When I put small SVHN data (50 samples to be exact) to a CNN, the model does not update and always gets stuck at outputting same class for all validation dataset.
This does not happen for CIFAR10 dataset, which works well with the same CNN architecture and data size.
Also, not just the small size of 50, a lot of times the update does not happen for sizes like 2000, 4000. And if I re-run the training from scratch for some times (after acknowledging the training failure), then model update after few re-run iterations.
Does anybody has encountered this for SVHN?
Thanks a lot