search

shenyangHuang / TGB

Temporal Graph Benchmark project repo
https://tgb.complexdatalab.com/
182 stars 36 forks source link

Model weights and optimizer state not reset between runs #88

Closed databaaz closed 3 months ago

databaaz commented 5 months ago

In examples/linkproppred, the models carry forward their optimal weights from one run to the next. reset_parameters() need to be called on the learnable modules of the model i.e. memory, gnn and link_pred in case of tgn. Similarly, the optimizer needs to be reinitialized at the beginning of each run.

shenyangHuang commented 5 months ago

Thanks for this point Shahbaz, in the paper we actually didn't use num_runs to run repeated trials (we do single runs (num_runs=1 and just explicitly set the single seed). Thus we didn't encounter this issue before. Yeah, we should update the script to reset the model weight etc.

Best, Andy

shenyangHuang commented 3 months ago

this is addressed in PR #94