Define Model Serialization Data

[ryanleary]

Serialized models must contain model metadata in addition to parameter weights.

Parameters

• Required: model parameter weights
• Optional: optimizer state

Model Architecture

• sample rate
• window size
• convolutional configuration
  • number of layers
  • stride
  • kernel size
  • activation
• rnn configuration
  • number of layers
  • rnn type
  • rnn size
  • activation
  • bidirectional

Training Configuration

• name/id of this model/experiment
• training set(s)/validation set(s) name/sha/identifier or
• labels
• batch size
• optimizer
  • optimizer configuration (lr / annealing / momentum etc)
• max norm
• noise/augmentation configuration
  • tbd

Training/Model Status

• epoch times
• WER/CER/Loss @ each epoch
• best epoch (shall match weights, but may not be current epoch)
• WER/CER of model

[ryanleary]

Done.