https://stackoverflow.com/questions/39274472/error-function-atomicadddouble-double-has-already-been-defined

I finally got it working with the help of @Robert Crovella's comment. I had to modify the file common.cuh from the DeepLab_v2 master branch in the following way:

#ifndef CAFFE_COMMON_CUH_
#define CAFFE_COMMON_CUH_

#include <cuda.h>

  #if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 600

  #else
  static __inline__ __device__ double atomicAdd(double *address, double val) {
    unsigned long long int* address_as_ull = (unsigned long long int*)address;
    unsigned long long int old = *address_as_ull, assumed;
    if (val==0.0)
      return __longlong_as_double(old);
    do {
      assumed = old;
      old = atomicCAS(address_as_ull, assumed, __double_as_longlong(val +__longlong_as_double(assumed)));
    } while (assumed != old);
    return __longlong_as_double(old);
  }


  #endif
#endif