Deepwalk is an algorithmic framework for representational learning on graphs. Given any graph, it can learn continuous feature representations for the nodes, which can then be used for various downstream machine learning tasks. Based on PGL, we reproduce distributed deepwalk algorithms and reach the same level of indicators as the paper.
The datasets contain two networks: BlogCatalog.
- paddlepaddle>=2.0rc
- pgl>=2.0
We adopt PaddlePaddle Fleet as our distributed training frameworks config.yaml is config file for deepwalk hyperparameter. In distributed CPU mode, we have 2 pservers and 2 trainers. We can use fleetrun to help you startup the parameter servers and model trainers.
For examples, train deepwalk mode on BlogCataLog dataset.
# train deepwalk in CPU mode.
python train.py
# train deepwalk in single GPU mode.
CUDA_VISIBLE_DEVICES=0 python train.py --use_cuda
# train deepwalk in multiple GPU mode.
CUDA_VISIBLE_DEVICES=0,1 fleetrun train_distributed_gpu.py
# train deepwalk in distributed CPU mode.
CPU_NUM=2 fleetrun --mode ps --worker_num 2 --server_num 2 train_distributed_cpu.py
# multiclass task example
python multi_class.py
- dataset: The citation dataset "BlogCatalog".
- conf: The model config file, default is
./config.yaml. - epoch: Number of training epoch.
| Dataset | model | Task | Metric | PGL Result | Reported Result |
|---|---|---|---|---|---|
| BlogCatalog | distributed deepwalk | multi-label classification | MacroF1 | 0.233 | 0.211 |