/pytorch_complex

A temporal module for PyTorch-ComplexTensor

Primary LanguagePython

pytorch_complex

Build Status codecov

A temporal python class for PyTorch-ComplexTensor

What is this?

A Python class to perform as ComplexTensor in PyTorch: Nothing except for the following,

class ComplexTensor: 
    def __init__(self, ...):
        self.real = torch.Tensor(...)
        self.imag = torch.Tensor(...)

Why?

PyTorch is great DNN Python library, except that it doesn't support ComplexTensor in Python level.

pytorch/pytorch#755

I'm looking forward to the completion, but I need ComplexTensor for now. I created this cheap module for the temporal replacement of it. Thus, I'll throw away this project as soon as ComplexTensor is completely supported!

Requirements

Python>=3.6
PyTorch>=1.0

Install

pip install git+https://github.com/kamo-naoyuki/pytorch_complex

How to use

Basic mathematical operation

import numpy as np
from torch_complex.tensor import ComplexTensor

real = np.random.randn(3, 10, 10)
imag = np.random.randn(3, 10, 10)

x = ComplexTensor(real, imag)
x.numpy()

x + x
x * x
x - x
x / x
x ** 1.5
x @ x  # Batch-matmul
x.conj()
x.inverse() # Batch-inverse

All are implemented with combinations of computation of RealTensor in python level, thus the speed is not good enough.

Functional

import torch_complex.functional as F
F.cat([x, x])
F.stack([x, x])
F.matmul(x, x)  # Same as x @ x
F.einsum('bij,bjk,bkl->bil', [x, x, x])

For DNN

Almost all methods that torch.Tensor has are implemented.

x.cuda()
x.cpu()
(x + x).sum().backward()