Announcing Hebel

I am publishing the GPU accelerated Deep Learning framework that I have been building for my own use on Github.

It is called Hebel (German for "lever") and supports training neural networks with various tricks, such as early stopping, momentum (regular and Nesterov), learning rate decay etc. Currently, only discriminative (single task or multi-task) neural networks of any architecture or depth are supported, but implementing other kinds of models such as Auto-Encoders is quite easy using the framework.

Hebel is fully GPU accelerated using PyCUDA and scikits-cuda in Python. It is not highly optimized, but already it is about 2-3 times faster than gnumpy.

Check the Github page for information and give it a try or help me develop it further. Documentation is yet a bit sparse, so please study the supplied examples and the code.

#deeplearning   #PyCUDA   #python   #neural_nets  
Shared publiclyView activity