Convolution bias that precedes batch normalization should be eliminated.

[hccho2]

For example,

self.conv1 = tf.keras.layers.Conv2D(filters=filter_num,
                                    kernel_size=(3, 3),
                                    strides=stride,
                                    padding="same")

shoud be

self.conv1 = tf.keras.layers.Conv2D(filters=filter_num,
                                    kernel_size=(3, 3),
                                    strides=stride,
                                    padding="same",use_bias=False)