| 610 | } |
| 611 | |
| 612 | void test_softmax() |
| 613 | { |
| 614 | using namespace dlib::tt; |
| 615 | print_spinner(); |
| 616 | const long nr = 3; |
| 617 | const long nc = 3; |
| 618 | resizable_tensor src(5,5,nr,nr), dest(5,5,nr,nc), gradient_input(5,5,nr,nc); |
| 619 | tt::tensor_rand rnd; |
| 620 | rnd.fill_uniform(src); |
| 621 | rnd.fill_uniform(dest); |
| 622 | // fill like this as a test of the assignment operator. |
| 623 | gradient_input = matrix_cast<float>(gaussian_randm(5,5*nr*nc, 2)); |
| 624 | |
| 625 | |
| 626 | |
| 627 | auto grad_src = [&](long idx) { |
| 628 | auto f = [&](float eps) { |
| 629 | const float old = src.host()[idx]; |
| 630 | src.host()[idx] += eps; |
| 631 | tt::softmax(dest, src); |
| 632 | float result = dot(gradient_input, dest); |
| 633 | src.host()[idx] = old; |
| 634 | return result; |
| 635 | }; |
| 636 | const float eps = 0.01; |
| 637 | return (f(+eps)-f(-eps))/(2*eps); |
| 638 | }; |
| 639 | |
| 640 | resizable_tensor src_grad; |
| 641 | src_grad.copy_size(src); |
| 642 | src_grad = 0; |
| 643 | |
| 644 | tt::softmax(dest, src); |
| 645 | softmax_gradient(src_grad, dest, gradient_input); |
| 646 | |
| 647 | auto grad_error = compare_gradients(src_grad, grad_src); |
| 648 | dlog << LINFO << "src error: " << grad_error; |
| 649 | DLIB_TEST(grad_error < 0.001); |
| 650 | |
| 651 | IF_DLIB_USE_CUDA( |
| 652 | resizable_tensor src1 = src; |
| 653 | resizable_tensor src2 = src; |
| 654 | resizable_tensor dest1, dest2; |
| 655 | dest1.copy_size(src); |
| 656 | dest2.copy_size(src); |
| 657 | cuda::softmax_all(dest1, src1); |
| 658 | cpu::softmax_all(dest2, src2); |
| 659 | DLIB_TEST_MSG(max(abs(mat(dest1)-mat(dest2))) < 1e-5, max(abs(mat(dest1)-mat(dest2)))); |
| 660 | ) |
| 661 | } |
| 662 | |
| 663 | void test_softmaxm() |
| 664 | { |
no test coverage detected