_mixture(bg_1d, fg_1d, sr, bg_lufs=-30.0, mix_lufs=-23.0):
    """Mono convenience wrapper. Returns (bg_norm, mix)."""
    bg_g, fg_g = find_lufs_mixture_gains(bg_1d, fg_1d, sr, bg_lufs, mix_lufs)
    bg_norm = (bg_1d * bg_g).astype(np.float32)
    mix     = (bg_norm + fg_g * fg_1d).astype(np.float32)
    return bg_norm, mix


# ------------------------- per-model wrappers -------------------------

def run_sa(args):
    """SA: native 44.1 kHz stereo. window=440320 samples (9.98s), hop=220160."""
    SA_ROOT = Path("/nfs/turbo/coe-ahowens-nobackup/dingqy/friendly-stable-audio-tools")
    sys.path.insert(0, str(SA_ROOT))
    from stable_audio_tools.models import create_model_from_config
    from stable_audio_tools.models.utils import load_ckpt_state_dict
    from stable_audio_tools.training import create_training_wrapper_from_config
    from stable_audio_tools.inference.generation import generate_diffusion_cond

    model_cfg_path = SA_ROOT / "stable_audio_tools/configs/model_configs/txt2audio/stable_audio_open_1_0_bg2fg_rebalance.json"
    mc = json.load(open(model_cfg_path))

    print("[sa] instantiating + loading wrapper...", flush=True)
    base = create_model_from_config(mc)
    wrapper = create_training_wrapper_from_config(mc, base)
    sd = load_ckpt_state_dict(args.ckpt)
    wrapper.load_state_dict(sd, strict=False)
    if getattr(wrapper, "diffusion_ema", None) is not None:
        wrapper.diffusion.model = wrapper.diffusion_ema.ema_model
    model = wrapper.diffusion.cuda().eval()

    SR = mc["sample_rate"]            # 44100
    CH = mc["audio_channels"]         # 2
    SAMPLE_SIZE = mc["sample_size"]   # 440320 = ~9.98s

    def sa_infer_window(bg_2d_np):
        """bg_2d_np shape [C=2, T=SAMPLE_SIZE] float32 in [-1,1] → fg [C, T]."""
        bg_t = torch.from_numpy(bg_2d_np).clamp(-1, 1)
        cond = [{
            "bg_audio": bg_t.unsqueeze(0),    # [1, C, T]
            "seconds_start": 0,
            "seconds_total": 10,
        }]
        with torch.no_grad(), torch.cuda.amp.autocast():
            fakes = generate_diffusion_cond(
                model, steps=args.steps, cfg_scale=args.cfg_scale,
                conditioning=cond, sample_size=SAMPLE_SIZE,
                seed=args.seed, disable_tqdm=True,
            )
        return fakes[0].cpu().float().numpy().clip(-1, 1)   # [C, T]

    out_dir = Path(args.out); out_dir.mkdir(parents=True, exist_ok=True)
    bg_files = sorted(Path(args.bg_dir).glob("*.wav"))
    print(f"[sa] {len(bg_files)} bg files -> {out_dir}", flush=True)

    for bg_path in bg_files:
        tag = bg_path.stem.replace("_bg", "")
        print(f"\n[sa] === {tag} ===", flush=True)
        # Load + resample to 44.1k stereo
        wav, sr_in = torchaudio.load(str(bg_path))
        if wav.shape[0] == 1:
            wav = wav.repeat(2, 1)
        elif wav.shape[0] != CH:
            wav = wav[:CH] if wav.shape[0] > CH else wav.mean(0, keepdim=True).repeat(CH, 1)
        if sr_in != SR:
            wav = torchaudio.functional.resample(wav, sr_in, SR)
        bg_full = wav.numpy().astype(np.float32)             # [C, T]
        L = bg_full.shape[-1]
        print(f"  bg loaded: shape={bg_full.shape} dur={L/SR:.2f}s", flush=True)

        if L <= SAMPLE_SIZE:
            print(f"  short → repeat-pad to {SAMPLE_SIZE/SR:.2f}s", flush=True)
            bg_pad = np.stack([repeat_pad(bg_full[c], SAMPLE_SIZE) for c in range(CH)])
            fg_pad = sa_infer_window(bg_pad)
            fg_full = fg_pad[:, :L]
        else:
            print(f"  long → overlap-add ({SAMPLE_SIZE//2/SR:.2f}s hop)", flush=True)
            hop = SAMPLE_SIZE // 2
            half = SAMPLE_SIZE // 2
            ramp_up = np.linspace(0.0, 1.0, half, endpoint=False, dtype=np.float32)
            ramp_dn = np.linspace(1.0, 0.0, SAMPLE_SIZE - half, endpoint=False, dtype=np.float32)
            win = np.concatenate([ramp_up, ramp_dn])
            starts = list(range(0, L - SAMPLE_SIZE + 1, hop))
            if starts[-1] + SAMPLE_SIZE < L:
                starts.append(L - SAMPLE_SIZE)
            fg_acc = np.zeros((CH, L), dtype=np.float32)
            w_acc  = np.zeros(L, dtype=np.float32)
            for k, s in enumerate(starts):
                bg_chunk = bg_full[:, s:s + SAMPLE_SIZE]
                fg_chunk = sa_infer_window(bg_chunk)         # [C, T]
                fg_acc[:, s:s + SAMPLE_SIZE] += fg_chunk * win[None, :]
                w_acc [s:s + SAMPLE_SIZE] += win
                print(f"    window {k+1}/{len(starts)}  start={s/SR:.2f}s", flush=True)
            w_acc = np.maximum(w_acc, 1e-6)
            fg_full = fg_acc / w_acc

        # Save bg / fg_pred FIRST so a downstream mixture/spec crash doesn't
        # discard the diffusion output.
        torchaudio.save(str(out_dir / f"{tag}_bg.wav"),
                        torch.from_numpy(bg_full), SR)
        torchaudio.save(str(out_dir / f"{tag}_fg_pred.wav"),
                        torch.from_numpy(fg_full).clamp(-1, 1), SR)

        try:
            # Hidingsound LUFS protocol: measure on mono channel-mean, apply
            # the resulting (bg_gain, fg_gain) to the stereo signals so bg ends
            # at exactly -30 LUFS and (bg+fg) at -23 LUFS in mixture.
            bg_mono = bg_full.mean(0); fg_mono = fg_full.mean(0)
            bg_g, fg_g = find_lufs_mixture_gains(bg_mono, fg_mono, SR)
            mix_stereo = (bg_full * bg_g + fg_full * fg_g).astype(np.float32)
            torchaudio.save(str(out_dir / f"{tag}_mixture.wav"),
                            torch.from_numpy(mix_stereo).clamp(-1, 1), SR)
            render_3panel_spec(bg_full.mean(0) * bg_g, fg_full.mean(0) * fg_g,
                               mix_stereo.mean(0),
                               sr=SR, title=f"[SA] {tag}",
                               out_path=str(out_dir / f"{tag}_spec.png"))
            print(f"  wrote {tag}_{{bg,fg_pred,mixture}}.wav + spec.png", flush=True)
        except Exception as e:
            print(f"  [warn] mixture/spec failed but bg+fg_pred saved: {e}", flush=True)


def run_frieren(args):
    """Frieren: 16 kHz mono. window=131072 samples (8.19s), hop=65536."""
    FR_ROOT = Path("/nfs/turbo/coe-ahowens-nobackup/dingqy/Frieren-V2A")
    sys.path.insert(0, str(FR_ROOT / "Frieren"))
    from cfm.util import instantiate_from_config
    from vocoder.bigvgan.models import VocoderBigVGAN
    from omegaconf import OmegaConf

    cfg = OmegaConf.load(FR_ROOT / "Frieren/configs/ldm_training/hidingsound_bg2fg_rebalance.yaml")
    print("[frieren] instantiating model...", flush=True)
    model = instantiate_from_config(cfg.model)
    sd = torch.load(args.ckpt, map_location="cpu", weights_only=False)
    state_dict = sd.get("state_dict", sd)
    model.load_state_dict(state_dict, strict=False)
    model = model.cuda().eval()
    vocoder = VocoderBigVGAN(str(FR_ROOT / "checkpoints/vocoder/bigvnat"), device="cuda")

    SR = 16000
    WINDOW = 131072       # 8.19s @ 16k
    HOP = WINDOW // 2

    def fr_infer_window(bg_1d_np):
        """bg [T=131072] float32 → fg [T=131072] float32 (vocoded)."""
        bg_t = torch.from_numpy(bg_1d_np).cuda().unsqueeze(0)              # [1, T]
        with torch.no_grad():
            bg_for_cond = bg_t.unsqueeze(-1)                               # [1, T, 1]
            cond = model.cond_stage_model(bg_for_cond)
            shape = (1, model.mel_dim, model.mel_length)
            z_pred, _ = model.sample_param_cfg(
                cond=cond, cfg_scale=args.cfg_scale, batch_size=1,
                timesteps=args.steps, solver="euler", shape=shape,
            )
            vae = getattr(model.first_stage_model, "vae", model.first_stage_model)
            mel_pred = vae.decode(z_pred)
        wav = vocoder.vocode(mel_pred[0].cpu().numpy())                   # ndarray [T_wav]
        wav = np.asarray(wav, dtype=np.float32)
        # vocoder may emit slightly different length; trim/pad to WINDOW
        if len(wav) >= WINDOW:
            return wav[:WINDOW]
        out = np.zeros(WINDOW, dtype=np.float32); out[:len(wav)] = wav
        return out

    out_dir = Path(args.out); out_dir.mkdir(parents=True, exist_ok=True)
    bg_files = sorted(Path(args.bg_dir).glob("*.wav"))
    print(f"[frieren] {len(bg_files)} bg files -> {out_dir}", flush=True)

    for bg_path in bg_files:
        tag = bg_path.stem.replace("_bg", "")
        print(f"\n[frieren] === {tag} ===", flush=True)
        wav, sr_in = torchaudio.load(str(bg_path))
        if wav.shape[0] > 1:
            wav = wav.mean(0, keepdim=True)
        if sr_in != SR:
            wav = torchaudio.functional.resample(wav, sr_in, SR)
        bg_1d = wav.squeeze(0).numpy().astype(np.float32)
        L = len(bg_1d)
        print(f"  bg loaded: dur={L/SR:.2f}s", flush=True)

        if L <= WINDOW:
            print(f"  short → repeat-pad to {WINDOW/SR:.2f}s", flush=True)
            bg_pad = repeat_pad(bg_1d, WINDOW)
            fg_pad = fr_infer_window(bg_pad)
            fg_1d  = fg_pad[:L]
        else:
            print(f"  long → overlap-add ({HOP/SR:.2f}s hop)", flush=True)
            fg_1d = overlap_stitch(bg_1d, fr_infer_window, WINDOW, HOP)

        # Save bg / fg_pred FIRST so a downstream crash (LUFS / spec) doesn't
        # discard the diffusion output that just took ~10 min to produce.
        torchaudio.save(str(out_dir / f"{tag}_bg.wav"),
                        torch.from_numpy(bg_1d).unsqueeze(0), SR)
        torchaudio.save(str(out_dir / f"{tag}_fg_pred.wav"),
                        torch.from_numpy(fg_1d.astype(np.float32)).clamp(-1, 1).unsqueeze(0), SR)

        try:
            _, mix_norm = make_mixture(bg_1d, fg_1d, SR)
            torchaudio.save(str(out_dir / f"{tag}_mixture.wav"),
                            torch.from_numpy(mix_norm.astype(np.float32)).clamp(-1, 1).unsqueeze(0), SR)
            render_3panel_spec(bg_1d, fg_1d, mix_norm,
                               sr=SR, n_fft=1024, hop_length=256, n_mels=80, fmax=8000,
                               title=f"[Frieren] {tag}",
                               out_path=str(out_dir / f"{tag}_spec.png"))
            print(f"  wrote {tag}_{{bg,fg_pred,mixture}}.wav + spec.png", flush=True)
        except Exception as e:
            print(f"  [warn] mixture/spec failed but bg+fg_pred saved: {e}", flush=True)


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--model", choices=["sa", "frieren"], required=True)
    ap.add_argument("--ckpt", required=True)
    ap.add_argument("--bg-dir", required=True)
    ap.add_argument("--out", required=True)
    ap.add_argument("--steps", type=int, default=None,
                    help="default: SA=100, Frieren=26")
    ap.add_argument("--cfg-scale", type=float, default=1.0)
    ap.add_argument("--seed", type=int, default=42)
    args = ap.parse_args()
    if args.steps is None:
        args.steps = 100 if args.model == "sa" else 26

    if args.model == "sa":
        run_sa(args)
    else:
        run_frieren(args)


if __name__ == "__main__":
    main()