yuxianmeng/matrix-capsules-pytorch

A Pytorch implementation of "Matrix Capsules with EM routing"

matrix-capsules-pytorch's Issues

Isn't here a bug？

Matrix-Capsules-pytorch/Capsules.py

Line 86 in bc9e26d

pose = pose.view(b,16,self.B,width_in,width_in).permute(0,2,3,4,1).contiguous() #b,B,12,12,16

pose's shape is (b, self.B, H, W), how can it be reshaped into ( b, 16, self.B, H, W) in the next line????? why does no one point this out? no error reported????

loss nan when r>1

Thanks for your work! However, when I set r = 2, the loss turned into nan. Have you met this problem? How to fix it?

Couldn't use full power of CUDA cores

When trying your demo on my 1050 Ti, this model only reaches 30% usage of CUDA cores and runs really slow. Meantime, another traditional CapsNet can reach 100% usage. Does this mean a performance problem of your model?

RuntimeError: zero-dimensional tensor (at position 0) cannot be concatenated

Hi,
I am on windows 10 with python 3 and conda install. I have just cloned and ran as suggested and I get this error RuntimeError: zero-dimensional tensor (at position 0) cannot be concatenated. Is this something you may have seen?
Thanks.

some code question

when Debug at here y_range = (max(floor((j-self.K)/self.stride)+1,0),min(j//self.stride+1,width_out))will print error TypeError: range() integer start argument expected, got float.

Overfitting with r>1

Hello. I'm trying to overfit to a toy batch with r=2. With batchsize >1 I am unable to overfit with r>1, though everything works with r=1. In particular, the network outputs the same results for all images in the batch.

Do you succeed in reproducing the author's result on smallNORB?

Hi, @shzygmyx

I am a contributor to another matrix CapsNet in TensorFlow. Are you able to reproduce the result suggested by the author of the paper?

some bugs here , I tried your code ,and .....

r.cuda() ?

https://github.com/shzygmyx/Matrix-Capsules-pytorch/blob/bc9e26dae9390bc96a5420a94cbe89432a1ad10a/Capsules.py#L199

there may be some mistakes in your code

1. for the loss funtion in 'class CapsNet' (train.py)
   you use mask = u.ge(0).float() to complete the max operation. but after a check about this function(tensor.ge) i think we should not us this function. because we want to preserve the value which >=0 and change the value which <0 into 0. If we use tensor.ge, we would turn all the value which is >=0 into 1. It's fault i think.
   2.loss = ((masku)2).sum()/b - m2 #float
   for this code, why we should masku and why we should - m**2. I could not find an introduction in the equ5 in the paper.

Thanks for your kind help!