Companion layers not resampled

[vinsis]

Thank you for this implementation of MADE. But I don't understand why companion layers are needed in the model.

Section 4.3 of the paper says:

instead of sampling the values of m^l(k) for all units and layers once and for all before training, we actually resample them for each training example or minibatch.

In your implementation it seems m^l(k) is chosen only once before training. If that's the case, how is it different from other linear layers?

Thanks in advance.