deepMIB MATLAB: datastoreio: pathlookup: emptyFolder error

Usage & Issues
#1

hello Everyone,
maybe some of you have come across this issue and can point me in the right direction-
i keep getting the error after pressing the “Train” button below.

mib_err
mib_err807×726 53.2 KB

i 've tried matlab and windows version, 2d and 3d networks.
preprocessing, changing depth and other settings. no luck.

the folders do get created in results folder, there is a file in prediction folder, also in validation folders for images and labels (in settings set to 0.25) but the training folders for images and labels are empty.

cheers,
GG

#2

Hi @GG_1,
how many datasets you have?
If you have only a single dataset, it may randomly placed to the validation folder rather than the training folder. If so, change fraction of images for validation to 0 or pick another value for the random seed.

Best regards,
Ilya

#3

thank you Ilya, for your quick response.

yes, we have only 1 set. to start trying things.
setting validation to 0, did help.
but i am concerned now, that if the network is trained and has nothing to use for validation, how can it actually be of a good quality/or other useful parameter?

thank you again

#4

The validation process is allowing you to see whether the network under- or overfits to your data. With that you have an independent measure of its performance, which is nice to report. In addition, it can also be used to automatically stop the training process upon reaching a certain criteria (Patience of validation stopping, in Training settings).
Otherwise, you may easy leave without it. The network will be trained against your data and will be able to predict the similar dataset.

If I have one dataset, I split it into several smaller datasets (MIB->Menu->File->Chopped images->Export) and use one for validation purposes.
As a note, if you are using 64 x 64 x 8 input patches, you may get better results with 3D anisotropic u-net, which starts with 2D convolutions to make it 32 x 32 x 8 and after that continue with 3D steps → 16 x 16 x 4 → 8x8x2…