About icarl_construct_exemplar_set function

[superduduguan]

Hi, I think the following codes seem not proper:

essentials_for_CIL/main_cifar.py

Lines 321 to 337 in c7eb378

	exemplar_set = []
	exemplar_features = [] # list of Variables of shape (feature_size,)
	exemplar_dist = []
	for k in range(int(m)):
	S = np.sum(exemplar_features, axis=0)
	phi = features
	mu = class_mean
	mu_p = 1.0/(k+1) * (phi + S)
	mu_p = mu_p / np.linalg.norm(mu_p)
	dist = np.sqrt(np.sum((mu - mu_p) ** 2, axis=1))
	idx = np.random.randint(0, features.shape[0])
	exemplar_dist.append(dist[idx])
	exemplar_set.append(images[idx])
	exemplar_features.append(features[idx])
	features[idx, :] = 0.0

After debugging I find that when features[idx, :] = 0.0 implemented, exemplar_features is also set to 0, thus S is always full of 0 at any time, making it meaningless. I am not sure whether it is a common problem or it is triggered due to my incompatible version of packages.
Besides, what if the random saving strategy get the same index? Is it right the purpose of setting the corresponding feature to 0, so that in the next phase the repetitive sample has a large distance and will be deleted? But still, what if class_mean itself is close to 0?

[haoweiz23]

I am confused about the meaning of S and mu_p, why not compute the distance between features and mu directly? Does anyone have idea about that?