search

mtogacar / COVID_19

5 stars 3 forks source link

Only net is used for feature extraction in SVM classifier and not netTransfer? #1

Open navinthenapster opened 3 years ago

navinthenapster commented 3 years ago

In the proposed design, you are doing transfer learning on the squeezenet and MobileNet. However when you are doing SVM, why you are extracting the training feature from net ( base graph) and not netTransfer.
It works fine for squeezenet however for mobilenet when we try to retraining, we can converting the fully-connected network to new_fc ( 1,numclassess -> kernel size) which causes the Logits to be removed. So we face error like the Logits layer is undefined when we go for SVM classifer. Can you give solution to overcome it.

navinthenapster commented 3 years ago

I went for netTransfer and did SVM classifier "pool10" and "new_fc" in the mobilenet. i got 2+2 features(2 classes) and which improved the accuracy.