xexiyong

I meet the same question; the strange is the size of model file almost increase once... while with vim shell open it, it seems no change.

and the tfa-nightly does not solve this problem

I found the tf.cond op which in crf_log_likelihood function is working on cpu thread, it causes this, delete it, training speed is better!

@bhack is this a bug in tf? the tf.cond in tf 1.4 is working in gpu thread, is it possible to recover this property?

https://github.com/tensorflow/addons/blob/53ef58f3aac9f3e1c36fca3bccaec73f67cd491c/tensorflow_addons/text/crf.py#L104 https://github.com/tensorflow/addons/blob/53ef58f3aac9f3e1c36fca3bccaec73f67cd491c/tensorflow_addons/text/crf.py#L201

> Can you log the device placement? https://www.tensorflow.org/guide/gpu#logging_device_placement this seems useless..