search

vlfeat / matconvnet

MatConvNet: CNNs for MATLAB
Other
1.4k stars 753 forks source link

CNN VGG-16 Training from scratch #939

Open Ellezon opened 7 years ago

Ellezon commented 7 years ago

Hi, I am trying to retrain all layers of the VGG-16 pretrained model from matlab with my data set. I set all weights{2} of the conv layers to zeros in the cnn_train script. However the error remains at 0.9 after 30 epochs using just 10 image categories. Can someone help me please?

lenck commented 7 years ago

Better approach would be to start with the examples for imagenet, which do provide a script to create a VGG networks. Zeroing out weights is not enough as the deployed networks are usually missing dropouts. Nevertheless, 10 categories seems like too little for the original VGG architecture...

Ellezon commented 7 years ago

I took your advice and created a new network however I am getting top1 error rates of over 40 .. do you have any solution please?

lenck commented 7 years ago

More data? Use a pre-trained network?

Ellezon commented 7 years ago

The error should be between 1 and 0 or am I wrong please?