DiverseMultiMNist config file.

[cinjon]

Hi, could you please post a config file for the DiverseMultiMNist?

[yaohungt]

You can modify the config based on Tables 10-12 in the paper.

[cinjon]

Is this correct?

{'params': {'backbone': {'kernel_size': 3,
   'output_dim': 1024,
   'input_dim': 3,
   'stride': 2,
   'padding': 1,
   'out_img_size': 18},
  'primary_capsules': {'kernel_size': 3,
   'stride': 2,
   'input_dim': 1024,
   'caps_dim': 64,
   'num_caps': 16,
   'padding': 0,
   'out_img_size': 8},
  'capsules': [{'type': 'CONV',
    'num_caps': 16,
    'caps_dim': 64,
    'kernel_size': 3,
    'stride': 1,
    'matrix_pose': True,
    'out_img_size': 6},
   {'type': 'FC', 'num_caps': 10, 'caps_dim': 64, 'matrix_pose': True}],
  'class_capsules': {'num_caps': 10, 'caps_dim': 64, 'matrix_pose': True}}}

[yaohungt]

It seems syntactically correct.

[cinjon]

Do you mind verifying that it is actually correct please? That would be quite helpful so that I don't spend a week debugging when it's not actually right.

[yaohungt]

I have no access to machines right now. But your config file is correct except that "True" should be "true".

We also conduct a variant of your config which is the vector-structured pose. This is done by setting 'matrix_pose' to false.

[cinjon]

With the config as is, it reaches only 73% accuracy on the diverse multi mnist test set. That's notably lower than the result in the paper.

It also is at 9.98m parameters, which is a tiny bit larger than what you shared. So I suspect it's off in some way from the config you used.

Any idea where it could be off or perhaps some other part that's different between the setup (for CIFAR) in the repo and the setup for DMM? Fwiw, it reaches ~99.99%/100% on the training set np.

[yaohungt]

I'm not sure your training/test match the setting in the paper or not. You can first try the vanilla Capsules (Dynamic/ EM Capsules) on multimnist, and see what you get.

Moreover, I don't think I reach ~99.99% accuracy on the training set.

[Sharut]

Hi @yaohungt , Could you please share your code for generating the Diverse Multi MNIST Dataset. When I tried to implement it in MNIST, Using the above config file I am able to cross 85% with the matrix pose itself whereas your number in the paper was 80%.

I really need to figure out where exactly I am messing up or is there something else which is causing this difference?