Wsrtjtyk

The essential and problem code of Capsule Net are all below.

The problem is when I switch Part I to Part II code, I will get incompatibable dimension matching error.

The difference, in my opinion , between two part is Part II code do not calculate one of dimension (input_num_capsule) for u_vecs.

Is it Keras cannot support exchanging two None dimensions?

If you want to try it yourself, please fork this code on Kaggle.

```
class Capsule(Layer):

.......



def call(self, u_vecs):

    if self.share_weights:

        u_hat_vecs = K.conv1d(u_vecs, self.W)

    else:

        # Part I ###########################

        ## `local_conv1d`' logic when set kernel_size=1 and stride=1

        ####################################

        # u_vecs: [batch_size, input_num_capsule, input_dim_capsule]



        # immediate value : [1, batch_size, input_dim_capsule] # slice_len = 1

        # concate immediate value, got X: [input_num_capsule, batch_size, input_dim_capsule] 



        # W : [input_num_capsule, input_dim_capsule, num_capsule * dim_capsule]



        # K.batch_dot(X, W) [input_num_capsule, batch_size, num_capsule * dim_capsule]



        # [batch_size, input_num_capsule, num_capsule * dim_capsule]

        u_hat_vecs = K.local_conv1d(u_vecs, self.W, [1], [1])



        # Part II ###################

        ### In my idea, this is identical to the `local_conv1d(u_vecs, self.W, [1], [1])`

        ### , but the first dim of `x_aggregate` is determined.

        ##############################

        u_vecs = K.permute_dimensions(u_vecs, (1, 0, 2))

        u_hat_vecs = K.batch_dot(u_vecs, self.W)

        u_hat_vecs = K.permute_dimensions(u_hat_vecs, (1, 0, 2))





    ....



   ```