- Add sweepEnd flag to the MinibatchData, returned from the
MinibatchSource.
- Add overloads TrainMinibatch, TestMinibatch that take a
Variable-MinibatchData map as their first parameter.
- Add sweepEnd to the Learner::Update function.
- Change the default epochSize to FullDataSweep in
TrainingParameterSchedule ctors.
- Allow minibatches to cross sweep boundaries: add a
flag (a config option) to specify if minibatches are
allowed to cross over the sweep boundaries (which
eliminates partial minibatches at the sweep ends).
- Update python API to pass MinibatchData as is into the
train_minibatch (without replacing it with corresponding ValuePtrs)
- Change the default epoch_size in parameter schedules to None,
in which case the entire data sweep is used as the scheduling unit.
- Add ReaderLib and python unit tests.
85c6aa8308