CODEWITHSUNDEEP
machine-learningcomputer-visionneural-networkcaffeconv-neural-network
Ali Sharifi B.
Ali Sharifi B.

Reputation: 595

How to loop over all the elements of a Blob variable in Caffe?

Working on a new loss layer in Caffe,

I have some values in diff_.cpu_data(), Lets name each element of that as Di :

Now, I want to compute this function for each Di :

enter image description here

and assign the result to the corresponding element in bottom[0]->mutable_cpu_diff() of the layer.

As you see, for the second term there is no need to loop over input and output variables ( diff_.cpu_data & bottom[0]->mutable_cpu_diff() respectively), while in the first term, I need to access the value of each element in input variable, then of course I need to assign the result of the function to the corresponding element of the output variable, If they were 2-D array, obviously I could do something like this:

enter image description here

But as you know, those variables are 4-D arrays and it is not clear for me how to do that.

Should I use Offset() function or something like that to loop over all the elements of those variables similar to this?

Could someone please explain it to me or refer me to a useful reference?

Thanks,

Upvotes: 1

Views: 468

Answers (1)

Shai
Shai

Reputation: 114796

First, you should store the result of the second term (1/N^2 \sum_i D_i) to a local variable. As you already know, this sum can be computed using caffe_cpu_dot.

So your code may look something like:

vector<Dtype> mult_data( diff_.count(), Dtype(1) );
const Dtype* Di = diff_.cpu_data();
Dtype const_sum = caffe_cpu_dot( diff_.count(), &mult_data[0], Di );
Dtype N = diff_.count();
const_sum /= N*N; // divide by N^2

Now you can loop over all items (assuming bottom[0]->count()==diff_.count()):

Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
for (int i=0; i<diff_.count(); i++) {
  bottom_diff[i] = 2*Di[i]/N - const_sum;
}

Alternatively, if you want something more "blas"-like:

caffe_copy(diff_.count(), Di, bottom_diff);  // copy Di to bottom_diff
caffe_scal(diff_.count(), Dtype(2.0/N), bottom_diff);  // bottom_diff is now (2/N)*Di
caffe_add_scalar(diff_.count(), -const_sum, bottom_diff);  // subtract the second term from all the elements.

Upvotes: 1

Related Questions