conv0 = tf.compat.v1.layers.Conv2D(
     filters=filters,
     kernel_size=kernel_size,
     strides=2,
     padding=('SAME' if strides == 1 else 'VALID'),
     use_bias=False,
     kernel_initializer=tf.variance_scaling_initializer(),
     data_format=data_format)

    # Use the image size without space-to-depth transform as the input of conv0.
    batch_size, h, w, channel = inputs.get_shape().as_list()
    conv0.build([