yolo4tiny works on tensorRT 🐬 🐬 🐬 🐬

Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>

include/tkDNN/NetworkRT.h

@@ -28,7 +28,7 @@ using namespace nvinfer1;
 #include "pluginsRT/ActivationMishRT.h"
 #include "pluginsRT/ReorgRT.h"
 #include "pluginsRT/RegionRT.h"
-//#include "pluginsRT/RouteRT.h"
+#include "pluginsRT/RouteRT.h"
 #include "pluginsRT/ShortcutRT.h"
 #include "pluginsRT/YoloRT.h"
 #include "pluginsRT/UpsampleRT.h"

include/tkDNN/pluginsRT/RouteRT.h

@@ -8,7 +8,9 @@ class RouteRT : public IPlugin {
 	*/
 
 public:
-	RouteRT() {
+	RouteRT(int groups, int group_id) {
+		this->groups = groups;
+		this->group_id = group_id;
 	}
 
 	~RouteRT(){
@@ -22,7 +24,7 @@ public:
 	Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override {
 		int out_c = 0;
 		for(int i=0; i<nbInputDims; i++) out_c += inputs[i].d[0];
-		return DimsCHW{out_c, inputs[0].d[1], inputs[0].d[2]};
+		return DimsCHW{out_c/groups, inputs[0].d[1], inputs[0].d[2]};
 	}
 
 	void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override {
@@ -34,6 +36,7 @@ public:
 		}
 		h = inputDims[0].d[1];
 		w = inputDims[0].d[2];
+		c /= groups;
 	}
 
 	int initialize() override {
@@ -56,8 +59,9 @@ public:
 		for(int i=0; i<in; i++) {
 			dnnType *input = (dnnType*)reinterpret_cast<const dnnType*>(inputs[i]);
 			int in_dim = c_in[i]*h*w;
-			checkCuda( cudaMemcpyAsync(dstData + offset, input, in_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream) );
+			int part_in_dim = in_dim / this->groups;
-			offset += in_dim;
+			checkCuda( cudaMemcpyAsync(dstData + offset, input + this->group_id*part_in_dim, part_in_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream) );
+			offset += part_in_dim;
 		}
 
 		return 0;
@@ -65,11 +69,13 @@ public:
 
 
 	virtual size_t getSerializationSize() override {
-		return (4+MAX_INPUTS)*sizeof(int);
+		return (6+MAX_INPUTS)*sizeof(int);
 	}
 
 	virtual void serialize(void* buffer) override {
 		char *buf = reinterpret_cast<char*>(buffer);
+		tk::dnn::writeBUF(buf, groups);
+		tk::dnn::writeBUF(buf, group_id);
 		tk::dnn::writeBUF(buf, in);
 		for(int i=0; i<MAX_INPUTS; i++)
 			tk::dnn::writeBUF(buf, c_in[i]);
@@ -83,4 +89,5 @@ public:
 	int in;
 	int c_in[MAX_INPUTS];
 	int c, h, w;
+	int groups, group_id;
 };

src/NetworkRT.cpp

@@ -449,12 +449,15 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Route *l) {
         // }
         // std::cout<<"\n";
     }
-    
-    IConcatenationLayer *lRT = networkRT->addConcatenation(tens, l->layers_n);
-    //IPlugin *plugin = new RouteRT();
-    //IPluginLayer *lRT = networkRT->addPlugin(tens, l->layers_n, *plugin);
-    checkNULL(lRT);
 
+    if(l->groups > 1){
+        IPlugin *plugin = new RouteRT(l->groups, l->group_id);
+        IPluginLayer *lRT = networkRT->addPlugin(tens, l->layers_n, *plugin);
+        checkNULL(lRT);
+        return lRT;
+    }
+    IConcatenationLayer *lRT = networkRT->addConcatenation(tens, l->layers_n);
+    checkNULL(lRT);
     return lRT;
 }
 
@@ -766,9 +769,9 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa
         r->w = readBUF<int>(buf);
         return r;
     } 
-/*
+
     if(name.find("Route") == 0) {
-        RouteRT *r = new RouteRT();
+        RouteRT *r = new RouteRT(readBUF<int>(buf),readBUF<int>(buf));
         r->in = readBUF<int>(buf);
         for(int i=0; i<RouteRT::MAX_INPUTS; i++)
             r->c_in[i] = readBUF<int>(buf);
@@ -777,7 +780,7 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa
         r->w = readBUF<int>(buf);
         return r;
     } 
-*/
+
     if(name.find("Deformable") == 0) {
         DeformableConvRT *r = new DeformableConvRT(readBUF<int>(buf), readBUF<int>(buf), readBUF<int>(buf),
                                                     readBUF<int>(buf), readBUF<int>(buf), readBUF<int>(buf),

tests/darknet/yolo4tiny.cpp

@@ -23,11 +23,11 @@ int main() {
     net->print();
 
     //convert network to tensorRT
-    // tk::dnn::NetworkRT *netRT = new tk::dnn::NetworkRT(net, net->getNetworkRTName(bin_path.c_str()));
+    tk::dnn::NetworkRT *netRT = new tk::dnn::NetworkRT(net, net->getNetworkRTName(bin_path.c_str()));
     
-    int ret = testInference(input_bins, output_bins, net, nullptr);
+    int ret = testInference(input_bins, output_bins, net, netRT);
     net->releaseLayers();
     delete net;
-    // delete netRT;
+    delete netRT;
     return ret;
 }