I made those charts “manually” by saving the terminal log from train.py, and then opening it in Excel as CSV and skipping the lines that did not have the epoch’s top1/top5 (since the number of iterations per epoch is constant). It wasn’t made automatically by matplotlib, ect.
When you train your model, what is the top1/top5 accuracy it ends up with? Are you using a different dataset, or replicating the PlantCLEF dataset.
Also note that ResNet is more accurate than GoogleNet, it is recommended to use ResNet for training new models.