Place a folder of images here for each class you want. The training scripts will automatically separate train/test/validation from each folder and will generate a new label for each folder. The labels will be written to the output model directory. Consult the .cmd and .sh training scripts in the training folder for full usage in the event that you are invoking the python scripts directly.
You should pre-resize your images to the target network input size. Otherwise, your training times will be increased by several hours due to the preprocessing expense and gaining nothing from it.
*nix users can just make symbolic links for each class here to avoid copying.
Windows users can, with an admin command prompt, create symbolic links as well with the MKLINK command. Like so:
mklink /J link_name C:\real\folder\path