How the does the prepocess work? [preprocess2,preprocess3]

[RobinY99]

It maybe a bit silly.

1. But I still wonder why it needs the mean and variance value for the preprocessing and how to get these values.
2. I guess random_flip can improve the performance.

[zwx8981]

@RobinY99 Sorry for the late reply.

1, The mean and variances values are statistical results of the dataset used to train CLIP models. This should be consistent between training and inference. Otherwise the prediction accuracy will be significantly degraded.

2,Yes, randomflip is a commonly used data augmentation technique that can help training models more effectively. Preprocess2 is used at inference phase, so we remove the random flip to achieve deterministic prediction behaviour.

[RobinY99]

Thanks for your patient answers 0.0 !