Hi @JackBrady, @zimmerrol
I am having difficulty reproducing Figure 4A of the article using the code provided. My procedure was to use the command line python train_model.py --data synth --num_slots --lam --dependent 0 --lr 1e-3 --num_iters 115000 as per the README, replacing the parameters of num_slots and lam as explained in the article, and finally plotting the values of Reconstruction Error, Compositional Contrast, and Slot Identifiability Score. Unfortunately, I did not obtain a graph similar to the one presented in the article (Image below), even when I normalized the Compositional Contrast according to section B.3.
I would like to know if you guys can provide more detailed information on how to reproduce the results or provide a simple code for reproduction?
Regards,
Bruno Silva
Hi @JackBrady, @zimmerrol
I am having difficulty reproducing Figure 4A of the article using the code provided. My procedure was to use the command line
python train_model.py --data synth --num_slots --lam --dependent 0 --lr 1e-3 --num_iters 115000as per the README, replacing the parameters ofnum_slotsandlamas explained in the article, and finally plotting the values of Reconstruction Error, Compositional Contrast, and Slot Identifiability Score. Unfortunately, I did not obtain a graph similar to the one presented in the article (Image below), even when I normalized the Compositional Contrast according to section B.3.I would like to know if you guys can provide more detailed information on how to reproduce the results or provide a simple code for reproduction?
Regards,
Bruno Silva