exp mask
Closed this issue · 1 comments
bbidong commented
JunweiLiang commented
Doesn't matter what val is actually. And the multiply is before tf.add. Think about this case: val is a vector of size 3, [2.4, -5.0, 0.23], and mask is [1, 1, 0] (we use 0 to mark unused positions). so we have (val + [0, 0, -1e30]), so the last dimension of val's would become a very negative number, no mater what its original value is (unless it is a very positive number, which is unlikely).