skeydan / convlstm Goto Github PK

I've been trying to use your convlstm into a real use case and for that I'm still suffering a bit with the tuning. I wanted to properly define a learning rate first, and this can be leading me to some issues. I'm always having Inf values for the loss after a while.

This way, I wanted to use luz package to define a learning rate, which has been simply impossible, because I think the returning value of the forward method of the convlstm is a list.

Do you mind having a look and perhaps coming up with any ideas?

input_dim = 1
hidden_dims = c(64, 1)
kernel_sizes = c(3, 3)
n_layers = 2
                           
 model <- convlstm %>%
    setup(
      loss = nn_mse_loss(),
      optimizer = torch::optim_adam,
      ) %>%
    set_hparams(n_layers = n_layers,
                input_dim = input_dim,
                hidden_dims = hidden_dims,
                kernel_sizes = kernel_sizes)

  rates_and_losses <- model %>% lr_finder(
    train_dl,
    # start_lr = 1e-4,
    # end_lr = 0.5,
    verbose = TRUE
  )

The error I get is always

Error in input$size() : attempt to apply non-function

My train_dl is a bit complex to reproduce, but you'll find the same error using your dummy dataset from the blogpost https://blogs.rstudio.com/ai/posts/2020-12-17-torch-convlstm/

Thanks a lot on advance!
Huriel Reichel