Open Dingel321 opened 5 months ago
For this case, could one then use the output of the flow as approximating the joint between param and data instead of the posterior?
Ah, here I mean the parameters in the training readme "THETA_SCALE" and "THETA_SHIFT". They are used to scale the theta, which is usually in units in the index, so from 0 to num_models to some number between -1 and 1. This helps with the convergence of the NN and might not be clear to a new user. Also, we can just get the number from the model file. So ni practice i choose them always THETA_SCALE = THETA_SHIFT = max_index / 2. In this case theta gets rescaled so that its always between -1 and 1. The formula for rescaling is theta_rescaled = theta / THETA_SCALE - THETA_SHIFT.
PS: But it is in interesting question!! Since normally we don't have the normalization
Instead of manually computing the normalization, one could also just automatically compute and save it in the NPE object.