Serialization
In torch: Tensors and Neural Networks with 'GPU' Acceleration

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  eval = identical(Sys.getenv("TORCH_TEST", unset = "0"), "1"),
  purl = FALSE
)

library(torch)

Torch tensors in R are pointers to Tensors allocated by LibTorch. This has one major consequence for serialization. One cannot simply use saveRDS for serializing tensors, as you would save the pointer but not the data itself. When reloading a tensor saved with saveRDS the pointer might have been deleted in LibTorch and you would get wrong results.

To solve this problem, torch implements specialized functions for serializing tensors to the disk:

torch_save(): to save tensors and models to the disk.
torch_load(): to load the models or tensors back to the session.

Please note that this format is still experimental and you shouldn't use it for long term storage.

Saving tensors

You can save any object of type torch_tensor to the disk using:

x <- torch_randn(10, 10)
torch_save(x, "tensor.pt")
x_ <- torch_load("tensor.pt")

torch_allclose(x, x_)

Saving modules

The torch_save and torch_load functions also work for nn_modules objects.

When saving an nn_module, all the object is serialized including the model structure and it's state.

module <- nn_module(
  "my_module",
  initialize = function() {
    self$fc1 <- nn_linear(10, 10)
    self$fc2 <- nn_linear(10, 1)
  },
  forward = function(x) {
    x %>% 
      self$fc1() %>% 
      self$fc2()
  }
)

model <- module()
torch_save(model, "model.pt")
model_ <- torch_load("model.pt")

# input tensor
x <- torch_randn(50, 10)
torch_allclose(model(x), model_(x))

Loading models saved in python

Currently the only way to load models from python is to rewrite the model architecture in R. All the parameter names must be identical.

You can then save the PyTorch model state_dict using:

torch.save(model, fpath, _use_new_zipfile_serialization=True)

You can then reload the state dict in R and reload it into the model with:

state_dict <- load_state_dict(fpath)
model <- Model()
model$load_state_dict(state_dict)

You can find working examples in torchvision. For example this is what we do for the AlexNet model.

Saving optimizer state

You can save the state of optimizers so you can continue training from the exact same position.

In order to this we use the state_dict() and load_state_dict() methods from the optimizer combined with torch_save:

model <- nn_linear(1, 1)
opt <- optim_adam(model$parameters)

train_x <- torch_randn(100, 1)
train_y <- torch_randn(100, 1)

loss <- nnf_mse_loss(model(train_x), train_y)
loss$backward()
opt$step()

# Now let's save the optimizer state
tmp <- tempfile()
torch_save(opt$state_dict(), tmp)

# And now let's create a new optimizer and load back
opt2 <- optim_adam(model$parameters)
opt2$load_state_dict(torch_load(tmp))