torch 0.10.0

torch v0.10.0 is now on CRAN. This version upgraded the underlying LibTorch to 1.13.1, and added support for Automatic Mixed Precision. As an experimental feature, we now also support pre-built binaries, so you can install torch without having to deal with the CUDA installation.

We are happy to announce that torch v0.10.0 is now on CRAN. In this blog post we highlight some of the changes that have been introduced in this version. You can check the full changelog here.

Automatic Mixed Precision

Automatic Mixed Precision (AMP) is a technique that enables faster training of deep learning models, while maintaining model accuracy by using a combination of single-precision (FP32) and half-precision (FP16) floating-point formats.

In order to use automatic mixed precision with torch, you will need to use the with_autocast context switcher to allow torch to use different implementations of operations that can run with half-precision. In general it’s also recommended to scale the loss function in order to preserve small gradients, as they get closer to zero in half-precision.

Here’s a minimal example, ommiting the data generation process. You can find more information in the amp article.

...
loss_fn <- nn_mse_loss()$cuda()
net <- make_model(in_size, out_size, num_layers)
opt <- optim_sgd(net$parameters, lr=0.1)
scaler <- cuda_amp_grad_scaler()

for (epoch in seq_len(epochs)) {
  for (i in seq_along(data)) {
    with_autocast(device_type = "cuda", {
      output <- net(data[[i]])
      loss <- loss_fn(output, targets[[i]])  
    })
    
    scaler$scale(loss)$backward()
    scaler$step(opt)
    scaler$update()
    opt$zero_grad()
  }
}

In this example, using mixed precision led to a speedup of around 40%. This speedup is even bigger if you are just running inference, i.e., don’t need to scale the loss.

Pre-built binaries

With pre-built binaries, installing torch gets a lot easier and faster, especially if you are on Linux and use the CUDA-enabled builds. The pre-built binaries include LibLantern and LibTorch, both external dependencies necessary to run torch. Additionally, if you install the CUDA-enabled builds, the CUDA and cuDNN libraries are already included..

To install the pre-built binaries, you can use:

options(timeout = 600) # increasing timeout is recommended since we will be downloading a 2GB file.
kind <- "cu117" # "cpu", "cu117" are the only currently supported.
version <- "0.10.0"
options(repos = c(
  torch = sprintf("https://storage.googleapis.com/torch-lantern-builds/packages/%s/%s/", kind, version),
  CRAN = "https://cloud.r-project.org" # or any other from which you want to install the other R dependencies.
))
install.packages("torch")

As a nice example, you can get up and running with a GPU on Google Colaboratory in less than 3 minutes!

Colaboratory running torch

Speedups

Thanks to an issue opened by @egillax, we could find and fix a bug that caused torch functions returning a list of tensors to be very slow. The function in case was torch_split().

This issue has been fixed in v0.10.0, and relying on this behavior should be much faster now. Here’s a minimal benchmark comparing both v0.9.1 with v0.10.0:

bench::mark(
  torch::torch_split(1:100000, split_size = 10)
)

With v0.9.1 we get:

# A tibble: 1 × 13
  expression      min  median `itr/sec` mem_alloc `gc/sec` n_itr  n_gc total_time
  <bch:expr> <bch:tm> <bch:t>     <dbl> <bch:byt>    <dbl> <int> <dbl>   <bch:tm>
1 x             322ms   350ms      2.85     397MB     24.3     2    17      701ms
# ℹ 4 more variables: result <list>, memory <list>, time <list>, gc <list>

while with v0.10.0:

# A tibble: 1 × 13
  expression      min  median `itr/sec` mem_alloc `gc/sec` n_itr  n_gc total_time
  <bch:expr> <bch:tm> <bch:t>     <dbl> <bch:byt>    <dbl> <int> <dbl>   <bch:tm>
1 x              12ms  12.8ms      65.7     120MB     8.96    22     3      335ms
# ℹ 4 more variables: result <list>, memory <list>, time <list>, gc <list>

Build system refactoring

The torch R package depends on LibLantern, a C interface to LibTorch. Lantern is part of the torch repository, but until v0.9.1 one would need to build LibLantern in a separate step before building the R package itself.

This approach had several downsides, including:

• Installing the package from GitHub was not reliable/reproducible, as you would depend on a transient pre-built binary.
• Common devtools workflows like devtools::load_all() wouldn’t work, if the user didn’t build Lantern before, which made it harder to contribute to torch.

From now on, building LibLantern is part of the R package-building workflow, and can be enabled by setting the BUILD_LANTERN=1 environment variable. It’s not enabled by default, because building Lantern requires cmake and other tools (specially if building the with GPU support), and using the pre-built binaries is preferable in those cases. With this environment variable set, users can run devtools::load_all() to locally build and test torch.

This flag can also be used when installing torch dev versions from GitHub. If it’s set to 1, Lantern will be built from source instead of installing the pre-built binaries, which should lead to better reproducibility with development versions.

Also, as part of these changes, we have improved the torch automatic installation process. It now has improved error messages to help debugging issues related to the installation. It’s also easier to customize using environment variables, see help(install_torch) for more information.

Final remarks

Thank you to all contributors to the torch ecosystem. This work would not be possible without all the helpful issues opened, PRs you created and your hard work.

If you are new to torch and want to learn more, we highly recommend the recently announced book ‘Deep Learning and Scientific Computing with R torch’.

If you want to start contributing to torch, feel free to reach out on GitHub and see our contributing guide.

The full changelog for this release can be found here.