Chiebot-Mirror
/
opencv
8
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Merge pull request #22840 from zihaomu:optimze_conv_memory_usage

Browse Source 
DNN: reduce the memory used in convolution layer

* reduce the memory in winograd and disabel the test when usage memory is larger than 2gb.

* remove VERY_LOG tag
pull/22933/head
Zihao Mu 2 years ago committed by GitHub
parent ab912329b6
commit 0a650b573b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 103 additions and 90 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 2
      modules/dnn/perf/perf_net.cpp
  2. 5
      modules/dnn/src/layers/convolution_layer.cpp
  3. 11
      modules/dnn/src/layers/fast_convolution/fast_convolution.cpp
  4. 6
      modules/dnn/test/test_backends.cpp
  5. 37
      modules/dnn/test/test_darknet_importer.cpp

2
modules/dnn/perf/perf_net.cpp
Unescape View File

@ -198,6 +198,7 @@ PERF_TEST_P_(DNNTestNetwork, Inception_v2_SSD_TensorFlow)
PERF_TEST_P_(DNNTestNetwork, YOLOv3) PERF_TEST_P_(DNNTestNetwork, YOLOv3)
{ {
applyTestTag(CV_TEST_TAG_MEMORY_2GB);
if (backend == DNN_BACKEND_HALIDE) if (backend == DNN_BACKEND_HALIDE)
throw SkipTestException(""); throw SkipTestException("");
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2020040000) // nGraph compilation failure #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2020040000) // nGraph compilation failure
@ -220,6 +221,7 @@ PERF_TEST_P_(DNNTestNetwork, YOLOv3)
PERF_TEST_P_(DNNTestNetwork, YOLOv4) PERF_TEST_P_(DNNTestNetwork, YOLOv4)
{ {
applyTestTag(CV_TEST_TAG_MEMORY_2GB);
if (backend == DNN_BACKEND_HALIDE) if (backend == DNN_BACKEND_HALIDE)
throw SkipTestException(""); throw SkipTestException("");
if (target == DNN_TARGET_MYRIAD) // not enough resources if (target == DNN_TARGET_MYRIAD) // not enough resources

5
modules/dnn/src/layers/convolution_layer.cpp
Unescape View File

@ -2112,8 +2112,11 @@ public:
int dilation_h = dilations[dilations.size() - 2]; int dilation_h = dilations[dilations.size() - 2];
int dilation_w = dilations.back(); int dilation_w = dilations.back();
// Winograd only works well on input h and w >12.
bool canUseWinograd = useWinograd && inputs[0].size[2] >= 12 && inputs[0].size[3] >= 12;
fastConv2dImpl = initFastConv2d(ngroups, K, C, Hk, Wk, stride_w, stride_h, dilation_w, fastConv2dImpl = initFastConv2d(ngroups, K, C, Hk, Wk, stride_w, stride_h, dilation_w,
dilation_h, pads_begin, pads_end, weightsMat, &biasvec[0], useWinograd); dilation_h, pads_begin, pads_end, weightsMat, &biasvec[0], canUseWinograd);
} }
if (fastConv2dImpl) if (fastConv2dImpl)

11
modules/dnn/src/layers/fast_convolution/fast_convolution.cpp
Unescape View File

@ -83,9 +83,7 @@ Ptr<FastConv2d> initFastConv2d(
weightsBufPtr[c*padded_ksize + k] = srcWeights[c*wstep + k]; weightsBufPtr[c*padded_ksize + k] = srcWeights[c*wstep + k];
}}); }});
} }
else else if(conv->conv_type == _FX_CONV_TYPE_WINOGRAD3X3) // winograd
{
if (conv->conv_type == _FX_CONV_TYPE_WINOGRAD3X3) // winograd
{ {
static const float ktm[8][3] = { static const float ktm[8][3] = {
{1.0f, 0.0f, 0.0f}, {1.0f, 0.0f, 0.0f},
@ -162,7 +160,8 @@ Ptr<FastConv2d> initFastConv2d(
} }
}}); }});
} }
else if (conv->conv_type == _FX_CONV_TYPE_GENERIC)
{
// The weights are packed as // The weights are packed as
// ngroups x (ceil((K/ngroups)/CONV_MR)*CONV_MR) x (Cg*Hk*Wk) x CONV_MR tensor // ngroups x (ceil((K/ngroups)/CONV_MR)*CONV_MR) x (Cg*Hk*Wk) x CONV_MR tensor
int Kg = K/ngroups, Cg = max(C/ngroups, 1); int Kg = K/ngroups, Cg = max(C/ngroups, 1);
@ -202,6 +201,8 @@ Ptr<FastConv2d> initFastConv2d(
} }
}}); }});
} }
else
CV_Error(CV_StsUnsupportedFormat, "Unknown convolution type.");
// store bias; append some zero's to make sure that // store bias; append some zero's to make sure that
// we can always read MR elements starting from any valid index // we can always read MR elements starting from any valid index
@ -271,7 +272,7 @@ void runFastConv2d(InputArray _input, OutputArray _output, const Ptr<FastConv2d>
CV_Assert(fusedAddMat.empty()); // Depthwise-Convolution layer should not be followed by Add layer. CV_Assert(fusedAddMat.empty()); // Depthwise-Convolution layer should not be followed by Add layer.
return runDepthwise(input, output, conv, minval, maxval, activ, ifMinMaxAct); return runDepthwise(input, output, conv, minval, maxval, activ, ifMinMaxAct);
} }
else if (conv->conv_type == _FX_CONV_TYPE_WINOGRAD3X3 && inputShape[2] >= 12 && inputShape[3] >= 12) // winograd else if (conv->conv_type == _FX_CONV_TYPE_WINOGRAD3X3) // winograd
{ {
CV_Assert(conv->weightsWinoBufPtr); CV_Assert(conv->weightsWinoBufPtr);
if (runWinograd63(input, fusedAddMat, output, conv, ntasks, minval, maxval, activ, ifMinMaxAct)) if (runWinograd63(input, fusedAddMat, output, conv, ntasks, minval, maxval, activ, ifMinMaxAct))

6
modules/dnn/test/test_backends.cpp
Unescape View File

@ -29,7 +29,7 @@ public:
void processNet(std::string weights, std::string proto, void processNet(std::string weights, std::string proto,
Mat inp, const std::string& outputLayer = "", Mat inp, const std::string& outputLayer = "",
std::string halideScheduler = "", std::string halideScheduler = "",
double l1 = 0.0, double lInf = 0.0, double detectionConfThresh = 0.2) double l1 = 0.0, double lInf = 0.0, double detectionConfThresh = 0.2, bool useWinograd = true)
{ {
checkBackend(); checkBackend();
l1 = l1 ? l1 : default_l1; l1 = l1 ? l1 : default_l1;
@ -49,6 +49,7 @@ public:
net.setInput(inp); net.setInput(inp);
net.setPreferableBackend(backend); net.setPreferableBackend(backend);
net.setPreferableTarget(target); net.setPreferableTarget(target);
net.enableWinograd(useWinograd);
if (backend == DNN_BACKEND_HALIDE && !halideScheduler.empty()) if (backend == DNN_BACKEND_HALIDE && !halideScheduler.empty())
{ {
halideScheduler = findDataFile(halideScheduler); halideScheduler = findDataFile(halideScheduler);
@ -347,7 +348,8 @@ TEST_P(DNNTestNetwork, SSD_VGG16)
} }
processNet("dnn/VGG_ILSVRC2016_SSD_300x300_iter_440000.caffemodel", processNet("dnn/VGG_ILSVRC2016_SSD_300x300_iter_440000.caffemodel",
"dnn/ssd_vgg16.prototxt", inp, "detection_out", "", scoreDiff, iouDiff); "dnn/ssd_vgg16.prototxt", inp, "detection_out", "", scoreDiff,
iouDiff, 0.2, false);
expectNoFallbacksFromIE(net); expectNoFallbacksFromIE(net);
} }

37
modules/dnn/test/test_darknet_importer.cpp
Unescape View File

@ -81,6 +81,7 @@ TEST(Test_Darknet, read_yolo_voc_stream)
Net net = readNetFromDarknet(cfgFile, weightsFile); Net net = readNetFromDarknet(cfgFile, weightsFile);
net.setInput(inp); net.setInput(inp);
net.setPreferableBackend(DNN_BACKEND_OPENCV); net.setPreferableBackend(DNN_BACKEND_OPENCV);
net.enableWinograd(false);
ref = net.forward(); ref = net.forward();
} }
// Import from bytes array. // Import from bytes array.
@ -92,6 +93,7 @@ TEST(Test_Darknet, read_yolo_voc_stream)
Net net = readNetFromDarknet(cfg.data(), cfg.size(), weights.data(), weights.size()); Net net = readNetFromDarknet(cfg.data(), cfg.size(), weights.data(), weights.size());
net.setInput(inp); net.setInput(inp);
net.setPreferableBackend(DNN_BACKEND_OPENCV); net.setPreferableBackend(DNN_BACKEND_OPENCV);
net.enableWinograd(false);
Mat out = net.forward(); Mat out = net.forward();
normAssert(ref, out); normAssert(ref, out);
} }
@ -178,7 +180,8 @@ public:
const std::vector<std::vector<int> >& refClassIds, const std::vector<std::vector<int> >& refClassIds,
const std::vector<std::vector<float> >& refConfidences, const std::vector<std::vector<float> >& refConfidences,
const std::vector<std::vector<Rect2d> >& refBoxes, const std::vector<std::vector<Rect2d> >& refBoxes,
double scoreDiff, double iouDiff, float confThreshold = 0.24, float nmsThreshold = 0.4) double scoreDiff, double iouDiff, float confThreshold = 0.24,
float nmsThreshold = 0.4, bool useWinograd = true)
{ {
checkBackend(); checkBackend();
@ -198,6 +201,7 @@ public:
findDataFile("dnn/" + weights, false)); findDataFile("dnn/" + weights, false));
net.setPreferableBackend(backend); net.setPreferableBackend(backend);
net.setPreferableTarget(target); net.setPreferableTarget(target);
net.enableWinograd(useWinograd);
net.setInput(inp); net.setInput(inp);
std::vector<Mat> outs; std::vector<Mat> outs;
net.forward(outs, net.getUnconnectedOutLayersNames()); net.forward(outs, net.getUnconnectedOutLayersNames());
@ -280,18 +284,19 @@ public:
const std::vector<int>& refClassIds, const std::vector<int>& refClassIds,
const std::vector<float>& refConfidences, const std::vector<float>& refConfidences,
const std::vector<Rect2d>& refBoxes, const std::vector<Rect2d>& refBoxes,
double scoreDiff, double iouDiff, float confThreshold = 0.24, float nmsThreshold = 0.4) double scoreDiff, double iouDiff, float confThreshold = 0.24,
float nmsThreshold = 0.4, bool useWinograd = true)
{ {
testDarknetModel(cfg, weights, testDarknetModel(cfg, weights,
std::vector<std::vector<int> >(1, refClassIds), std::vector<std::vector<int> >(1, refClassIds),
std::vector<std::vector<float> >(1, refConfidences), std::vector<std::vector<float> >(1, refConfidences),
std::vector<std::vector<Rect2d> >(1, refBoxes), std::vector<std::vector<Rect2d> >(1, refBoxes),
scoreDiff, iouDiff, confThreshold, nmsThreshold); scoreDiff, iouDiff, confThreshold, nmsThreshold, useWinograd);
} }
void testDarknetModel(const std::string& cfg, const std::string& weights, void testDarknetModel(const std::string& cfg, const std::string& weights,
const cv::Mat& ref, double scoreDiff, double iouDiff, const cv::Mat& ref, double scoreDiff, double iouDiff,
float confThreshold = 0.24, float nmsThreshold = 0.4) float confThreshold = 0.24, float nmsThreshold = 0.4, bool useWinograd = true)
{ {
CV_Assert(ref.cols == 7); CV_Assert(ref.cols == 7);
std::vector<std::vector<int> > refClassIds; std::vector<std::vector<int> > refClassIds;
@ -318,7 +323,7 @@ public:
refBoxes[batchId].push_back(box); refBoxes[batchId].push_back(box);
} }
testDarknetModel(cfg, weights, refClassIds, refScores, refBoxes, testDarknetModel(cfg, weights, refClassIds, refScores, refBoxes,
scoreDiff, iouDiff, confThreshold, nmsThreshold); scoreDiff, iouDiff, confThreshold, nmsThreshold, useWinograd);
} }
}; };
@ -396,7 +401,7 @@ TEST_P(Test_Darknet_nets, YoloVoc)
{ {
SCOPED_TRACE("batch size 1"); SCOPED_TRACE("batch size 1");
testDarknetModel(config_file, weights_file, ref.rowRange(0, 3), scoreDiff, iouDiff); testDarknetModel(config_file, weights_file, ref.rowRange(0, 3), scoreDiff, iouDiff, 0.24, 0.4, false);
} }
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2022010000) #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2022010000)
@ -410,7 +415,7 @@ TEST_P(Test_Darknet_nets, YoloVoc)
#endif #endif
{ {
SCOPED_TRACE("batch size 2"); SCOPED_TRACE("batch size 2");
testDarknetModel(config_file, weights_file, ref, scoreDiff, iouDiff, 0.24, nmsThreshold); testDarknetModel(config_file, weights_file, ref, scoreDiff, iouDiff, 0.24, nmsThreshold, false);
} }
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2022010000) #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2022010000)
@ -599,7 +604,7 @@ TEST_P(Test_Darknet_nets, YOLOv3)
{ {
applyTestTag( applyTestTag(
CV_TEST_TAG_LONG, CV_TEST_TAG_LONG,
(target == DNN_TARGET_CPU ? CV_TEST_TAG_MEMORY_1GB : CV_TEST_TAG_MEMORY_2GB), CV_TEST_TAG_MEMORY_2GB,
CV_TEST_TAG_DEBUG_VERYLONG CV_TEST_TAG_DEBUG_VERYLONG
); );
@ -656,7 +661,7 @@ TEST_P(Test_Darknet_nets, YOLOv3)
{ {
SCOPED_TRACE("batch size 1"); SCOPED_TRACE("batch size 1");
testDarknetModel(config_file, weights_file, ref.rowRange(0, N0), scoreDiff, iouDiff); testDarknetModel(config_file, weights_file, ref.rowRange(0, N0), scoreDiff, iouDiff, 0.24, 0.4, false);
} }
#if defined(INF_ENGINE_RELEASE) #if defined(INF_ENGINE_RELEASE)
@ -674,7 +679,7 @@ TEST_P(Test_Darknet_nets, YOLOv3)
{ {
SCOPED_TRACE("batch size 2"); SCOPED_TRACE("batch size 2");
testDarknetModel(config_file, weights_file, ref, scoreDiff, iouDiff); testDarknetModel(config_file, weights_file, ref, scoreDiff, iouDiff, 0.24, 0.4, false);
} }
} }
@ -682,7 +687,7 @@ TEST_P(Test_Darknet_nets, YOLOv4)
{ {
applyTestTag( applyTestTag(
CV_TEST_TAG_LONG, CV_TEST_TAG_LONG,
(target == DNN_TARGET_CPU ? CV_TEST_TAG_MEMORY_1GB : CV_TEST_TAG_MEMORY_2GB), CV_TEST_TAG_MEMORY_2GB,
CV_TEST_TAG_DEBUG_VERYLONG CV_TEST_TAG_DEBUG_VERYLONG
); );
@ -756,7 +761,7 @@ TEST_P(Test_Darknet_nets, YOLOv4)
{ {
SCOPED_TRACE("batch size 1"); SCOPED_TRACE("batch size 1");
testDarknetModel(config_file, weights_file, ref.rowRange(0, N0), scoreDiff, iouDiff); testDarknetModel(config_file, weights_file, ref.rowRange(0, N0), scoreDiff, iouDiff, 0.24, 0.4, false);
} }
{ {
@ -792,7 +797,7 @@ TEST_P(Test_Darknet_nets, YOLOv4)
} }
#endif #endif
testDarknetModel(config_file, weights_file, ref, scoreDiff, iouDiff); testDarknetModel(config_file, weights_file, ref, scoreDiff, iouDiff, 0.24, 0.4, false);
} }
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2022010000) #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2022010000)
@ -877,7 +882,7 @@ TEST_P(Test_Darknet_nets, YOLOv4x_mish)
{ {
applyTestTag( applyTestTag(
CV_TEST_TAG_LONG, CV_TEST_TAG_LONG,
(target == DNN_TARGET_CPU ? CV_TEST_TAG_MEMORY_1GB : CV_TEST_TAG_MEMORY_2GB), CV_TEST_TAG_MEMORY_2GB,
CV_TEST_TAG_DEBUG_VERYLONG CV_TEST_TAG_DEBUG_VERYLONG
); );
@ -939,7 +944,7 @@ TEST_P(Test_Darknet_nets, YOLOv4x_mish)
{ {
SCOPED_TRACE("batch size 1"); SCOPED_TRACE("batch size 1");
testDarknetModel(config_file, weights_file, ref.rowRange(0, N0), scoreDiff, iouDiff); testDarknetModel(config_file, weights_file, ref.rowRange(0, N0), scoreDiff, iouDiff, 0.24, 0.4, false);
} }
{ {
@ -958,7 +963,7 @@ TEST_P(Test_Darknet_nets, YOLOv4x_mish)
} }
#endif #endif
testDarknetModel(config_file, weights_file, ref, scoreDiff, iouDiff); testDarknetModel(config_file, weights_file, ref, scoreDiff, iouDiff, 0.24, 0.4, false);
} }
} }

Write Preview
Loading…
Cancel
Save