infocusp / tf_cnnvis

CNN visualization tool in TensorFlow
MIT License
781 stars 209 forks source link

Deconvolution is super slow #70

Open kitovyj opened 5 years ago

kitovyj commented 5 years ago

Hello all.

Thank you for a great tool. I wonder, is there any way to speed up the computations? Layer deconvolution takes ~7 seconds in my case and I want to calculate deconvolutions for around ~10k examples, the network is of 7 layers.

BR, Alexander

falaktheoptimist commented 5 years ago

You could run the experiment in batches. Using a large batch size would be useful here to speed up the computations. We haven't experimented extensively with this, but the outputs of a batch would get written to disk together in an image. Let us know if this helps.

kitovyj commented 5 years ago

Hmm, when I feed a batch with multiple inputs, the tool still generates a single image for each layer... I don't understand the principle here.

falaktheoptimist commented 5 years ago

Hi @kitovyj We are working on this. We'll roll out a version where you should be able to insert multiple images simultaneously. You are right - we revisited the code and in its current state it would return just the first image output.