Repository of General-purpose Library of ML/AI Methods.

The important part of successful neural network solution is a choice of an appropriate architecture. We provide tools automating this process, including semi-manual tools and fully automatic search. The semi-manual solution provides tools for automatic evaluation of a set of user defined architectures. The fully automatic search is based on evolutionary optimisation that finds a suitable network for a given problem.

Generative adversial networks (GANs) are used to expand available database of disc photographs. Different loss function-based architectures such as DCGAN and LSGAN are employed. Both unconditional and conditional configurations are available. The scripts also have distributed versions that can run on a GPU cluster.

D. Coufal, F. Hakl, P. Vidnerová. The Czech Academy of Sciences, Institute of Computer Science

deep neural networks, generative adverisal networks, conditional generation, generative algorithms, neural architecture search, model selection, evolutionary algorithms, multiobjective optimisation

|__ NAS  (Neural Architecture Search tools) 
|     |
|     |__ semi_manual   
|     |            |__ data         (data preprocessing for pytorch) 
|     |            |
|     |            |__ net_search   (scripts for automatic network evaluation)
|     |            |
|     |            |__ examples     (examples of config files)
|     |
|     |__ automatic 
|               |__ auto_nas    (scripts for automatic architecture search)
|               |
|               |__ examples  (examples of config files)
|                
|__ GANs
      |__ dcgan    (Deep Convolutional GAN) 
      |
      |__ lsgan    (Least Squares GAN)

See the individual subdirectories for details on the individual parts and corresponding user instructions.

This work was partially supported by the TAČR grant TN01000024 and institutional support of the Institute of Computer Science RVO 67985807.