Thx for your job. I wonder what's suitable to train model, specially in what type of gpu and how big of gpu memory. I trained a few times in quadro 6000 but the log said sigkill problem in second epoch. about cofced

[nievuelo]

I mean, the first epoch seems work well, but in the second epoch and during the mid time of the seonc batch like 4660/5033, my thread will be killed and the pycharm tips are like:"interrupted by signal 9: sigkill" I check my log in linux and it said cpu out of memory, But my main memory of linux is 64g, I think it might be enough. I don't know what causes this kind of error.

[Nicozwy]

I think that the CPU memory is too small because we have to load many reports.
Have you tried loading a small portation of reports for each claim?

[nievuelo]

Thanks a lot for your generous help. I will star this project.
But I mean it's not the storage capacity 199g, it's main memory capacity 64g. So I revised the eval_exp_fc5 from report each claim=12->report_each_claim=6, and change the report _each_claim from 30 to 15 in train_exp_fc5_liar_raw2. But it seems still throw kill sig problem. But the difference is before I change the parameter, it would crash during training time but after I revised it, it would crash during evaluation time. But I just had changed the parameter of report_each_claim in evaluate_model func from eval_exp_fc5. So could you help me to figure it out?

[nievuelo]

Or in other words， how big of the required memory is?

[Nicozwy]

Hi, @nievuelo . I can not figure it out with limited information, but you can try this code on the other machine with GPU 3090 because we have successfully tested it.