GRU and LSTM are running very slow in TF

d2l-ai / d2l-en

Interactive deep learning book with multi-framework code, math, and discussions. Adopted at 500 universities from 70 countries including Stanford, MIT, Harvard, and Cambridge.

https://D2L.ai

Other

23.23k stars 4.27k forks source link

GRU and LSTM are running very slow in TF #1823

Open astonzhang opened 3 years ago

astonzhang commented 3 years ago

Task "Evaluating ./chapter_recurrent-modern/gru.md" on CPU [2], GPU [3] is running for 00:19:02 ask "Evaluating ./chapter_recurrent-modern/lstm.md" on CPU [3], GPU [1] is finished in 00:20:03

http://ci.d2l.ai/blue/organizations/jenkins/d2l-en/detail/master/935/pipeline

0.16.6 release: http://ci.d2l.ai/blue/organizations/jenkins/d2l-en/detail/PR-1799/7/pipeline

terrytangyuan commented 3 years ago

Is this a regression or it's always been slow?

astonzhang commented 3 years ago

I think it's always. Can you look into it?

terrytangyuan commented 3 years ago

I took a look but couldn't find obvious mistakes yet. Did we get a chance to bump up TF version (trying to see if it's any low-level API bug that's fixed recently?