xml2pytorch

Using xml to define pytorch neural networks

What can it Do

With xml2pytorch, you can easily define neural networks in xml, and then declare them in pytorch.

RNN and LSTM are not supported currently.

Installation

Environment

OS independent. Python3. (Not tested on Python2, but it should work.)

Install requirements

torch>=0.4.1 numpy>=1.15.1

Installing by pip3

pip3 install xml2pytorch

Quick Start

How to declare the CNN defined by a xml file

import torch
import xml2pytorch as xm

# declare the net defined in .xml
net = xm.convertXML(xml_filename)    

# input a random tensor
x = torch.randn(1, 3, 32, 32)
y = net(x)
print(y)

How to define a simple CNN in xml

<graph>
	<net>
		<layer>
			<net_style>Conv2d</net_style>
			<in_channels>3</in_channels>
			<out_channels>6</out_channels>
			<kernel_size>5</kernel_size>
		</layer>	
		<layer>
			<net_style>ELU</net_style>
		</layer>	
		<layer>
			<net_style>MaxPool2d</net_style>
			<kernel_size>2</kernel_size>
			<stride>2</stride>
			<activation>logsigmoid</activation>
		</layer>
		<layer>
			<net_style>Conv2d</net_style>
			<in_channels>6</in_channels>
			<out_channels>16</out_channels>
			<kernel_size>5</kernel_size>
			<activation>relu</activation>
		</layer>	
		<layer>
			<net_style>MaxPool2d</net_style>
			<kernel_size>2</kernel_size>
			<stride>2</stride>
			<activation>relu</activation>
		</layer>
		<layer>
			<net_style>reshape</net_style>
			<dimensions>[-1, 16*5*5]</dimensions>
		</layer>
		<layer>
			<net_style>Linear</net_style>
			<in_features>400</in_features> 
			<out_features>120</out_features>
			<activation>tanh</activation>
		</layer>
		<layer>
			<net_style>Linear</net_style>
			<in_features>120</in_features> 
			<out_features>84</out_features>
			<activation>sigmoid</activation>
		</layer>
		<layer>
			<net_style>Linear</net_style>
			<in_features>84</in_features>
			<out_features>10</out_features>
			<activation>softmax</activation>
		</layer>
	</net>
</graph>