Fix OpenCV compilation. maxpool_layer on GPU uses cuDNN.

src/darknet.c

@@ -370,6 +370,8 @@ int main(int argc, char **argv)
     gpu_index = find_int_arg(argc, argv, "-i", 0);
     if(find_arg(argc, argv, "-nogpu")) {
         gpu_index = -1;
+        printf("\n Currently Darknet doesn't support -nogpu flag. If you want to use CPU - please compile Darknet with GPU=0 in the Makefile, or compile darknet_no_gpu.sln on Windows.\n");
+        exit(-1);
     }
 
 #ifndef GPU

src/data.c

@@ -336,8 +336,8 @@ void fill_truth_detection(char *path, int num_boxes, float *truth, int classes,
         // if truth (box for object) is smaller than 1x1 pix
         char buff[256];
         if (id >= classes) {
-            printf("\n Wrong annotation: class_id = %d. But class_id should be [from 0 to %d] \n", id, classes);
+            printf("\n Wrong annotation: class_id = %d. But class_id should be [from 0 to %d] \n", id, (classes-1));
-            sprintf(buff, "echo %s \"Wrong annotation: class_id = %d. But class_id should be [from 0 to %d]\" >> bad_label.list", labelpath, id, classes);
+            sprintf(buff, "echo %s \"Wrong annotation: class_id = %d. But class_id should be [from 0 to %d]\" >> bad_label.list", labelpath, id, (classes-1));
             system(buff);
             getchar();
             ++sub;

src/image.c

@@ -21,6 +21,8 @@
 #include "http_stream.h"
 #endif
 #include "http_stream.h"
+
+#define CV_RGB(r, g, b) cvScalar( (b), (g), (r), 0 )
 #endif
 
 extern int check_mistakes;

src/layer.h

@@ -307,6 +307,7 @@ struct layer{
     cudnnConvolutionFwdAlgo_t fw_algo;
     cudnnConvolutionBwdDataAlgo_t bd_algo;
     cudnnConvolutionBwdFilterAlgo_t bf_algo;
+    cudnnPoolingDescriptor_t poolingDesc;
     #endif
     #endif
 };

src/maxpool_layer.c

@@ -47,7 +47,27 @@ maxpool_layer make_maxpool_layer(int batch, int h, int w, int c, int size, int s
     l.indexes_gpu = cuda_make_int_array(output_size);
     l.output_gpu  = cuda_make_array(l.output, output_size);
     l.delta_gpu   = cuda_make_array(l.delta, output_size);
-    #endif
+#ifdef CUDNN
+    cudnnStatus_t maxpool_status;
+    maxpool_status = cudnnCreatePoolingDescriptor(&l.poolingDesc);
+
+    maxpool_status = cudnnSetPooling2dDescriptor(
+        l.poolingDesc,
+        CUDNN_POOLING_MAX,
+        CUDNN_PROPAGATE_NAN,    // CUDNN_PROPAGATE_NAN, CUDNN_NOT_PROPAGATE_NAN
+        l.size,
+        l.size,
+        0, //l.pad,
+        0, //l.pad,
+        l.stride,
+        l.stride);
+
+    cudnnCreateTensorDescriptor(&l.srcTensorDesc);
+    cudnnCreateTensorDescriptor(&l.dstTensorDesc);
+    cudnnSetTensor4dDescriptor(l.srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.c, l.h, l.w);
+    cudnnSetTensor4dDescriptor(l.dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.out_c, l.out_h, l.out_w);
+#endif // CUDNN
+    #endif  // GPU
 	l.bflops = (l.size*l.size*l.c * l.out_h*l.out_w) / 1000000000.;
     fprintf(stderr, "max          %d x %d / %d  %4d x%4d x%4d   ->  %4d x%4d x%4d %5.3f BF\n", size, size, stride, w, h, c, l.out_w, l.out_h, l.out_c, l.bflops);
     return l;

src/maxpool_layer_kernels.cu

@@ -86,6 +86,33 @@ __global__ void backward_maxpool_layer_kernel(int n, int in_h, int in_w, int in_
 
 extern "C" void forward_maxpool_layer_gpu(maxpool_layer layer, network_state state)
 {
+
+#ifdef CUDNN
+    if (!state.train) {// && layer.stride == layer.size) {
+        // cudnnPoolingBackward
+        cudnnStatus_t maxpool_status;
+
+        float alpha = 1, beta = 0;
+        maxpool_status = cudnnPoolingForward(
+            cudnn_handle(),
+            layer.poolingDesc,
+            &alpha,
+            layer.srcTensorDesc,
+            state.input,
+            &beta,
+            layer.dstTensorDesc,
+            layer.output_gpu);
+
+        //maxpool_status = cudnnDestroyPoolingDescriptor(poolingDesc);
+        //cudnnDestroyTensorDescriptor(layer.srcTensorDesc);
+        //cudnnDestroyTensorDescriptor(layer.dstTensorDesc);
+
+        return;
+    }
+#endif
+
+
+
     int h = layer.out_h;
     int w = layer.out_w;
     int c = layer.c;