mainly tf work, may do some pytorch related  

show how to write and read TFRecord(tensorflow standard dataa format, Example and SequenceExample, sparse_to_dense, dynamic code)
might move to ./exp as this is for demo purpose

reading libsvm format then do text classification

reading libsvm format then do text regression

reading libsvm format then do text binary classification, evaluate by auc

this is self contained mlp classification example showing
how to read sparse TFRecord and train a mlp classifier

this is implmentaion of pointer network, dealing sorting problem

right now actually seq2seq related applications root

image-caption related work now support

bow,
rnn,
cnn(TODO)  

show_and_tell   show_attend_and_tell(TODO)
Input with both image and text(TODO)

show_and_tell supported similary as im2txt, but here we also support discriminant method
we support both ingraph/outgraph beam search
will follow google/seq2seq method but now here just works though beam search is not dynamic
training is dynamic also support sampled softmax
support directly deal with image like im2txt(using inception v3, allow distort images) and also support use pre calc image feature as image input(faster train speed but can not distort images any more, experiment show distort not very useful?)
support using Example while im2txt use SequenceExample, also support use SequenceExample incase you want do bucket batch for rnn decode train
use melt for training the code will be much shorter and handel all training details and auto handel summary ops
support <train + validate(random) + fixed validate + predict evaluate> all in one mode, see below graph, will help experiment a lot, while im2txt you need seperate process to do validation

app with long text as input(like image title, ct0) and predict shortter summary text(like click query)
supporting method:
seq2seq
seq2seq_attetion
seq2seq_attetion_copy(TODO)

common seq2seq codes used for image-caption, text-sum and other applications

common util for deepiu

common lib

common tensorflow related lib, you can view it similar as tf.contrib

like tf.supervisor, make train and test flow easier,
main functions include, application only make graph and pass train_ops and evaluate_ops to flow
flow will do model save, log save(auto add all one dimentional shape tensors to tensorboard), show elapsed time ...

wrapper for reading tfrcords support shuffle_then_decode and decode_then shuffle, support libsvm sparse decode