cv-net-library 0.0.2

A library of complex-valued neural networks

cv_net_library

---

The idea of this library is just to implement Complex layers () so that everything else stays the same as any PyTorch code.

Installation

Using [PIP]

Only use complex-valued neural networks library: pip install cv-net-library==0.0.2

Using GitHub

Useful if you want to modify the source code and view the relevant tests: address

---

To Use

Import the cvnn base class. then Import related classes or functions from the corresponding module.

The functions in cv_net has the same function as the corresponding function in Pytorch, and the prefix Complex is added before the original function. as the following modules.

layer

• ComplexLayers

  ComplexConv2d ComplexConv1d ComplexConv3d ComplexFlatten ComplexConvTransposed2d ComplexLinear

• ComplexDropout

  ComplexDropout2D ComplexDropout ComplexDropoutRespectively

• ComplexPooling

  ComplexAvgPool1D ComplexAvgPool2D ComplexAvgPool3D ComplexPolarAvgPooling2D ComplexMaxPool2D ComplexUnPooling2D

• ComplexUpSampooling

  ComplexUpSampling ComplexUpSamplingBilinear2d ComplexUpSamplingNearest2d

function

• ComplexBatchNorm

  ComplexBatchNorm ComplexBatchNorm1d ComplexBatchNorm2d

activation

• ComplexActivations

  complex_relu complex_elu complex_exponential complex_sigmoid complex_tanh complex_hard_sigmoid complex_leaky_relu complex_selu complex_softplus complex_softsign complex_softmax modrelu zrelu complex_cardioid sigmoid_real softmax_real_with_abs softmax_real_with_avg softmax_real_with_mult softmax_of_softmax_real_with_mult softmax_of_softmax_real_with_avg softmax_real_by_parameter softmax_real_with_polar georgiou_cdbp complex_signum mvn_activation apply_pol pol_tanh pol_sigmoid pol_selu

loss

• ComplexLoss

  ComplexAverageCrossEntropy ComplexAverageCrossEntropyAbs ComplexMeanSquareError ComplexAverageCrossEntropyIgnoreUnlabeled ComplexWeightedAverageCrossEntropy ComplexWeightedAverageCrossEntropyIgnoreUnlabeled

Example

# Make a A-ConvNets 
import torch.nn as nn
import cv_net_library
from cv_net_library.activation.ComplexActivation import complex_relu, complex_softmax
from cv_net_library.layer.ComplexLayers import ComplexLinear, ComplexConv2d
from cv_net_library.layer.ComplexDropout import ComplexDropout2D
from cv_net_library.layer.ComplexPooling import ComplexMaxPool2D
from cv_net_library.layer.ComplexUpSampling import ComplexUpSamplingBilinear2d
from cv_net_library.layer.ComplexLayers import ComplexLinear, ComplexConv2d
from cv_net_library.function.ComplexBatchNorm import ComplexBatchNorm2d, ComplexBatchNorm1d
from cv_net_library.loss.ComplexLoss import ComplexAverageCrossEntropy, ComplexAverageCrossEntropyAbs

class ComplexNet(nn.Module):

    def __init__(self):
        super(ComplexNet, self).__init__()
        self.conv1 = ComplexConv2d(1, 16, 13, 1)
        self.bn2d1 = ComplexBatchNorm2d(16, track_running_stats=False)
        self.maxpool1 = ComplexMaxPool2D(2, 2)
        self.conv2 = ComplexConv2d(16, 32, 13, 1)
        self.bn2d2 = ComplexBatchNorm2d(32, track_running_stats=False)
        self.maxpool2 = ComplexMaxPool2D(2, 2)
        self.conv3 = ComplexConv2d(32, 64, 12, 1)
        self.bn2d3 = ComplexBatchNorm2d(64, track_running_stats=False)
        self.maxpool3 = ComplexMaxPool2D(2, 2)
        self.dropout1 = ComplexDropout2D(p=0.5)
        self.conv4 = ComplexConv2d(64, 128, 10, 1)
        self.bn2d4 = ComplexBatchNorm2d(128, track_running_stats=False)
        self.conv5 = ComplexConv2d(128, 7, 6, 1)
        self.bn2d5 = ComplexBatchNorm2d(7, track_running_stats=False)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn2d1(x)
        x = complex_relu(x)
        x = self.maxpool1(x)
        x = self.conv2(x)
        x = self.bn2d2(x)
        x = complex_relu(x)
        x = self.maxpool2(x)
        x = self.conv3(x)
        x = self.bn2d3(x)
        x = complex_relu(x)
        x = self.maxpool3(x)
        x = self.dropout1(x)
        x = self.conv4(x)
        x = self.bn2d4(x)
        x = complex_relu(x)
        x = self.conv5(x)
        x = self.bn2d5(x)
        x = x.view(x.shape[0], -1)
        x = complex_softmax(x, 1)
        return x

---

Update Log

0.0.2 fix bug, Added pip installation mode, verified the example.

0.0.1 : The original upload includes a variety of complex-valued neural networks modules.

---