0

I'd like to know, How many epochs you typically run for training mozilla deep speech, with its default architecture. (no changes from GitHub version). What would be the decent number of epochs to run, especially considering it takes quite a bit of time for each epoch on the infrastructure available?((Tesla K80 single GPU))

Also on the other hand, With early stopping (default criterion), what would be that number? Appreciate your suggestions/advice.

Training data : CV

Rajesh D
  • 311
  • 1
  • 2
  • 12

1 Answers1

1

A good rule of thumb is to let it run as long as the model learns, i.e. as long as the validation error is below the training error. You can monitor this at runtime since its printed out to Console or automate this with e.g. the parameter dropout_rate. If you set it to something like 0.1 it will stop when the validation error gets very close to the training error.

It of course also depends on other parameter settings and your data.

  • n_hidden and epoch are the main parameters that affect model accuracy
  • n_hidden and train_batch_size affect speed(if the n_hidden is smaller and train_batch_size is larger, the speed will be faster).
TinaW
  • 989
  • 1
  • 18
  • 28