[How-To] Quantize my own model in Tensorflow using this approach?

[mahimairaja]

No description provided.

[jundaf2]

This dynamic quantization simply leverage the fact that
(1) the global max of the matrix P is not necessarily the max value of each line of matrix P
(2)you know the max and min value in the softmax computation of a line on the fly -- it's the inherit property of softmax, i.e. the numerators in every line is between [0,1]
, so you can leverage this fact without passing addtional global quantization information of matrix $P$.

Thanks.