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
You can reuse your favorite python packages such as numpy, scipy and Cython to
extend PyTorch when needed.
Contents of the PyTorch image
This container has the PyTorch framework installed and ready to use. The
pytorch python module is installed as part of a Python 3.5 Conda environment in
/opt/conda/envs/pytorch-py35.
Both the compiled pytorch libraries and the Python 3.5 environment are included in $PATH. As
a result, running python from the command line executes a Python 3.5
interpreter by default.
/opt/pytorch contains the complete source of this version of PyTorch.
Running PyTorch
You can choose to use PyTorch as provided by NVIDIA, or you can choose to
customize it. Run pytorch as you would any python program: run a python script,
open interactive session in ipython or jupyter notebook. Start your scripts
or interactive sessions with
import torch
Customizing PyTorch
You can customize PyTorch one of two ways:
(1) Modify the version of the source code in this container and run your
customized version, or (2) use docker build to add your customizations on top
of this container if you want to add additional packages.
NVIDIA recommends option 2 for ease of migration to later versions of the
PyTorch container image.
For more information, see
https://docs.docker.com/engine/reference/builder/ for
a syntax reference. Several example Dockerfiles are provided in the container
image in
/workspace/docker-examples.
Suggested Reading
For more information about pytorch, see
- pytorch documentation http://pytorch.org/docs,
- pytorch tutorials https://github.com/pytorch/tutorials
- pytorch examples https://github.com/pytorch/examples
- a collection of links to pytorch tutorials, examples, projects and paper implementations https://github.com/ritchieng/the-incredible-pytorch
