Fine-tuning EfficientNetB0 for 104 flower classes: validation vs held-out test scores, honestly

I built a GPU transfer-learning workflow for 104-class flower recognition and tried to report it...