Changelogs

0.1.7

Interface with external modules:

• Swapped out register_backward_hook with register_full_backward_hook() in HookModule callback

• Added torch.nn.Module.preserveDevice() context manager

Callbacks:

• Changed Callbacks.suspend() to take in names only

• Added Callbacks.suspendClasses() to take in classes

• Changed Callbacks.suspendEvaluation() to Callbacks.suspendEval(), also added functionality for white and black lists

• Ripped out complex Callback.__getattr__() mechanism (that automatically searches the parent learner for missing attributes) to improve speed

• Added k1lib.callbacks.core callbacks CoreNormal and CoreRNN

• Added more plot decorators for MemoryProfiler, including backward-forward separator with their text

• Changed LossLandscape callback so that it follows the common flow structure. Before, it implements a whole new cbs(...) schema that essentially duplicated the loop inside of k1lib.Learner

Learner:

• Implemented grace stop handling for Learner. Before, if someone throws CancelRunException while executing a batch, then it will immediately jump to the end of the run. Changed it so that it will be caught in the batch loop, execute checkpoints cancelBatch and endBatch and rethrow it. Same thing with CancelEpochException

• Moved loss function from being a core functionality inside Learner to an external callback, so that adapting multiple loss functions are possible

• Added Learner.__call__()

Bioinfo cli:

• Added isValue, intersection, AA_, infinite, dereference cli tools

• Modified cat to be able to take in file name by pipe

• Added new joining operator “+” between init.BaseCli

Style changes:

• Changed Callbacks.learner to Callbacks.l

• Changed Callback.learner to Callbacks.l

• Removed k1lib.eqn from k1lib.imports’s namespace

0.1.6

I don’t really keep track of things 0.1.6 and before, so nothing here really.