Although, Python and PyTorch can be installed
directly from the R console, before start running
rTorch
, I would recommend testing PyTorch
first in a new Python or Anaconda environment. Check if PyTorch and
Torchvision packages are imported alright. The advantage of doing it
this way is that you could have already in advance a PyTorch version
that you are sure is working alright.
If you prefer to install PyTorch and its dependencies, that can be
done from rTorch through one its functions
install_pytorch()
.
This function is public and can be invoked with
rTorch::install_pytorch()
.
This function will allow you to indicate (i) the Python version; (ii)
the PyTorch version; (iii) the name of the conda environment; (iv) which
channel (stable
or nightly
); (v) if you
require CUDA (GPU) computation; (vi) additional packages such as
matplotlib
, pandas
; (vii) more.
install_pytorch(
method = c("conda", "virtualenv", "auto"),
conda = "auto",
version = "default",
envname = "r-torch",
extra_packages = NULL,
restart_session = TRUE,
conda_python_version = "3.6",
pip = FALSE,
channel = "stable",
cuda_version = NULL,
dry_run = FALSE,
...
)
If you prefer do it manually, use this example:
Create a conda environment with
conda create -n my-torch python=3.7 -y
Activate the new environment with
conda activate my-torch
Inside the new environment, install PyTorch and related packages with:
conda install python=3.6 pytorch torchvision matplotlib pandas -c pytorch
Note: If you you don’t specify a version,
conda
will install the latest PyTorch. As of this writing (August-September 2020), the latest PyTorch version is 1.6.
Alternatively, you could create and install a conda environment a specific PyTorch version with:
conda create -n my-torch python=3.6 pytorch=1.3 torchvision matplotlib pandas -c pytorch -y
conda
will resolve the dependencies and versions of the
other packages automatically, or let you know your options.
Note. matplotlib
and
pandas
are not really necessary, but I was asked if
matplotlib
or pandas
would work in PyTorch.
Then, I decided to put them for testing and experimentation. They both
work.
In rTorch there is an automatic detection of
Python built in in the package that will ask you to install
Miniconda
first if you don’t have any Python installed in
your machine. For instance, in macOS
, Miniconda will be
installed under
PREFIX=/Users/user_name/Library/r-miniconda
.
After Miniconda is installed, you could proceed to install the flavor or PyTorch you want, and the packages you want, with a command like this:
rTorch:::install_conda(package="pytorch=1.4", envname="r-torch", conda="auto", conda_python_version = "3.6", pip=FALSE, channel="pytorch", extra_packages=c("torchvision", "cpuonly", "matplotlib", "pandas"))
The command above will install the stable
PyTorch 1.4 version on Python 3.6,
including three additional packages: torchvision
,
cpuonly
, matplotlib
and
pandas.
NOTE. My experience with
Miniconda
is spotty and not 100% reliable, specially in macOS. I would strongly recommend using full conda for your PyTorch installation.