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

Last update: Nov 02, 2022

Related tags

Deep Learning IIRNet

Overview

IIRNet

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

Usage

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

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`

Citation

 @article{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={arXiv:2110.03691},
    year={2021}}

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

Related tags

Overview

IIRNet

Usage

Filter design

Training

Evaluation

Filter methods

Citation

Owner

Christian J. Steinmetz

Exploring the link between uncertainty estimates obtained via "exact" Bayesian inference and out-of-distribution (OOD) detection.

Repo for CReST: A Class-Rebalancing Self-Training Framework for Imbalanced Semi-Supervised Learning

Implementation of Retrieval-Augmented Denoising Diffusion Probabilistic Models in Pytorch

ICML 21 - Voice2Series: Reprogramming Acoustic Models for Time Series Classification

Rule-based Customer Segmentation

HODEmu, is both an executable and a python library that is based on Ragagnin 2021 in prep.

Code for Talking Face Generation by Adversarially Disentangled Audio-Visual Representation (AAAI 2019)

RaceBERT -- A transformer based model to predict race and ethnicty from names

The code uses SegFormer for Semantic Segmentation on Drone Dataset.

Implementing a simplified copy of Shazam application from scratch using MinHashing and LSH.

J.A.R.V.I.S is an AI virtual assistant made in python.

Neural-net-from-scratch - A simple Neural Network from scratch in Python using the Pymathrix library

Consensus Learning from Heterogeneous Objectives for One-Class Collaborative Filtering

A collection of resources on GAN Inversion.

CDTrans: Cross-domain Transformer for Unsupervised Domain Adaptation

Code reproduce for paper "Vehicle Re-identification with Viewpoint-aware Metric Learning"

An Efficient Training Approach for Very Large Scale Face Recognition or F²C for simplicity.

PyTorch Code for the paper "VSE++: Improving Visual-Semantic Embeddings with Hard Negatives"

SegTransVAE: Hybrid CNN - Transformer with Regularization for medical image segmentation

A Moonraker plug-in for real-time compensation of frame thermal expansion