KellyHwong
commented
4 years ago Hi there, First of all, thank you for this invaluable repo.
My question is this: the
policy_v0in the codebase does not match any of the autoaugment policies from the paper. Where does it come from? Thanks in advance for your help.
The policy defined in autoaugment.py from line 41 seems not same with any policy mentioned in the paper AutoAugment: Learning Augmentation Strategies from Data.
Actually if the policy is made for ImageNet, policy in the paper is still not consistent with the repo, refer the paper and find: Table 9. AutoAugment policy found on reduced ImageNet. Operation 1 Operation 2 Sub-policy 0 (Posterize,0.4,8) (Rotate,0.6,9) Sub-policy 1 (Solarize,0.6,5) (AutoContrast,0.6,5) Sub-policy 2 (Equalize,0.8,8) (Equalize,0.6,3) Sub-policy 3 (Posterize,0.6,7) (Posterize,0.6,6) Sub-policy 4 (Equalize,0.4,7) (Solarize,0.2,4) Sub-policy 5 (Equalize,0.4,4) (Rotate,0.8,8) Sub-policy 6 (Solarize,0.6,3) (Equalize,0.6,7) Sub-policy 7 (Posterize,0.8,5) (Equalize,1.0,2) Sub-policy 8 (Rotate,0.2,3) (Solarize,0.6,8) Sub-policy 9 (Equalize,0.6,8) (Posterize,0.4,6) Sub-policy 10 (Rotate,0.8,8) (Color,0.4,0) Sub-policy 11 (Rotate,0.4,9) (Equalize,0.6,2) Sub-policy 12 (Equalize,0.0,7) (Equalize,0.8,8) Sub-policy 13 (Invert,0.6,4) (Equalize,1.0,8) Sub-policy 14 (Color,0.6,4) (Contrast,1.0,8) Sub-policy 15 (Rotate,0.8,8) (Color,1.0,2) Sub-policy 16 (Color,0.8,8) (Solarize,0.8,7) Sub-policy 17 (Sharpness,0.4,7) (Invert,0.6,8) Sub-policy 18 (ShearX,0.6,5) (Equalize,1.0,9) Sub-policy 19 (Color,0.4,0) (Equalize,0.6,3) Sub-policy 20 (Equalize,0.4,7) (Solarize,0.2,4) Sub-policy 21 (Solarize,0.6,5) (AutoContrast,0.6,5) Sub-policy 22 (Invert,0.6,4) (Equalize,1.0,8) Sub-policy 23 (Color,0.6,4) (Contrast,1.0,8) Sub-policy 24 (Equalize,0.8,8) (Equalize,0.6,3)
Hi there, First of all, thank you for this invaluable repo.
My question is this: the
policy_v0in the codebase does not match any of the autoaugment policies from the paper. Where does it come from? Thanks in advance for your help.