digits for multi-label data => python data layer?

[RSly]

Hi,

I am trying to train a simple network with two sets of label masks for two different tasks. The goal is to define two separate loss functions that optimize two score maps for each label mask.

My question is very simple:) How can I get this data + labels into digits?

1. I could see that hdf5 can be helpful but is not supported largely in digits...
2. in caffe, one can use a python data layer as the one from FCN-8s in here https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/nyud_layers.py and https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/nyud-fcn32s-color-d/trainval.prototxt

So any chance one can do the same in digits? I guess, we need an option in the 'Select Dataset' to be 'others'

[RSly]

Hi @gheinrich You have a good experience with the data ingest plug-in/methods.
Do you think it can be of interest to have a python data layer for digits? It seems to me that python layer can be much more flexible than the current plug-in method. Also one can easily use the existing python code from say fcn-8s, etc. for multi-label tasks