ComplexCNN

pytorch implementation of complex convolutional neural network Reference: https://arxiv.org/pdf/1705.09792.pdf

For whom?

When using complex numbers as a domain of a neural network (such as speech enhancement) deep complex networks can be very effective.
Phase-Aware Speech Enhancement with Deep Complex U-Net is a great example. Use this as a building block of complex number targeted architecture.

Usage

0. Install Package

pip install complexcnn

1. Preprocess Input

# Suppose X is a complex vector shape of [batch,channel,axis1,axis2]
X = np.stack((X.real,X.imag),axis=1) # shape: [batch,2,channel,axis1,axis2]
X = torch.Tensor(X).to(device)

2. ComplexConv Module

Parameters

Same as Pytorch Conv2d Parameters

• in_channel (required)
• out_channel (required)
• kernel_size (required)
• stride (default: 1)
• padding (default: 0)
• dilation (default: 1)
• groups (default: 1)
• bias (default: True)

Input

• Tensor (Type: torch.Tensor) shape: (batchsize, 2, input channel, axis1, axis2)

Output

• Tensor (Type: torch.Tensor) shape: (batchsize, 2, output channel, axis1, axis2)

from complexcnn.modules import ComplexConv

## Parameters Below are totally random
input_channel = 3
output_channel = 24
kernel_size = (5,5)

complex_conv = ComplexConv(input_channel, output_channel, kernel_size)

Y = complex_conv(X)