Open rafa-cxg opened 2 years ago
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 https://github.com/SHTUPLUS/PySGG/issues/3 )
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