Model Reference

The model must be specified via a .json file. This file is validated via a jsonschema.

The model definition must consist of a list of modules. Each module has an attribute “module_name”, which can be any string, and an array of layers.

Each layer must have a type (Conv2d, Linear, etc.) and a unique name. Depending on the type, the layer has certain required and optional parameters. The possible layer types and their arguments are described in the following, while required parameters are bold.

Layer reference

Conv2d

Does a 2d-image convolution. Expects 3d-input and returns 3d-output.

Parameters:

Name	Type	Description
out_channels	int	#channels in the output image
in_channels	int	#channels in the input image
kernel_size	string	tuple (int,int) describing kernel height and width
stride	int	stride of the convolution, default=1
padding	string	padding, either same or valid, default=valid

Conv2dTranspose

Does a 2d-image transposed convolution (also known as deconvolution), used to increase the image size in upsampling tasks.

Parameters: see Conv2d

Linear

A layer of linear perceptrons.

Parameters:

Name	Type	Description
in_features	int	# of input features
out_features	int	# of output features

Flatten

Flattens the input into 1d.

Parameters: none

ReLU

ReLU activation layer.

Parameters: none