Sockeye

Added
- Added Tensorboard logging for all parameter values and gradients as histograms/distributions. The logged values
correspond to the current batch at checkpoint time.

Changed
- Tensorboard logging now is done with the MXNet compatible 'mxboard' that supports logging of all kinds of events
(scalars, histograms, embeddings, etc.). If installed, training events are written out to Tensorboard compatible
even files automatically.

Removed
- Removed the `--use-tensorboard` argument from `sockeye.train`. Tensorboard logging is now enabled by default if
`mxboard` is installed.

Changed
- Change default target vocab name in model folder to `vocab.trg.0.json`
- Changed serialization format of top-k lexica to pickle/Numpy instead of JSON.
- `sockeye-lexicon` now supports two subcommands: create & inspect.
The former provides the same functionality as the previous CLI.
The latter allows users to pass source words to the top-k lexicon to inspect the set of allowed target words.

Added
- Added ability to choose a smaller `k` at decoding runtime for lexicon restriction.

Added
- Added a flag `--strip-unknown-words` to `sockeye.translate` to remove any `<unk>` symbols from the output strings.

Added
- Added a flag `--fixed-param-names` to prevent certain parameters from being optimized during training.
This is useful if you want to keep pre-trained embeddings fixed during training.
- Added a flag `--dry-run` to `sockeye.train` to not perform any actual training, but print statistics about the model
and mode of operation.

Changed
- `sockeye.evaluate` can now handle multiple hypotheses files by simply specifying `--hypotheses file1 file2...`.
For each metric the mean and standard deviation will be reported across files.

Added
- Optionally store the beam search history to a `json` output using the `beam_store` output handler.

Changed
- Use stack operator instead of expand_dims + concat in RNN decoder. Reduces memory usage.