Change CenterNet input dimension.

This commit changes the image input dimension, it updates the
CenterNet detection class.

Signed-off-by: Davide Sapienza <sapienza.dav@gmail.com>

include/tkDNN/CenternetDetection.h

@@ -73,14 +73,14 @@ class CenternetDetection {
         //processing
         float toll = 0.000001;
         int K = 100;
-        int width = 56;        // TODO
+        int width = 128;//56;        // TODO
     
     
     public:
         dnnType *rt_out[4];
 
-        float inp_height = 224;//512;
+        float inp_height = 512;//224;//512;
-        float inp_width = 224;//512;
+        float inp_width = 512;//224;//512;
     
         int classes = 80;
         int num = 0;

src/CenternetDetection.cpp

@@ -18,7 +18,7 @@ bool CenternetDetection::init(std::string tensor_path) {
     std::cout<<(tensor_path).c_str()<<"\n";
     netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str() );
     
-    dim = tk::dnn::dataDim_t(1, 3, 224, 224, 1);
+    dim = tk::dnn::dataDim_t(1, 3, 512, 512, 1);
     const char *coco_class_name_[] = {
             "person", "bicycle", "car", "motorcycle", "airplane", 
             "bus", "train", "truck", "boat", "traffic light", "fire hydrant",
@@ -43,9 +43,12 @@ bool CenternetDetection::init(std::string tensor_path) {
     checkCuda(cudaMallocHost(&input, sizeof(dnnType)*netRT->input_dim.tot()));
     checkCuda(cudaMalloc(&input_d, sizeof(dnnType)*netRT->input_dim.tot()));
 
-    dim_hm = tk::dnn::dataDim_t(1, 80, 56, 56, 1);
+    // dim_hm = tk::dnn::dataDim_t(1, 80, 56, 56, 1);
-    dim_wh = tk::dnn::dataDim_t(1, 2, 56, 56, 1);
+    // dim_wh = tk::dnn::dataDim_t(1, 2, 56, 56, 1);
-    dim_reg = tk::dnn::dataDim_t(1, 2, 56, 56, 1);
+    // dim_reg = tk::dnn::dataDim_t(1, 2, 56, 56, 1);
+    dim_hm = tk::dnn::dataDim_t(1, 80, 128, 128, 1);
+    dim_wh = tk::dnn::dataDim_t(1, 2, 128, 128, 1);
+    dim_reg = tk::dnn::dataDim_t(1, 2, 128, 128, 1);
 
     checkCuda( cudaMalloc(&topk_scores, dim_hm.c * K *sizeof(float)) );
     checkCuda( cudaMalloc(&topk_inds_, dim_hm.c * K *sizeof(int)) );      
@@ -93,6 +96,8 @@ bool CenternetDetection::init(std::string tensor_path) {
 
     mean << 0.408, 0.447, 0.47;
     stddev << 0.289, 0.274, 0.278;
+    // mean << 0.485, 0.456, 0.406;
+    // stddev << 0.229, 0.224, 0.225;
 }
 
 void CenternetDetection::testdog() {
@@ -104,8 +109,8 @@ void CenternetDetection::testdog() {
     imageORIG.convertTo(imageF, CV_32FC3, 1/255.0); 
     sz = imageF.size();
     std::cout<<"image: "<<sz.width<<", "<<sz.height<<std::endl;
-    resize(imageF, imageF, cv::Size(256, 256));
+    resize(imageF, imageF, cv::Size(512, 512));
-    const int cropSize = 224;
+    const int cropSize = 512;
     const int offsetW = (imageF.cols - cropSize) / 2;
     const int offsetH = (imageF.rows - cropSize) / 2;
     const cv::Rect roi(offsetW, offsetH, cropSize, cropSize);

src/DeformConv2d.cpp

@@ -31,16 +31,19 @@ void DeformConv2d::initCUDNN() {
     // kernel ones
     
     checkCuda( cudaMalloc(&ones_d1, (height_ones*width_ones)*sizeof(dnnType)) );
-    float aus1[height_ones*width_ones];
+    dnnType *aus1;
+    checkCuda( cudaMallocHost(&aus1, (height_ones*width_ones)*sizeof(dnnType)) );
     for(int i=0; i<height_ones*width_ones; i++)
         aus1[i]=1.0f;
     checkCuda( cudaMemcpy(ones_d1, aus1, (height_ones*width_ones)*sizeof(dnnType), cudaMemcpyHostToDevice) );
-    
+    checkCuda( cudaFreeHost(aus1) );
     checkCuda( cudaMalloc(&ones_d2, dim_ones*sizeof(dnnType)) );
-    float aus2[dim_ones];
+    dnnType *aus2;
+    checkCuda( cudaMallocHost(&aus2, dim_ones*sizeof(dnnType)) );
     for(int i=0; i<dim_ones; i++)
         aus2[i]=1.0f;
     checkCuda( cudaMemcpy(ones_d2, aus2, (dim_ones)*sizeof(dnnType), cudaMemcpyHostToDevice) );
+    checkCuda( cudaFreeHost(aus2) );
     checkCuda( cudaDeviceSynchronize() );
 }
 

tests/dla34_cnet/dla34_cnet.cpp

@@ -104,7 +104,7 @@ int main()
 {
 
     // Network layout
-    tk::dnn::dataDim_t dim(1, 3, 224, 224, 1);
+    tk::dnn::dataDim_t dim(1, 3, 512, 512, 1);
     tk::dnn::Network net(dim);
     tk::dnn::Layer *last1, *last2, *last3, *last4;
     tk::dnn::Layer *base1, *base2, *base3, *base4, *base5, *base6, *ida1, *ida2_1, *ida2_2, *ida3_1, *ida3_2, *ida3_3, *idaup_1, *idaup_2;

tests/resnet101_cnet/resnet101_cnet.cpp

@@ -187,7 +187,7 @@ int main()
 {
 
     // Network layout
-    tk::dnn::dataDim_t dim(1, 3, 224, 224, 1);
+    tk::dnn::dataDim_t dim(1, 3, 512, 512, 1);
     tk::dnn::Network net(dim);
 
     tk::dnn::Conv2d conv1(&net, 64, 7, 7, 2, 2, 3, 3, conv1_bin, true);