Why AMC Pruner use the training sub-set as the validation set?

[twmht]

Hi,

I recently read the code about AMC Pruner, but found out that it splits training set into sub training set and a validation set (https://github.com/microsoft/nni/blob/master/examples/model_compress/pruning/amc/amc_search.py#L73).

But why not use the original validation set?

In the white paper they said the reason is to prevent from overfitting, But I can't understand and get convinced.

Any idea?

[linbinskn]

Yes, splitting training set into train_loader and val_loader is to be aligned with AMC paper. The AMC author uses 5k images from cifar10 training images and 3k images from imagenet training images as validation dataset. You can check the evaluation part of that paper.