Partial re-initialisation of conv layer for finetuning

[HoldenCaulfieldRye]

For finetuning, is there a way of copying a conv layer and re-initialising only a subset of its kernel maps?

Justification: when doing transfer learning aka finetuning, it can be useful to freeze backprop (on lower layers) if the source training set is small, otherwise good filters can degenerate and be lost as the net overfits.
On the other hand, sometimes the network lacks expressive power, and it helps to enable backprop and randomly (re-)initialise the weights.

So what if the target task requires a few significantly different (high level) filters to the ones learned on the source task? Would be nice to re-initialise just a few kernel maps.

[shelhamer]

You can edit model parameters to your heart's content. Selectively replacing some convolutional filter kernels with random initialization is certainly possible. See the editing model parameters example on the Caffe site. Note that in the Python interface all Caffe model parameters are mutable: you can alter them and save the new model. Once the new weights are saved, you can finetune from it.

Please continue the discussion and ask future usage questions on the caffe-users mailing list. As of the latest release we prefer to keep issues reserved for Caffe development. Thanks!