Ideas for Benchmarking Performance in Mac M1

[reshinthadithyan]

❓ Questions & Help

Benchmark the performance of Pytorch Geometric in Mac M1 Processor[Any processor/environment] for the matter of fact.

cc @rusty1s

[rusty1s]

Besides benchmarking on Mac M1 Processors, it would be great to have a general test bed to benchmark common GNNs on various datasets. We recently released

• standard GNN models
• a profile package

which we can directly make use of to implement this idea.

One idea for this is the following:

We provide a general benchmark function, in which the user inputs a GNN model, a data object/data loader, and a mapping from data attributes to model forward inputs, and receives general profiling information such as average runtime and peak memory consumption:

model = GCN(...)
dataset = TUDataset(..., name='ENZYMES')
loader = DataLoader(dataset, batch_size=128)

benchmark(model, loader, lambda: data: data.x, data.edge_index)

[reshinthadithyan]

Makes sense, I'll work on this.