search

mlcommons / training

Reference implementations of MLPerf™ training benchmarks
https://mlcommons.org/en/groups/training
Apache License 2.0
1.57k stars 548 forks source link

[GNN] Reference implementation for GNN node classification #700

Closed LiSu closed 3 months ago

LiSu commented 4 months ago

In this PR we (Alibaba, Intel & Nvidia) propose a GNN training benchmark, which is a multi-class node classification task in a heterogenous graph using the IGB Heterogeneous Dataset named IGBH-Full. The task is carried out using a GAT model based on the Relational Graph Attention Networks paper.

github-actions[bot] commented 4 months ago

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

LiSu commented 4 months ago

recheck

Elnifio commented 4 months ago

Is it possible to also update the file name from gnn_node_classification to graph_neural_network so that the folder name refers to the domain name (GNN) but not the overly detailed task name (node classification), which is similar to all other benchmarks?

LiSu commented 4 months ago

Is it possible to also update the file name from gnn_node_classification to graph_neural_network so that the folder name refers to the domain name (GNN) but not the overly detailed task name (node classification), which is similar to all other benchmarks?

The folder is renamed as graph_neural_network

Elnifio commented 3 months ago

As discussed in the MLLogging PR, could we also add gradient accumulation step (1 in our current case) and optimizer name (Adam in our case) to MLLog outputs?

LiSu commented 3 months ago

As discussed in the MLLogging PR, could we also add gradient accumulation step (1 in our current case) and optimizer name (Adam in our case) to MLLog outputs?

Added gradient accumulation step and optimizer name to MLLog outputs ;-)

Elnifio commented 3 months ago

As discussed in the MLLogging PR, could we also add gradient accumulation step (1 in our current case) and optimizer name (Adam in our case) to MLLog outputs?

Added gradient accumulation step and optimizer name to MLLog outputs ;-)

Just noticed that the checker is asking for "adam" instead of "Adam". Could we have this small fix checked in so that the reference is consistent with the compliance checker?

LiSu commented 3 months ago

As discussed in the MLLogging PR, could we also add gradient accumulation step (1 in our current case) and optimizer name (Adam in our case) to MLLog outputs?

Added gradient accumulation step and optimizer name to MLLog outputs ;-)

Just noticed that the checker is asking for "adam" instead of "Adam". Could we have this small fix checked in so that the reference is consistent with the compliance checker?

Fixed in the last commit.