Training process is too slow

[hongzimao]

Training process is too slow

[hongzimao]

Thanks for carefully reading our paper. The 1.5 sec is for the very initial training rounds in the curriculum learning process (section 5.3 challenge 1, presentation slides 22 - 25). For simplicity the training code here didn't start with the very short episode length. You can reduce the initial episode length by reducing num_stream_jobs (curriculum learning is also controlled by reset_prob, reset_prob_decay, reset_prob_min, num_stream_jobs_grow and num_stream_jobs_max)

Nonetheless, ~30s per round checks out with what we observed for this codebase. In our experience 200 iterations (~16 hours) should give you something better than the baseline schemes (we usually let the program run overnight). Here's an example trained model #12

The 160s GPU time is a little weird. We observed GPU runtime to be on par with the CPU time. Honestly we didn't optimize the code efficiency that much (we just did the basic things like making sure the device memory is enough to hold all the sparse matrices for GNNs). If you find ways to optimize the code in any method, please do let use know! We can perhaps include your update in a pull request. Thanks!