| 656 | } |
| 657 | |
| 658 | void batch_normalize_conv_gradient( |
| 659 | const double eps, |
| 660 | const tensor& gradient_input, |
| 661 | const tensor& means, |
| 662 | const tensor& invstds, |
| 663 | const tensor& src, |
| 664 | const tensor& gamma, |
| 665 | tensor& src_grad, |
| 666 | tensor& gamma_grad, |
| 667 | tensor& beta_grad |
| 668 | ) |
| 669 | { |
| 670 | DLIB_CASSERT(src.k() == (long)means.size()); |
| 671 | DLIB_CASSERT(src.k() == (long)invstds.size()); |
| 672 | DLIB_CASSERT(src.k() == (long)gamma.size()); |
| 673 | DLIB_CASSERT(src.k() == (long)gamma_grad.size()); |
| 674 | DLIB_CASSERT(src.k() == (long)beta_grad.size()); |
| 675 | DLIB_CASSERT(have_same_dimensions(gradient_input, src)); |
| 676 | DLIB_CASSERT(have_same_dimensions(gradient_input, src_grad)); |
| 677 | DLIB_CASSERT(eps > 0); |
| 678 | |
| 679 | const float in_scale = 1; |
| 680 | const float out_scale = 1; |
| 681 | const float in_scale_params = 1; |
| 682 | const float out_scale_params = 0; |
| 683 | |
| 684 | CHECK_CUDNN(cudnnBatchNormalizationBackward( |
| 685 | context(), |
| 686 | CUDNN_BATCHNORM_SPATIAL, |
| 687 | &in_scale, |
| 688 | &out_scale, |
| 689 | &in_scale_params, |
| 690 | &out_scale_params, |
| 691 | descriptor(src), |
| 692 | src.device(), |
| 693 | descriptor(gradient_input), |
| 694 | gradient_input.device(), |
| 695 | descriptor(src_grad), |
| 696 | src_grad.device(), |
| 697 | descriptor(gamma), |
| 698 | gamma.device(), |
| 699 | gamma_grad.device(), |
| 700 | beta_grad.device(), |
| 701 | eps, |
| 702 | means.device(), |
| 703 | invstds.device())); |
| 704 | } |
| 705 | |
| 706 | // ------------------------------------------------------------------------------------ |
| 707 | // ------------------------------------------------------------------------------------ |
nothing calls this directly
no test coverage detected