MA8701

This repository contains various resources that may help in MA8701. In particular, we start by setting up a python environment that can be used for estimating deep learning models on the epic cluster.

Login

We use ssh for login, and the correct login server is epic.hpc.ntnu.no

Running

ssh $NTNU_USERNAME@epic.hpc.ntnu.no

should let you log into the cluster, provided that you are enrolled in MA8701.

Modules

The compute clusters use EasyBuild for managing multiple, possibly incompatible environments.

This makes it easy for us to load pre-configured environments that have tools and libraries ready to use.

We load a module by running

module load $MODULE_NAME

The $MODULE_NAME will in most cases be name that semantically makes sense.

To see the available modules, simply run

module spider

And to search for a specific module, run

module spider $MODULE_NAME

Before loading a module, it is a good idea to make sure that the environment is clean. Therefore, run

module purge

before loading any modules.

In this repository, the file load_tf_modules.sh contains what is needed for loading a tensorflow and pytorch module. To load these modules, run

source load_tf_modules.sh

Note: We can see what core module are available by running

module available

Setting up the python environment

To make sure that everything is as expected, we should inspect what the tensorflow module contains and that it works as expected. For a detailed intro, look at this page.

First, we load the modules that we need:

source load_tf_modules.sh

Then we can run an interactive python environment, ipython. In ipython, we check that tensorflow and pytorch is installed and that we can access a gpu

import tensorflow as tf
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
print(tf.__version__)

The output should indicate that a GPU is available and which version of tensorflow we have available. For torch, we do

import torch
print(torch.cuda.is_available())
print(torch.__version__)

Creating a virtual environment

The modules we have loaded contain the main libraries we need for estimating models using either tensorflow or pytorch. We may, however, need more libraries, and we want full control of what we use when we run experiments. For this we use python virtualenv. To create a virtual environment, we run

virtualenv model_env

To use this environment instead of the "global" environment, we run

source model_env/bin/activate

A problem with this environment is that we don't have access to all the pre-installed packages in the modules we have loaded with module load. To create an environment that has access to these packages, we create a virtualenv using the following command instead

virtualenv --system-site-packages model_env

When in this environment, we can install any package we need and still have access to what the module has loaded.

To delete a virtualenv (it is common make mistakes the first times), we delete the folder in which it has been created.

rm -rf model_env/

It is recommended to create scripts that create the virtualenvs rather than making them interactively. We do this in order to create reproducible environments. An example is given in create_venv.sh. Here we install the newest version of keras.

Scheduling long-running code through SLURM-scripts

For a detailed information on scheduling jobs using slurm, see this page. We will only do a quick summary here, to get started with scheduling jobs.

To submit a long-running job, we use the sbatch command:

sbatch some_shell_script.sh

will schedule the a job using the commands within some_shell_script.sh.

Inside a slurm shell-script, there are some special headers that will serve as configuration for the scheduled job. These headers are prefixed by #SBATCH.

Typical confugration parameters are which computational resources to use, as well as output logging, and naming of the job(s).

There are some general guidelines available here.

For using GPUs, the following are required

#SBATCH --gres=gpu:$N_GPUs

where $N_GPUS is the number of gpus to require (typically 1).

The #SBATCH --time=XX:YY:ZZ configuration sets the time limit for the job in XX hours, YY minutes and ZZ seconds. When the time limit has hit, the job will be killed. It's wise to set reasonable time limits for scheduling purposes.

Here, we have provided a file schedule_job.sh, which contains an end-to-end training example on the epic cluster, using the image classifier in simple_image_classifier/omniglot_classifier.py.

Controlling experiments using configuration files

When running multiple experiments to create a good model, it is necessary to have control over the configuration of the experiment. In this repository, we present one way of controlling the experiment: configuration files.

The configuration files should be simple to create and contain all information about parameters you want to vary in the experiments. Examples are: optimizer configuration, model configurations (e.g. number of layers in a neural net).

Here, we store configuration files as simple json files. For example:

{
  "model_id": "dogsvscats.h5",
  "model_type": "simple_convnet",
  "weights": "imagenet",
  "batch_size": 16,
  "optimizer": {"type": "RMSprop",
    "params": {"lr": 0.001}},
  "train_path": "/lustre1/projects/fs_ma8701_1/dogsvscats/train",
  "valid_path": "/lustre1/projects/fs_ma8701_1/dogsvscats/validation"
}

This configuration contains information about where to find data, what optimizer to use, and what model.

Data folder

The avito dataset is available at /lustre1/projects/fs_ma8701_1/avito on the epic cluster. On /lustre1/projects/fs_ma8701_1 there are some other datasets also available, for experimentation purposes.

Mirroring external structure locally

When creating experiments, it's extremely valuable to be able to test that new code works, without having to rely on running on a server. A reasonable way to mimic the external structure is to

• Use the same virtualenv
• Use the same folder structure as on the server (but with less data).

On the epic cluster, we have the data folder /lustre1/projects/fs_ma8701_1. Here, I have put some example data, and will include the entire dataset for the project soon.

It is convenient to create a /lustre1/projects/fs_ma8701_1 folder locally, to avoid dealing with extra command line arguments when doing fast incremental code changes.

For jupyter fans

If you prefer using tools like jupyter notebooks for creating experiments, a good alternative with GPU/TPU support is google colab. In order to use this, you will need a google drive account.

A tutorial on fine tuning an image classifier on google colab can be found here.

Note that it may be tricky to use all the data in the avito dataset from google colab. Using the columnar data should, however, be easy.

Tutorials

Here is a tutorial that covers training a model in Keras that uses both tabular data and image data: https://www.pyimagesearch.com/2019/02/04/keras-multiple-inputs-and-mixed-data/