Model File Format

A model describes a neural network. It is a tar file which contains some other files:

• model.yml: A YAML file which describes the neural network
• input_semantics.csv: A file which has exactly one line for every input neuron. It can be used to describe semantically what the inputs mean. If no semantics are wished, the line could contain input [LINE NR].
• output_semantics.csv: Similar to input semantics.

model.yml

One example for a model.yml is

type: mlp
layers:
- W:
    filename: W0.hdf5
    size:
    - 167
    - 500
  activation: sigmoid
  b:
    filename: b0.hdf5
    size:
    - 500
- W:
    filename: W1.hdf5
    size:
    - 500
    - 500
  activation: sigmoid
  b:
    filename: b1.hdf5
    size:
    - 500
- W:
    filename: W2.hdf5
    size:
    - 500
    - 369
  activation: softmax
  b:
    filename: b2.hdf5
    size:
    - 369

The content might be different for other model types. Each layer has a weight matrix W and a bias vector b. They are stored in HDF5 files.