potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-29 07:13:02 +04:00
Code Packages Releases Wiki Activity

optimize out scaleLayer & concatLayer whenever possible

Browse Source 
fixed problem in concat layer by disabling memory re-use in layers with multiple inputs

trying to fix the tests when Halide is used to run deep nets

another attempt to fix Halide tests

see if the Halide tests will pass with concat layer fusion turned off

trying to fix failures in halide tests; another try

one more experiment to make halide_concat & halide_enet tests pass

continue attempts to fix halide tests

moving on

uncomment parallel concat layer

seemingly fixed failures in Halide tests and re-enabled concat layer fusion; thanks to dkurt for the patch
This commit is contained in:
Vadim Pisarevsky
2017-07-04 17:23:47 +03:00
parent 431e2e6d68
commit 0488d9bdb2
5 changed files with 337 additions and 62 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/convolution_layer.cpp
+48 -11
View File
@@ -148,6 +148,7 @@ public:
std::vector<float> reluslope;
Ptr<ActivationLayer> activ;
Ptr<BatchNormLayer> bnorm;
Ptr<ScaleLayer> scaleLayer;
MatShape computeColRowShape(const MatShape &inpShape, const MatShape &outShape) const
{
@@ -202,6 +203,9 @@ public:
bool setBatchNorm(const Ptr<BatchNormLayer>& layer )
{
// for now the scale layer followed by the batch norm cannot be fused, only vice versa.
if( !scaleLayer.empty() )
return false;
bnorm = layer;
// we will need to re-compute the weights with the batch
// norm coefficients taken into account
@@ -209,6 +213,15 @@ public:
return !bnorm.empty();
}
bool setScale(const Ptr<ScaleLayer>& layer)
{
scaleLayer = layer;
// we will need to re-compute the weights with the scaling
// coefficients taken into account
weightsMat.release();
return !scaleLayer.empty();
}
virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs)
{
#ifdef HAVE_HALIDE
@@ -678,32 +691,56 @@ public:
biasvec[k] = biasMat.at<float>(k);
}
if( !bnorm.empty() )
if( !bnorm.empty() || !scaleLayer.empty() )
{
Mat scale, shift;
bnorm->getScaleShift(scale, shift);
Mat scale, shift, scale2, shift2;
const float *scaleptr = 0, *shiftptr = 0;
const float *scaleptr2 = 0, *shiftptr2 = 0;
CV_Assert( scale.isContinuous() && shift.isContinuous() &&
scale.type() == CV_32F && shift.type() == CV_32F &&
scale.total() == (size_t)outCn &&
shift.total() == (size_t)outCn );
if( !bnorm.empty() )
{
bnorm->getScaleShift(scale, shift);
CV_Assert( scale.isContinuous() && shift.isContinuous() &&
scale.type() == CV_32F && shift.type() == CV_32F &&
scale.total() == (size_t)outCn &&
shift.total() == (size_t)outCn );
scaleptr = scale.ptr<float>();
shiftptr = shift.ptr<float>();
}
if( !scaleLayer.empty() )
{
scale2 = scaleLayer->blobs[0];
CV_Assert( scale2.isContinuous() && scale2.type() == CV_32F &&
scale2.total() == (size_t)outCn );
scaleptr2 = scale2.ptr<float>();
if( scaleLayer->hasBias )
{
shift2 = scaleLayer->blobs[1];
CV_Assert( shift2.isContinuous() && shift2.type() == CV_32F &&
shift2.total() == (size_t)outCn );
shiftptr2 = shift2.ptr<float>();
}
}
for( int i = 0; i < outCn; i++ )
{
float s = scale.at<float>(i);
float delta = shift.at<float>(i);
float s1 = scaleptr ? scaleptr[i] : 1.f;
float delta1 = shiftptr ? shiftptr[i] : 0.f;
float s2 = scaleptr2 ? scaleptr2[i] : 1.f;
float delta2 = shiftptr2 ? shiftptr2[i] : 0.f;
float* w_i = weightsMat.ptr<float>(i);
int j, wcols = weightsMat.cols;
for( j = 0; j < wcols; j++ )
w_i[j] *= s;
w_i[j] *= (s1*s2);
biasvec[i] = biasvec[i]*s + delta;
biasvec[i] = biasvec[i]*(s1*s2) + (delta1*s2 + delta2);
}
}
biasvec[outCn] = biasvec[outCn+1] = biasvec[outCn-1];
}
reluslope.clear();
if( activ )
{
Ptr<ReLULayer> activ_relu = activ.dynamicCast<ReLULayer>();