Just to re-confirm my understanding: I know that for fine-tuning an NGC model “sampling=auto” in JSON config will use the data distribution to balance batch when training. What happens during validation at the end of the epoch? Will a similar sampling schema be used or will the full validation set be only used once? Thanks!
This option is in the pipeline so you can configure it for training and/or validation separately. However, you should not use it for validation as you want the pipeline and transforms to be deterministic