Fine tuning VGG_FACE model

[Mallika Agarwal]

Hello,

Completely new to Caffe, so please pardon me.

I want to fine-tune the VGG_FACE model for my own dataset. Now I went through many links and tutorials, but they edit the solver.prototxt to edit the path to the dataset. However, when I download the source for vgg_face, I do not have any such file. I only have the deploy.prototxt (in which i have edited the last two layers, hopefully, correctly) and the the VGG_FACE.caffemodel file. (Attaching the folder structure image).

I don't understand how to input the path of my dataset (which i have created in the lmdb formated) to the finetuning command.

ANY help would be highly appreciated.

Thanks in anticipation,

[Przemek D]

You might want to take a look at existing caffe models and examples. Deploy proto is not suitable for training - you want to use a train_val proto: see eg. this example to see the difference; this will also show you how to supply your own data to the network. Other examples (like this one) will better illustrate using LMDB for training.

[Mallika Agarwal]

Okay, thanks a lot.

Also, what am I supposed to pass for the mean_file parameter? Am I supposed to create one for my own dataset?

Thanks,

[Przemek D]

Mean file allows you to center your data on zero value. You don't have to use it, but if you want to, you must create it yourself from your dataset. Basically you sum up all your images as matrices and take a mean of them. Nvidia DIGITS makes this process much easier - see their repository, particularly this file to see how it's done.