Release v1.3.0: MEMM, gradient and bug fixes
- Added support for Maximum Entropy Markov Models.
- Added use of atomic operation in gradient computation
- Improved RProp numerical stability
- Fix bug with unseen features in raw mode (thanks to Jurgen Van Gael)
- Fix bug discarding some features in maxent mode (thanks to George Foster)
- Switched code to stdint, should resolve some issue with size_t on exotic systems.
Release v1.2.0: RPROP and features improvements
- Switch from splay-trees to critbit-tries in features index
- Add RPROP+ and RPROP- variants of the algorithm
- Add a new projection scheme for RPROP with l1
- Maxent now work with sequences
- Add spaces matching in regexp
- Fix a few small bugs
Release v1.1.3: Some small improvements
- New option --jobsize for fine grained multi threading
- Improved SGD index construction : a lot faster
- Fix a small bug in sparse multi-threaded gradient
Release v1.1.2: Bug fix release
- Fix a small bug in L-BFGS/OWL-QN, should improve a bit convergence speed in some case.
- Fix a bug in multi-thread job system thanks to Alexander Fraser, should fix error rates and training speed on large dataset.
- Fix two small memory leaks.
- Some improvment in quark database handling.
Release v1.1.1: Mainly multi-threading improvements
- RPROP algorithm is now fully multi-threaded.
- Error rate estimation during training is now multi-threaded.
- Better jobs scheduling in multi-threaded gradient.
- Multi-threading code can be disabled (compilation on Windows should be simpler).
- Fixed bug in L1 optimization with RPROP (should improve stability).
Release v1.1.0: A few new features
- Added maxent mode.
- Added decoding through posteriors, this should improve accuracy at the price of computational time.
- Added the RPROP optimization algorithm.
- Added absolute indexing in patterns.
- Changed the scored output format as the posterior decoding provide normalized score at each position. The output is now compatible with CRF++.
- Some code cleanup.
Release v1.0.2: Mainly a bug fix version.
- Fixed some memory leaks, thanks to David Keeler
- Fixed argument processing to be more user friendly
- Fixed small bug in model compaction
- Added reading of raw files
- Spell corrections in man page
Release v1.0.0: Initial public version.