Saving/Loading tabnet models about tabnet HOT 12 CLOSED

It actually looks like it's still not working for me. Not sure what the deal is. The workaround I'm using is this serialization process:

# Serialization
write_rds(wflw_tabnet_fit, "models/wflw_tabnet_fit.rds")

torch::torch_save(wflw_tabnet_fit$fit$fit$fit$fit$network, "models/torch_network")

# Loading
torch_network <- torch::torch_load("models/torch_network")

wflw_fit <- read_rds("models/wflw_tabnet_fit.rds")

wflw_fit$fit$fit$fit$fit$network <- torch_network

wflw_fit

from tabnet.

you can save/load the underlying parnsip model - see https://blogs.rstudio.com/ai/posts/2021-02-11-tabnet/

fitted_model <- wf %>% fit(train)

# access the underlying parsnip model and save it to RDS format
# depending on when you read this, a nice wrapper may exist
# see https://github.com/mlverse/tabnet/issues/27  
fitted_model$fit$fit$fit %>% saveRDS("saved_model.rds")

from tabnet.

Ok, for me it's giving an Error in cpp_Tensor_storage() problem when I try to load the model using read_rds().

write_rds(wflw_tabnet_fit$fit$fit$fit, "models/tabnet_fit.rds")
read_rds("models/tabnet_fit.rds")
# Error in cpp_Tensor_storage(self$ptr) : external pointer is not valid

from tabnet.

Typically there's a serialization function that allows us to read the underlying model. That's what I normally have to do with TensorFlow and MXNet.

from tabnet.

Ok, I've got it working. I had to save the "nn_module" using torch::torch_save(), which is in the 'network' part of the fitted workflow. This could probably be wrapped into a save/load tabnet function.

# Serialization
write_rds(wflw_tabnet_fit, "models/wflw_tabnet_fit.rds")

torch::torch_save(wflw_tabnet_fit$fit$fit$fit$fit$network, "models/torch_network")

# Loading
torch_network <- torch::torch_load("models/torch_network")

wflw_fit <- read_rds("models/wflw_tabnet_fit.rds")

wflw_fit$fit$fit$fit$fit$network <- torch_network

wflw_fit

from tabnet.

In theory, saveRDS should just work. The model object contains a version of the serialized weights, so we reload it as needed. We test this here:

There could be a bug depending if you are using CRAN or dev versions of torch.

from tabnet.

Ok, so the problem was that I was not using the development versions of the torch and tabnet packages. I can save workflows that contain fitted tabnet models with readr::write_rds(). Thanks for your help!!

Solution:

# GITHUB INSTALLATIONS ----
# TORCH: 
remotes::install_github("mlverse/torch")
# TABNET: 
remotes::install_github("mlverse/tabnet")

from tabnet.

OK! I'll investigate, I think I might have broken this with recent changes in torch.

from tabnet.

Ok, for what it's worth, I'm getting pretty good results with tabnet. I'm doing some tests now, and I'll be showcasing this project during my Learning Lab 51 - Torch for Tabular Data. The saving/loading is a small inconvenience.

from tabnet.

I could not reproduce this issue any longer with dev versions... The test should be more robust with: aa1f85f
Let me know if this persists.

from tabnet.

@mdancho84 You might be using readr::save_rds instead of base::saveRDS, I can confirm base::saveRDS and base::readRDS works.

from tabnet.

Ok thank you!

from tabnet.