Some confusion about the speed of DGCNN

From now on, we recommend using our discussion forum (https://github.com/rusty1s/pytorch_geometric/discussions) for general questions.

❓ Questions & Help

Thank you for the code. I have some confusions about the speed of DGCNN. I found that the speed is much slower than this implementation. I used the batch parameter as you said in this issue. However, the speed is still very slow (about 6s a batch in inference). Besides, to my surprise, the speed of training (about 2s a batch) is faster than inference. Could you know how to slove it?

[rusty1s]

A batch of None refers to batch_size==1.

It can be more computationally efficient to out-source the kNN calculation to the CPU (dependent on your batch_size and num_nodes). This is just something I would test in practice via runtime measurement. We are also working on a better num_workers interface for batched computation.

[rusty1s]

Thanks for providing the example code. Indeed, the major bottleneck of DGCNN is the dynamic edge creation via knn before each layer execution. It currently amounts to about 99% of the runtime of your script. We have plans to support a faster knn implementation in the future, but for now, there is nothing I can do about it.

Your linked repository computes the knn graph via pairwise-distance computation followed by a topk filtering. This approach can be, in fact, very fast, since it will trade runtime with memory consumption. In fact, the code will compute a potentially huge pair-wise distance matrix of shape [num_nodes, num_nodes], leading to better parallelization. Your linked repository has further differences: It expects examples to have the same number of points, which leads to faster runtime, but cannot be applied to examples of varying size (which are likely to happen in real-world scenarios).

To answer your questions:

1. They are in fact equal since the bottleneck is the knn computation rather than neural network execution.
2. 1D-Convolutions can’t be faster than simple matrix multiplication, since they will default to matrix multiplication under the hood
3. See my previous comment.

If you want to speed up your model, I suggest to utilize the knn implementation in the linked repository, and utilize EdgeConv instead of DynamicEdgeConv. Keep in mind that you need to reshape your feature matrix to [batch_size, num_nodes, num_features] though.