How to replicate the results shown in the paper?

[MathsShen]

Very nice work! But I'm wondering what the experimental settings are that can replicate the results shown in your paper (which is ~90% MCC performance).

If I understand it the right way, I run the following command which is exactly the settings that you described in the paper (i.e. M=40, L=1000, DimOfHiddenLayers={50, 100, 200}, #layers={3, 4, 5, 6}, lr={1e-2, 1e-3}, n=d=5, NumberOfMixingLayers=3.)

nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 50 -d 4 -l 1e-3 -c > log2.txt 2>&1 &
nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 50 -d 5 -l 1e-3 -c > log3.txt 2>&1 &
nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 50 -d 6 -l 1e-3 -c > log4.txt 2>&1 &
nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 100 -d 3 -l 1e-3 -c > log5.txt 2>&1 &
nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 100 -d 4 -l 1e-3 -c > log6.txt 2>&1 &
nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 100 -d 5 -l 1e-3 -c > log7.txt 2>&1 &
nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 100 -d 6 -l 1e-3 -c > log8.txt 2>&1 &
nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 200 -d 3 -l 1e-3 -c > log9.txt 2>&1 &
nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 200 -d 4 -l 1e-3 -c > log10.txt 2>&1 &
nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 200 -d 5 -l 1e-3 -c > log11.txt 2>&1 &
nohup python main.py -x "1000_40_5_5_3_1_gauss_xtanh_u_f" -g 200 -d 6 -l 1e-3 -c > log12.txt 2>&1 &

They all gave me the poor results (~40% MCC), which is far from ~90%, however. Do you mind providing the exact experimental settings for reproducing the results you described in the paper?

Thanks a lot!

[MathsShen]

(Continued with the last post)

The following two are exemplary results I obtained:

{"id": 1, "elbo": 1.4046511459350586, "perf": 0.4658019185241058, "metadata": {"path": "data/tcl_1000_40_5_5_3_1_gauss_xtanh_u.npz", "file": "data/tcl_1000_40_5_5_3_1_gauss_xtanh_u.npz", "latent_dim": 5, "batch_size": 64, "epochs": 20, "max_iter": null, "hidden_dim": 50, "depth": 3, "lr": 0.001, "seed": 1, "cuda": true, "preload": false, "anneal": false, "log_freq": 25, "nps": 1000, "ns": 40, "n": 40000, "data_dim": 5, "aux_dim": 40}}

{"id": 2, "elbo": 8.97379638671875, "perf": 0.3602966198650229, "metadata": {"path": "data/tcl_1000_40_5_5_3_1_gauss_xtanh_u_n.npz", "file": "data/tcl_1000_40_5_5_3_1_gauss_xtanh_u_n.npz", "latent_dim": 5, "batch_size": 64, "epochs": 20, "max_iter": null, "hidden_dim": 50, "depth": 3, "lr": 0.001, "seed": 1, "cuda": true, "preload": false, "anneal": false, "log_freq": 25, "nps": 1000, "ns": 40, "n": 40000, "data_dim": 5, "aux_dim": 40}}

Could you please show me your detailed settings for ~0.90-0.95 MCC as stated in the paper?