hmmm, do you have some reproducible code using the Linear tranformation?
You can mostly ignore the forward_val method - it is not a user facing method. We use it to pass the numpy value around (instead of the theano tensor) for creating test values.