quantize model
Last updated
Was this helpful?
Last updated
Was this helpful?
quantization
continuous range to finite range of discreet values
2.568 -> 1,2,3
time analogy
24 bits / 8 bits / 1 bit per pixel
不同於 normalization
自己的誤解 : normalization (scale 完後仍是 continous)
quantization 不是, 為 discreet (離散)
quantization
unsinged int 8 bit max = 255
image 0~255 / 255.0 = 0~1
dynamic range?
範圍大: 解析度差 (0, 0.49 都變成 0)
8 bit quantize * scale = floating range
ex: floating range 100 * 255.0
scale > 0.5 大
scale < 0.01 以下, 還不錯準的 model
scale 可用於表達 dynamic range, 藉此判斷 quantization 效果好不好
offset = - zero point (通常我們會去 - offset)
offset
zero point
可以用 y = ax + b 去想, a: scale, b: offset
x_quant = (x_ori - offset) * scale
bit-width b, 例如 8 bit = 255 = 2 ^8 - 1
symmetric quantization (對稱)
注意重點在公式 x_ori = x_quant * scale
asymmetric quantization (非對稱)
公式 x_ori = (x_quant - zero point)* scale
x_ori_min = - sacle * zeropoint (= quant的0)
zeropint = quant 後固定位置 = s(x_ori_0 - zero point), x_ori_0 = zero point
優點 0 可以被固定, 更精準
post training quantization
