Open myurushkin opened 7 years ago
@myurushkin Are you using length penalty weight > 0.0 for beam search? You may disable it in your training for now as we found it doesn't utilize GPU well. A fix will be pushed to the future version of tensorflow.
Yep, I have used one of the default configurations: wmt16_gnmt_4_layer.json length_penalty_weight=1 in this case. I'll set this parameter to 0. Thanks!
@oahziur is there a TF issue to track this? Thanks!
@okuchaiev This should be fixed in tensorflow version 1.4
Hi there!
My GPU: gtx1070, TF 1.3
Is it ok that evaluation of test/dev datasets is performed on CPU mostly? GPU loading: 20% CPU loading: >90%
In my case training doesn't take much time in comparision with evaluation. Is it ok?
Best regards, Mikhail