It would be good to add a small number like epsilon = 0.000000001 or something to the pdistfcm function and whenever you have a denominator that could be 0.

def pdistfcm(cntr, data):
    out = np.zeros(shape=(cntr.shape[0], data.shape[0]))
    for k in range(cntr.shape[0]):
        out[k] = np.sqrt(np.sum((np.power(data-cntr[k], 2)).T, axis=0)) + 0.000000000001 #add eps to avoid divide by 0
    return out

Hello,

First of all, thank you very much for sharing this project. Just one suggestion: could you transform your project to large scale kmeans with GPU capacity? I think it will be quicker when we choose a big number of clusters. What do you think?

Reference:
https://github.com/src-d/kmcuda

Hi, thank you for sharing. What should I change with these codes if I want to segment such a brain MRI (181*217, bmp)which is shown as follow? Looking forward to your reply. Thanks.

Equation 21 in the PFCM paper has an exponent of 2/(m-1) which SHOULD correspond to 2/(expo-1), but you have 2/(nc-1) in the code. I believe this is a typo in both the predict and pfcm functions.