capsnet | Reimplementation of Capsule net from https | Machine Learning library

I have a dataset with more than 4000 images and 3 classes, and I'm reusing a code for capsule neural network with 10 classes but I modified it to 3 classes, when I'm running the model the following error occurs at the last point of the first epoch (44/45):

I am learning capsnet now, and trying to transfer the code from local computer to colab. The code runs well on my local computer, but raise an error when I try it on colab. ValueError: Inconsistent shapes: saw (1152, 10, 1, 10, 16) but expected (1152, 10, 1, 16).

When I try other matching like [3,1], I will get the following error. In this case, x's dimension backs to 4 and x[3] == y[2]. ValueError: Can not do batch_dot on inputs with shapes (1152, 10, 1, 8) and (1152, 10, 8, 16) with axes=[3, 1]. x.shape[3] != y.shape[1] (8 != 10).

I locate the reason of this error on the function tf.scan. And I installed tensorflow 1.13 on my computer. But I don't know how to fix it. Please help me.

Here is the code.

It worked fine when I used fit, but when I used fit_generator, I got a problem.

I used the call back method to find the confusion matrix at the end of each train epoch.

However, the accuracy obtained from the confusion matrix and the validation accuracy output from keras differ.

My code is below.

I am training a Capsule Network with both encoder and decoder part. It works perfectly fine with all the classes (10 classes) of the MNIST data set. But when I am extracting a single class say (class 0 or class 5) and then training the capsule network, the reconstruction of the image is very poor.

Where do I need to change the network setting, or do I have an error in my data preparation?

I tried:

1. I changed the total class from 10 (for ten digits to 1 for 1 digit and even for 2 for 2 digits).
2. When I am using the default MNIST dataset, I am getting no error or tensor size, but when I am extracting a particular class and then passing it into the network, I am facing issues like a) Dimensional Issues b) Float tensor warning.

I fixed these things but manually adding a dimension and converting the data to data.float().cuda() tensor. I did this for both the case i.e when I am using the 10 Digit Capsules and when I am using the 1 Digit Capsules for training a single class digit.

But after this, the network is running fine, but I am getting really blurred and poor reconstructions. While when I am training the whole MNIST dataset without extracting any class and passing it to the network, it doesn't throw any error and the reconstruction works really fine.

I would love to share the more detail and other parts of the code -

I'm new in Tensorflow and Keras. I trying to run code in Keras. I have labels that convert to one hot encoding whit 2 depth and after that convert it to Keras layers format. now when I rum my code I got this error:

Traceback (most recent call last): File "/Users/vahid/Documents/Thesis/Code/new_code/CapsNet_model.py", line 224, in CapsNet((None,28,28,1),2,3,trainImg,trainLbl) File "/Users/vahid/Documents/Thesis/Code/new_code/CapsNet_model.py", line 164, in CapsNet x_recon = k.layers.Dense(512, activation='relu')(masked) File "/Users/vahid/TensorProject/lib/python3.6/site-packages/keras/engine/base_layer.py", line 443, in call previous_mask = _collect_previous_mask(inputs) File "/Users/vahid/TensorProject/lib/python3.6/site-packages/keras/engine/base_layer.py", line 1311, in _collect_previous_mask mask = node.output_masks[tensor_index] AttributeError: 'Node' object has no attribute 'output_masks'

and the part of the code that I run and got error:

I was looking at this capsnet code on github

And I can't find what does the line no. 116 mean?