I think that when updating weights, should not target all nodes.

[cuiguohua]

Hello, first of all, thank you for sharing, and gave me a lot of help.

def update_weights(self, inputVec, sigma, lrate, bmu):

	"""Update the node Weights.

	Args:
		inputVec (np.array): A weights vector whose distance drives the direction of the update.
		sigma (float): The updated gaussian sigma.
		lrate (float): The updated learning rate.
		bmu (somNode): The best matching unit.
	"""

	dist=self.get_nodeDistance(bmu)
	gauss=np.exp(-dist*dist/(2*sigma*sigma))  # I think gauss will always > 0
	if gauss>0:
		for i in range(len(self.weights)):
			self.weights[i] = self.weights[i] - gauss*lrate*(self.weights[i]-inputVec[i])

In somNode::update_weights() , expression ' gauss > 0 ' will always be true.
So throughout the training process, the weights of all nodes will be changed.

I read some literature. The literature says that the node weights near BMU should be updated, and this neighborhood is gradually reduced, eventually containing only BMU itself.
So, I think we should change ' gauss >0' to 'gauss>x ' (0<x<1)

Thank you again for sharing and looking forward to your reply.

[fcomitani]

Hello @cuiguohua,

the radius of the Gaussian is determined by the sigma which is reduced at every step by update_sigma, thus effectively reducing the neighborhood set.

You are however right about the Gaussian being always > 0, unless the sigma gets infinitely small.
I haven't looked at the code in a while and I'm not sure why I put that condition. Possibly it was just a debugging device which forgot to remove.

Just let me think it over and do a couple of extra tests. If I still can't find a justification for its presence I will remove it in the next version.

Thanks.

[fcomitani]

The redundant condition was removed, I'm closing this issue.