Pretrained models & Labels/binaryproto (noobie question)

[Dourado]

Hi there!
As mentioned on subject, this is a really noobie question since I probably didn't notice what I am missing here.

When I browse through pretrained models on zoo, I see that they usually provide the caffemodel, deploy, solver and the train_val files. However, I based my implementation on the C++ Classifier and it uses both label and mean (binaryproto) files to classify images. How do I get/generate the mean (binaryproto) file? Is the label file the same one that the model was designed?

Btw, my own model was trained on nvidia DIGITS as it generated all the files for the C++ classifier.

Thank you so much.

[Przemek D]

The train_val and deploy files define the network architecture. solver defines method of training of that network. caffemodel contains weights of a trained model. Those define a particular net.
label and mean files on the other hand are related to the dataset the network was (or is going to be) trained. Label simply contains names of classes in a dataset and mean is a mean image of all images in it. Which label/mean files will you use only depends on your particular task. If you want to use a model to classify images related to the ones that the model was trained on - you probably want to use label/mean of the original dataset. If you want to fine-tune on a new set, you should obtain new label/mean files for it. Caffe has a tool to create a mean file of a compiled dataset (that is, in LMDB database), see compute_image_mean.
Additionally, if you're using DIGITS, your dataset job folder should contain both label and mean files.