Add CenterNet based on Resnet101, TensorRT not implemented.

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

include/tkDNN/Layer.h

@@ -12,6 +12,7 @@ enum layerType_t {
     LAYER_DENSE,
     LAYER_CONV2D,
     LAYER_DECONV2D,
+    LAYER_DEFORMCONV2D,
     LAYER_ACTIVATION,
     LAYER_FLATTEN,
     LAYER_MULADD,
@@ -49,6 +50,7 @@ public:
             case LAYER_DENSE:       return "Dense";
             case LAYER_CONV2D:      return "Conv2d";
             case LAYER_DECONV2D:    return "DeConv2d";
+            case LAYER_DEFORMCONV2D:return "DeformConv2d";
             case LAYER_ACTIVATION:  return "Activation";
             case LAYER_FLATTEN:     return "Flatten";
             case LAYER_MULADD:      return "MulAdd";
@@ -78,7 +80,7 @@ class LayerWgs : public Layer {
 
 public:
     LayerWgs(Network *net, int inputs, int outputs, int kh, int kw, int kt,
-             std::string fname_weights, bool batchnorm = false); 
+             std::string fname_weights, bool batchnorm = false, bool additional_bias = false); 
     virtual ~LayerWgs();
 
     int inputs, outputs;
@@ -87,6 +89,10 @@ public:
     dnnType *data_h, *data_d;
     dnnType *bias_h, *bias_d;
 
+    // additional bias for DCN
+    bool additional_bias;
+    dnnType *bias2_h, *bias2_d;
+
     //batchnorm
     bool batchnorm;
     dnnType *power_h;
@@ -194,6 +200,31 @@ public:
 };
 
 
+/**
+    Deformable Convolutionl 2d layer
+*/  
+class DeformConv2d : public LayerWgs {
+
+public:
+    DeformConv2d( Network *net, int out_ch, int deformable_group, int kernelH, int kernelW,
+                int strideH, int strideW, int paddingH, int paddingW,
+                std::string d_fname_weights, std::string fname_weights, bool batchnorm);
+    virtual ~DeformConv2d();
+    virtual layerType_t getLayerType() { return LAYER_DEFORMCONV2D; };
+
+    virtual dnnType* infer(dataDim_t &dim, dnnType* srcData);
+    tk::dnn::Conv2d *preconv; 
+    int out_ch;
+    int deformableGroup;
+    int kernelH, kernelW, strideH, strideW, paddingH, paddingW;
+protected:
+
+    cudnnTensorDescriptor_t biasTensorDesc;
+
+    void initCUDNN();
+
+};
+
 /**
     Flatten layer
     is actually a matrix transposition

include/tkDNN/Network.h

@@ -32,7 +32,7 @@ struct dataDim_t {
 };
 
 class Layer;
-const int MAX_LAYERS = 256;
+const int MAX_LAYERS = 512;
 
 class Network {
 

include/tkDNN/kernels.h

@@ -33,4 +33,17 @@ void modulated_deformable_im2col_cuda(cudaStream_t stream,
                                 const int pad_h, const int pad_w, const int stride_h, const int stride_w,
                                 const int dilation_h, const int dilation_w,
                                 const int deformable_group, float *data_col);
+
+void dcn_v2_cuda_forward(float *input, float *weight,
+                         float *bias, float *ones,
+                         float *offset, float *mask,
+                         float *output, float *columns,
+                         int kernel_h, int kernel_w,
+                         const int stride_h, const int stride_w,
+                         const int pad_h, const int pad_w,
+                         const int dilation_h, const int dilation_w,
+                         const int deformable_group, 
+                         const int in_n, const int in_c, const int in_h, const int in_w,
+                         const int out_n, const int out_c, const int out_h, const int out_w,
+                         const int dst_dim, cudaStream_t stream = cudaStream_t(0));
 #endif //KERNELS_H