| 1699 | // ------------------------------------------------------------------------------------ |
| 1700 | |
| 1701 | void softmax( |
| 1702 | tensor& dest, |
| 1703 | const tensor& src, |
| 1704 | operation_mode mode |
| 1705 | ) |
| 1706 | { |
| 1707 | DLIB_CASSERT(have_same_dimensions(dest, src)); |
| 1708 | if (src.size() == 0) return; |
| 1709 | |
| 1710 | const float alpha = 1; |
| 1711 | const float beta = 0; |
| 1712 | |
| 1713 | if (mode == operation_mode::CHANNEL_WISE) |
| 1714 | { |
| 1715 | CHECK_CUDNN(cudnnSoftmaxForward(context(), |
| 1716 | CUDNN_SOFTMAX_ACCURATE, |
| 1717 | CUDNN_SOFTMAX_MODE_CHANNEL, |
| 1718 | &alpha, |
| 1719 | descriptor(src), |
| 1720 | src.device(), |
| 1721 | &beta, |
| 1722 | descriptor(dest), |
| 1723 | dest.device())); |
| 1724 | } |
| 1725 | else if (mode == operation_mode::PLANE_WISE) |
| 1726 | { |
| 1727 | const long num_samples = src.num_samples(); |
| 1728 | const long num_channels = src.k(); |
| 1729 | const size_t plane_size = src.nr() * src.nc(); |
| 1730 | |
| 1731 | for (long s = 0; s < num_samples; ++s) |
| 1732 | { |
| 1733 | for (long k = 0; k < num_channels; ++k) |
| 1734 | { |
| 1735 | auto src_slice = src.device() + (s * num_channels + k) * plane_size; |
| 1736 | auto dest_slice = dest.device() + (s * num_channels + k) * plane_size; |
| 1737 | auto a_src_slice = alias_tensor(src.nr(), src.nc())(src, (s * num_channels + k) * plane_size); |
| 1738 | auto a_dest_slice = alias_tensor(dest.nr(), dest.nc())(dest, (s * num_channels + k) * plane_size); |
| 1739 | |
| 1740 | CHECK_CUDNN(cudnnSoftmaxForward(context(), |
| 1741 | CUDNN_SOFTMAX_ACCURATE, |
| 1742 | CUDNN_SOFTMAX_MODE_CHANNEL, |
| 1743 | &alpha, |
| 1744 | descriptor(a_src_slice), |
| 1745 | src_slice, |
| 1746 | &beta, |
| 1747 | descriptor(a_dest_slice), |
| 1748 | dest_slice)); |
| 1749 | } |
| 1750 | } |
| 1751 | } |
| 1752 | } |
| 1753 | |
| 1754 | void softmax_gradient( |
| 1755 | tensor& grad, |
nothing calls this directly
no test coverage detected