search

arcelien / pba

Efficient Learning of Augmentation Policy Schedules
https://bair.berkeley.edu/blog/2019/06/07/data_aug/
Apache License 2.0
504 stars 86 forks source link

Fix probability and magnitude order mismatch in policy #28

Closed prob1995 closed 4 years ago

prob1995 commented 4 years ago

Fix the probability and magnitude order mismatch in policy which was mentioned by @chichan01 and rerun the notebook to get the fixed figure. However, I found that the older figure (wrong one) is more similar to the paper (Figure 4.). You can easily check it by observing magnitude(sum) bar which was around 8 in the max and the fixed one is around 80 in the max.

I think it makes more sense to be around 80 because the range of magnitude is 0~9 in integer, which is easily exceed 8 in sum for 15 operation.

arcelien commented 4 years ago

Thanks for the fix. Figure 4 plotted normalized magnitude values where the range for each operation was adjusted to [0,1], and after accounting for this I believe the plots are similar.

It looks like you included your original commit which was already merged and is causing a conflict. Could you remove that commit from the PR?

prob1995 commented 4 years ago

Sure! Will do this.

prob1995 commented 4 years ago

@arcelien I have created another PR to fix this. This PR can be closed. Thanks.