kchua/handful-of-trials

Possible issue in normalization

amini2nt opened this issue · 1 comments

amini2nt commented

handful-of-trials/dmbrl/modeling/utils/TensorStandardScaler.py

Line 45 in e1a62f2

sigma[sigma < 1e-12] = 1.0

If the sigma is < 1e-12, you are setting sigma to 1 which does not do any scaling. Shouldn't it be

sigma[sigma < 1e-12] = 1e-12 ?

kchua commented

This was intentional, the intuition being that dimensions where sigma < 1e-12 over the dataset have negligible variation and essentially deterministic and should thus not be normalized.