Commit 5f3fa4fc authored by MAIRAL Julien's avatar MAIRAL Julien

update

parent 61f58132
......@@ -55,8 +55,8 @@ All of these results are for a single model with no data augmentation (I never f
## Q: Why do the results seem to be slightly better than in the original paper?
There are several reasons for that:
* This is a reimplementation which features different architectures than the original paper.
* Whereas the results of the NIPS paper were obtained by optimizing parameters on a validation set, before reporting results on the test set.
* This is a reimplementation which features slightly different architectures than the original paper.
* Whereas the results of the NIPS paper were obtained by optimizing hyper-parameters on a validation set, before reporting results on the test set (which results in about 0.4% loss in accuracy compared to optimizing hyper-parameters on the test set), this is not the case here. If you want to report the accuracy of the CKNs in your papers, please use the numbers provided in the NIPS paper (unless you are yourself doing experiments without a proper train/val or cross-validation split).
[1]: https://hal.inria.fr/hal-01387399/document
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment