IIRNet

Direct design of biquad filter cascades with deep learning by sampling random polynomials.

Click here to watch the paper video explanation.

Citation

If you use any of our code in your work please consider citing us.

  @inproceedings{colonel2021iirnet,
    title={Direct design of biquad filter cascades with deep learning by sampling random polynomials},
    author={Colonel, Joseph and Steinmetz, Christian J. and Michelen, Marcus and Reiss, Joshua D.},
    booktitle={ICASSP},
    year={2022}}

Usage

git clone https://github.com/csteinmetz1/IIRNet.git
cd IIRNet
pip install -e .

Filter design

Start designing filters with just a few lines of code. In this example (demos/basic.py ) we create a 32nd order IIR filter to match an arbitrary response that we define over a few points. Internally, this specification will be interpolated to 512 points.

import torch
import numpy as np
import scipy.signal
import matplotlib.pyplot as plt
from iirnet.designer import Designer

# first load IIRNet with pre-trained weights
designer = Designer()

n = 32  # Desired filter order (4, 8, 16, 32, 64)
m = [0, -3, 0, 12, 0, -6, 0]  # Magnitude response specification
mode = "linear"  # interpolation mode for specification
output = "sos"  # Output type ("sos", or "ba")

# now call the designer with parameters
sos = designer(n, m, mode=mode, output=output)

# measure and plot the response
w, h = scipy.signal.sosfreqz(sos.numpy(), fs=2)

# interpolate the target for plotting
m_int = torch.tensor(m).view(1, 1, -1).float()
m_int = torch.nn.functional.interpolate(m_int, 512, mode=mode)

fig, ax = plt.subplots(figsize=(6, 3))
plt.plot(w, 20 * np.log10(np.abs(h)), label="Estimation")
plt.plot(w, m_int.view(-1), label="Specification")
# .... more plotting ....

See demos/basic.py for the full script.

Training

We provide a set of shell scripts that will launch training jobs that reproduce the experiments from the paper in configs/. These should be launched from the top level after installing.

./configs/train_hidden_dim.sh
./configs/filter_method.sh
./configs/filter_order.sh

Evaluation

Running the evaluation will require both the pre-trained models (or models you trained yourself) along with the HRTF and Guitar cabinet datasets. These datasets can be downloaded as follows:

First, change to the data directory and then run the download script.

cd data
./dl.sh

Note, you may need to install 7z if you don't already have it. brew install p7zip on macOS

Next download the pre-trained checkpoints if you haven't already.

mkdir logs
cd logs 
wget https://zenodo.org/record/5550275/files/filter_method.zip
wget https://zenodo.org/record/5550275/files/filter_order.zip
wget https://zenodo.org/record/5550275/files/hidden_dim.zip

unzip filter_method.zip
unzip filter_order.zip
unzip hidden_dim.zip

rm filter_method.zip
rm filter_order.zip
rm hidden_dim.zip

Now you can run the evaluation on checkpoints from the three different experiments as follows.

python eval.py logs/filter_method --yw --sgd --guitar_cab --hrtf --filter_order 16
python eval.py logs/hidden_dim --yw --sgd --guitar_cab --hrtf --filter_order 16

For the filter order experiment we need to run the eval script across all models for every order.

python eval.py logs/filter_order --guitar_cab --hrtf --filter_order 4
python eval.py logs/filter_order --guitar_cab --hrtf --filter_order 8
python eval.py logs/filter_order --guitar_cab --hrtf --filter_order 16
python eval.py logs/filter_order --guitar_cab --hrtf --filter_order 32
python eval.py logs/filter_order --guitar_cab --hrtf --filter_order 64

Note: Requires PyTorch >=1.8

Filter methods

ID	Sampling method	Name
(A)	Normal coefficients	normal_poly
(B)	Normal biquads	normal_biquad
(C)	Uniform disk	uniform_disk
(D)	Uniform magnitude disk	uniform_mag_disk
(E)	Characteristic	char_poly
(F)	Uniform parametric	uniform_parametric