search

FenTechSolutions / CausalDiscoveryToolbox

Package for causal inference in graphs and in the pairwise settings. Tools for graph structure recovery and dependencies are included.
https://fentechsolutions.github.io/CausalDiscoveryToolbox/html/index.html
MIT License
1.08k stars 198 forks source link

problem about the SAM model #61

Closed Wangt-CN closed 3 years ago

Wangt-CN commented 4 years ago

Hi, thanks for your great work, I have tried the SAM model on a 90 variables graph estimation. However, after an hour, I found the loss become to nan (It's NOT always nan, sometimes seemed regularly)

Is that no problem? and need I change the lr or any parameters? (I just use the default parameters)

commond: obj = SAM(gpus=3, njobs=6,nruns=16,batchsize=1024) output = obj.predict(data)

output: 397/11000 [2:52:25<76:44:58, 26.06s/it, disc=nan, gen=nan, regul_loss=nan, tot=nan] 402/11000 [2:52:16<75:41:45, 25.71s/it, disc=6.07, gen=-0.992, regul_loss=0.49, tot=-78.9]

ritik99 commented 4 years ago

Hello @Wangt-CN ,

Would it be possible for you to share the dataset you run the model on as well?

Thanks.

diviyank commented 4 years ago

Hi @Wangt-CN, Try setting the DAG constraint to 0 to see if the training stabilizes ; We noticed that the No-Tears might bring some instability in the loss. Best, Diviyan

diviyank commented 3 years ago

The datasets used in the paper are available in the cdt.data.load_dataset function, please check https://fentechsolutions.github.io/CausalDiscoveryToolbox/html/data.html#cdt.data.load_dataset

I will close this issue for inactivity, don't hesitate to reopen it if the issue persists or if you have more questions.

Best, Diviyan