CODEWITHSUNDEEP
caffematrix-multiplicationcblasinner-product
ahmad navid ghanizadeh
ahmad navid ghanizadeh

Reputation: 139

How to do dot product between matrices in caffe?

In inner product layer, I need to multiply (top_diff * bottom_data) .* (2*weight). First we calculate (result = top_diff * bottom_data) as matrix multiplication in caffe_cpu_gemm and then do a dot product between weight and result.

More explanation is defined as follow:

const Dtype* weight = this->blobs_[0]->cpu_data();
     if (this->param_propagate_down_[0]) {
            const Dtype* top_diff = top[0]->cpu_diff();
            const Dtype* bottom_data = bottom[0]->cpu_data();
caffe_cpu_gemm<Dtype>(CblasTrans, CblasNoTrans, N_, K_, M_, (Dtype)1.,
        top_diff, bottom_data, (Dtype)1., this->blobs_[0]->mutable_cpu_diff());
}

For more understanding, I checked math_function.c. It is implemented as follows:

template<>
void caffe_cpu_gemm<float>(const CBLAS_TRANSPOSE TransA,
    const CBLAS_TRANSPOSE TransB, const int M, const int N, const int K,
    const float alpha, const float* A, const float* B, const float beta,
    float* C) {
  int lda = (TransA == CblasNoTrans) ? K : M;
  int ldb = (TransB == CblasNoTrans) ? N : K;
  cblas_sgemm(CblasRowMajor, TransA, TransB, M, N, K, alpha, A, lda, B,
      ldb, beta, C, N);
}

I think I should perform multiplication (result = top_diff * bottom_data) in caffe_cpu_gemm() and after that do dot product with weight. how should I do?!

Many thanks!!!! Any advice would be appreciated!

Upvotes: 1

Views: 891

Answers (1)

Bharat
Bharat

Reputation: 2179

If you just want to perform dot product between two matrices, you can use the following function to multiply matrices on CPU,

void caffe_mul<float>(const int n, const float* a, const float* b, float* y)

If you want to do the same operation on a GPU, use this template

void caffe_gpu_mul<float>(const int N, const float* a, const float* b, float* y)

a and b are you matrices and c will contain the final result. N is the total number of elements in your matrix.

You can also use the 'Eltwise' layer, which already does this.

Upvotes: 1

Related Questions