Calling theano.grad on the graph built from theano.scalar.clip
is not working, so we have to use theano.tensor.clip instead.