scripts/generation/fx_transfer_b2.py

#!/usr/bin/env python3
"""
FX-Encoder Style Transfer for HAIM B2
Minimal inference: loads FXencoder + MixFXcloner directly, no legacy deps needed.

Usage:
    python fx_transfer.py \
        --input /path/to/ai_track.wav \
        --reference /path/to/human_track.wav \
        --output /path/to/output.wav
"""
import argparse
import sys
import os
from pathlib import Path
from collections import OrderedDict

import torch
import torchaudio
import soundfile as sf
import numpy as np

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

WEIGHTS_DIR = Path(__file__).parent / "FXEncoder" / "weights"

# Default configs from FXEncoder/inference/configs.yaml
CFG_ENCODER = {
    "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_CONVERTER = {
    "condition_dimension": 2048,
    "nblocks": 14,
    "dilation_growth": 2,
    "kernel_size": 15,
    "channel_width": 128,
    "stack_size": 15,
    "causal": False,
}

SAMPLE_RATE = 44100
SEGMENT_LENGTH = SAMPLE_RATE * 10  # 10 seconds per segment


def load_models(device):
    enc = FXencoder(CFG_ENCODER).to(device)
    conv = TCNModel(
        nparams=CFG_CONVERTER["condition_dimension"],
        ninputs=2, noutputs=2,
        nblocks=CFG_CONVERTER["nblocks"],
        dilation_growth=CFG_CONVERTER["dilation_growth"],
        kernel_size=CFG_CONVERTER["kernel_size"],
        channel_width=CFG_CONVERTER["channel_width"],
        stack_size=CFG_CONVERTER["stack_size"],
        cond_dim=CFG_CONVERTER["condition_dimension"],
        causal=CFG_CONVERTER["causal"],
    ).to(device)

    # Load weights (trained with DDP, strip 'module.' prefix)
    for name, model, path in [
        ("FXencoder", enc, WEIGHTS_DIR / "FXencoder.pt"),
        ("MixFXcloner", conv, WEIGHTS_DIR / "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(f"  Loaded {name}: {path.name}")

    return enc, conv


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


@torch.no_grad()
def transfer(enc, conv, input_wav, ref_wav, device, segment_length=SEGMENT_LENGTH):
    """
    Transfer the mixing style of ref_wav onto input_wav.
    Process in segments to handle long tracks.
    """
    # Extract FX embedding from reference (use whole track, averaged)
    ref = ref_wav.unsqueeze(0).to(device)  # [1, 2, T]

    # Segment reference and average embeddings
    ref_len = ref.shape[2]
    embeddings = []
    for start in range(0, ref_len, segment_length):
        seg = ref[:, :, start:start + segment_length]
        if seg.shape[2] < segment_length:
            seg = torch.nn.functional.pad(seg, (0, segment_length - seg.shape[2]))
        emb = enc(seg)
        embeddings.append(emb)
    fx_embedding = torch.mean(torch.stack(embeddings), dim=0)  # [1, 2048]

    # Apply style to input, segment by segment
    inp = input_wav.unsqueeze(0).to(device)
    inp_len = inp.shape[2]
    output_segments = []

    for start in range(0, inp_len, segment_length):
        seg = inp[:, :, start:start + segment_length]
        actual_len = seg.shape[2]
        if actual_len < segment_length:
            seg = torch.nn.functional.pad(seg, (0, segment_length - seg.shape[2]))

        out = conv(seg, fx_embedding)
        out = out[:, :, :actual_len]
        output_segments.append(out.cpu())

    return torch.cat(output_segments, dim=2).squeeze(0)


def main():
    parser = argparse.ArgumentParser(description="FX-Encoder Mixing Style Transfer")
    parser.add_argument("--input", required=True, help="AI track (input to transform)")
    parser.add_argument("--reference", required=True, help="Human track (style source)")
    parser.add_argument("--output", required=True, help="Output path")
    args = parser.parse_args()

    device = "cuda" if torch.cuda.is_available() else "cpu"
    print(f"Device: {device}")

    print("Loading models...")
    enc, conv = load_models(device)

    print(f"Input (AI): {args.input}")
    print(f"Reference (Human): {args.reference}")

    input_wav = load_audio(args.input)
    ref_wav = load_audio(args.reference)

    print("Transferring mixing style...")
    output_wav = transfer(enc, conv, input_wav, ref_wav, device)
    output_wav = torch.clamp(output_wav, -1.0, 1.0)

    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
    sf.write(args.output, output_wav.numpy().T, SAMPLE_RATE)
    print(f"Saved: {args.output}")


if __name__ == "__main__":
    main()