Sacred is a python infrastructure tool, to help with running machine-learning experiments. It helps to configure, organize, log and reproduce experiments.


Brainstorm is a deep-learning framework with a focus on modularity and support for Recurrent Neural Networks.


LSTM implementation in numpy

A Jupyter notebook with a reference implementation of a Long Short-Term Memory recurrent neural network in numpy alongside all the formulas. It is meant for educational purposes and not optimized for speed or efficiency.

TIMIT preprocessing

A Jupyter notebook that reads and preprocesses the TIMIT corpus. It can extract MFCCs or raw signals and supports the most common train/test splits. This is the preprocessing used for the LSTM: A Search Space Odyssey paper and it closely mimics the preprocessing used by many others.


The code accompanying the paper Binding via Reconstruction Clustering.


The now discontinued predecessor of brainstorm. A deep learning framework that we used for the A Clockwork RNN and LSTM: A Search Space Odyssey papers.


A very early prototype of a browser-based interface for querying sacred experiments stored in a MongoDB.


A pure-python reimplementation of the FANOVA framework for assesing hyperparameter importance (see this paper). At this point it is largely undocumented.