Neural Network Architecture Editor With Code Generation

This paper discusses the particular properties of modern architectures of neural networks, that lead to complication in their design and implementation. A neural network is a machine learning algorithm, that can be represented as a fixed computational graph with learnable weights on its edges. Modern neural networks have 2 main distinct features – their computation graphs are non-linear and they often consist of repeated blocks. Representing these structures in a programming language is difficult and prone to errors, so a solution to simplify this process was proposed – to develop a graphical editor, which would allow to create, edit, share and generate code for a given Deep Learning framework. An interactive visual representation of the network’s architecture as a 2-D graph should greatly simplify the process of designing a neural network. A description of the developed software, its features and internal implementation is given, a comparison with alternative solutions is provided as well.