Number recognition with SVHN dataset(full version)

I want to build a LPR model and need to test some of my hypothesis and ideas, so i started with SVHN full dataset(pruned to 4 digit numbers.)
My simplest model is resnet-18 split before FC layers, and added 2 Hidden layers of 256 on top of that and will finally 4 different heads of 11 classes ,concated and thus resulting in output shape - (batch_size,4,11).
Here 2nd dimension(4) represent each of the 4 number,
3rd dimension(11) represents – 10 (0-9)numbers + 1 no number.
I have used Imagelist here,but dont know which label function to use. I tried to use .label_from_df and .label_from_func and returned array like [9,0,10,10](label as 90), but the API takes it as multilabel classification i believe, outputting one-hot encoded(len - 11).
Can anyone please guide me in the right direction!!