Fixed training, calculation mAP and anchors for Yolo v3

AlexeyAB

2018-03-28 3fc3fd0f1f1924acb7195fe98fe68648e4056a09

 src/yolo_layer.c

@@ -378,9 +378,26 @@
        return;
    }

    cuda_pull_array(l.output_gpu, state.input, l.batch*l.inputs);
    forward_yolo_layer(l, state);
    //cuda_pull_array(l.output_gpu, state.input, l.batch*l.inputs);
   float *in_cpu = calloc(l.batch*l.inputs, sizeof(float));
   cuda_pull_array(l.output_gpu, in_cpu, l.batch*l.inputs);
   float *truth_cpu = 0;
   if (state.truth) {
      int num_truth = l.batch*l.truths;
      truth_cpu = calloc(num_truth, sizeof(float));
      cuda_pull_array(state.truth, truth_cpu, num_truth);
   }
   network_state cpu_state = state;
   cpu_state.net = state.net;
   cpu_state.index = state.index;
   cpu_state.train = state.train;
   cpu_state.truth = truth_cpu;
   cpu_state.input = in_cpu;
   forward_yolo_layer(l, cpu_state);
    //forward_yolo_layer(l, state);
    cuda_push_array(l.delta_gpu, l.delta, l.batch*l.outputs);
   free(in_cpu);
   if (cpu_state.truth) free(cpu_state.truth);
}

void backward_yolo_layer_gpu(const layer l, network_state state)