Unintuitive --normalize_features option in benchmark scripts

[gau-nernst]

🐛 Describe the bug

In the reference benchmark script for GAT here, the --normalize-features option is probably not behaving what it is intended to do. As noted by the official Python argparse documentation

The bool() function is not recommended as a type converter. All it does is convert empty strings to False and non-empty strings to True. This is usually not what is desired.

In other words,

python gat.py --dataset=Cora --normalize_features=False  # this will silently evaluate to normalize_features=True
python gat.py --dataset=Cora --normalize_features=0  # same problem as above, since "0" is a string
python gat.py --dataset=Cora --normalize_features="" # the only way to turn off feature normalization correctly

One alternative would be

parser.add_argument('--no-normalize_features', action="store_true")
...

dataset = get_planetoid_dataset(args.dataset, not args.no_normalize_features)

It is not a huge bug, but I think it would improve usability for the users, as well as providing better reference scripts.

Environment

• PyG version: 2.0.4
• PyTorch version: 1.11.0
• OS: macOS (ARM)
• Python version: 3.10
• CUDA/cuDNN version: NA
• How you installed PyTorch and PyG (conda, pip, source): PyTorch through conda, PyG through pip
• Any other relevant information (e.g., version of torch-scatter): NA

[rusty1s]

Thanks! It is fixed in #4967.

[gau-nernst]

Thanks for the speedy fix, and also fixing other bool arguments!

I notice that you miss out not for GCN and SGC scripts

pytorch_geometric/benchmark/citation/gcn.py

Line 41 in 845e5d8

dataset = get_planetoid_dataset(args.dataset, args.no_normalize_features)

pytorch_geometric/benchmark/citation/sgc.py

Line 37 in 845e5d8

dataset = get_planetoid_dataset(args.dataset, args.no_normalize_features)

The rest are good!

[rusty1s]

Oh shit, I was too fast. Thanks for checking!