Skip to content


The MNIST dataset is a collection of 60,000 handwritten digits widely used for training statistical, Machine Learning (ML) and Deep Learning (DL) models. The MNIST MLCube® example demonstrates how data scientists, ML and DL researchers and developers can distribute their ML projects (including training, validation and inference code) as MLCube cubes. MLCube establishes a standard to package user workloads, and provides unified command line interface. In addition, MLCube provides a number of reference runners - python packages that can run cubes on different platforms including Docker, Singularity, KubeFlow and several others.


A data scientist has been working on a machine learning project. The goal is to train a simple neural network to classify the collection of 60,000 small images into 10 classes.


The source files for this MNIST example can be found on GitHub in MLCube Example repository.

MNIST training code

Training an ML model is a process involving multiple steps such as downloading data, analyzing and cleaning data, splitting data into train/validation/test data sets, running hyperparameter optimization experiments and performing final model testing. MNIST dataset is a relatively small and well studied dataset that provides standard train/test split. In this simple example a developer needs to implement two steps - (1) downloading data and (2) training a model. We call these steps as tasks. Each task requires several parameters, such as a URL of the data set that we need to download, location on a local disk where the data set will be downloaded to, path to a directory that will contain training artifacts such as log files, training snapshots and Machine Learning models. We can characterize these two tasks in the following way:

  • Download task:
    • Inputs: A yaml file (data_config) with two parameters - dataset URI and dataset hash.
    • Outputs: Directory to serialize the data set (data_dir) and directory to serialize log files (log_dir).
  • Training task:
    • Inputs: Directory with MNIST data set (data_dir), training hyper-parameters defined in a file (train_config).
    • Outputs: Directory to store training results (model_dir) and directory to store log files (log_dir).

We have intentionally made all input/output parameters to be file system artifacts. By doing so, we support reproducibility. Instead of command line arguments that can easily be lost, we store them in files. There are many ways to implement the MNIST example. For simplicity, we assume the following:

  • We use one python file.
  • Task name (download, train) is a command line positional parameter.
  • Both tasks write logs, so it makes sense to add a parameter that defines a directory for log files.
  • The download task accepts additional data directory parameter.
  • The train task accepts such parameters as data and model directories, path to a file with hyper-parameter.
  • Configurable hyper-parameters are: (1) optimizer name, (2) number of training epochs and (3) global batch size.

Then, our implementation could look like this. Parse command line arguments and identify a task to run. If it is the download task, call a function that downloads data sets. If it is the train task, train a model. This is sort of single entrypoint implementation where we run one script asking to perform various tasks. We run our script ( in the following way:

python download --data_config=PATH --data_dir=PATH --log_dir=PATH
python train --train_config=PATH --data_dir=PATH --model_dir=PATH --log_dir=PATH

MLCube implementation

Packaging our MNIST training script as a MLCube is done in several steps. We will be using a directory-based cube where a directory is structured in a certain way and contains specific files that make it MLCube compliant. We need to create an empty directory on a local disk. Let's assume we call it mnist and we'll use {MLCUBE_ROOT} to denote a full path to this directory. This is called an MLCube root directory. At this point this directory is empty:


Build location

The MLCube root directory will contain project source files, resources required for training, other files to recreate run time (such as requirements.txt, docker and singularity recipes etc.). We need to copy two files: that implements training and requirements.txt that lists python dependencies. By doing so, we are enforcing reproducibility. A developer of this MLCube wants to make it easier to run their training workload in a great variety of environments including universities, commercial companies, HPC-friendly organizations such as national labs. One way to achieve it is to use container runtime such as docker or singularity. So, we'll provide both docker file and singularity recipe that we'll put into the MLCube root directory as well. Thus, we'll make this directory a build context. For reasons that we will explain later, we also need to add .dockerignore file (that contains single line - workspace/). The MLCube directory now looks like:

A good test at this point would be to ensure that project is runnable from the build directory, and docker and singularity images can be built.

MLCube definition file

At this point we are ready to create a cube definition file. This is the first definition file that makes a folder to be an MLCube folder. This is a YAML file that provides information such as name, author, version, named as mlcube.yaml and located in the cube root directory . The most important section is the one that lists what tasks are implemented in this cube:

# Name of this MLCube.
name: mnist
# Brief description for this MLCube.
description: MLCommons MNIST MLCube example
# List of authors/developers. 
  - name: "First Second"
    email: ""
    org: "Company Inc."

# Platform description. This is where users can specify MLCube resource requirements, such as 
# number of accelerators, memory and disk requirements etc. The exact structure and intended 
# usage of information in this section is work in progress. This section is optional now.
  accelerator_count: 0
  accelerator_maker: NVIDIA
  accelerator_model: A100-80GB
  host_memory_gb: 40
  need_internet_access: True
  host_disk_space_gb: 100

# Configuration for docker runner (additional options can be configured in system settings file).
  image: mlcommons/mnist:0.0.1

# Configuration for singularity runner (additional options can be configured in system settings 
# file).
  image: mnist-0.0.1.simg

# Section where MLCube tasks are defined.
  # `Download` task. It has one input and two output parameters.
      inputs: {data_config: data.yaml}
      outputs: {data_dir: data/, log_dir: logs/}
  # `Train` task. It has two input and two output parameters.
      inputs: {data_dir: data/, train_config: train.yaml}
      outputs: {log_dir: logs/, model_dir: model/}
At this point, the directory looks like:


The workspace is a directory inside cube (workspace) where, by default, input/output file system artifacts are stored. There are multiple reasons to have one. One is to formally have default place for data sets, configuration and log files etc. Having all these parameters in one place makes it simpler to run cubes on remote hosts and then sync results back to users' local machines.

We need to be able to provide URI and hash of the MNIST dataset, collection of hyper-parameters and formally define a directory to store logs, models and MNIST data set. To do so, we create the directory tree workspace/, and then create two files with the following content (data.yaml):

hash: 731c5ac602752760c8e48fbffcf8c3b850d9dc2a2aedcf2cc48468fc17b673d1
and train.yaml:
optimizer: "adam"
train_epochs: 5
batch_size: 32
At this point, the cube directory looks like:

MNIST MLCube directory structure summary

  workspace/          # Default location for data sets, logs, models, parameter files.
    data.yaml         #   URI and hash of MNIST dataset.
    train.yaml        #   Train hyper-parameters.
  .dockerignore       # Docker ignore file that prevents workspace directory to be sent to docker server.
  Dockerfile          # Docker recipe.
  mlcube.yaml         # MLCube definition file.            # Python source code training simple neural network using MNIST data set.
  requirements.txt    # Python project dependencies.
  Singularity.recipe  # Singularity recipe.

Running MNIST MLCube

We need to set up the Python virtual environment. These are the steps outlined in the Introduction section except we do not clone GitHub repository with the example MLCube cubes.

Create python environment.

conda create -n mlcube python=3.8
conda activate mlcube
virtualenv --python=3.8 .mlcube
source .mlcube/bin/activate

Install MLCube Docker and Singularity runners.

pip install mlcube mlcube-docker mlcube-singularity


Before running MNIST cube below, it is probably a good idea to remove tasks' outputs from previous runs that are located in the workspace directory. All files and directories except data.yaml and train.yaml files can be removed.

Docker Runner

Configure MNIST cube (this is optional step, docker runner checks if image exists, and if it does not, runs configure phase automatically):

mlcube configure --mlcube=. --platform=docker

Run two tasks - download (download data) and train (train tiny neural network):

mlcube run --mlcube=. --platform=docker --task=download
mlcube run --mlcube=. --platform=docker --task=train

Singularity Runner

Configure MNIST cube:

mlcube configure --mlcube=. --platform=singularity

Run two tasks - download (download data) and train (train tiny neural network):

mlcube run --mlcube=. --platform=singularity --task=download
mlcube run --mlcube=. --platform=singularity --task=train