search

gaudetcj / DeepQuaternionNetworks

Code for Deep Quaternion Networks
MIT License
55 stars 11 forks source link

Runtime Performance #3

Open ahundt opened 6 years ago

ahundt commented 6 years ago

When I try running the mnist example, runtime performance for training looks like it will be about 2 hours on CPU for a single epoch. For comparison Keras' mnist example takes 2 minutes on CPU for each epoch. Is this 60x difference expected?

gaudetcj commented 6 years ago

Yes the parameters from the mnist tests are those used by Hinton in his comparison to CNNs against Capsule Nets. I believe it is quite a bit bigger than the mnist from the keras examples. I have never run any of this code on CPU but it should not be much different from running a regular CNN on CPU unless you are using the quaternion batch normalization which still increase time by a good bit.

On Fri, Jun 15, 2018, 11:26 PM Andrew Hundt notifications@github.com wrote:

When I try running the mnist example, runtime performance for training looks like it will be about 2 hours on CPU for a single epoch. For comparison Keras' mnist example takes 2 minutes on CPU for each epoch. Is this 60x difference expected?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/gaudetcj/DeepQuaternionNetworks/issues/3, or mute the thread https://github.com/notifications/unsubscribe-auth/AGvq1PFKBKCHgnX12dPFLUcBlzU8jqMCks5t9IkOgaJpZM4UqWIM .