Finetuning: how to adapt/retrain a model to new categories?

[roshansingh]

I want to retrain imagenet with newly defined categories (lets say around 50). How can I edit the various prototxt files to achieve this? Also can I refer any doc to understand all the parameters used in training prototxt file.

[shelhamer]

See the Caffe presentation (slide 15) linked from the README for how to finetune the model to a new set of classes. "Finetuning" is the method of taking a pretrained network, replacinga part, and then continuing the training on new data.

You only need to edit the last classifier layer to finetune.

Also can I refer any doc to understand all the parameters used in training prototxt file.

The parameters are mostly derived from the Krizhevsky model [1]. For a description of Caffe model parameters, read the comments in caffe.proto or the (better documented) dev caffe.proto.

[1] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 2012.