All Results are 0.0 for ogbg-code2 example

[shuoyinn]

I tried to reproduce result of ogbg-code2, but at every epoch, I got 0.0 for precision, recall and F1 score and for train, validation and test, as the following image shows.

The issue occurred when running in env where torch==1.10.0+cu113, torch-geometric==2.0.4 and ogb==1.3.3

I try running in env where torch==1.8.0+cu111 and torch-geometric==2.0.4 training and evaluating seemed correctly

=====Epoch 1 Training... Iteration: 100%|█████████████████████████████████████████████████████| 3188/3188 [06:03<00:00, 8.78it/s] Average training loss: 3.1347407830255096 Evaluating... Iteration: 100%|█████████████████████████████████████████████████████| 3188/3188 [02:30<00:00, 21.16it/s] Iteration: 100%|███████████████████████████████████████████████████████| 179/179 [00:09<00:00, 17.99it/s] Iteration: 100%|███████████████████████████████████████████████████████| 172/172 [00:09<00:00, 18.35it/s] {'Train': {'precision': 0.11241788732670548, 'recall': 0.05771350526060361, 'F1': 0.07280211499814168}, 'Validation': {'precision': 0.09887364684226672, 'recall': 0.04849786534824006, 'F1': 0.06207377717216873}, 'Test': {'precision': 0.0993560537026912, 'recall': 0.05024431846124028, 'F1': 0.06371399366888705}}

I just want to know why?

[shuoyinn]

ps: torch==1.10.2+cu113 and torch==1.11.0+cu113 both is ok, too

Then maybe it is only 1.10.0 that incurred this bug?

[weihua916]

That's interesting. Maybe you can see what's fixed from 1.10.0 to 1.10.2

[weihua916]

Closing this issue for now, just make sure people use 1.10.2