lgnn training time
yCobanoglu opened this issue · 2 comments
yCobanoglu commented
Generating the lgnn adjacencies for the hierachy takes forever. np.dot(QQ, QQ) in file load.py line 19 is the bottleneck. This is a very large matmul 6000 times (for 6000 training samples) and only possible to perform on the cpu (due to vram constraints) ?
How long does it take to train the lgnn ? How did you pull it off for the paper ?
zhengdao-chen commented
It's been a few years since I trained the models and I don't remember exactly how long it took. The experiments did take a while but would eventually finish... Are these graphs generated under the same configurations as described in the paper, for example?
yCobanoglu commented
Ok good to know thanks !