Development platform for
modern machine learning

Explore all features
Real-time team collaboration
Analyse and track experiments
Debug and optimize models
Manage server and environments

Major features at a glance

Click on a feature to get more information.


  • Monitor
    Monitor all running jobs aka experiments in real-time, watch metrics like loss/accuracy, additional information, hyper-parameters and much more.
  • Track & Compare
    All jobs are tracked in Git as new commit history viewable, sortable and exportable in our interface. You can compare all information of multiple jobs side-by-side.
  • Reproduce
    Since every job aka experiment is tracked in its own Git commit history, you have full transparency and best reproducibility.


  • Debug
    In each experiment, you get valuable insights of your model. With our interactive model debugger, you get very detailed insights of all of your layers in real-time (in work).
  • Optimise
    With our hyper-parameter optimization, you can find better hyper-parameters fully automated and distributed across multiple servers.
  • Design
    A deep neural network designer helps you to build quickly prototypes or helps you to understand new architectures.


  • Server Cluster
    Add your servers and build a cluster to monitor all their hardware utilization and information in one interface. The cluster as a whole gives resources to jobs and assigns automatically.
  • Job scheduler
    With your connected servers, you can quickly create new jobs aka experiments on different servers directly through our interface without fighting with dozens of open SSH terminals.
  • Docker container
    Every jobs can run in a Docker container, making it possible to completely make the environment reproducible.


  • Organisation support
    If you work in an organisation with multiple team members, you can create an organisation account and work together on models/experiments easier.
  • Real-time web application
    Since our main application AETROS Trainer is primarly a HTML5 web application, you see all jobs aka experiments and models in real-time across multiple devices.
  • Email/Slack notification
    Get notified when an experiment failed or finishes via e-mail or Slack integration.


  • Framework independent
    Our Python SDK is suitable for all kind of Python scripts you want to monitor and analyse. We are not limited to a particular machine learning framework.
  • On-Premises installation
    Our whole AETROS Trainer application with its web interface, API and Git server is available as docker images, so you have all data in your own infrastructure.
  • Full Git integration
    We store all experiments data in Git, along with all of its outputs. We also support external Git repositories, like you're used to with CI/Build tools.

Monitor your

Supervise key performance indicators
and custom metrics in real time of your model
created in the model designer or your
custom model (framework independent).

Also see all information you need to replicate
the experiment: Used hyperparameters, environment,
own additional information, files, etc.

Track all experiments

See all current running training jobs aka experiments or history to see which
model and hyper-parameters performed best. With custom channels you can add as many metrics as you need.

Compare experiments

Compare every aspect of multiple experiments in very detail.
You can attach custom additional information at each job (e.g. which datasource and split ratio you used)
and compare side by side. Metrics as graphs, hyperparameters and uploaded files can be compared as well.

Server Cluster

You can easily connect your servers and create a cluster that provides resources like CPU, memory and GPUs. Each job defines the resource requirements it needs and AETROS finds the best free server automatically. Since all jobs are encapsulated in Docker container, all assigned resources are quaranteed and GPUs exclusive. More information in our docs Server Cluster.

# Connect any server with AETROS Trainer.
$ aetros server marcj/alpha
Server connected to as marcj/alpha

# run commands from your working machine on
# connected server and monitor in AETROS Trainer.
$ aetros run --server=marcj/alpha 'python'

Automatic hyperparameter optimization

With our fully automated and scaleable hyperparameter optimization based on Hyperop you find easier and faster better hyperparameters for your model. More information in our documentation.

Deep neural network

Model designer

You can quickly create a deep neural network model using our model designer. For example a regression or image classification, without deep understanding of neural networks.

You can train the model completely on your own server infrastructure.
Python code based on the Keras framework is automatically generated and versioned.


Datasets in AETROS Trainer allows you to quickly create small datasets and link your created model with your own data.

You data stays in your data center and never reaches AETROS.
Use built-in feature image augmentation to prevent overfitting of your model.


With activated insights you get automatically a snapshot of all layers of your model per epoch and other insights like a confusion matrix.

Simple deployment

You can access your trained model via our API. The only thing you need to do is to train your model and tag your job with a version number.

You can also download the state of your trained model and run it in your own infrastructure.

curl -F "input=@/path_to/local/image.jpg"
{ "prediction": [ { "class": 8, "prediction": 0.13521507382393 }, { "class": 9, "prediction": 0.1149984151125 }, { "class": 4, "prediction": 0.10820455849171 }, ], "times": { "prediction": 0.1549129486084, "prepare_fetch_input": 0.1016149520874 } }

Video demo

Simple Keras integration

Are you ready?

Improve your current machine learning workflow
or start digging into machine learning using our neural network designer.

Register now