show image by plt with axis off

squash function for keep the length of capsules between 0 - 1 :arg cap_input: total input of capsules, with shape:

reconstruct from digit capsules with 3 fully connected layer :param digit_caps: digit capsules with shape [None, 16]

set up accuracy

build the graph of the network

get training summary

save checkpoint

test trained model on specific dataset split

compute tweak matrix for experiment of capsule unit representations

add the margin loss and reconstruction loss :arg digit_caps: output of digit capsule layer, shape [None, 10, 16]

dynamic routing between capsules :arg primary_caps: primary capsules with shape [None, 1, 32 x 6 x 6, 1, 8] l

doing dynamic routing with for loop as static implementation :arg proir: log prior for scaling with shape [10, num_caps]

doing dynamic routing with tf.while_loop :arg proir: log prior for scaling with shape [10, num_caps] cap_pred

advertisal test

capsule unit representation tweaking test

evaluation architecture

parse input arguments

parse input arguments

do image reconstruction and representations

training process :arg data: (images, labels)

creat architecture of the whole network

initial class with mnist dataset

compute a capsule, conv op with kernel: 9x9, stride: 2, padding: VALID, output channels: 8 per capsule. As described