SigPy

https://travis-ci.com/mikgroup/sigpy.svg?branch=master

SigPy is a package for signal processing, with emphasis on iterative methods. It is built to operate directly on numpy arrays on CPU and cupy arrays on GPU. SigPy also provides several domain-specific submodules: sigpy.plot for multi-dimensional array plotting, sigpy.mri for MRI iterative reconstruction, and sigpy.learn for dictionary learning.

Installation

SigPy requires Python version >= 3.5. The core module depends on numba, numpy, PyWavelets, scipy, and tqdm.

Additional features can be unlocked by installing the appropriate packages. To enable the plotting functions, you will need to install matplotlib. To enable CUDA support, you will need to install cupy. And to enable MPI support, you will need to install mpi4py.

Via `conda`

For general users, we recommend installing SigPy through conda:

conda install -c frankong sigpy
# (optional for plot support) conda install matplotlib
# (optional for CUDA support) conda install cupy
# (optional for MPI support) conda install mpi4py

Via `pip`

SigPy can also be installed through pip:

pip install sigpy
# (optional for plot support) pip install matplotlib
# (optional for CUDA support) pip install cupy
# (optional for MPI support) pip install mpi4py

Installation for Developers

If you want to contribute to the SigPy source code, we recommend you install it with pip in editable mode:

cd /path/to/sigpy
pip install -e .

To run tests and contribute, we recommend installing the following packages:

pip install coverage flake8 sphinx

and run the script run_tests.sh.

Features

CPU/GPU Signal Processing Functions

SigPy provides signal processing functions with a unified CPU/GPU interface. For example, the same code can perform a CPU or GPU convolution on the input array device:

# CPU convolve
x = numpy.array([1, 2, 3, 4, 5])
y = numpy.array([1, 1, 1])
z = sigpy.convolve(x, y)

# GPU convolve
x = cupy.array([1, 2, 3, 4, 5])
y = cupy.array([1, 1, 1])
z = sigpy.convolve(x, y)

In addition, the function interfaces are geared towards signal processing applications. For example, sigpy.convolve supports mode={'valid', 'full'} options and actually performs convolution instead of cross-correlation like in all machine learning packages.

Iterative Algorithms

SigPy also provides convenient abstractions and classes for iterative algorithms. A compressed sensing experiment can be implemented in four lines using SigPy:

# Given some observation vector y, and measurement matrix mat
A = sigpy.linop.MatMul([n, 1], mat)  # define forward linear operator
proxg = sigpy.prox.L1Reg([n, 1], lamda=0.001)  # define proximal operator
x_hat = sigpy.app.LinearLeastSquares(A, y, proxg=proxg).run()  # run iterative algorithm

Users can easily define their own linear operator Linop and proximal operator Prox. Different iterative algorithm Alg can be selected and extended as well.

PyTorch Interoperability

Want to do machine learning without giving up signal processing? SigPy has convenient functions to convert arrays and linear operators into PyTorch Tensors and Functions. For example, given a cupy array x, and a Linop A, we can convert them to Pytorch:

x_torch = sigpy.to_pytorch(x)
A_torch = sigpy.to_pytorch_function(A)

The conversion has no copying, and the resulting Tensor and Function can be backpropagated. Users can easily mix wavelet transforms and Fourier transforms with neural networks.

jschoormans/sigpy