I've rearranged the output so that we can easily read it.
For each experiment in experiments/ there is now a classification and a augmentation folder containing
respectively best_model.txt, best_params.json, report_in.json and report_out.json.
For classification the results are on the classifier being trained on real data, for augmentation on generated.
report_in means the output is on the prediction of "in-domain" data (for classification this is real test data), and report_out means the output is on "out of domain" (for classification this is real test data - the generator wasn't trained on that data).
It follows that augmentation/report_in.json isn't interesting since that's just training on the test data (but it's there for completeness, just ignore it).
The output is json so it can be easily read and plot etc... (I am working on the plots now).
I've rearranged the output so that we can easily read it. For each experiment in
experiments/there is now aclassificationand aaugmentationfolder containing respectivelybest_model.txt,best_params.json,report_in.jsonandreport_out.json.For classification the results are on the classifier being trained on real data, for augmentation on generated. report_in means the output is on the prediction of "in-domain" data (for classification this is real test data), and report_out means the output is on "out of domain" (for classification this is real test data - the generator wasn't trained on that data).
It follows that augmentation/report_in.json isn't interesting since that's just training on the test data (but it's there for completeness, just ignore it).
The output is json so it can be easily read and plot etc... (I am working on the plots now).