Reproducing run_sort results

[eisenjulian]

Hello there, I really loved the paper, thank you so much for sharing and open-sourcing the code! I have been trying to reproduce some of the results and but I am having trouble with the run_sort ones. Getting between 10 and 20% absolute difference in prob any correct and prob all correct.

I suspect I am not picking the hyper-params correctly, but tried to do everything as described in the paper. Currently using the following:

NUM_EPOCHS=200
LR=1e-4
L=4 # Number of images to be concatenated
M=20 # Batch size
TAUS = [1, 2, 4, 6, 8, 16]

Any ideas if any of those is wrong? Thanks again