Torchbearer

Added
- Added a verbose level (options are now 0,1,2) which will print progress for the entire fit call, updating every epoch. Useful when doing dynamic programming with little data.
- Added support for dictionary outputs of dataloader
- Added abstract superclass for building TensorBoardX based callbacks
Changed
- Timer callback can now also be used as a metric which allows display of specified timings to printers and has been moved to metrics.
- The loss_criterion is renamed to criterion in `torchbearer.Model` arguments.
- The criterion in `torchbearer.Model` is now optional and will provide a zero loss tensor if it is not given.
- TensorBoard callbacks refactored to be based on a common super class
- TensorBoard callbacks refactored to use a common `SummaryWriter` for each log directory
Deprecated
Removed
Fixed
- Standard deviation calculation now returns 0 instead of complex value when given very close samples

Added
- Added a on_validation_criterion callback hook
- Added a DatasetValidationSplitter which can be used to create a validation split if required for datasets like Cifar10 or MNIST
- Added simple timer callback
Changed
Deprecated
Removed
Fixed
- Fixed a bug where checkpointers would not save the model in some cases
- Fixed a bug with the ROC metric causing it to not work

Added
- Added a decorator API for metrics which allows decorators to be used for metric construction
- Added a default_for_key decorator which can be used to associate a string with a given metric in metric lists
- Added a decorator API for callbacks which allows decorators to be used for simple callback construction
- Added a add_to_loss callback decorator which allows quicker constructions of callbacks that add values to the loss
Changed
- Changed the API for running metrics and aggregators to no longer wrap a metric but instead receive input
Deprecated
Removed
Fixed

Added
- Added a flag (step_on_batch) to the LR Scheduler callbacks which allows for step() to be called on each iteration instead of each epoch
- Added on_sample_validation and on_forward_validation calls for validation callbacks
- Added GradientClipping callback which simply clips the absolute gradient of the model parameters
Changed
- Changed the order of the arguments to the lambda function in the EpochLambda metric for consistency with pytorch and other metrics
- Checkpointers now create directory to savepath if it doesn't exist
- Changed the 'on_forward_criterion' callback method to 'on_criterion'
- Changed epoch number in printer callbacks to be consistent with the rest of torchbearer
Deprecated
Removed
Fixed
- Fixed tests which were failing as of version 0.1.2
- Fixed validation_steps not being added to state
- Fixed checkpointer bug when path contained only filename and no directory path
- Fixed console printer bug not printing validation statistics
- Fixed console printer bug calling final_metrics before they existed in state

Added
- Added support for tuple outputs from generators, torchbearer expects output to be length 2. Specifically, x, y = next() is possible, where x and y can be tuples of arbitrary size or depth
- Added support for torch dtypes in torchbearer Model.to(...)
- Added pickle_module and pickle_protocol to checkpointers for consistency with torch.save
Changed
- Changed the learning rate scheduler callbacks to no longer require an optimizer and to have the proper arguments
Deprecated
Removed
Fixed
- Fixed an issue in GradientNormClipping which raised a warning in PyTorch >= 0.4