scripts/generation/run_fx_transfer_b2.py

#!/usr/bin/env python3
"""FX-Encoder B2: Style transfer ACE-Step tracks with random MTG reference."""
import os, sys, json, random, time
from pathlib import Path

sys.path.insert(0, str(Path(__file__).parent / "FXEncoder" / "mixing_style_transfer"))
from networks.architectures import FXencoder, TCNModel

import torch
import torchaudio
import soundfile as sf
import numpy as np
from collections import OrderedDict

ACESTEP_DIR = Path("/ssd_data/dataset/haim_dataset/fake/acestep/samples")
MTG_DIR = Path("/ssd_data/dataset/haim_dataset/real/MTG")
OUT_DIR = Path("/ssd_data/dataset/haim_dataset/B_hybrid/B2_human_mastered_ai_dsp")
WEIGHTS = Path("/ssd_data/dataset/haim_dataset/FXEncoder/weights")
TARGET = 6000
SR = 44100
SEG_LEN = SR * 10

CFG_ENC = {
    "channels": [16,32,64,128,256,256,512,512,1024,1024,2048,2048],
    "kernels": [25,25,15,15,10,10,10,10,5,5,5,5],
    "strides": [4,4,2,2,2,2,2,2,2,2,1,1],
    "dilation": [1,1,1,1,1,1,1,1,1,1,1,1],
    "bias": True, "norm": "batch", "conv_block": "res", "activation": "relu",
}
CFG_CONV = {
    "condition_dimension": 2048, "nblocks": 14, "dilation_growth": 2,
    "kernel_size": 15, "channel_width": 128, "stack_size": 15, "causal": False,
}

def load_audio(path):
    wav, sr = torchaudio.load(str(path))
    if wav.shape[0] == 1: wav = wav.repeat(2, 1)
    elif wav.shape[0] > 2: wav = wav[:2, :]
    if sr != SR: wav = torchaudio.functional.resample(wav, sr, SR)
    return wav

def main():
    device = "cuda" if torch.cuda.is_available() else "cpu"

    # Load models
    enc = FXencoder(CFG_ENC).to(device)
    conv = TCNModel(nparams=2048, ninputs=2, noutputs=2,
                    nblocks=14, dilation_growth=2, kernel_size=15,
                    channel_width=128, stack_size=15, cond_dim=2048, causal=False).to(device)

    for name, model, path in [("enc", enc, WEIGHTS/"FXencoder.pt"), ("conv", conv, WEIGHTS/"MixFXcloner.pt")]:
        ckpt = torch.load(str(path), map_location=device)
        state = OrderedDict()
        for k, v in ckpt["model"].items():
            state[k[7:] if k.startswith("module.") else k] = v
        model.load_state_dict(state)
        model.eval()
    print("Models loaded")

    # Get file lists
    ai_files = sorted(ACESTEP_DIR.glob("*.mp3"))
    ref_files = sorted(MTG_DIR.glob("*.mp3"))
    random.seed(int(time.time()))

    OUT_DIR.mkdir(parents=True, exist_ok=True)
    meta_path = OUT_DIR / "metadata.jsonl"

    existing = len(list(OUT_DIR.glob("*.wav")))
    if existing >= TARGET:
        print(f"Already at {existing}/{TARGET}")
        return
    print(f"Existing: {existing}, processing {TARGET - existing} more")

    meta_f = open(meta_path, "a", encoding="utf-8")
    done = existing
    t0 = time.time()

    with torch.no_grad():
        for i, ai_path in enumerate(ai_files):
            if done >= TARGET:
                break
            fname = f"B2_fx_{ai_path.stem}.wav"
            out_path = OUT_DIR / fname
            if out_path.exists():
                done += 1
                continue

            for attempt in range(3):
              try:
                ref_path = random.choice(ref_files)
                inp = load_audio(ai_path)
                ref = load_audio(ref_path)

                # Extract FX embedding from reference
                ref_gpu = ref.unsqueeze(0).to(device)
                embs = []
                for s in range(0, ref_gpu.shape[2], SEG_LEN):
                    seg = ref_gpu[:, :, s:s+SEG_LEN]
                    if seg.shape[2] < SEG_LEN:
                        seg = torch.nn.functional.pad(seg, (0, SEG_LEN - seg.shape[2]))
                    embs.append(enc(seg))
                fx_emb = torch.mean(torch.stack(embs), dim=0)

                # Apply to input
                inp_gpu = inp.unsqueeze(0).to(device)
                T = inp_gpu.shape[2]
                out_segs = []
                for s in range(0, T, SEG_LEN):
                    seg = inp_gpu[:, :, s:s+SEG_LEN]
                    actual = seg.shape[2]
                    if actual < SEG_LEN:
                        seg = torch.nn.functional.pad(seg, (0, SEG_LEN - actual))
                    out = conv(seg, fx_emb)
                    out_segs.append(out[:, :, :actual].cpu())

                result = torch.clamp(torch.cat(out_segs, dim=2).squeeze(0), -1, 1)
                sf.write(str(out_path), result.numpy().T, SR)

                meta = {
                    "track_id": f"B2_fx_{ai_path.stem}",
                    "filename": fname,
                    "input_source": ai_path.name,
                    "reference_source": ref_path.name,
                    "method": "fx_encoder_style_transfer",
                }
                # Per-track JSON
                with open(OUT_DIR / f"B2_fx_{ai_path.stem}.json", "w", encoding="utf-8") as jf:
                    json.dump(meta, jf, ensure_ascii=False, indent=2)
                meta_f.write(json.dumps(meta, ensure_ascii=False) + "\n")
                meta_f.flush()
                done += 1

                if done % 50 == 0:
                    elapsed = time.time() - t0
                    eta = (TARGET - done) * elapsed / max(done - existing, 1)
                    print(f"[{done}/{TARGET}] ETA: {eta/3600:.1f}h")

              except Exception as e:
                print(f"Error {ai_path.name} (attempt {attempt+1}): {e}")
                torch.cuda.empty_cache()
                if attempt < 2:
                    continue
                break

    meta_f.close()
    print(f"Done: {done}/{TARGET}")

if __name__ == "__main__":
    main()