Currently, the MDNOutput class just takes the unweighted average of all its constituent response distribution's point estimates. This doesn't make any sense, since a MDN also outputs weights that mix the response distributions. Leads to weird behavior like one individual gaussian getting most of the weight, but the point estimate being VERY far from that gaussian.
Currently, the
MDNOutput
class just takes the unweighted average of all its constituent response distribution's point estimates. This doesn't make any sense, since a MDN also outputs weights that mix the response distributions. Leads to weird behavior like one individual gaussian getting most of the weight, but the point estimate being VERY far from that gaussian.