All Versions
3
Latest Version
Avg Release Cycle
70 days
Latest Release
2704 days ago

Changelog History

  • v0.02.1 Changes

    May 22, 2014

    05-22-2014

    • โž• Added setup.py
    • โž• Added to PyPi
  • v0.02 Changes

    May 08, 2014

    05-08-2014

    • ๐Ÿ Windows compatibility (Thanks to @Wainberg)
    • โฌ†๏ธ CUDA 4.x is no longer supported, please upgrade to CUDA 5 or CUDA 6
    • All initialization is now handled through hebel.init(). No need to initialize PyCUDA separately anymore.
    • LogisticLayer has been renamed to SoftmaxLayer. LogisticLayer now does binary classification while SoftmaxLayer is for multiclass classification.
    • Framework for cross-validation.
    • When ProgressMonitor has save_interval=None, then only the currently best model is serialized. If it is a positive integer, then regular snapshots of the model are stored with that frequency.
  • v0.01 Changes

    January 02, 2014

    01-01-2014

    • โœ‚ Removed dependency on scikits.cuda (this should make Hebel compatible with Windows, but I couldn't test that yet)

    • Serious speed-ups by avoiding freeing and reallocating memory for temporary objects. Previously, many temporary gpuarrays were reallocated in every single minibatch and then discarded, which was very inefficient. By using persistent objects for temporary objects across minibatches and some other improvements such as doing more computations in-place, a roughly 2x speed-up could be realised.