PyTorch

PyTorch is a Python package that provides two high-level features:

• Tensor computation (like NumPy) with strong GPU acceleration
• Deep neural networks built on a tape-based autograd system

PyTorch has a distant connection with Torch, but for all practical purposes you can treat them as separate packages.

Installation

Latest available wheels

To see the latest version of PyTorch that we have built:

[name@server ~]$ avail_wheels torch*

For more information on listing wheels, see listing available wheels.

Pre-build

The preferred option is to install it using the python wheel that we compile, as follows:

1. Load a python module, either python/2.7, python/3.5, or python/3.6

2. Create and start a virtual environment.

3. Install PyTorch in the virtual environment with pip install. For both GPU and CPU support:

(venv) [name@server ~] pip install numpy torch_gpu

If you only need CPU support:

(venv) [name@server ~] pip install numpy torch_cpu

Extra

In addition to torch_cpu or torch_gpu, you can install torchvision and torchtext:

(venv) [name@server ~] pip install numpy six torch_cpu torchvision torchtext

Anaconda

Attention: While Conda works well in a desktop environment, it tends to create more problems than it solves on a cluster. We encourage users to instead use virtual environments and the binary packages that we provide through Python wheels as documented on the Python page.

You can install PyTorch using Anaconda. First install Anaconda, and then install PyTorch in a conda environment as follows. Note that Anaconda has a tendency to update its versions often, which may result in the instructions below not working in the future.

1. Load the Miniconda 2 or Miniconda 3 module.

[name@server ~]$ module load miniconda3

2. Create a new conda virtual environment.

[name@server ~]$ conda create --name pytorch

3. When conda asks you to proceed, type y.

4. Activate the newly created conda virtual environment.

[name@server ~]$ source activate pytorch

5. Install PyTorch in the conda virtual environment.

[name@server ~]$ conda install pytorch torchvision cuda80  -c soumith

Here, we instruct conda to use the soumith channel to retrieve the packages from the release channel belonging to the main PyTorch developer, Soumith Chintala. This guarantees you will have the latest version.

Job submission

Once the setup is completed, you can submit a PyTorch job with

[name@server ~]$ sbatch pytorch-test.sh

Here is an example of a job submission script using the python wheel, with a virtual environment in $HOME/pytorch:

File : pytorch-test.sh

#!/bin/bash
#SBATCH --gres=gpu:1       # Request GPU "generic resources"
#SBATCH --cpus-per-task=6  # Cores proportional to GPUs: 6 on Cedar, 16 on Graham.
#SBATCH --mem=32000M       # Memory proportional to GPUs: 32000 Cedar, 64000 Graham.
#SBATCH --time=0-03:00
#SBATCH --output=%N-%j.out
module load python/3.6
source $HOME/pytorch/bin/activate
python ./pytorch-test.py

and here is an example of a job submission script using Anaconda:

File : pytorch-test.sh

#!/bin/bash
#SBATCH --gres=gpu:1       # Request GPU "generic resources"
#SBATCH --cpus-per-task=6  # Cores proportional to GPUs: 6 on Cedar, 16 on Graham.
#SBATCH --mem=32000M       # Memory proportional to GPUs: 32000 Cedar, 64000 Graham.
#SBATCH --time=0-03:00
#SBATCH --output=%N-%j.out
module load miniconda3
source activate pytorch
python ./pytorch-test.py

The Python script pytorch-test.py has the form

File : pytorch-test.py

import torch
x = torch.Tensor(5, 3)
print(x)
y = torch.rand(5, 3)
print(y)
# let us run the following only if CUDA is available
if torch.cuda.is_available():
    x = x.cuda()
    y = y.cuda()
    print(x + y)