Closed lingfanyu closed 4 years ago
Hi:
The hyper-parameter on the paper is --n_hid 256 --n_layer 3 . If the memory permits, the higher sample_width, and sample_depth are the better.
I tried such setting:
'' python train_paper_field.py --n_hid 256 --n_layer 3 --batch_size 512 --sample_width 256 ''
and get the result:
Best Test NDCG: 0.4434 Best Test MRR: 0.5164
Thanks for the reply!
In the paper, for paper_venue classification on graph_CS, the reported NDCG is .473±.054, and MRR is .288±.088 (0.5164 might be a typo?). What was the sample_width and sample_depth to achieve this result? And how much CPU/GPU memory did you have?
Oh sorry. The result I send is for paper_field. As reported, sample_width=256 and sample_depth=6.
The GPU is a single K80, and memory is 220 GB.
Hi, could you share the hyper-parameters you used to train paper-venue and paper-field for graph_CS which can reproduce results in the paper?
I downloaded graph_CS.pk and ran
python train_paper_venue.py
which used default parameters in the code. The results I got was: Best Test NDCG: 0.4332 Best Test MRR: 0.2547Thanks!