项目文件夹

文件

T

文件历史

Quan (Andy) Gan d41d07d0f6 [Doc and bugfix] Add docs and user guide and update tutorial for sampling pipeline (#3774 )

* huuuuge update

* remove

* lint

* lint

* fix

* what happened to nccl

* update multi-gpu unsupervised graphsage example

* replace most of the dgl.mp.process with torch.mp.spawn

* update if condition for use_uva case

* update user guide

* address comments

* incorporating suggestions from @jermainewang

* oops

* fix tutorial to pass CI

* oops

* fix again

Co-authored-by: Xin Yao <xiny@nvidia.com>

2022-02-28 02:21:51 +08:00

scripts

[Example]Add sparse embedding in GATNE-T example to support big graph (#2234 )

2020-10-10 17:22:45 +08:00

src

[Doc and bugfix] Add docs and user guide and update tutorial for sampling pipeline (#3774 )

2022-02-28 02:21:51 +08:00

README.md

[Feature] Add left normalizer for GCN (#3114 )

2021-07-13 12:49:35 +08:00

requirements.txt

[Model] GATNE-T (#1470 )

2020-05-05 22:35:25 -07:00

README.md

Representation Learning for Attributed Multiplex Heterogeneous Network (GANTE)

Paper link: https://arxiv.org/abs/1905.01669
Author's code repo: https://github.com/THUDM/GATNE. Note that only GATNE-T is implemented here.

Requirements

requirements

pip install -r requirements.txt

Also requires PyTorch 1.7.0+.

Datasets

To prepare the datasets:

```
mkdir data
cd data
```
Download datasets from the following links:
Unzip the datasets

Training

Run with following (available dataset: "example", "youtube", "amazon")

python src/main.py --input data/example

To run on "twitter" dataset, use

python src/main.py --input data/twitter --eval-type 1 --gpu 0

For a big dataset, use sparse to avoid cuda out of memory in backward

python src/main_sparse.py --input data/example --gpu 0

If you have multiple GPUs, you can also accelerate training with DistributedDataParallel

python src/main_sparse_multi_gpus.py --input data/example --gpu 0,1

It is worth noting that DistributedDataParallel will cause more cuda memory consumption and a certain loss of preformance.

Results

All the results match the official code with the same hyper parameter values, including twiiter dataset (auc, pr, f1 is 76.29, 76.17, 69.34, respectively).

	auc	pr	f1
amazon	96.88	96.31	92.12
youtube	82.29	80.35	74.63
twitter	72.40	74.40	65.89
example	94.65	94.57	89.99