Getting Started

In this tutorial, we show how to start working with

  • Install CLI client;

  • Understand core concepts; and

  • Start developing on GPU in Jupyter Notebooks.

Installing CLI

To start working with CLI, you have two options:

  • Install CLI on your machine and run neuro login

The first option is recommended for exporing the platform. The second option, though, is better when working on your projects.

Linux and Mac OS instructions CLI requires Python 3 installed (recommended: 3.7, required: >=3.6). We suggest using Anaconda Python 3.7 Distribution.

pip install -U neuromation
neuro login

Windows instructions

While there are several options to make CLI work on Windows, we highly recommend using Anaconda Python 3.7 Distribution with default installation settings.

When you have it up and running, run the following commands in Conda Prompt:

conda install -c conda-forge make
conda install -c conda-forge git
pip install -U neuromation
pip install -U certifi
neuro login

Understanding core concepts

On Core level, one works with jobs, environments, and storage. To be more specific, one runs a job (an execution unit) in a given environment (Docker container) on a given preset (a combination of CPU, GPU, and memory resources allocated for this job) with several parts of storage (block or object storage) available (attached).

Let us show several examples.

Run a job on CPU which prints “Hello, World!” and shuts down:

neuro run --preset cpu-small --name test ubuntu echo Hello, World!

Upon execution of this command you’ll see an output like this:

Job ID: job-2b743322-f53a-4211-be4e-5d493e6cc770 Status: pending
Name: test
Http URL:
neuro status test # check job status
neuro logs test # monitor job stdout
neuro top test # display real-time job telemetry
neuro exec test bash # execute bash shell to the job
neuro kill test # kill job
Status: pending Creating
Status: pending Scheduling
StatStatus: pending ContainerCreating
Status: succeeded
Terminal is attached to the remote job, so you receive the job's output.
Use 'Ctrl-C' to detach (it will NOT terminate the job), or restart the
job with `--detach` option.
Hello, World!

Run a job in default environment (neuromation/base) on GPU which prints checks if CUDA is available in this environment:

neuro run -preset gpu-small --name test neuromation/base python -c "import torch; print(torch.cuda.is_available());"

Check the presets you can use:

neuro config show

Create a directory demo in your platform storage root:

neuro mkdir -p storage:demo

Run a job which mounts demo directory on storage to /demo directory in the job container and creates a file in it:

neuro run --preset cpu-small --name test --volume storage:demo:/demo:rw ubuntu "echo Hello >> /demo/hello.txt"

Check that the file is on storage:

neuro ls storage:demo

Developing on GPU

While one can run a Jupyter Notebooks session with one command in a command line or with one click in web UI, we recommend project-based development. To simplify the process, we provide the project template, which is a part of Toolbox, and provides the folder structure and integrations with several recommended tools.

Initializing a project

To initialize a new project from the template, run:

neuro project init

This command asks several questions about your project:

project_name [Name of the project]: Neuro Tutorial
project_slug [neuro-tutorial]:
code_directory [modules]:

You can press Enter if you agree with the suggested choice.

To navigate to the project directory, run:

cd neuro-tutorial

Project structure

After you execute the command mentioned above, you get the following structure:

├── config/ <- configuration files for various integrations
├── data/ <- training and testing datasets (we do not keep it under source control)
├── notebooks/ <- Jupyter notebooks
├── modules/ <- source code of models
├── results/ <- training artifacts
├── .gitignore <- default .gitignore for a Python ML project
├── <- autogenerated template reference
├── Makefile <- various ML development tasks (see `make help`)
├── <- autogenerated informational file
├── apt.txt <- list of system packages to be installed in the training environment
├── requirements.txt <- list of Python dependencies to be installed in the training environment
└── setup.cfg <- linter settings (Python code quality checking)

The template contains Makefile, which guarantees the contract between the above-shown structure, the base environment which we provide, and manipulations with storage and jobs. For example, through upload-* and download-* commands, sub-folders on your local machine are synced with sub-folders on persistent platform storage, and those sub-folders are synced with the corresponding sub-folders in job containers.

Setting up the environment and running Jupyter

To set up the project environment, run:

make setup

Upon execution of this command, system packages from apt.txt and pip dependencies from requirements.txt are installed in the base environment, which already contains CUDA support and the most popular ML/AI frameworks, like Tensorflow and Pytorch.

To start a Jupyter Notebooks session on GPU, run:

make jupyter

This command open Jupyter Notebooks interface in your default browser.

Now, when you edit notebooks, they update on your platform storage. To download them locally (for example, to put them under version control), run:

make download-notebooks

Don’t forget to terminate your job when you no longer need it (the files won’t disappear after that):

make kill-jupyter

To check how much GPU and CPU hours you have left, run:

neuro config show-quota