`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.

You can save any object of type `torch_tensor`

to the disk using:

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

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.

```
<- nn_module(
module "my_module",
initialize = function() {
$fc1 <- nn_linear(10, 10)
self$fc2 <- nn_linear(10, 1)
self
},forward = function(x) {
%>%
x $fc1() %>%
self$fc2()
self
}
)
<- module()
model torch_save(model, "model.pt")
<- torch_load("model.pt")
model_
# input tensor
<- torch_randn(50, 10)
x torch_allclose(model(x), model_(x))
```

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:

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

You can find working examples in `torchvision`

. For example this is what we do for the AlexNet model.

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`

:

```
<- nn_linear(1, 1)
model <- optim_adam(model$parameters)
opt
<- torch_randn(100, 1)
train_x <- torch_randn(100, 1)
train_y
<- nnf_mse_loss(model(train_x), train_y)
loss $backward()
loss$step()
opt
# Now let's save the optimizer state
<- tempfile()
tmp torch_save(opt$state_dict(), tmp)
# And now let's create a new optimizer and load back
<- optim_adam(model$parameters)
opt2 $load_state_dict(torch_load(tmp)) opt2
```