I need to sum all the elements of a matrix in caffe,
But as I noticed, the caffe wrapper of the cblas functions ('math_functions.hpp'
& 'math_functions.cpp'
) is using cblas_sasum
function as caffe_cpu_asum
that computes the sum of the absolute values of elements in a vector.
Since I'm a newbie in cblas, I tried to find a suitable function to get rid of absolute there, but it seems that there is no function with that property in cblas.
Any suggestion?
There is a way to do so using cblas functions, though it is a bit of an awkward way.
What you need to do is to define an "all 1" vector, and then do a dot product between this vector and your matrix, the result is the sum.
Let myBlob
be a caffe Blob whose elements you want to sum:
vector<Dtype> mult_data( myBlob.count(), Dtype(1) );
Dtype sum = caffe_cpu_dot( myBlob.count(), &mult_data[0], myBlob.cpu_data() );
This trick is used in the implementation of "Reduction"
layer.
To make this answer both GPU compliant, one need to allocate a Blob
for mult_data
and not a std::vector
(because you need it's pgu_data()
):
vector<int> sum_mult_shape(1, diff_.count());
Blob<Dtype> sum_multiplier_(sum_mult_shape);
const Dtype* mult_data = sum_multiplier_.cpu_data();
Dtype sum = caffe_cpu_dot( myBlob.count(), &mult_data[0], myBlob.cpu_data() );
For GPU, (in a '.cu'
source file):
vector<int> sum_mult_shape(1, diff_.count());
Blob<Dtype> sum_multiplier_(sum_mult_shape);
const Dtype* mult_data = sum_multiplier_.gpu_data();
Dtype sum;
caffe_gpu_dot( myBlob.count(), &mult_data[0], myBlob.gpu_data(), &sum );
Summation of all the elements of an array is simple enough to be implemented by a single for-loop. You only need to use proper compile options to vectorise it with SIMD instructions.
For Blob in caffe, you could use .cpu_data()
to get the raw pointer of the array and then use for-loop.