This is a simple tutorial on running hyperparameter search with Oríon on Pytorch’s MNIST example

Installation and setup

Make sure Oríon is installed (Installing Oríon).

Then install pytorch and torchvision and clone the PyTorch examples repository:

$ pip3 install torch torchvision
$ git clone

Adapting the code for Oríon

To use Oríon with any code we need to do two things

  1. import the orion.client.report_objective helper function

  2. call report_objective on the final objective output to be minimized (e.g. final test error rate)

After cloning pytorch examples repository, cd to mnist folder:

$ cd examples/mnist

1. At the top of the file, below the imports, add one line of import for the helper function orion.client.report_objective():

from orion.client import report_objective
  1. We need the test error rate so we’re going to add a line to the function test() to return it

return 1 - (correct / len(test_loader.dataset))

Finally, we get back this test error rate and call report_objective to return the final objective value to Oríon. Note that report_objective is meant to be called only once because Oríon only looks at 1 'objective' value per run.

    test_error_rate = test(model, device, test_loader)



Once the script is adapted, optimizing the hyper-parameters with Oríon is rather simple. Normally you would call the script the following way.

$ python --lr 0.01

To use it with Oríon, you simply need to prepend the call with orion hunt -n <experiment name> and specify the hyper-parameter prior distributions.

$ orion hunt -n orion-tutorial python --lr~'loguniform(1e-5, 1.0)'

This commandline call will sequentially execute python --lr=<value> with random values sampled from the distribution loguniform(1e-5, 1.0). We support all distributions from scipy.stats, plus choices() for categorical hyper-parameters (similar to numpy’s choice function).

Experiments are interruptible, meaning that you can stop them either with <ctrl-c> or with kill signals. If your script is not resumable automatically then resuming an experiment will restart your script from scratch.

You can resume experiments using the same commandline or simply by specifying the name of the experiment.

$ orion hunt -n orion-tutorial

Note that experiment names are unique, you cannot create two different experiment with the same name.

You can also register experiments without executing them.

$ orion hunt --init-only -n orion-tutorial python --lr~'loguniform(1e-5, 1.0)'


When an experiment reaches its termination criterion, basically max-trials, it will print the following statistics if Oríon is called with -v or -vv.

{'best_evaluation': 0.05289999999999995,
 'best_trials_id': 'b7a741e70b75f074208942c1c2c7cd36',
 'duration': datetime.timedelta(0, 49, 751548),
 'finish_time': datetime.datetime(2018, 8, 30, 1, 8, 2, 562000),
 'start_time': datetime.datetime(2018, 8, 30, 1, 7, 12, 810452),
 'trials_completed': 5}

[{'name': '/lr', 'type': 'real', 'value': 0.012027705702344259}]

These results can be printed in terminal later on with the command info or fetched using the library API.

$ orion info -n orion-tutorial


We use test error rate for sake of simplicity, because the script does not contain validation dataset loader as-is, but we should never optimize our hyper-parameters on the test set and instead always use a validation set.

Oríon will always minimize the objective so make sure you never try to optimize something like the accuracy of the model unless you are looking for very very bad models.

You can also report results of types 'gradient' and 'constraint' for algorithms which require those parameters as well, or 'statistic' for metrics to be saved with the trial. See report_results() for more details.


When preparing a script for hyper-parameter optimization, we recommend first testing with debug mode. This will use an in-memory database which will be flushed at the end of execution. If you don’t use --debug you will likely quickly fill your database with broken experiments.

$ orion --debug hunt -n orion-tutorial python --lr~'loguniform(1e-5, 1.0)'