Using model weights for training/continuing-training with a modified network architecture

[Mohit Jain]

Hello,

     I want to know how does one use learnt features from a model to retrain a slightly modified model. To be precise, I have a CNN (deploy.prototxt attached) with which I train upon the CIFAR-10 dataset. This model uses a softmax classifier at the top of the network to finally classify images to the 10 categories. No I want to use the features learnt from this model and use it to train another model which has a SVM classifier at the top instead of the softmax. How can I go about doing so? Code/PseudoCode is highly appreciated.

Similar question asked previously : https://groups.google.com/forum/#!searchin/caffe-users/Using$20learnt$20weights$20with$20different$20network/caffe-users/tIhUqcW0Pck/z758qO29ynoJ

Regards,
Mohit

[Mohit Jain]

Found the solution!

For possible future questioners :Turns out there is a very good example that comes along with caffe itself (caffe-root/examples/net_surgery.ipynb) which performs all such model update operations.

Regards,
Mohit