Minor fix for shortcut multilayer if it doesn't use weights

src/blas.c

@@ -79,7 +79,7 @@ void shortcut_multilayer_cpu(int size, int src_outputs, int batch, int n, int *o
     // nweights - l.n or l.n*l.c or (l.n*l.c*l.h*l.w)
     const int layer_step = nweights / (n + 1);    // 1 or l.c or (l.c * l.h * l.w)
     int step = 0;
-    if (weights) step = src_outputs / layer_step; // (l.c * l.h * l.w) or (l.w*l.h) or 1
+    if (nweights > 0) step = src_outputs / layer_step; // (l.c * l.h * l.w) or (l.w*l.h) or 1
 
     int id;
     #pragma omp parallel for
@@ -148,7 +148,7 @@ void backward_shortcut_multilayer_cpu(int size, int src_outputs, int batch, int
     // nweights - l.n or l.n*l.c or (l.n*l.c*l.h*l.w)
     const int layer_step = nweights / (n + 1);    // 1 or l.c or (l.c * l.h * l.w)
     int step = 0;
-    if (weights) step = src_outputs / layer_step; // (l.c * l.h * l.w) or (l.w*l.h) or 1
+    if (nweights > 0) step = src_outputs / layer_step; // (l.c * l.h * l.w) or (l.w*l.h) or 1
 
     int id;
     #pragma omp parallel for

src/blas_kernels.cu

@@ -687,7 +687,7 @@ __global__ void shortcut_multilayer_kernel(int size, int src_outputs, int batch,
     // nweights - l.n or l.n*l.c or (l.n*l.c*l.h*l.w)
     const int layer_step = nweights / (n + 1);    // 1 or l.c or (l.c * l.h * l.w)
     int step = 0;
-    if (weights_gpu) step = src_outputs / layer_step; // (l.c * l.h * l.w) or (l.w*l.h) or 1
+    if (nweights > 0) step = src_outputs / layer_step; // (l.c * l.h * l.w) or (l.w*l.h) or 1
 
     int src_id = id;
     const int src_i = src_id % src_outputs;
@@ -762,7 +762,7 @@ __global__ void backward_shortcut_multilayer_kernel(int size, int src_outputs, i
     // nweights - l.n or l.n*l.c or (l.n*l.c*l.h*l.w)
     const int layer_step = nweights / (n + 1);    // 1 or l.c or (l.c * l.h * l.w)
     int step = 0;
-    if (weights_gpu) step = src_outputs / layer_step; // (l.c * l.h * l.w) or (l.w*l.h) or 1
+    if (nweights > 0) step = src_outputs / layer_step; // (l.c * l.h * l.w) or (l.w*l.h) or 1
 
     int src_id = id;
     const int src_i = src_id % src_outputs;
@@ -836,7 +836,8 @@ extern "C" void backward_shortcut_multilayer_gpu(int src_outputs, int batch, int
     float **layers_delta_gpu, float *delta_out, float *delta_in, float *weights_gpu, float *weight_updates_gpu, int nweights, float *in, float **layers_output_gpu, WEIGHTS_NORMALIZATION_T weights_normalizion)
 {
     const int layer_step = nweights / (n + 1);    // 1 or l.c or (l.c * l.h * l.w)
-    const int step = src_outputs / layer_step; // (l.c * l.h * l.w) or (l.w*l.h) or 1
+    int step = 0;
+    if (nweights > 0) step = src_outputs / layer_step; // (l.c * l.h * l.w) or (l.w*l.h) or 1
     //printf(" nweights = %d, n = %d, layer_step = %d, step = %d \n", nweights, n, layer_step, step);
 
     //printf(" src_outputs = %d, batch = %d, n = %d \n", src_outputs, batch, n);

src/darknet.c

@@ -466,6 +466,8 @@ int main(int argc, char **argv)
 
     show_opencv_info();
 
+    init_cpu();
+
     if (0 == strcmp(argv[1], "average")){
         average(argc, argv);
     } else if (0 == strcmp(argv[1], "yolo")){

src/parser.c

@@ -1544,7 +1544,6 @@ void save_shortcut_weights(layer l, FILE *fp)
 #endif
     int num = l.nweights;
     fwrite(l.weights, sizeof(float), num, fp);
-
 }
 
 void save_convolutional_weights(layer l, FILE *fp)
@@ -1822,10 +1821,6 @@ void load_convolutional_weights(layer l, FILE *fp)
 
 void load_shortcut_weights(layer l, FILE *fp)
 {
-    if (l.binary) {
-        //load_convolutional_weights_binary(l, fp);
-        //return;
-    }
     int num = l.nweights;
     int read_bytes;
     read_bytes = fread(l.weights, sizeof(float), num, fp);

src/shortcut_layer.c

@@ -186,6 +186,7 @@ void backward_shortcut_layer(const layer l, network_state state)
 
 void update_shortcut_layer(layer l, int batch, float learning_rate_init, float momentum, float decay)
 {
+    if (l.nweights > 0) {
         float learning_rate = learning_rate_init*l.learning_rate_scale;
         //float momentum = a.momentum;
         //float decay = a.decay;
@@ -194,6 +195,7 @@ void update_shortcut_layer(layer l, int batch, float learning_rate_init, float m
         axpy_cpu(l.nweights, -decay*batch, l.weights, 1, l.weight_updates, 1);
         axpy_cpu(l.nweights, learning_rate / batch, l.weight_updates, 1, l.weights, 1);
         scal_cpu(l.nweights, momentum, l.weight_updates, 1);
+    }
 }
 
 #ifdef GPU
@@ -238,6 +240,7 @@ void backward_shortcut_layer_gpu(const layer l, network_state state)
 
 void update_shortcut_layer_gpu(layer l, int batch, float learning_rate_init, float momentum, float decay)
 {
+    if (l.nweights > 0) {
         float learning_rate = learning_rate_init*l.learning_rate_scale;
         //float momentum = a.momentum;
         //float decay = a.decay;
@@ -253,13 +256,14 @@ void update_shortcut_layer_gpu(layer l, int batch, float learning_rate_init, flo
         //if (l.clip) {
         //    constrain_gpu(l.nweights, l.clip, l.weights_gpu, 1);
         //}
+    }
 }
 
 void pull_shortcut_layer(layer l)
 {
     cuda_pull_array_async(l.weights_gpu, l.weights, l.nweights);
     CHECK_CUDA(cudaPeekAtLastError());
-    cudaStreamSynchronize(get_cuda_stream());
+    CHECK_CUDA(cudaStreamSynchronize(get_cuda_stream()));
 }
 
 void push_shortcut_layer(layer l)