In stanford CNN course, it is stated that we can use the available Convnet for transfer learning. (http://cs231n.github.io/transfer-learning/)
I saw that some guys use VGGnet for transfer learning. The basic idea is to use the feature maps from the last conversational layer as features. All densely connected layers are retrained using the training data.
My question is: as we know that in VGGnet densely connected layers contain the large proportion of the parameters, is it possible to train the VGGNet model mentioned above? With such a large number of parameters, I guess that the model is easily overfitted.
If anyone have experience in transfer learning using VGGNet, please give me some helping comments. Thanks.
You must doing transfer learning in your own dataset:
1-apply pretrain VGG on your dataset
2-fine tune weights and bias
you can refer this link in MATLAB softwer guide:
Transfer Learning and Fine-Tuning of Convolutional Neural Networks
I did transfer learning with VGG-19 and VGG-16. I fine tuned the CNNs in 4th and 5th block of the models using small learning rate (1e-4), it did not overfitting.
I recommend you not to fine-tune all of CNNs if you have a limited number of data and use small learning rate.
Good luck
- Badges
- Science topic
- Similar topics
- Biological Science
- Bioecology
- Systems Ecology
- Landscape Ecology
- Connectivity