seq2seq-chatbot v0.2

Major new features:

1. Support for transfer learning via pre-trained embeddings

Embeddings can be loaded and optionally frozen during training. This is especially helpful when training on smaller datasets.

2. New web UI for chat including speech recognition and text-to-speech

(Requires flask)

3. New trained model trained_model_v2

Minor enhancements:

1. Both SGD & Adam optimizers are available for training

Other optimizers may be supported in the future.

2. Checkpoints can be backed up automatically at predefined loss thresholds during training

This is useful for testing out the model at different points throughout the training process.

3. Vocabulary supports both normalized and non-normalized words

Normalized means automatic conversion of "don't" to "do not", etc.

4. Auto-truncation of the dataset during processing is more intelligent

Lines of dialog are truncated at the last instance of sentence-terminating punctuation before the max length threshold.

5. Automatic punctuation is supported for chat.

Periods and question marks can be added automatically in certain contexts if the user wishes to be lazy and omit them during chat. This can improve conversation quality if the model was trained using punctuation.

6. Beam search & sampling decoders can now be enabled/disabled & customized on the fly during chat

--beamwidth=N
--enablesampling / --disablesampling

7. Batch files were added to help make training & chatting easier

(on windows only for now)

8. The cleaned dialog corpus can be saved to disk along with the model

And so can the training set question/answer pairs.