Regression with Caffe - does creating an lmdb and changing the loss from softmax to L2 would work?

[Gil Levi]

Hi,

I'd like to train/fine-tune a net for age estimation. 

It's straightforward to create an lmdb for the problem (just feed the ready-to-use shell script with a txt files with path_to_images and labels). It's also straightforward to take an off-the-shelf net (say, VGG_S) and change the loss from softmax to L2. 

Is that really all that is needed to finetune one of the pretrained models on a regression task? Am I missing something here?

Thanks!

Gil