Personal framework for conducting experiments with tensorflow

This was designed and implemented before tensflow layer interface came out.

a pytorch like interface framework, including:

• boilerpipe codes of logging, model save/load, weight dumping, partial freezing weight
• advanced ops like convolutional rnn cell, beam search
• abstract interface for supervised training, policy gradient training, structured learning training
• also support horovod as backend for paralleling training

the asbtract pytorch like inteface

• AbstractModule: module as base component
• AbstractModel: subclass of module for supervised training, with additional interface of loss, build_trn_tst_graph, build_tst_graph and so on
• AbstractPGModel: subclass of AbstractModel for policy gradient training, with additional inteface of rollout (sampling) phase
• AbstractStructModel: subclass of AbstractModel for structured learning training, with addition interface of scoring (x,y) phase

• Reader: data loader interface

• TrnTst: boilerpipe codes for supervised training, such as building graph, running session, saving/loading model, validation
• PGTrnTst: boilerpipe codes for policy gradient training
• StructTrnTst: boilerpipe codes for structured learning

implementation of some common modules used in various experiments

• encoder/vlad.py: VLAD pooling
• encoder/conv_rnn.py: convolutional RNN
• encoder/birnn.py: bidirectional RNN
• gradient/poincare.py: get gradient for embedding on Poincare manifold
• gradient/lorentz.py: get gradient for embedding in Lorentz space

common utility functions