Keras: [Suggestion for improvement] Snapshot Ensemble
The paper "Snapshot Ensembles : Train 1, Get M for Free" proposes a technique to be able to train an ensemble of neural networks without additional training cost.
One implementation is at https://github.com/gaohuang/SnapshotEnsemble
It would be great to add this into Keras. :)
stale
jaybosamiya
👍4
>All comments
@jaybosamiya I've implemented the technique in my repository https://github.com/titu1994/Snapshot-Ensembles.
I don't have a large enough GPU to train WRN-34-4 or DenseNet-100, so it currently has the weights for WRN-16-4 instead. It has been trained on CIFAR-10 for now, but I will be training it on CIFAR-100 later as well. Sadly, I can't train the model on the Tiny ImageNet but it's pretty simple for someone to do that.
Edit: CIFAR 100 weights for WRN-16-4 are up as well
titu1994
on 7 Nov 2016
👍5
Was this page helpful?
0 / 5 - 0 ratings
Related issues
KeironO
路
3Comments
harishkrishnav
路
3Comments
oweingrod
路
3Comments
farizrahman4u
路
3Comments
anjishnu
路
3Comments
Most helpful comment
@jaybosamiya I've implemented the technique in my repository https://github.com/titu1994/Snapshot-Ensembles.
I don't have a large enough GPU to train WRN-34-4 or DenseNet-100, so it currently has the weights for WRN-16-4 instead. It has been trained on CIFAR-10 for now, but I will be training it on CIFAR-100 later as well. Sadly, I can't train the model on the Tiny ImageNet but it's pretty simple for someone to do that.
Edit: CIFAR 100 weights for WRN-16-4 are up as well