8x faster graph normalization (5408a47f) · Commits · 钟慕尧 / deepchem

2 changes: 1. numpy matrix inversion is 2x faster than scipy inversion 2. For diagonal matrices, np.sqrt and scipy.linalg.sqrtm give the same result because there are no interaction terms. But np.sqrt is 250x faster.

contrib/torch/pytorch_graphconv.py

+2 −3

Original line number	Diff line number	Diff line
		@@ -6,7 +6,6 @@ import torch.optim as optim
		import random
		import numpy as np
		from sklearn.metrics import roc_auc_score
		import scipy


		def symmetric_normalize_adj(adj):
		@@ -24,8 +23,8 @@ def symmetric_normalize_adj(adj):
		adj = adj[:n_atoms, :n_atoms]
		degree = np.sum(adj, axis=1)
		D = np.diag(degree)
		D_sqrt = scipy.linalg.sqrtm(D)
		D_sqrt_inv = scipy.linalg.inv(D_sqrt)
		D_sqrt = np.sqrt(D)
		D_sqrt_inv = np.linalg.inv(D_sqrt)
		sym_norm = D_sqrt_inv.dot(adj)
		sym_norm = sym_norm.dot(D_sqrt_inv)
		new_adj = np.zeros(orig_shape)

Admin message