Minor version change 0.3.x

Feature additions
- Added 5 new metrics
- Improved RecurrentCuDNNWrapper (now works without seq_len)
- Added CyclicNoamLR schedule
- Added cross entropy loss
- Added normaliser_sources function for model specification
- Added command line device option
- Samplers work for higher-dimensional arrays

Breaking changes
- Loss branching changes
- *_data_source and normaliser_source breaking changes
- Change analysis_for_train_batch default behaviour
- Remove BaseVAE.loss default definition

Bug fixes
- viz/plotting: undefined metrics
- viz/synthesis: MLP now detaches seq_len
- experiment_builder: num_data_workers default, plateau sanity check, plotting after training
- losses: fixed loss wrapper argument order
- metrics: fixed hidden option
- utils: detach_batched_seqs handling of seq_len, detach_batch_seqs specification of squeezing, get_epoch_from_checkpoint_path bug

API changes galore! New normaliser API, new analysis_kwargs option, n…

…ew loss API, new metrics, new sequence utilities.

Moved to pytorh>1.2

Increased version to v0.1.0, updates since v0.0.1 are technically not…

… backward compatible, but the code is still in early testing so will stay in v0.x

Initial release of Morgana, see docs at https://zackhodari.github.io/…

…morgana/