Transfer learning: can't justify small accuracy during model training

Question

[Setting context]

Working on a CIFAR-10 similar task, but instead of that data I use custom data with 15 classes, namely:

classes = [\'b