How to reduce the memory use in big graph

coco11563 · May 28, 2020, 9:25am

I’m trying to run relgraphcov and link prediction example
the dataset I used is FB15K

entities: 14951

relations: 1345

edges: 483142

my problem came when the model comes to eval phase and the pytorch need to allocate 275GB cuda memory or cpu memory

I see that you guys added some new feature in DGL 3.0 (some awesome fuse msg passing)
and in your code (in example) the information passing function is self-define bdd-norm function or self define basis-norm funciton

is there any way to reduce the memory usage? (like modify the message passing function to obey some rules or other)

BarclayII · May 31, 2020, 4:00am

We have two RGCN examples: one implemented with homogeneous graph and another with heterogeneous graph. Which example were you running?

Thanks.

coco11563 · June 6, 2020, 2:18am

github.com

dmlc/dgl/blob/master/examples/pytorch/rgcn/link_predict.py

"""
Modeling Relational Data with Graph Convolutional Networks
Paper: https://arxiv.org/abs/1703.06103
Code: https://github.com/MichSchli/RelationPrediction

Difference compared to MichSchli/RelationPrediction
* Report raw metrics instead of filtered metrics.
* By default, we use uniform edge sampling instead of neighbor-based edge
  sampling used in author's code. In practice, we find it achieves similar MRR
  probably because the model only uses one GNN layer so messages are propagated
  among immediate neighbors. User could specify "--edge-sampler=neighbor" to switch
  to neighbor-based edge sampling.
"""

import argparse
import numpy as np
import time
import torch
import torch.nn as nn
import torch.nn.functional as F

This file has been truncated. show original

this code
the dataset is fb15k

and I see u provided a wrapped function to load fb15k data
can you tell me how did you guys split the rdf data to test set\ train set \ validation set?

BarclayII · June 6, 2020, 4:37pm

I ran this command on nightly build and it ran successfully with 3GB CPU memory and 6.6GB GPU memory:

python3 link_predict.py -d FB15k --gpu 0 --eval-protocol=filtered

Could you tell us what your DGL version was?