potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-28 23:03:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #21372 from zihaomu:dnn_quantize_per_tensor

Browse Source 
Add per_tensor_quantize to int8 quantize

* add per_tensor_quantize to dnn int8 module.

* change api flag from perTensor to perChannel, and recognize quantize type and onnx importer.

* change the default to hpp
This commit is contained in:
Zihao Mu
2022-07-06 00:14:42 +08:00
committed by GitHub
parent 16b5fd4bf2
commit a80fcacd90
13 changed files with 160 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/convolution_layer.cpp
+22 -12
View File
@@ -2226,26 +2226,36 @@ public:
Mat weightsQuantized(weightsMat.rows, weightsMat.cols, CV_8S);
Mat biasQuantized(1, numOutput, CV_32S);
Mat outputMultiplier(1, numOutput, CV_32F);
double realMin, realMax, weightsScale;
bool perChannel = params.get<bool>("per_channel", true);
for( int i = 0; i < numOutput; i++ )
if (perChannel) // per-Channel quantization.
{
// Quantize weights
cv::minMaxIdx(weightsMat.row(i), &realMin, &realMax);
realMin = std::min(realMin, 0.0);
realMax = std::max(realMax, 0.0);
weightsScale = (realMax == realMin) ? 1.0 : std::max(-realMin, realMax)/127;
weightsMat.row(i).convertTo(weightsQuantized.row(i), CV_8S, 1.f/weightsScale);
for (int i = 0; i < numOutput; i++)
{
double weightsScale = getWeightScale(weightsMat.row(i));
// Quantize biases
weightsMat.row(i).convertTo(weightsQuantized.row(i), CV_8S, 1.f/weightsScale);
float biasScale = inputScale * weightsScale;
biasQuantized.at<int>(i) = cvRound(biasvec[i]/biasScale) - inputZp*(cv::sum(weightsQuantized.row(i))[0]);
outputMultiplier.at<float>(i) = biasScale / outputScale;
}
}
else // per-Tensor quantization.
{
double weightsScale = getWeightScale(weightsMat);
weightsMat.convertTo(weightsQuantized, CV_8S, 1.f/weightsScale);
float biasScale = inputScale * weightsScale;
biasQuantized.at<int>(i) = (int)std::round(biasvec[i]/biasScale) - inputZp*(cv::sum(weightsQuantized.row(i))[0]);
// Store multiplier
outputMultiplier.at<float>(i) = biasScale / outputScale;
for (int i = 0; i < numOutput; i++)
{
biasQuantized.at<int>(i) = cvRound(biasvec[i]/biasScale) - inputZp*(cv::sum(weightsQuantized.row(i))[0]);
outputMultiplier.at<float>(i) = biasScale / outputScale;
}
}
params.blobs.clear();
params.set("per_channel", perChannel);
params.blobs.push_back(weightsQuantized.reshape(1, shape(blobs[0])));
params.blobs.push_back(biasQuantized);
params.blobs.push_back(outputMultiplier);