node2vec example gives the same loss always

[MohamedEldafrawi2]

Hi,
I was trying to run the node2vec example but I am always getting the same loss regardless how many epochs i have.
Epoch: 01, Loss: 1.3863, Acc: 0.1300
Epoch: 02, Loss: 1.3863, Acc: 0.1300
Epoch: 03, Loss: 1.3863, Acc: 0.1300
Epoch: 04, Loss: 1.3863, Acc: 0.1300
Epoch: 05, Loss: 1.3863, Acc: 0.1300
Epoch: 06, Loss: 1.3863, Acc: 0.1300
Epoch: 07, Loss: 1.3863, Acc: 0.1300
Epoch: 08, Loss: 1.3863, Acc: 0.1300
Epoch: 09, Loss: 1.3863, Acc: 0.1300
Epoch: 10, Loss: 1.3863, Acc: 0.1300
Epoch: 11, Loss: 1.3863, Acc: 0.1300
Epoch: 12, Loss: 1.3863, Acc: 0.1300
Epoch: 13, Loss: 1.3863, Acc: 0.1300
Epoch: 14, Loss: 1.3863, Acc: 0.1300
Epoch: 15, Loss: 1.3863, Acc: 0.1300
Epoch: 16, Loss: 1.3863, Acc: 0.1300
Epoch: 17, Loss: 1.3863, Acc: 0.1300
Epoch: 18, Loss: 1.3863, Acc: 0.1300
Epoch: 19, Loss: 1.3863, Acc: 0.1300
Epoch: 20, Loss: 1.3863, Acc: 0.1300
Epoch: 21, Loss: 1.3863, Acc: 0.1300
Epoch: 22, Loss: 1.3863, Acc: 0.1300
Epoch: 23, Loss: 1.3863, Acc: 0.1300
Epoch: 24, Loss: 1.3863, Acc: 0.1300

Could you please tell me what is the reason behind that?

[EdisonLeeeee]

I guess there must be something wrong with your training process, e.g., the model parameters are not updated during training. Could you please show more details so that I can dig into it?

[MohamedEldafrawi2]

I was trying to debug it and I found out that the Embedding layer always returns zeros! inside (Loss) method >>Node2Vec class

[EdisonLeeeee]

Interesting. I can indeed reproduce this bug only on the Windows machine. Will figure it out and report back.

[MohamedEldafrawi2]

Thanks!

[EdisonLeeeee]

This is a bug caused by PyTorch Dataloader when num_workers>0, as also mentioned in pytorch/pytorch#51344. To resolve this bug, just set num_workers=0 in line 22, i.e.,

 loader = model.loader(batch_size=128, shuffle=True, num_workers=0)

CC @rusty1s

[MohamedEldafrawi2]

Now it works!
Thanks a lot.