lhotseWorkspace/opusTransform.py

import os
# Prevent PyTorch from spawning hundreds of OpenMP threads and locking up your i7
os.environ["OMP_NUM_THREADS"] = "1"
os.environ["MKL_NUM_THREADS"] = "1"

import torch
import torchaudio
from pathlib import Path
import concurrent.futures
import time

def save_opus_async(waveform_cpu, out_path):
    """
    Runs entirely on a background CPU thread.
    Compresses the tensor to Opus and writes to disk.
    """
    try:
        # Ensure subdirectories exist if your source folder has nested folders
        out_path.parent.mkdir(parents=True, exist_ok=True)
        torchaudio.save(str(out_path), waveform_cpu, 16000, format="opus")
    except Exception as e:
        print(f"Failed to encode {out_path}: {e}")

def main():
    # Enforce single-threading for PyTorch operations
    torch.set_num_threads(1)
    
    # Force the FFmpeg backend globally for reading any weird audio extensions
    try:
        torchaudio.set_audio_backend("ffmpeg")
        print("Backend explicitly set to FFmpeg.")
    except Exception as e:
        print(f"Note: Could not explicitly set ffmpeg backend (might be default in your version): {e}")

    # Set up exact paths
    input_dir = Path("/mnt/jonathanDisk/ivritData")
    output_dir = Path("/mnt/jonathanDisk/ivritDataOpus")
    output_dir.mkdir(parents=True, exist_ok=True)

    # Initialize GPU
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Matrix math will execute on: {device}")
    
    # Dictionary to cache GPU resamplers dynamically based on input sample rates
    gpu_resamplers = {}

    # Scan for all standard audio formats recursively
    print("Scanning directory for audio files...")
    exts = ["*.wav", "*.flac", "*.mp3", "*.m4a", "*.ogg"]
    files = []
    for ext in exts:
        files.extend(list(input_dir.rglob(ext)))
        
    print(f"Found {len(files)} files to process.")

    # Launch a ThreadPool for the Opus encoding
    # We use ThreadPool instead of ProcessPool here because torchaudio.save 
    # releases the Python GIL (C++ level I/O), making threads incredibly fast with zero memory overhead.
    max_workers = 12
    executor = concurrent.futures.ThreadPoolExecutor(max_workers=max_workers)
    
    start_time = time.time()
    
    for i, file_path in enumerate(files):
        try:
            # 1. Fast FFmpeg Load (CPU)
            waveform, sr = torchaudio.load(file_path)
            
            # 2. Lightning Fast GPU Resample
            if sr != 16000:
                # Create/cache a resampler if we encounter a new sample rate
                if sr not in gpu_resamplers:
                    gpu_resamplers[sr] = torchaudio.transforms.Resample(orig_freq=sr, new_freq=16000).to(device)
                
                waveform_gpu = waveform.to(device)
                with torch.inference_mode():
                    resampled_gpu = gpu_resamplers[sr](waveform_gpu)
                waveform_16k_cpu = resampled_gpu.to("cpu")
            else:
                waveform_16k_cpu = waveform
                
            # 3. Map the output path to mirror the input directory structure
            rel_path = file_path.relative_to(input_dir)
            out_path = output_dir / rel_path.with_suffix('.opus')
            
            # 4. Fire and forget to the background CPU Opus encoders
            executor.submit(save_opus_async, waveform_16k_cpu, out_path)
            
            # Status tracking
            if i % 5000 == 0 and i > 0:
                print(f"Dispatched {i}/{len(files)} files to CPU encoders...")
                
        except Exception as e:
            print(f"Error loading {file_path}: {e}")

    print("Finished streaming files through GPU. Waiting for final CPU workers to flush Opus files to disk...")
    
    # This prevents the script from exiting until the final Opus file is written
    executor.shutdown(wait=True)
    
    total_time = time.time() - start_time
    print(f"Complete! Converted {len(files)} files to 16kHz Opus in {total_time:.2f} seconds.")

if __name__ == '__main__':
    main()