model VCTREE training too slow

[rafa-cxg]

when i train vctree predictor, i find that train 50 iters need 30mins!, which normally been between 2-5min. I do not know why.

Thank you

[Scarecrow0]

Hi,
For vctree, I simply adopt the implementation from Scene-Graph-Benchmark. It does slow than other model.
You can try to comment the torch.autograd.set_detect_anomaly(True) in line 58 of tools/relation_train_net.py. It may slow down backward in some versions of Pytorch. (Refer from this issue #3 )