When I put a simple keras.layers.Concatenate into the relay keras frontend, its attribute axis is set to 1 instead of -1. The same issue happens for softmax. Should this happen?
The reason for this is that Keras by default uses NHWC layout while most TVM operations are better optimized for NCHW. The relay frontend for Keras automatically converts formats to NCHW (which use axis=1 to indicate channels).