Which Tensorflow API to use for Seq2Seq

Question

This year Google released 5 different packages for seq2seq:

• seq2seq (stated that this is a common goal, but inactive)
• nmt (active, but should be just about NMT perhaps)
• legacy_seq2seq (clearly legacy).
• contrib / seq2seq (possibly not complete)
• tensor2tensor (similar goal, also active development)

Which package is really worth using for implementation? It seems that they are all different, but none of them are stable enough.

Answer 1

- , ? OCR Encoder-Decoder. legacy_seq2seq ( ), , , .

• https://github.com/google/seq2seq: script, , , , (, OCR), ,

• https://github.com/tensorflow/tensor2tensor: , , . . - . ​​, Image Caption, . , , - txt- > txt image- > txt, . OCR. , (, CNN ).

• https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/seq2seq: , , , seq2seq , TF. Attention, Sequence Loss .. , . CNN, RNN, Bi Uni RNN, .. , .

• https://github.com/tensorflow/nmt: , tf.contrib.seq2seq library

:

• , , tensor2tensor. .
• , , seq2seq, tf.contrib.seq2seq.