Posit AI Weblog: torch 0.10.0

We’re pleased to announce that torch v0.10.0 is now on CRAN. On this weblog submit we
spotlight among the adjustments which were launched on this model. You’ll be able to
examine the total changelog here.

Automated Combined Precision

Automated Combined Precision (AMP) is a method that permits quicker coaching of deep studying fashions, whereas sustaining mannequin accuracy by utilizing a mixture of single-precision (FP32) and half-precision (FP16) floating-point codecs.

In an effort to use automated combined precision with torch, you’ll need to make use of the with_autocast
context switcher to permit torch to make use of completely different implementations of operations that may run
with half-precision. Normally it’s additionally really helpful to scale the loss perform in an effort to
protect small gradients, as they get nearer to zero in half-precision.

Right here’s a minimal instance, ommiting the info technology course of. You will discover extra data within the amp article.

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

for (epoch in seq_len(epochs)) {
  for (i in seq_along(knowledge)) {
    with_autocast(device_type = "cuda", {
      output <- web(knowledge[[i]])
      loss <- loss_fn(output, targets[[i]])  
    })
    
    scaler$scale(loss)$backward()
    scaler$step(choose)
    scaler$replace()
    choose$zero_grad()
  }
}

On this instance, utilizing combined precision led to a speedup of round 40%. This speedup is
even greater if you’re simply working inference, i.e., don’t have to scale the loss.

Pre-built binaries

With pre-built binaries, putting in torch will get rather a lot simpler and quicker, particularly if
you’re on Linux and use the CUDA-enabled builds. The pre-built binaries embrace
LibLantern and LibTorch, each exterior dependencies essential to run torch. Moreover,
when you set up the CUDA-enabled builds, the CUDA and
cuDNN libraries are already included..

To put in the pre-built binaries, you should 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")

issue opened by @egillax, we might discover and repair a bug that induced
torch capabilities returning an inventory of tensors to be very sluggish. The perform in case
was torch_split().

This situation has been mounted in v0.10.0, and counting on this habits ought to be a lot
quicker now. Right here’s a minimal benchmark evaluating each 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.

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.

recently announced guide ‘Deep Studying and Scientific Computing with R torch’.

If you wish to begin contributing to torch, be at liberty to succeed in out on GitHub and see our contributing guide.

The complete changelog for this launch will be discovered here.