Need help with Milano?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

144 Stars 18 Forks Apache License 2.0 23 Commits 0 Opened issues


Milano is a tool for automating hyper-parameters search for your models on a backend of your choice.

Services available


Need anything else?

Contributors list

License Documentation


(This is a research project, not an official NVIDIA product.)



Milano (Machine learning autotuner and network optimizer) is a tool for enabling machine learning researchers and practitioners to perform massive hyperparameters and architecture searches.

You can use it to: * Tune your model on a cloud backend of your choice * Benchmark Auto-ML algorithms (see how to add new search algorithm)

Your script can use any framework of your choice, for example, TensorFlow, PyTorch, Microsoft Cognitive Toolkit etc. or no framework at all. Milano only requires minimal changes to what your script accepts via command line and what it returns to stdout.

Currently supported backends: * Azkaban - on a single multi-GPU machine or server with Azkaban installed * AWS - Amazon cloud using GPU instances * SLURM - any cluster which is running SLURM


  • Linux
  • Python 3
  • Ensure you have Python version 3.5 or later with packages listed in the
  • Backend with NVIDIA GPU

How to Get Started

  1. Install all dependencies with the following command pip install -r requirements.txt.
  2. Follow this mini-tutorial for local machine or this mini-tutorial for AWS


We provide a script to convert the csv file output into two kinds of graphs:

  • Graphs of each hyperparameter with the benchmark (e.g. valid perplexity)
  • Color graphs that show the relationship between any two hyperparameters and the benchmark

To run the script, use:

python3 --file [the name of the results csv file] 
                     --n [the number of samples to visualize]
                     --subplots [the number of subplots to show in a plot]
                     --max [the max value of benchmark you care about]

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.