GitLab

An interesting feature for the neural networks module would be the possibility to define a few hyperparameters in a .json file in order to create different model versions. For example, this settings file could define the number of convolutional layers to be used, the number of filters in each layer, the learning rate, etc.

Here is an example used in Google's seq2seq framework for tensorflow: https://github.com/google/seq2seq/blob/master/example_configs/nmt_small.yml