search

QUVA-Lab / e2cnn_experiments

Experiment for General E(2)-Equivariant Steerable CNNs
Other
24 stars 4 forks source link

Differences between EXP and SFCNN models #2

Closed blazejdolicki closed 2 years ago

blazejdolicki commented 2 years ago

Hello,

I've tried reproducing your results, but some parts are unclear to me. For example, in mnist_bench.sh you seem to be using the EXP model for equivariant networks while in mnist_final.sh you use SFCNN model. Could you explain how do these two models differ? And what these abbreviations stand for? I couldn't find anything about that in the paper.

Thanks in advance!

Gabri95 commented 2 years ago

Hey!

the difference is that the EXP model is the smaller architecture we used to generate the large table with different groups / non-linearities while te SFCNN model is a larger architecture we used to reproduce the results from Learning Steerable Filters for Rotation Equivariant CNNs.

In other words, EXP is the model used in Tab 3 while SFCNN is the model used in Tab 5 of our paper (arxiv version)

Hope this helps

Gabriele

blazejdolicki commented 2 years ago

Thanks, that clears things up!