1

Currently, I am working on my thesis which is built on LSTM networks and I am using PyTorch library. However I am struggling to solve the conceptual problem of archiving trained models.

To make the question more clear; I am saving models in a archicture that I can give this form /models/model-with-loss-2.634221 as an example. But with this form, it is hard to determine which is which. I tried use more detailed form like 1-layered-100-epoch-128-batchsize-...-etc, but it is also hard to read and determine.

What is your way that you think is most productive to handle such operation?

By the way I am not sure this is the correct network ask this question on, you can drop an comment if it is not.

Brian Spiering
  • 23,131
  • 2
  • 29
  • 113
Faruk
  • 155
  • 1
  • 7

1 Answers1

1

One option is to give each model a unique identifier (e.g., a hash value or nickname). Then store all the metadata in another file.

Another option is using the PyTorch torch-model-archiver feature.

Brian Spiering
  • 23,131
  • 2
  • 29
  • 113