How to just continue training

[GitHubGeniusOverlord]

Hello!

I scanned to all the documents, but could not find anything on this. There is the callback, luz_callback_resume_from_checkpoint, which allows to continue from one certain checkpoint. But what if I just want to continue training from the latest state? How can this be done?

Maybe we could add an example for this to the documentation. Thank You!

[dfalbel]

We currently don't have a nice way of doing it. That's something we would like to have, but it still didn't have a good consensus of what should be considered 'continuing training' for example:

• Should we also keep optimizers state, learning rate decay state? Or just the model weights.
• Should we concatenate metrics?
• What happens with mutability of the weights, should we copy them to a new object before starting training, so we don't modify them in place?

You can do something like this though, if you don't care about optimizer state. Otherwise the recommended way is to use a the resume from checkpoint callback and checkpoint the model state that you want to recover from.

1. Train a model:

library(luz)
library(torch)

model <- nn_module(
  initialize = function() {
    self$linear <- nn_linear(10, 1)
  },
  forward = function(x) {
    self$linear(x)
  }
)

fitted <- model %>% 
  setup(loss = nnf_mse_loss, optimizer = optim_adam) %>% 
  set_hparams() %>% 
  set_opt_hparams(lr = 0.01) %>% 
  fit(list(
    torch_randn(10, 10),
    torch_randn(100, 1)
  ))

2. Create a wrapper model and train it:

model2 <- nn_module(
  initialize = function(model) {
    self$model <- model
  },
  forward = function(x) {
    self$model(x)
  }
)

fitted <- model2 %>% 
  setup(loss = nnf_mse_loss, optimizer = optim_adam) %>% 
  set_hparams(model = fitted$model) %>% 
  set_opt_hparams(lr = 0.01) %>% 
  fit(list(
    torch_randn(10, 10),
    torch_randn(100, 1)
  ))

[GitHubGeniusOverlord]

I guess, this is outdated, as there are updates in luz on how to continue training. Closing therefore.