Mocha <https://github.com/pluskid/Mocha.jl> is a Deep Learning framework
for Julia. Two major changes in v0.0.5:
- CUDA backend now available on Mac
- Mocha now handles general ND-tensors (originally everything was
restricted to 4D tensor), making it much easier to deal with non-vision data
Detailed change log:
- Infrastructure
- *{Breaking Changes}* cuDNN 6.5 R2 (Release Candidate) (@JobJob)
- cuDNN 6.5 R2 is *NOT* backward compatible with 6.5 R1
- Forward convolution speed up
- Pooling with padding is supported
- Mac OS X is supported
- 4D-tensor -> ND-tensor
- Mocha is now capable of handling general ND-tensor
- Except that (for now) ConvolutionLayer and PoolingLayer still
requires the inputs to be 4D
- The generalization is *almost* backward compatible, except
- The interface for ReshapeLayer changed b/c the target shape
needs to be ND, instead of 4D now
- Parameters added for some layers to allow the user to specify
which dimension to operate on
- The output of InnerProductLayer is now 2D-tensor instead of 4D
- Unit-tests are expanded to cover test cases for ND-tensor when
applicable
- Interface
- print a constructed Net to get a brief overview of the geometry of
input/output blobs in each layers
- Documentation
- Setup the Roadmap Ticket
<https://github.com/pluskid/Mocha.jl/issues/22>,
duscussions/suggestions are welcome
- Update everything to reflect 4D -> ND tensor changes
- Document for parameter norm constraints
- Developer's Guide for blob and layer API
Best,
pluskid
