[GNN] Reference implementation for GNN node classification

[LiSu]

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]

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

[LiSu]

recheck

[Elnifio]

Could we have a library version pinned here so that we can reproduce the exact results even after a few months / years? This would also apply for all other libraries, such as torch_geometric in line 14.

Additionally, it would be very helpful for users if we can integrate these into a Dockerfile, so that all they need to do is to run docker build -f Dockerfile . instead of manually following all the steps.

[LiSu]

Added guidelines for building Docker image using the Dockerfile

[Elnifio]

Minor correction: the validation fraction should be set as 0.005 - $1 \over 40$ of the original validation dataset size. Since the original validation dataset size is ${1 \over 5} = 0.2$ of the full trainable nodes, we should use ${1 \over 5} \times {1 \over 40} = {1 \over 200} = 0.005$ of the full trainable nodes, instead of ${1 \over 40} = 0.025$ of the full trainable nodes, so that the eval cost would be at most 10% for the most basic single-node configuration, similar to other benchmarks.

[LiSu]

Thanks for the correction!

[Elnifio]

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]

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]

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]

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]

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]

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.