The original Resnet paper states that

We adopt batch-normalization (BN) [16] right after each convolution and
before activation

But in your implementation (in the residual unit function) the ordering seems to be different.

 # Layer 1
        batch_norm_op_1 = BatchNormalization()(residual_input_data)
        activation_op_1 = Activation('relu')(batch_norm_op_1)
        conv_op_1 = Conv2D(filters=filter1,
                         kernel_size=(1,1),
                         strides=(1,1),
                         padding='same')(activation_op_1)

Great implementation though. Very readable 👍

Thank you for sharing this repository! While recompiling the notebook Residual Attention Network Implementation - Cats vs Dogs Example.ipynb , I got an error in line model = ResidualAttentionNetwork((IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_CHANNELS), 1, activation='sigmoid').build_model() : AttributeError: 'int' object has no attribute 'value'. You have build_model() function in class ResidualAttentionNetwork (Code/ResidualAttentionNetwork.py) and I do not understand why I got this error. Have you ever encountered such an error?

Hey! There seems to be some issue with the code here:

Element-wise Addition

    if identity_x.shape[-1].value != conv_op_3.shape[-1].value:
        filter_n = conv_op_3.shape[-1].value

I tried replacing this with the updated set of code but it shows same error:

Element-wise Addition

    if identity_x.shape.as_list()[-1] != conv_op_3.shape.as_list()[-1]:
        filter_n = conv_op_3.shape.as_list()[-1]

Downgarding tensorflow again effects the other modules and creates issue while implementation. Please help !

TIA