It seems like there are two bugs in the master branch:

1. unlike the other training parameters, the training method, method, is not exported when executing reset_all() (line 616);
2. the bound on the number of popping images in the "Example predictions on Test set" section does not work: images keep popping until Chrome eats all the memory (lines 452-453).