tf.nn.bidirectional_dynamic_rnn \ tf.contrib.training.bucket_by_sequence_length \ tf.nn.ctc_loss理解

版权声明：本文为博主原创文章，未经博主允许不得转载。 https://blog.csdn.net/ssmixi/article/details/80825301

https://www.zhihu.com/question/42057513?sort=created
https://www.zhihu.com/question/52200883

https://stackoverflow.com/questions/38383988/seq2seq-use-of-buckets-in-tensorflow-tutorial

https://danijar.com/variable-sequence-lengths-in-tensorflow/ (tensorflow rnn源码作者博客讲dynamic_rnn)

http://www.wildml.com/2016/08/rnns-in-tensorflow-a-practical-guide-and-undocumented-features/

cuda多GPU使用：
https://blog.csdn.net/jasonzzj/article/details/59057287

tf.data.dataset:
https://stackoverflow.com/questions/46965098/how-does-one-move-data-to-multiple-gpu-towers-using-tensorflows-dataset-api

https://stackoverflow.com/questions/47580828/is-making-multiple-shards-of-your-data-with-multiple-threads-minimize-the-traini (重要)

https://stackoverflow.com/questions/50227336/using-the-tf-data-api-with-distributed-tf

CTC
http://www.tbluche.com/ctc_and_blank.html
http://www.a2ialab.com/lib/exe/fetch.php?media=presentations:icdar2015_trainingcosts_slides.pdf
https://zhuanlan.zhihu.com/p/28133530