Why the the set of sub-policies is different from that in paper?

[buaaMars]

Thanks for your greak work! I have some questions. There is a sub-policy of no oprations in paper Learning Data Augmentation Strategies for Object Detection, but the v0 set in your code which is used as the policy for the data augment of Retina Net, replaces this sub-policy with [('Sharpness', 0.0, 8), ('ShearX_BBox', 0.4, 0)]. Why they are not same? Which one is better? What are v1, v2, and v3 for? Which one is best? Why there are three operations in one sub-policy at v2 instead of two in the paper? In Figure 7 of the paper, you got a result of 39.0 mAP. Why it says that

Expected accuracy on ResNet 50 with using autoaugment: 38.67

at https://github.com/tensorflow/tpu/blob/a60be5d11ec7c1cc3972f1f09d435c3ebf024250/models/official/detection/configs/yaml/retinanet_autoaugment.yaml#L6?

Thanks a lot!

[BarretZoph]

Hello,

They are the same actually. Since sharpness has a probability of 0.0 it will never be applied and shearX with a magnitude of 0 will not move the image. We kept it like this as this is how controller predicted the policy.

We found v0 to work best, but the other 3 still worked well without much accuracy degradation. We put them in as others might find them useful for different models/detection datasets as they might transfer better.

In the v2 policy we experimented with having 3 operations per sub-policy instead of 2, but found that it did not help the performance.

Yes we are working on closing this discrepancy. This was caused by some changes in the code from when we ran it to when it was opensourced.