Quantization Error about susi HOT 4 CLOSED

Hi @aspassiani, could you give me an example with small example data?

from susi.

Hi @felixriese, not quite sure the best example to give, but if I were to set weights_per_datapoint and X as random matrixes,

 weights_per_datapoint  =[[8,5,0,4,5,7,0,2,0,2]
                          [4,0,7,4,8,0,4,2,2,8]
                          [5,3,1,2,7,9,9,8,8,7]
                          [8,6,8,7,7,5,4,7,1,1]
                          [6,2,3,5,7,5,9,5,9,0]]

X = [[8,9,7,8,6,3,0,4,1,6]
    [3,2,1,2,5,7,0,0,3,3]
    [0,7,5,1,4,5,7,5,8,5]
    [9,1,1,7,5,9,8,9,3,3]
    [2,7,7,2,3,3,3,3,3,3]]

then,

quantization_errors = np.linalg.norm(np.subtract(weights_per_datapoint, X)) 

would give me a single value of 25.729360660537214 making return np.mean(quantization_errors) unnecessary, whereas,

quantization_errors = np.linalg.norm(np.subtract(weights_per_datapoint, X), axis = 1 ) 

would return a vector of values, [10.9087,12.2066,10,11.0905,13.0767] and a mean of 11.45650021348017.

from susi.

@aspassiani Thanks for the data! Makes sense. I added a fix in the PR #40. Please have a look

from susi.

The revised quantization error looks correct, thanks!

from susi.