Very cool. I'm surprised ResNet50 had so much of an advantage over ResNet34 with only 660 samples.
I'm also very curious what the performance of some of the newer architectures, i.e. capsulenets would look like.
Ideally, you have a held-out test which the model hasn't seen and only used after model has trained and tuned on the dev and validation sets. Often bad experimental models will repeatedly use the test set in fine-tuning an existing model which may result in your model learning about the test set rendering the test set useless.
In the practice real world results may vary as your training and test data may not represent the actual distribution of the real world data.