zfjsail / gae-pytorch Goto Github PK

In VAE, sampling is z_mean + torch.exp(0.5 * z_log_var) * epsilon , but why is z_mean + torch.exp( z_log_var) in VGAE, does it cause anything different?

When I train with Cora dataset, I get the following error in binary_cross_entropy_with_logits. Shouldn't pos_weight be a Tensor? Thanks!

Traceback (most recent call last):
  File "train.py", line 83, in <module>
    gae_for(args)
  File "train.py", line 62, in gae_for
    norm=norm, pos_weight=pos_weight)
  File "/gae-pytorch/gae/optimizer.py", line 7, in loss_function
    cost = norm * F.binary_cross_entropy_with_logits(preds, labels, pos_weight=pos_weight)
  File "/anaconda3/lib/python3.6/site-packages/torch/nn/functional.py", line 2077, in binary_cross_entropy_with_logits
    return torch.binary_cross_entropy_with_logits(input, target, weight, pos_weight, reduction_enum)
TypeError: binary_cross_entropy_with_logits(): argument 'pos_weight' (position 4) must be Tensor, not numpy.float64

KLD = -0.5 / n_nodes * torch.mean(torch.sum(1 + 2 * logvar - mu.pow(2) - logvar.exp().pow(2), 1))
/ n_nodes should be removed or
torch.mean → torch.sum

In the code in utils.py, the function get_roc_score() has this line:

labels_all = np.hstack([np.ones(len(preds)), np.zeros(len(preds))])

The second argument should be np.zeros(len(preds_neg))

I was using your code to understand GVAE and found this small mistake. Thank you.

self.dc = InnerProductDecoder(dropout, act=lambda x: x)
Why not use act=torch.sigmoidhere?

How can I implement a mini-batch version of GVAE?

I found that in train.py mu.data.numpy() is used to get hidden_emb, but it would get None when using GCNModelAE as model, hidden_emb should be got from model.encode() instead.

Can I use this framework on my own custom datasets?

Also, I'm experiencing an issue, when I try to run:
python gae/train.py

I get:

Traceback (most recent call last):
File "gae/train.py", line 12, in
from gae.model import GCNModelVAE
ModuleNotFoundError: No module named 'gae'

Yet this command works interactively.

May I know why adding self loop to adj_train to get adj_label?

adj_label = adj_train + sp.eye(adj_train.shape[0])