HW1 solutions Discretized Mixture of Logistics Parameter initialization confusion about deepul

[wilson1yan]

I believe the initializations made it faster to train. logits are mixture parameters and initialized to be uniform. means are initialized to be uniformly distributed across the entire range of possible values (e.g. evenly spaced bins from 0-255 if we're doing rgb values). log_scales are just initialized randomly, but could also be initialized to unit variance. Overall, I think the means initialization helped the most and the logits and log_scales initializations could be zeros or randn.

[cangozpi]

Thank you very much !