Features:
- Major revision of Tensor library resulting in serious performance gains
- No more Java Tensor implementation in favour of simpler code
- Change build system from ant to gradle
- Fixed issues with native compilation not being performed correctly, especially with cutorch
- Improved performance for minibatches
- Improved memory footprint by reusing tensors where possible & triggering garbage collection
- Added various new modules, a.o. FullConvolution, BatchNormalization, ...
- Added various new composite modules, a.o. Inception, Residual, ...
- Extended Convolution to include temporal & volumetric operation
- MaxPooling has indices as optional second output, which can be passed to new MaxUnpooling module
- More efficient loading of datasets
- Added descriptors for datasets, allowing to define custom datasets without coding
- Added dataset adapters for data augmentation, a.o. flip, rotate, crop, ...
- Separated evaluator into generic & classification
- Added more stop criteria for a learn job based on error
- Added options for learning rate decay in SGD
- Perform validation while learning
- Automatically save best weights based on validation error
- Better error handling for jobs (especially on NaN)
- Easier resubmission for jobs, incl. changing parameters
- Fixed issue with URLs redirecting indefinitely
- UI improvements, such as ability to remove modules, job-specific URLs, etc.
- Extended JSON RPC
- Clean up of run configurations
- Various bug fixes
- Updated documentation to reflect new features in more detail
- Added contributor license agreement