The integration tests depend on config files I generated in a hurry, using default-like hyperparameters. The natural thing to do, though, is to conduct some degree of hyperparameter tuning for each dataset/architecture pair, and then insert the obtained hyperparameters in the configs, and then regenerate the test data. This is both documentary of what "reasonable" hyperparameters are for these tasks, but also I hypothesize that "reasonable" hyperparameters are better for testing, because small changes are more likely to cause detectable regressions in a well-tuned system.
This should be done after #330 is resolved; it also needs #329.
The integration tests depend on config files I generated in a hurry, using default-like hyperparameters. The natural thing to do, though, is to conduct some degree of hyperparameter tuning for each dataset/architecture pair, and then insert the obtained hyperparameters in the configs, and then regenerate the test data. This is both documentary of what "reasonable" hyperparameters are for these tasks, but also I hypothesize that "reasonable" hyperparameters are better for testing, because small changes are more likely to cause detectable regressions in a well-tuned system.
This should be done after #330 is resolved; it also needs #329.