| 741 | } |
| 742 | |
| 743 | void test_softmax_all() |
| 744 | { |
| 745 | using namespace dlib::tt; |
| 746 | print_spinner(); |
| 747 | const long nr = 3; |
| 748 | const long nc = 3; |
| 749 | resizable_tensor src(5,5,nr,nc), dest(5,5,nr,nc), gradient_input(5,5,nr,nc); |
| 750 | tt::tensor_rand rnd; |
| 751 | rnd.fill_uniform(src); |
| 752 | rnd.fill_uniform(dest); |
| 753 | // fill like this as a test of the assignment operator. |
| 754 | gradient_input = matrix_cast<float>(gaussian_randm(5,5*nr*nc, 2)); |
| 755 | |
| 756 | |
| 757 | |
| 758 | auto grad_src = [&](long idx) { |
| 759 | auto f = [&](float eps) { |
| 760 | const float old = src.host()[idx]; |
| 761 | src.host()[idx] += eps; |
| 762 | tt::softmax_all(dest, src); |
| 763 | float result = dot(gradient_input, dest); |
| 764 | src.host()[idx] = old; |
| 765 | return result; |
| 766 | }; |
| 767 | const float eps = 0.01; |
| 768 | return (f(+eps)-f(-eps))/(2*eps); |
| 769 | }; |
| 770 | |
| 771 | resizable_tensor src_grad; |
| 772 | src_grad.copy_size(src); |
| 773 | src_grad = 0; |
| 774 | |
| 775 | tt::softmax_all(dest, src); |
| 776 | softmax_all_gradient(src_grad, dest, gradient_input); |
| 777 | |
| 778 | auto grad_error = compare_gradients(src_grad, grad_src); |
| 779 | dlog << LINFO << "src error: " << grad_error; |
| 780 | DLIB_TEST(grad_error < 0.001); |
| 781 | |
| 782 | IF_DLIB_USE_CUDA( |
| 783 | resizable_tensor src1 = src; |
| 784 | resizable_tensor src2 = src; |
| 785 | resizable_tensor dest1, dest2; |
| 786 | dest1.copy_size(src); |
| 787 | dest2.copy_size(src); |
| 788 | cuda::softmax_all(dest1, src1); |
| 789 | cpu::softmax_all(dest2, src2); |
| 790 | DLIB_TEST_MSG(max(abs(mat(dest1)-mat(dest2))) < 1e-5, max(abs(mat(dest1)-mat(dest2)))); |
| 791 | ) |
| 792 | } |
| 793 | |
| 794 | void test_mish() |
| 795 | { |
no test coverage detected