SW:PyTorch

PyTorch

Description

PyTorch is deep learning framework that puts Python first.

• Homepage: https://www.pytorch.org/

Access

PyTorch is open to all HPRC users.

Options include:

• the PyTorch modules that were built at HPRC that have been optimized for our modern HPRC clusters
• the less efficient Anaconda that was built elsewhere for hardware/CPUs from 10 years ago

If PyTorch performs anything like TensorFlow, then you will likely want to stick with the CUDA versions (whether as a module or in Anaconda) to save SUs.

PyTorch modules (recommended)

For a list of modules that have been optimized for HPRC clusters, run:

mla | grep PyTorch

The ones with 'cuda' in the toolchain (e.g. PyTorch/1.3.1-fosscuda-2019b-Python-3.7.4) will likely provide the best performance. You will find a much newer collection on terra (ada is due to be retired in 2020).

Anaconda packages

[Note: the prebuilt virtual environments (VEs) described below are for a fairly old version of PyTorch. If you insist upon using Anaconda, then you are probably better off creating your own VE with a newer version of PyTorch.]

TAMU HPRC currently supports the user of Pytorch though the Anaconda modules. There are a variety of Anaconda modules available on Ada and Terra.

While several versions of Anaconda have some Pytorch environment installed, it is simplest to use exactly the versions in the following sections.

You can learn more about the module system on our SW:Modules page.

You can explore the available Anaconda environments on a per-module basis using the following:

[NetID@ada ~]$ module load Anaconda/[SomeVersion]
[NetID@ada ~]$ conda info --envs

Pytorch on Ada

A single version 0.2.0 of PyTorch is currently available on Ada. This version supports both CPUs and GPUs. To use GPUs, the program should run on a node with one or more GPUs.

To load this version which uses python 3.6.3:

[NetID@ada ~]$ module load Anaconda/3-5.0.0.1
[NetID@ada ~]$ source activate pytorch-0.2.0
[NetID@ada ~]$ [run your Python program accessing Pytorch]
[NetID@ada ~]$ source deactivate

This version can be run on any of the 64GB or 256GB compute nodes.

Pytorch on Terra

A single version 0.4.0 of PyTorch is currently available on Terra. This version supports both CPUs and GPUs. Your program using GPUs should run on GPU nodes.

To load this version which uses python 3.6.5:

[NetID@terra ~]$ module load Anaconda/3-5.0.0.1
[NetID@terra ~]$ source activate pytorch-0.4.0
[NetID@terra ~]$ [run your Python program accessing Pytorch]
[NetID@terra ~]$ source deactivate

Example Pytorch Script

As with any job on the system, Pytorch should be used via the submission of a job file. Scripts using Pytorch are written in Python, and thus Pytorch scripts should not be written directly inside a job file or entered in the shell line by line. Instead, a separate file for the Python/Pytorch script should be created, which can then be executed by the job file.

To create a new script file, simply open up the text editor of your choice.

Below is an example script (entered in the text editor of your choice) from http://pytorch.org/tutorials/beginner/pytorch_with_examples.html:

import torch
dtype = torch.FloatTensor
# dtype = torch.cuda.FloatTensor # Uncomment this to run on GPU
# N is batch size; D_in is input dimension;
# H is hidden dimension; D_out is output dimension.
N, D_in, H, D_out = 64, 1000, 100, 10
# Create random input and output data
x = torch.randn(N, D_in).type(dtype)
y = torch.randn(N, D_out).type(dtype)
# Randomly initialize weights
w1 = torch.randn(D_in, H).type(dtype)
w2 = torch.randn(H, D_out).type(dtype)
learning_rate = 1e-6
for t in range(500):
    # Forward pass: compute predicted y
    h = x.mm(w1)
    h_relu = h.clamp(min=0)
    y_pred = h_relu.mm(w2)
    # Compute and print loss
    loss = (y_pred - y).pow(2).sum()
    print(t, loss)
    # Backprop to compute gradients of w1 and w2 with respect to loss
    grad_y_pred = 2.0 * (y_pred - y)
    grad_w2 = h_relu.t().mm(grad_y_pred)
    grad_h_relu = grad_y_pred.mm(w2.t())
    grad_h = grad_h_relu.clone()
    grad_h[h < 0] = 0
    grad_w1 = x.t().mm(grad_h)
    # Update weights using gradient descent
    w1 -= learning_rate * grad_w1
    w2 -= learning_rate * grad_w2

It is recommended to save this script with a .py file extension, but not necessary.

Once saved, the script can be tested on a login node by entering:

[NetID@terra ~]$ python testscript.py