About the performance on VLCS dataset about jigendg HOT 4 CLOSED

Hi, sorry for the late replies - I no longer have access to all my files, so I'll have to go by memory.

@yunkaiyu I would trust the splits defined inside the JigenDG/data/txt_lists/ for the VLCS splits

@dailenson I think the specific parameters should be: 30 epochs, batch size 128, l r0.001 jigsaw classes 30 jigsaw weight0.5, bias0.8

from jigendg.

Hi, how many times did you run the experiment? We run each experiments multiple times and report the average, that's why you might have found a higher score on a specific run.

from jigendg.

hi, I want to ask what is your script to train the VLCS dataset? I want to train the model on VLCS dataset, but I don't know the parameters settings.

from jigendg.

Hi, may i confirm the train-test split protocol? The VLCS dataset is divided to training set (70%) and a test set (30%) by random selection as reported in your work. However, training set: validation set(crossval): test set $\approx$ 9:1:9 from your VLCS copy. So do we need to re-divide the VLCS (full)? Thanks!

from jigendg.