Any import from `torch_geometric` loads `matplotlib` (and a lot more)

🐛 Bug

Because torch_geometric imports * from .nn, which imports * from .models, which imports all of the models, a lot of unnecessary dependencies get loaded whenever anything is imported from torch_geometric.

I came across this behavior because it was filling my logging with matplotlib messages, but even with logging disabled it would be good for performance if heavy libraries that do file I/O like matplotlib are only loaded if they are needed. This can be particularly important for start-up performance on clusters with networked filesystems.

To Reproduce

import logging
logging.basicConfig(level=logging.DEBUG)

import torch
from torch_geometric.data import Data

d = Data(x=torch.zeros(10, 3))
print(d)

Gives many pages of matplotlib debug logging, followed by the expected

Data(x=[10, 3])

Expected behavior

That matplotlib (and other libraries that aren’t necessary for basic functions) won’t be loaded for any import from torch_geometric, giving an output of just

Data(x=[10, 3])

Environment

• OS: Linux
• Python version: 3.8.5 / Anaconda
• PyTorch version: 1.7.1
• CUDA/cuDNN version: N/A
• GCC version: N/A
• Any other relevant information: N/A

[rusty1s]

You are right! This should be fixed in master. Thanks!

[Linux-cpp-lisp]

Fully resolved in #2199.