semi-supervised-paper-implementation
semi-supervised-paper-implementation copied to clipboard
May i know how you estimate the ZCA matrix?
Hello,
I found your ZCA matrix and thanks for your upload.
I know how to calculate ZCA matrix, but for CIFAR10 or 100 the matrix is huge and my hardware cannot support this.
So I want to know, how you calculate the matrix? by brutal force(means you have some super hardware) or is there some tricks to estimate the ZCA matrix for the whole dataset?
really thanks if you can give me some hint
I just rent an AWS instance and calculated the ZCA matrix of CIFAR10: the result is different from yours:
my one(calculated based on the whole dataset, i.e. 50K images): [[ 3.16225757e+00 -1.93212111e-05 -1.86190303e-05 ... -5.32311332e-06 -5.54466758e-06 -5.91178032e-06] [-1.93212111e-05 3.16225810e+00 -1.90527165e-05 ... -5.17075653e-06 -5.37499166e-06 -5.64742587e-06] [-1.86190303e-05 -1.90527165e-05 3.16225821e+00 ... -5.08378483e-06 -5.26505783e-06 -5.50406835e-06] ... [-5.32311332e-06 -5.17075653e-06 -5.08378483e-06 ... 3.16226195e+00 -1.54090321e-05 -1.49700770e-05] [-5.54466758e-06 -5.37499166e-06 -5.26505783e-06 ... -1.54090321e-05 3.16226180e+00 -1.56390237e-05] [-5.91178032e-06 -5.64742587e-06 -5.50406835e-06 ... -1.49700770e-05 -1.56390237e-05 3.16226137e+00]] size = (3072, 3072)
yours: [[ 4.6536251e+01 -1.6922934e+01 -4.2759528e+00 ... 3.2285467e-01 -1.3231681e-01 -1.8512923e-01] [-1.6922934e+01 6.5734108e+01 -2.5042648e+01 ... -6.3035327e-01 2.3367840e-01 1.9456828e-01] [-4.2759528e+00 -2.5042648e+01 7.5144257e+01 ... 2.4870060e-01 1.9999361e-02 -2.3955733e-02] ... [ 3.2285476e-01 -6.3035333e-01 2.4870069e-01 ... 7.9308556e+01 -3.2609344e+01 1.0282707e+00] [-1.3231690e-01 2.3367813e-01 1.9999307e-02 ... -3.2609344e+01 7.7910210e+01 -2.8092228e+01] [-1.8512926e-01 1.9456822e-01 -2.3955926e-02 ... 1.0282707e+00 -2.8092228e+01 5.6274792e+01]] size = (3072, 3072)
based on the fomula, the ZCA matrix should be a diagnal matrix, and i think your matrix is not correct.