Can't Split edges into pos and neg edge_index

[atale480]

I'm trying to build Adversarially Regularised Graph Auto-Encoder. For this I'm using Princeton Segmentation Benchmark as dataset. When I look at the documentation of PyTorch_Grometric for Adversarially Regularised Graph Auto-Encoder I came to know I need pos_edge_index and neg_edge_index. For this I thought of using torch_geometric.utils.train_test_splitedgesmethod. But As this method is going to be deprecated soon and as mentioned in that documentation I'm using torch_geometric.transforms.RandomLinkSplit method. But when I tried to used this method I got following error.

Here's code I wrote

dataset = MyOwnDataset(root = 'data/MeshData/')
dataset = dataset.shuffle()
transform = RandomLinkSplit(is_undirected=True)
train_data, val_data, test_data = transform(dataset)

Error I got

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Input In [11], in <cell line: 2>()
      1 transform = RandomLinkSplit(is_undirected=True)
----> 2 train_data, val_data, test_data = transform(dataset)

File ~/miniforge3/envs/pytorch/lib/python3.8/site-packages/torch_geometric/transforms/random_link_split.py:140, in RandomLinkSplit.__call__(self, data)
    138 else:
    139     rev_edge_types = [None]
--> 140     stores = [data._store]
    141     train_stores = [train_data._store]
    142     val_stores = [val_data._store]

AttributeError: 'MyOwnDataset' object has no attribute '_store'

I've used basic method while creating own dataset like raw_paths, raw_dir, processed_dir, raw_file_names, processed_file_names.
I don't know what is _store attribute. Can anyone let me know what is that? What is that I missed?
Also I want to know if it is possible with decoder that I pass encoder output to decoder and decoder will output 1 dimensional vector? if yes then how to do that?

Environment

• PyG version: 2.1.0.post1
• PyTorch Version: 1.13.0.dev20220707
• PyTorch Scatter Version: 2.0.9
• PyTorch Sparse Version: 0.6.14
• OS: MacOS Ventura 13.1
• Python Version: 3.8.13
• Installation Source( conda, pip): pip

[wsad1]

RandomLinkSplit needs a Data , which is a single graph, object as input.

dataset = MyOwnDataset(root = 'data/MeshData/')
dataset = dataset.shuffle()
from torch_geometric.data import Batch
# Convert from n graphs to one large graph with n disconnected components
data = Batch.from_data_list(dataset[:])
transform = RandomLinkSplit(is_undirected=True)
train_data, val_data, test_data = transform(dataset)

[rusty1s]

Sorry for the confusion. It looks like I was wrong: ÀRGVA` doesn't need any negative edges. Here is a full example on how to use. All you need to change is iterate over your data in mini-batches.

[atale480]

I followed that example but got stuck
Code

import torch_geometric.transforms as T
transform = T.Compose([T.RandomLinkSplit(num_val=0.1, num_test=0.1, is_undirected=True,
                                         split_labels=True, add_negative_train_samples=False),])
dataset = MyOwnDataset(root = 'data/MeshData/', transform=transform)
dataset = dataset.shuffle()
print(dataset[2])
train_data, val_data, test_data = dataset

Output

(Data(x=[5614, 3], edge_index=[2, 26940], y=[1], pos_edge_label=[13470], pos_edge_label_index=[2, 13470]),
 Data(x=[5614, 3], edge_index=[2, 26940], y=[1], pos_edge_label=[1683], pos_edge_label_index=[2, 1683], neg_edge_label=[1683], neg_edge_label_index=[2, 1683]),
 Data(x=[5614, 3], edge_index=[2, 30306], y=[1], pos_edge_label=[1683], pos_edge_label_index=[2, 1683], neg_edge_label=[1683], neg_edge_label_index=[2, 1683]))
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Input In [15], in <cell line: 5>()
      3 dataset = MyOwnDataset(root = 'data/MeshData/', transform=transform)
      4 dataset = dataset.shuffle()
----> 5 train_data, val_data, test_data = dataset

ValueError: too many values to unpack (expected 3)

As you see this is creating 3 data object for each graph. If I try to use train_data, val_data, test_data = transform(batch) then again I get error that dataset has no attribute _store

When I get data in train, val and test with pos and neg edge_label_index then only I can use data loader and then iterate model in mini-batches. Also I tried following code for data loader after getting above error then I did not get pos and neg edge_label_index in that when I tried to print DataLoader.

train_dataset = dataset[:int(len(dataset)*0.8)]
test_dataset = dataset[int(len(dataset)*0.1):]

train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)

Maybe I'm not able to understand what is written GitHub example. But Question I'm getting is that how to split data in train test test and val for given batch size along with pos and neg edge_label_index of each (i.e. pos and neg edge_label_index for train split, val split and test split)

[atale480]

I followed that example but got stuck Code

import torch_geometric.transforms as T
transform = T.Compose([T.RandomLinkSplit(num_val=0.1, num_test=0.1, is_undirected=True,
                                         split_labels=True, add_negative_train_samples=False),])
dataset = MyOwnDataset(root = 'data/MeshData/', transform=transform)
dataset = dataset.shuffle()
print(dataset[2])
train_data, val_data, test_data = dataset

Output

(Data(x=[5614, 3], edge_index=[2, 26940], y=[1], pos_edge_label=[13470], pos_edge_label_index=[2, 13470]),
 Data(x=[5614, 3], edge_index=[2, 26940], y=[1], pos_edge_label=[1683], pos_edge_label_index=[2, 1683], neg_edge_label=[1683], neg_edge_label_index=[2, 1683]),
 Data(x=[5614, 3], edge_index=[2, 30306], y=[1], pos_edge_label=[1683], pos_edge_label_index=[2, 1683], neg_edge_label=[1683], neg_edge_label_index=[2, 1683]))
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Input In [15], in <cell line: 5>()
      3 dataset = MyOwnDataset(root = 'data/MeshData/', transform=transform)
      4 dataset = dataset.shuffle()
----> 5 train_data, val_data, test_data = dataset

ValueError: too many values to unpack (expected 3)

As you see this is creating 3 data object for each graph. If I try to use train_data, val_data, test_data = transform(batch) then again I get error that dataset has no attribute _store

When I get data in train, val and test with pos and neg edge_label_index then only I can use data loader and then iterate model in mini-batches. Also I tried following code for data loader after getting above error then I did not get pos and neg edge_label_index in that when I tried to print DataLoader.

train_dataset = dataset[:int(len(dataset)*0.8)]
test_dataset = dataset[int(len(dataset)*0.1):]

train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)

Maybe I'm not able to understand what is written GitHub example. But Question I'm getting is that how to split data in train test test and val for given batch size along with pos and neg edge_label_index of each (i.e. pos and neg edge_label_index for train split, val split and test split)

My dataset contain 380 graphs and I want to divide 380 graph in train test and val with some batch size and along with this I need edge_label_index for each.

[rusty1s]

I think you would need to call this via train_data, val_data, test_data = zip(dataset).

[atale480]

Thanks for support.
I wrote simple code to get train, val and test data as follows

train_data = []
val_data = []
test_data = []
for i in dataset:
    a,b,c = i
    train_data.append(a)
    val_data.append(b)
    test_data.append(c)