Quantize

MIN_COMBINED Mode Example

Assume the input is type float and has a possible range of [0.0, 6.0] and the output type is quint8 ([0, 255]). The min_range and max_range values should be specified as 0.0 and 6.0. Quantizing from float to quint8 will multiply each value of the input by 255/6 and cast to quint8.

If the output type was qint8 ([-128, 127]), the operation will additionally subtract each value by 128 prior to casting, so that the range of values aligns with the range of qint8.

If the mode is 'MIN_FIRST', then this approach is used:

num_discrete_values = 1 << (# of bits in T)
 range_adjust = num_discrete_values / (num_discrete_values - 1)
 range = (range_max - range_min) * range_adjust
 range_scale = num_discrete_values / range
 quantized = round(input * range_scale) - round(range_min * range_scale) +
   numeric_limits<T>::min()
 quantized = max(quantized, numeric_limits<T>::min())
 quantized = min(quantized, numeric_limits<T>::max())
 

SCALED mode Example

`SCALED` mode matches the quantization approach used in `QuantizeAndDequantize{V2|V3}`.

If the mode is `SCALED`, the quantization is performed by multiplying each input value by a scaling_factor. The scaling_factor is determined from `min_range` and `max_range` to be as large as possible such that the range from `min_range` to `max_range` is representable within values of type T.

const int min_T = std::numeric_limits<T>::min();
   const int max_T = std::numeric_limits<T>::max();
   const float max_float = std::numeric_limits<float>::max();
 
   const float scale_factor_from_min_side =
       (min_T * min_range > 0) ? min_T / min_range : max_float;
   const float scale_factor_from_max_side =
       (max_T * max_range > 0) ? max_T / max_range : max_float;
 
   const float scale_factor = std::min(scale_factor_from_min_side,
                                       scale_factor_from_max_side);
 

min_range = min_T / scale_factor;
       max_range = max_T / scale_factor;
 

So we will quantize input values in the range (-10, 9.921875) to (-128, 127).

The input tensor can now be quantized by clipping values to the range `min_range` to `max_range`, then multiplying by scale_factor as follows:

result = round(min(max_range, max(min_range, input)) * scale_factor)
 

narrow_range (bool) attribute

If true, we do not use the minimum quantized value. i.e. for int8 the quantized output, it would be restricted to the range -127..127 instead of the full -128..127 range. This is provided for compatibility with certain inference backends. (Only applies to SCALED mode)

axis (int) attribute

An optional `axis` attribute can specify a dimension index of the input tensor, such that quantization ranges will be calculated and applied separately for each slice of the tensor along that dimension. This is useful for per-channel quantization.

If axis is specified, min_range and max_range

if `axis`=None, per-tensor quantization is performed as normal.

ensure_minimum_range (float) attribute

Ensures the minimum quantization range is at least this value. The legacy default value for this is 0.01, but it is strongly suggested to set it to 0 for new uses.