# Polyaxon

> Track, schedule, and scale deep learning experiments on Kubernetes

- Category: Training & Fine-Tuning / Experiment Tracking
- Language: Python
- Repository: https://github.com/polyaxon/polyaxon
- Website: https://polyaxon.com
- Canonical page: https://ai-tldr.dev/tools/polyaxon/

## In short

Polyaxon is an open-source platform for building, training, and monitoring deep learning applications at scale. It runs on Kubernetes and deploys into any data center, cloud provider, or a hosted managed service, supporting major frameworks such as TensorFlow, MXNet, Caffe, and Torch. It sits in the Experiment Tracking category of the Training & Fine-Tuning tools and is open-source, written in Python.

## Overview

Polyaxon is an open-source platform for building, training, and monitoring deep learning applications at scale. It runs on Kubernetes and deploys into any data center, cloud provider, or a hosted managed service, supporting major frameworks such as TensorFlow, MXNet, Caffe, and Torch.

It is aimed at data science and ML teams that need to track experiments, schedule jobs, and share GPU servers as self-service resources. Polyaxon manages workloads through container and node management, so a team can reproduce runs and keep results organized as the number of experiments grows.

As an experiment-tracking and MLOps tool, Polyaxon covers the full loop: a CLI to run and track experiments, a dashboard to browse projects and runs, plus built-in support for distributed jobs, parallel execution, hyperparameter tuning, and DAG-based pipelines.

## What it does

- Tracks logs and resource usage for experiments, with reproducible runs across teams
- Runs on Kubernetes and deploys to any data center, cloud provider, or Polyaxon's hosted service
- Supports major deep learning frameworks including TensorFlow, MXNet, Caffe, and Torch
- Distributed training support for TensorFlow, PyTorch, MPI, Horovod, Spark, and Dask
- Hyperparameter tuning with grid search, random search, Hyperband, Bayesian optimization, and Hyperopt
- DAG flow engine for running container-native ML pipelines with dependencies between operations

## Getting started

Install the CLI with pip, deploy Polyaxon onto a Kubernetes cluster, then create a project and run your first experiment.

1. Install the CLI

   Install the Polyaxon command-line interface with pip.

```bash
# Install Polyaxon CLI
pip install -U polyaxon
```

2. Deploy to Kubernetes

   Create a namespace, add the Helm charts repo, deploy Polyaxon from a config file, and forward the API port.

```bash
# Create a namespace
kubectl create namespace polyaxon

# Add Polyaxon charts repo
helm repo add polyaxon https://charts.polyaxon.com

# Deploy Polyaxon
polyaxon admin deploy -f config.yaml

# Access API
polyaxon port-forward
```

3. Create a project and run an experiment

   Create a project, then upload your code and start an experiment while tracking logs and resources.

```bash
# Create a project
polyaxon project create --name=quick-start --description='Polyaxon quick start.'

# Upload code and start experiments
polyaxon run -f experiment.yaml -u -l
```

4. Open the dashboard

   Launch the Polyaxon dashboard to browse your projects and experiments.

```bash
polyaxon dashboard -y
```

## When to use it

- Teams sharing GPU servers who need to schedule and queue training jobs as self-service resources
- Tracking experiment logs, metrics, and resource usage to keep deep learning runs reproducible
- Running hyperparameter searches such as grid, random, Hyperband, or Bayesian optimization at scale
- Orchestrating multi-step ML pipelines and distributed training jobs on Kubernetes via DAGs
