Machine Learning

This page explains how to set up the tensorflow python module for accelerating machine learning using the GPUs installed on nodes in the GPU partition. If you have other software you prefer to use for machine learning you are free to try it out instead, but this page uses tensorflow as an example.

Begin by logging in to the HPC onto the gruffalo server as normal.

Next request an interactive session on a gpu enabled compute node using srun:

srun --partition=gpu --gpus=1 --mem=8G --cpus-per-task=4 --pty bash

Once connected to the compute node (assuming one is available), create a new conda environment (you can pick any name for the environment) and activate it:

conda create --name tensorflow-gpu-env
conda activate tensorflow-gpu-env

Next install the tensorflow-gpu package:

conda install tensorflow-gpu

We will now test the install to confirm that it can use the GPU to speed up training. Begin by starting python:

python

At the python prompt enter the following commands:

import tensorflow as tf
print( tf.constant( tf.__version__) )

This should print a long list of messages confirming that it has found a GPU to use. Next we will train a model on some of the included test data that comes with the package:

mnist = tf.keras.datasets.mnist
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()
train_images = train_images.reshape(60000, 28, 28, 1)
test_images = test_images.reshape(10000, 28, 28, 1)
train_images, test_images = train_images/255, test_images/255
model = tf.keras.Sequential([
  tf.keras.layers.Conv2D(32, (3,3), activation='relu', input_shape=(28,28,1)), tf.keras.layers.Conv2D(64,
  (3,3), activation='relu'), tf.keras.layers.MaxPooling2D(2,2), tf.keras.layers.Dropout(0.25),
  tf.keras.layers.Flatten(), tf.keras.layers.Dense(128, activation='relu'), tf.keras.layers.Dropout(0.5),
  tf.keras.layers.Dense(10, activation='softmax')
])
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

We are now ready to run the training, which should take about 30 seconds to run using GPU acceleration:

model.fit(train_images, train_labels, batch_size=128, epochs=15, verbose=1,
          validation_data=(test_images, test_labels))

In production use you will obviously not have to create the conda environment or install tensorflow-gpu as you’ve now set those up in the above example. To run your own code you only need to use the srun commands to access the gpu node and then conda activate tensorflow-gpu-env.