Currently, when we run an experiment, the output is stored in a folder that is defined by the name attribute in Experiment. If we run the same experiment again with the same output folder, the folder is partially overwritten, which can lead to confusing situations.
Example: I ran an experiment for sensitive attribute A with four categories: 1, 2, 3 and 4. So, 4 different folders were created for each of those categories. Then, I ran the same experiment with sensitive attribute B with two categories, 1 and 2. Since the output folder name was the same for both experiments, the second experiment overwrote the folders for categories 1 and 2 but retained the categories 3 and 4 from the last experiment. This led to incorrect plots when plotting the results.
Ideally, we should have an attribute overwrite that's False by default. If it's False, the experiment should throw an error if the given folder name already exists. If it's True, we should delete the contents of the output folder before running the experiment to prevent data from two experiments being mixed up.
Currently, when we run an experiment, the output is stored in a folder that is defined by the
nameattribute inExperiment. If we run the same experiment again with the same output folder, the folder is partially overwritten, which can lead to confusing situations.Example: I ran an experiment for sensitive attribute
Awith four categories:1,2,3and4. So, 4 different folders were created for each of those categories. Then, I ran the same experiment with sensitive attributeBwith two categories,1and2. Since the output folder name was the same for both experiments, the second experiment overwrote the folders for categories1and2but retained the categories3and4from the last experiment. This led to incorrect plots when plotting the results.Ideally, we should have an attribute
overwritethat'sFalseby default. If it'sFalse, the experiment should throw an error if the given folder name already exists. If it'sTrue, we should delete the contents of the output folder before running the experiment to prevent data from two experiments being mixed up.