HopSkipJump Attack Performance

[ClarktheDarkShark]

I am currently running the HopSkipJump Attack with a ResNet model on CIFAR10 dataset. I added a counter to each place in the ART hop_skip_jump.py code that I could find an 'estimator.predict' model query to count the model of queries for comparison. Something is not quite adding up right and I cannot figure it out. The paper lists a median L2 = 0.21 for CIFAR10 on a ResNet classifier for 5k model queries, but I am getting much better results.

I would assume that adding += 1 to the counter each time a prediction is conducted will correctly count the queries, but I was hoping to double check that I am not missing something, and also ensure that I correctly understand the parameters. I believe the best way to limit the queries is to use the 'max_iter' parameter, but I do not quite understand what the 'max_eval' parameter is for.

Thanks.