public static (double[] scale, double bias) QuantizeWeights(bool isConv2d, Tensor <float> weights, K210ConvLayerConfig config, int weightsBits)
{
#if CHANNEL_WISE
var kernels = weights.ToDenseTensor().Buffer.Span;
var channels = weights.Dimensions[isConv2d ? 0 : 1];
var channelSize = weights.Dimensions.GetSize() / channels;
var totalRange = Quantizer.GetRange(kernels);
var scales = new double[channels];
for (int i = 0; i < channels; i++)
{
double s;
var buffer = kernels.Slice(i * channelSize, channelSize);
var range = Quantizer.GetRange(buffer);
var s1 = totalRange.Max / range.Max;
var s2 = totalRange.Min / range.Min;
s = (s1 < 0 || s2 < 0) ? Math.Max(s1, s2) : Math.Min(s1, s2);
Debug.Assert(s > 0);
for (int j = 0; j < buffer.Length; j++)
{
buffer[j] = (float)(buffer[j] * s);
}
scales[i] = s;
}
(var scale, var bias) = Quantizer.GetRange(kernels).GetScaleBias(weightsBits);
(var mul, var shift) = Quantizer.ExtractValueAndShift(bias, 24, 15);
config.Weights = Quantizer.Quantize(kernels, scale, bias, weightsBits);
config.ArgX = (int)Math.Round(mul);
config.ShiftX = shift;
for (int i = 0; i < scales.Length; i++)
{
scales[i] *= scale;
}
return(scales, bias);
#else
var buffer = weights.ToDenseTensor().Buffer.Span;
(var scale, var bias) = GetRange(buffer).GetScaleBias();
(var mul, var shift) = ExtractValueAndShift(bias, 24, 15);
config.Weights = Quantize(buffer, scale, bias);
config.ArgX = (int)Math.Round(mul);
config.ShiftX = shift;
return(Enumerable.Repeat(scale, weights.Dimensions[0]).ToArray(), bias);
#endif
}
