I'm wondering, what is the most correct way to evaluate models with U-IDS and P-IDS on small datasets?
The output dimension of InceptionV3 is 2048, which means that for smaller datasets (< 1024 images) the SVM problem is ill-posed and in absense of regularization term each run may lead to a different solution.
Should we SVD the Inception features for all the data before feeding to SVM?
Or the better option is to train SVM multiple times for same set of predictions and then take the mean value?
Hi!
I'm wondering, what is the most correct way to evaluate models with U-IDS and P-IDS on small datasets?
The output dimension of InceptionV3 is 2048, which means that for smaller datasets (< 1024 images) the SVM problem is ill-posed and in absense of regularization term each run may lead to a different solution.
Should we SVD the Inception features for all the data before feeding to SVM?
Or the better option is to train SVM multiple times for same set of predictions and then take the mean value?