microsoft / ADBench

Benchmarking various AD tools.

MIT License

107 stars 40 forks source link

Tensorflow module for GMM, LSTM and BA #122

Closed mikhailnikolaev closed 5 years ago

mikhailnikolaev commented 5 years ago

GMM, LSTM and BA objective Tensorflow differentiation test classes are implemented, using Tensorflow eager execution. Known issue: low efficiency Note: Tensorflow 2.0 is used

Graphs

Jacobian

Jacobian ÷ objective

Objective

Objective ÷ manual

GMM (1k) Objective ÷ manual - Release Graph

mikhailnikolaev commented 5 years ago

@awf, I have updated graphs. Now we have one point in BA (before all was terminated) and one point in GMM 10k (before all was terminated). Point count in GMM 1k is increased. LSTM hasn't been changed at all, so, for it all graphs are the same

awf commented 5 years ago

@awf, I have updated graphs. Now we have one point in BA (before all was terminated) and one point in GMM 10k (before all was terminated). Point count in GMM 1k is increased. LSTM hasn't been changed at all, so, for it all graphs are the same

Thanks! Where are the new graphs?

mikhailnikolaev commented 5 years ago

@awf, I have updated graphs. Now we have one point in BA (before all was terminated) and one point in GMM 10k (before all was terminated). Point count in GMM 1k is increased. LSTM hasn't been changed at all, so, for it all graphs are the same

Thanks! Where are the new graphs?

Sorry, it seems that I've forgotten to update PR descrption last time. Now they are reattached

awf commented 5 years ago

Great, thanks. And the 10K instances - are they running out of memory or time?

mikhailnikolaev commented 5 years ago

Great, thanks. And the 10K instances - are they running out of memory or time?

Of time