ValueError: Can not do batch_dot on inputs with shapes!

[akashshingha850]

Hello, @XifengGuo greetings. I was using your capsnet implementation for a multi-label classification where model output 3 in number for each label target. However, I've faced the following error.

Each three output has a different number of classes. For example, the labels are A, B, and C. Label A has 168 different class, whereas B has 11 and C has 7. A basic set-up has given below:

  def CapsNet:

          ............... more code ......................

            # Layer 3: Capsule layer. Routing algorithm works here.
           digitcaps_0 = CapsuleLayer(num_capsule=168, 
                                   dim_capsule=16, 
                                   routings=routings,
                                   name='digitcaps_0')(primarycaps)
    
           digitcaps_1 = CapsuleLayer(num_capsule=11, 
                                   dim_capsule=16, 
                                   routings=routings,
                                   name='digitcaps_1')(primarycaps)
    
           digitcaps_2 = CapsuleLayer(num_capsule=7, 
                                   dim_capsule=16, 
                                   routings=routings,
                                   name='digitcaps_2')(primarycaps)

           # Layer 4
           out_caps_0 = Length(name='out_caps_gra')(digitcaps_0)
           out_caps_1 = Length(name='out_caps_vow')(digitcaps_1)
           out_caps_2 = Length(name='out_caps_cons')(digitcaps_2)

           # Models for training 
           train_model = models.Model(inputs=[x], outputs=[out_caps_0, 
                                                    out_caps_1, 
                                                    out_caps_2])  
    
           return train_model


# build the model
model = CapsNet(input_shape=[64, 64, 1], routings=3)

And it threw the following error. The logs

ValueError                                Traceback (most recent call last)
<ipython-input-23-9f7ff279bacc> in <module>
      1 import tensorflow as tf
      2 
----> 3 model = CapsNet(input_shape=[64, 64, 1], routings=3)

<ipython-input-22-1735b9121899> in CapsNet(input_shape, routings)
    109                                    dim_capsule=16,
    110                                    routings=routings,
--> 111                                    name='digitcaps_gra')(primarycaps)
    112 
    113     digitcaps_vow = CapsuleLayer(num_capsule=11, 

c:\users\devlix\.conda\envs\tflite\lib\site-packages\keras\engine\base_layer.py in __call__(self, inputs, **kwargs)
    487             # Actually call the layer,
    488             # collecting output(s), mask(s), and shape(s).
--> 489             output = self.call(inputs, **kwargs)
    490             output_mask = self.compute_mask(inputs, previous_mask)
    491 

<ipython-input-10-d94539d78fd5> in call(self, inputs, training)
    154                 # then matmal: [dim_capsule] x [input_num_capsule, dim_capsule]^T -> [input_num_capsule].
    155                 # b.shape=[batch_size, num_capsule, input_num_capsule]
--> 156                 b += K.batch_dot(outputs, inputs_hat, [2, 3])
    157         # End: Routing algorithm -----------------------------------------------------------------------#
    158 

c:\users\devlix\.conda\envs\tflite\lib\site-packages\keras\backend\tensorflow_backend.py in batch_dot(x, y, axes)
   1497                          str(x_shape) + ' and ' + str(y_shape) +
   1498                          ' with axes=' + str(axes) + '. x.shape[%d] != '
-> 1499                          'y.shape[%d] (%d != %d).' % (axes[0], axes[1], d1, d2))
   1500 
   1501     # backup ndims. Need them later.

ValueError: Can not do batch_dot on inputs with shapes (None, 168, 168, 18432, 16) and (None, 168, None, 18432, 16) with axes=[2, 3]. x.shape[2] != y.shape[3] (168 != 18432).

Any catch? TIA.

[innat]

Downgrade Keras version to 2.2.4.