Olympus

Decades of machine learning research at your fingertips.

Features

• Deterministic Blocks
• Reproducible baselines for a variety of tasks
• Integrated Hyperparameter Optimizer (Orion)
• Experiment Tracking
• Model Zoo
• Pretrained Models
• Multi GPU training
• Automatic Checkpointing
• Mixed precision Available

Baselines

Run any baselines in a few lines of code

$ pip install olympus
$ export OLYMPUS_DATA_PATH=/fast
$ olympus --devices 0 classification --batch-size 32 --epochs 10 --dataset mnist --model resnet18
{
  "train_accuracy": 0.6458333333333334,
  "train_loss": 2.109870990117391,
  "elapsed_time": 9,
  "sample_count": 960,
  "epoch": 9,
  "adversary_accuracy": 0.3020833333333333,
  "adversary_loss": 2.234758218129476,
  "adversary_distortion": 0.2575291295846303,
  "validation_accuracy": 0.5986421725239617,
  "validation_loss": 2.108673614815782
}
{
  "temperature.gpu": 34.083333333333336,
  "utilization.gpu": 10.333333333333334,
  "utilization.memory": 0.0,
  "memory.total": 32480.0,
  "memory.free": 31672.833333333332,
  "memory.used": 807.1666666666666
}

Deterministic Blocks

Writing a full pipeline has never been easier, even when optimizing over hyper parameters !

.. literalinclude:: ../examples/hpo_simple.py
   :language: python
   :linenos:

Install

pip install git+git://github.com/mila-iqia/olympus.git

with fANOVA

sudo apt-get install swig
# pip install pyrex
pip install fanova