A code base for creating and running sequence models of language. Including language modeling, definition modeling, and common encoder-decoder stuffs. Required python 3.6.
- python 3.6
- tensorflow 1.1
- numpy 1.12
- nltk 3.2.4
- six 1.10
- dstruct.py: basic data structures and tuples
- generator.py: functions data reader and batch generator
- graph.py: functions to create various types of graphs
- model.py: runnable models from configuration
- run.py: functions to train and evaluate model (data + model -> result)
- util.py: utility functions (dictionary, array, logging, and cmd arguments)
- Value Network and A2C
- TD(lambda)
- Bootstrap last state if not terminal
- Option to randomly select sequences of the same encode input
- Option to select reward function from CLI
- Take advantage of tf.Session.make_callable (need benchmark)
- Compile my own TensorFlow
- Use tf.summary for tensorboard.
- It would be nice if we do not need to fecth state and feed it back in when training a langauge model (sentence dependent).
- Ensure the indices of
</s>and<s>are 0 and 1, and index of_is 0 for char-level data