Why is the norm not equal to `2` in `find_params`? about obc HOT 3 CLOSED

Hi, this part of the code and its hyper-parameters are mostly adapted from BRECQ. The purpose of this is to find a narrower quantization grid (relative to the min-max which includes all outliers) by minimizing the p-norm. Setting norm = 2.4 gives slightly more weight to outliers in this search than for norm = 2.

from obc.

Ok thank you. Another question - where exactly is the quantization happening?

• The quantize function is simply a chain of affine tranform, clamping and reverse affine transform. It does not reduce the number of bits to store the weights.
• The bits variable is simply used to calculate maxq which is used to calculate the scale.

I am unable to find where a fp32 value is stored in a bits number of bits.

from obc.

I guess you could add an extra line in the quantize function like so:

def quantize(x, scale, zero, maxq):
    q = torch.clamp(torch.round(x / scale) + zero, 0, maxq)
    q = convert_q_to_nbits(q, bits)
    return scale * (q - zero)

But I get why you didn't do it. Closing the issue now.

from obc.