Re: [FFmpeg-devel] [PATCH V2 1/3] dnn: add layer pad which is equivalent to tf.pad
LGTM. Pushed, thanks! Em dom, 28 de jul de 2019 às 22:59, Guo, Yejun escreveu: > > the reason to add this layer first is that vf_sr uses it in its > tensorflow model, and the next plan is to update the python script > to convert tf.pad into native model. > > Signed-off-by: Guo, Yejun > --- > libavfilter/dnn/Makefile | 1 + > libavfilter/dnn/dnn_backend_native_layer_pad.c | 211 > + > libavfilter/dnn/dnn_backend_native_layer_pad.h | 40 + > 3 files changed, 252 insertions(+) > create mode 100644 libavfilter/dnn/dnn_backend_native_layer_pad.c > create mode 100644 libavfilter/dnn/dnn_backend_native_layer_pad.h > > diff --git a/libavfilter/dnn/Makefile b/libavfilter/dnn/Makefile > index 1d12ade..83938e5 100644 > --- a/libavfilter/dnn/Makefile > +++ b/libavfilter/dnn/Makefile > @@ -1,5 +1,6 @@ > OBJS-$(CONFIG_DNN) += dnn/dnn_interface.o > OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native.o > +OBJS-$(CONFIG_DNN) += > dnn/dnn_backend_native_layer_pad.o > > DNN-OBJS-$(CONFIG_LIBTENSORFLOW) += dnn/dnn_backend_tf.o > > diff --git a/libavfilter/dnn/dnn_backend_native_layer_pad.c > b/libavfilter/dnn/dnn_backend_native_layer_pad.c > new file mode 100644 > index 000..5417d73 > --- /dev/null > +++ b/libavfilter/dnn/dnn_backend_native_layer_pad.c > @@ -0,0 +1,211 @@ > +/* > + * Copyright (c) 2019 Guo Yejun > + * > + * This file is part of FFmpeg. > + * > + * FFmpeg is free software; you can redistribute it and/or > + * modify it under the terms of the GNU Lesser General Public > + * License as published by the Free Software Foundation; either > + * version 2.1 of the License, or (at your option) any later version. > + * > + * FFmpeg is distributed in the hope that it will be useful, > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU > + * Lesser General Public License for more details. > + * > + * You should have received a copy of the GNU Lesser General Public > + * License along with FFmpeg; if not, write to the Free Software > + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 > USA > + */ > + > +#include > +#include "libavutil/avassert.h" > +#include "dnn_backend_native_layer_pad.h" > + > +static int before_get_buddy(int given, int paddings, LayerPadModeParam mode) > +{ > +if (mode == LPMP_SYMMETRIC) { > +return (2 * paddings - 1 - given); > +} else if (mode == LPMP_REFLECT) { > +return (2 * paddings - given); > +} else { > +av_assert0(!"should not reach here"); > +return 0; > +} > +} > + > +static int after_get_buddy(int given, int border, LayerPadModeParam mode) > +{ > +if (mode == LPMP_SYMMETRIC) { > +int offset = given - border; > +return (border - 1 - offset); > +} else if (mode == LPMP_REFLECT) { > +int offset = given - border; > +return (border - 2 - offset); > +} else { > +av_assert0(!"should not reach here"); > +return 0; > +} > +} > + > +void dnn_execute_layer_pad(const float *input, float *output, const > LayerPadParams *params, int number, int height, int width, int channel) > +{ > +int32_t before_paddings; > +int32_t after_paddings; > + > +// suppose format is > +int new_number = number + params->paddings[0][0] + > params->paddings[0][1]; > +int new_height = height + params->paddings[1][0] + > params->paddings[1][1]; > +int new_width = width + params->paddings[2][0] + params->paddings[2][1]; > +int new_channel = channel + params->paddings[3][0] + > params->paddings[3][1]; > + > +int c_stride = channel; > +int wc_stride = c_stride * width; > +int hwc_stride = wc_stride * height; > + > +int new_c_stride = new_channel; > +int new_wc_stride = new_c_stride * new_width; > +int new_hwc_stride = new_wc_stride * new_height; > + > +// copy the original data > +for (int n = 0; n < number; n++) { > +for (int h = 0; h < height; h++) { > +for (int w = 0; w < width; w++) { > +const float *src = input + n * hwc_stride + h * wc_stride + > w * c_stride; > +float *dst = output + (n + params->paddings[0][0]) * > new_hwc_stride > ++ (h + params->paddings[1][0]) * > new_wc_stride > ++ (w + params->paddings[2][0]) * > new_c_stride > ++ params->paddings[3][0]; > +memcpy(dst, src, channel * sizeof(float)); > +} > +} > +} > + > +// handle the first dimension > +before_paddings = params->paddings[0][0]; > +after_paddings = params->paddings[0][1]; > +for (int n = 0; n < before_paddings; n++) { > +float *dst = output + n * new_hwc_stride; > +if (params->mode ==
[FFmpeg-devel] [PATCH V2 1/3] dnn: add layer pad which is equivalent to tf.pad
the reason to add this layer first is that vf_sr uses it in its tensorflow model, and the next plan is to update the python script to convert tf.pad into native model. Signed-off-by: Guo, Yejun --- libavfilter/dnn/Makefile | 1 + libavfilter/dnn/dnn_backend_native_layer_pad.c | 211 + libavfilter/dnn/dnn_backend_native_layer_pad.h | 40 + 3 files changed, 252 insertions(+) create mode 100644 libavfilter/dnn/dnn_backend_native_layer_pad.c create mode 100644 libavfilter/dnn/dnn_backend_native_layer_pad.h diff --git a/libavfilter/dnn/Makefile b/libavfilter/dnn/Makefile index 1d12ade..83938e5 100644 --- a/libavfilter/dnn/Makefile +++ b/libavfilter/dnn/Makefile @@ -1,5 +1,6 @@ OBJS-$(CONFIG_DNN) += dnn/dnn_interface.o OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native.o +OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_pad.o DNN-OBJS-$(CONFIG_LIBTENSORFLOW) += dnn/dnn_backend_tf.o diff --git a/libavfilter/dnn/dnn_backend_native_layer_pad.c b/libavfilter/dnn/dnn_backend_native_layer_pad.c new file mode 100644 index 000..5417d73 --- /dev/null +++ b/libavfilter/dnn/dnn_backend_native_layer_pad.c @@ -0,0 +1,211 @@ +/* + * Copyright (c) 2019 Guo Yejun + * + * This file is part of FFmpeg. + * + * FFmpeg is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * FFmpeg is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#include +#include "libavutil/avassert.h" +#include "dnn_backend_native_layer_pad.h" + +static int before_get_buddy(int given, int paddings, LayerPadModeParam mode) +{ +if (mode == LPMP_SYMMETRIC) { +return (2 * paddings - 1 - given); +} else if (mode == LPMP_REFLECT) { +return (2 * paddings - given); +} else { +av_assert0(!"should not reach here"); +return 0; +} +} + +static int after_get_buddy(int given, int border, LayerPadModeParam mode) +{ +if (mode == LPMP_SYMMETRIC) { +int offset = given - border; +return (border - 1 - offset); +} else if (mode == LPMP_REFLECT) { +int offset = given - border; +return (border - 2 - offset); +} else { +av_assert0(!"should not reach here"); +return 0; +} +} + +void dnn_execute_layer_pad(const float *input, float *output, const LayerPadParams *params, int number, int height, int width, int channel) +{ +int32_t before_paddings; +int32_t after_paddings; + +// suppose format is +int new_number = number + params->paddings[0][0] + params->paddings[0][1]; +int new_height = height + params->paddings[1][0] + params->paddings[1][1]; +int new_width = width + params->paddings[2][0] + params->paddings[2][1]; +int new_channel = channel + params->paddings[3][0] + params->paddings[3][1]; + +int c_stride = channel; +int wc_stride = c_stride * width; +int hwc_stride = wc_stride * height; + +int new_c_stride = new_channel; +int new_wc_stride = new_c_stride * new_width; +int new_hwc_stride = new_wc_stride * new_height; + +// copy the original data +for (int n = 0; n < number; n++) { +for (int h = 0; h < height; h++) { +for (int w = 0; w < width; w++) { +const float *src = input + n * hwc_stride + h * wc_stride + w * c_stride; +float *dst = output + (n + params->paddings[0][0]) * new_hwc_stride ++ (h + params->paddings[1][0]) * new_wc_stride ++ (w + params->paddings[2][0]) * new_c_stride ++ params->paddings[3][0]; +memcpy(dst, src, channel * sizeof(float)); +} +} +} + +// handle the first dimension +before_paddings = params->paddings[0][0]; +after_paddings = params->paddings[0][1]; +for (int n = 0; n < before_paddings; n++) { +float *dst = output + n * new_hwc_stride; +if (params->mode == LPMP_CONSTANT) { +for (int i = 0; i < new_hwc_stride; i++) { +dst[i] = params->constant_values; +} +} +else { +int buddy = before_get_buddy(n, before_paddings, params->mode); +float *src = output + buddy * new_hwc_stride; +memcpy(dst, src,