namkyeong / bgrl_pytorch

Pytorch implementation of "Large-Scale Representation Learning on Graphs via Bootstrapping"

BGRL_Pytorch

Implementation of Large-Scale Representation Learning on Graphs via Bootstrapping.

A PyTorch implementation of "Large-Scale Representation Learning on Graphs via Bootstrapping" paper, accepted in ICLR 2021 Workshop

Hyperparameters for training BGRL

Following Options can be passed to train.py

--layers: or -l: one or more integer values specifying the number of units for each GNN layer. Default is 512 256.
usage example :--layers 512 256

--aug_params: or -p: four float values specifying the hyperparameters for graph augmentation (p_f1, p_f2, p_e1, p_e2). Default is 0.2 0.1 0.2 0.3.
usage example : --aug_params 0.2 0.1 0.2 0.3

params	WikiCS	Am.Computers	Am.Photos	Co.CS	Co.Physics
p_f1	0.2	0.2	0.1	0.3	0.1
p_f2	0.1	0.1	0.2	0.4	0.4
p_e1	0.2	0.5	0.4	0.3	0.4
p_e2	0.3	0.4	0.1	0.2	0.1
embedding size	256	128	256	256	128
encoder hidden size	512	256	512	512	256
predictor hidden size	512	512	512	512	512

• Hyperparameters are from original paper

Experimental Results

WikiCS	Am.Computers	Am.Photos	Co.CS	Co.Physics
79.50	88.21	92.76	92.49	94.89

Codes borrowed from

Codes are borrowed from BYOL and SelfGNN

name	Implementation Code	Paper
Bootstrap Your Own Latent	Implementation	paper
SelfGNN	Implementation	paper

bgrl_pytorch's Issues

cora，citeseer，PubMed都有这个情况

Traceback (most recent call last):
File "train.py", line 176, in
main()
File "train.py", line 172, in main
train_eval(args)
File "train.py", line 166, in train_eval
trainer.train()
File "train.py", line 44, in train
dev_best, dev_std_best, test_best, test_std_best = self.evaluate()
File "train.py", line 139, in evaluate
logits, loss = classifier(self._embeddings[self._train_mask], self._labels[self._train_mask])
File "/home/ps/anaconda3/envs/bgrl/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/ssd/export/jiajie/BGRL_Pytorch-main/models.py", line 125, in forward
loss = self.cross_entropy(logits, y)
File "/home/ps/anaconda3/envs/bgrl/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/home/ps/anaconda3/envs/bgrl/lib/python3.8/site-packages/torch/nn/modules/loss.py", line 1150, in forward
return F.cross_entropy(input, target, weight=self.weight,
File "/home/ps/anaconda3/envs/bgrl/lib/python3.8/site-packages/torch/nn/functional.py", line 2846, in cross_entropy
return torch._C._nn.cross_entropy_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index, label_smoothing)
RuntimeError: Expected target size [1, 6], got [1, 3327]

About the PPI performance

Hi, thank you for the great work. I wonder whether there is code for reproducing the PPI performance of BGRL (70.49 reported in Table 6).

RuntimeError: running_mean should contain 256 elements not 512

Dear BGRL authors,

I try to run your open source code, but encounter RuntimeError: running_mean should contain 256 elements, not 512.

Could you please check the code and see why this issue happened?

Best!

Unable to process dataset

Dear authors,

Except the wikiCS, other datasets encounter an issue: AttributeError: 'GlobalStorage' object has no attribute 'train_mask'.
Do you have any suggestions? I suppose the issue comes from pyg