Julius (PyTorch DSP Library)

Warnings

• gotcha The `resample_frac` function is highly optimized for fractional changes in sample rates where the `old_sr` and `new_sr` reduce to a small irreducible fraction (e.g., 2000 to 3000 simplifies to 2:3). Performance can degrade significantly if the sample rates do not simplify to small integers (e.g., 20001 to 30001).
  Fix: Be mindful of the ratio between `old_sr` and `new_sr`. For arbitrary ratios where performance is critical, consider pre-filtering or alternative resampling methods if `julius` is slow in specific cases.
• gotcha The `julius.fftconv` modules (e.g., `FFTConv1d`, `fft_conv1d`) are optimized for convolutions with *large kernels* (typically >= 128) and a stride of 1. For smaller kernels or different strides, `torch.nn.Conv1d` might be faster or more memory efficient. Dilation and groups are not supported by `julius.fftconv`.
  Fix: Benchmark both `julius.fftconv` and `torch.nn.functional.conv1d` for your specific use case to determine the faster option. Avoid `julius.fftconv` if dilation or groups are required.
• deprecated As of January 2021, the `julius` implementation of resampling has been officially integrated into `torchaudio`. Users primarily focused on resampling may consider using `torchaudio.functional.resample` for potentially more robust or integrated solutions within the PyTorch ecosystem.
  Fix: Evaluate `torchaudio.functional.resample` as an alternative to `julius.resample_frac` for new projects or if migrating existing `torchaudio` pipelines.
• breaking Version 0.2.2 (released January 2021) introduced changes to filter normalization in `lowpass` and `resample` and switched from zero padding to replicate padding. This could subtly alter output signals, especially at the edges or for very low frequencies, compared to previous versions.
  Fix: If migrating from versions prior to 0.2.2, re-evaluate output consistency, particularly for edge artifacts or filter characteristics, and adjust parameters like `rolloff` or `zeros` if necessary.

Install

• pip install julius Install stable version

Imports

• julius

  import julius

• resample_frac

  import julius
  signal = torch.randn(1000)
  resampled_signal = julius.resample_frac(signal, old_sr=44100, new_sr=16000)

  While 'resample_frac' exists in 'julius.resample', the top-level 'julius.resample_frac' is the commonly documented and used entry point for the functional API.
• ResampleFrac

  from julius.resample import ResampleFrac
  resampler = ResampleFrac(old_sr=44100, new_sr=16000)

  Used for class-based resampling, e.g., in torch.nn.Module.
• fft_conv1d

  import julius
  x = torch.randn(1, 1, 1024)
  w = torch.randn(1, 1, 256)
  y = julius.fftconv.fft_conv1d(x, w)

  Functional API for FFT-based 1D convolution.

Quickstart

This quickstart demonstrates the core resampling functionality using `julius.resample_frac` and a basic FFT-based convolution with `julius.fftconv.fft_conv1d`. The resampling function is designed for efficiency when sample rates form a fraction with a small numerator and denominator after GCD reduction.

import julius
import torch

# Create a dummy audio signal (batch_size, channels, time)
signal = torch.randn(2, 1, 44100) # 2 batches, 1 channel, 44100 samples (1 second at 44.1kHz)

old_sample_rate = 44100
new_sample_rate = 16000

# Resample the signal
resampled_signal = julius.resample_frac(signal, old_sr=old_sample_rate, new_sr=new_sample_rate)

print(f"Original signal shape: {signal.shape}")
print(f"Resampled signal shape: {resampled_signal.shape}")
# For FFT-based convolution:
x = torch.randn(1, 1, 2048) # Input (batch, channels, time)
w = torch.randn(1, 1, 512) # Kernel (out_channels, in_channels, kernel_size)
y = julius.fftconv.fft_conv1d(x, w)
print(f"FFT Conv output shape: {y.shape}")