hli2020 / nn_encapsulation Goto Github PK

In table 4 an error of 4.55% is reported (this is also the top result from table 3s ablative analysis). When training on cifar10 using the default settings (calling main.py without setting any args) an error of 11.93% can be reproduced. However it doesn't appear that EncapNet is doing any routing with these settings, and without routing between capsules EncapNet appears to be a deep CNN with grouped convolutions to for capsules. --withCapRoute looks to turn on capsule routing for modules 3 and 4. Are any other settings needed to get an error of 4.55% on cifar10. Is the augmentation code used for EncapNet++ include in nn_encapsulation.

First, thanks for your work. The approach is novel and I'm really interesting in your design of master branch and aide branch. I also read your implementation and if i'm right, it would be mentioned in class capConvRoute3:
the master branch is generated from main_cap and the aide branch is generated from res_cap
And also in your paper, you said that

... one is a main mapping that directly
receives knowledge from its lower counterpart i, whose spatial location is the
same as j’s; another is a side mapping which sums up all the remaining lower
capsules, whose spatial location is different from j’s.

However, in your code, the res_cap is generated by

nn.Conv2d(ch_num_in, ch_num_out, kernel_size=ksize[0]+4, stride=stride, padding=pad[0]+2)

, which don't have any pre-processing to ignore the lower capsule in the same spatial location.

Is this for the convenience of implementation? Or if I have a wrong understanding.

Thanks.