diff mbox

[FFmpeg-devel] libavfilter: Add multiple padding methods in FFmpeg dnn native mode.

Message ID 2e26f947.5d6c2.16a96cad3fb.Coremail.xwmeng@pku.edu.cn
State New
Headers show

Commit Message

xwmeng@pku.edu.cn May 8, 2019, 9:33 a.m. UTC 
This patch is for the support of derain filter project in GSoC. It adds supports for the following operations: 




 (1) Conv padding method: "SAME", "VALID" and "SAME_CLAMP_TO_EDGE"




These operations are all needed in derain filter. As we discussed before, the "SAME_CLAMP_TO_EDGE" method is the same as dnn native padding method in the current implementation. And the sr model generation code should be changed if mutiple padding method supports added. So I sent a PR (https://github.com/HighVoltageRocknRoll/sr/pull/4)to the original sr repo(https://github.com/HighVoltageRocknRoll/sr).



From c0724bb304a6f4c3ca935cccda5b810e5c4eceb1 Mon Sep 17 00:00:00 2001
From: Xuewei Meng <xwmeng@pku.edu.cn>
Date: Wed, 8 May 2019 17:32:30 +0800
Subject: [PATCH] Add multiple padding method in dnn native

Signed-off-by: Xuewei Meng <xwmeng@pku.edu.cn>
---
 libavfilter/dnn_backend_native.c | 52 ++++++++++++++++++++++++--------
 libavfilter/dnn_backend_native.h |  3 ++
 2 files changed, 43 insertions(+), 12 deletions(-)

Comments

Liu Steven May 8, 2019, 11:34 a.m. UTC | #1 
> 在 2019年5月8日，下午5:33，xwmeng@pku.edu.cn 写道：
> 
> 
> 
> 
> This patch is for the support of derain filter project in GSoC. It adds supports for the following operations: 
> 
> 
> 
> 
> (1) Conv padding method: "SAME", "VALID" and "SAME_CLAMP_TO_EDGE"
> 
> 
> 
> 
> These operations are all needed in derain filter. As we discussed before, the "SAME_CLAMP_TO_EDGE" method is the same as dnn native padding method in the current implementation. And the sr model generation code should be changed if mutiple padding method supports added. So I sent a PR (https://github.com/HighVoltageRocknRoll/sr/pull/4)to the original sr repo(https://github.com/HighVoltageRocknRoll/sr).
Cannot sure Sergey Lavrushkin is maintaining that repo, maybe Pedro Arthur can tick he.

> 
> 
> 
> From c0724bb304a6f4c3ca935cccda5b810e5c4eceb1 Mon Sep 17 00:00:00 2001
> From: Xuewei Meng <xwmeng@pku.edu.cn>
> Date: Wed, 8 May 2019 17:32:30 +0800
> Subject: [PATCH] Add multiple padding method in dnn native
> 
> Signed-off-by: Xuewei Meng <xwmeng@pku.edu.cn>
> ---
> libavfilter/dnn_backend_native.c | 52 ++++++++++++++++++++++++--------
> libavfilter/dnn_backend_native.h |  3 ++
> 2 files changed, 43 insertions(+), 12 deletions(-)
> 
> diff --git a/libavfilter/dnn_backend_native.c b/libavfilter/dnn_backend_native.c
> index 70d857f5f2..b7c0508d91 100644
> --- a/libavfilter/dnn_backend_native.c
> +++ b/libavfilter/dnn_backend_native.c
> @@ -59,6 +59,12 @@ static DNNReturnType set_input_output_native(void *model, DNNData *input, DNNDat
>                 return DNN_ERROR;
>             }
>             cur_channels = conv_params->output_num;
> +
> +            if(conv_params->padding_method == VALID){
> +                int pad_size = conv_params->kernel_size - 1;
> +                cur_height -= pad_size;
> +                cur_width -= pad_size;
> +            }
>             break;
>         case DEPTH_TO_SPACE:
>             depth_to_space_params = (DepthToSpaceParams *)network->layers[layer].params;
> @@ -75,6 +81,10 @@ static DNNReturnType set_input_output_native(void *model, DNNData *input, DNNDat
>         if (network->layers[layer].output){
>             av_freep(&network->layers[layer].output);
>         }
> +
> +        if(cur_height <= 0 || cur_width <= 0)
> +            return DNN_ERROR; 
> +        
>         network->layers[layer].output = av_malloc(cur_height * cur_width * cur_channels * sizeof(float));
>         if (!network->layers[layer].output){
>             return DNN_ERROR;
> @@ -157,13 +167,14 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename)
>                 ff_dnn_free_model_native(&model);
>                 return NULL;
>             }
> +            conv_params->padding_method = (int32_t)avio_rl32(model_file_context);
>             conv_params->activation = (int32_t)avio_rl32(model_file_context);
>             conv_params->input_num = (int32_t)avio_rl32(model_file_context);
>             conv_params->output_num = (int32_t)avio_rl32(model_file_context);
>             conv_params->kernel_size = (int32_t)avio_rl32(model_file_context);
>             kernel_size = conv_params->input_num * conv_params->output_num *
>                           conv_params->kernel_size * conv_params->kernel_size;
> -            dnn_size += 16 + (kernel_size + conv_params->output_num << 2);
> +            dnn_size += 20 + (kernel_size + conv_params->output_num << 2);
>             if (dnn_size > file_size || conv_params->input_num <= 0 ||
>                 conv_params->output_num <= 0 || conv_params->kernel_size <= 0){
>                 avio_closep(&model_file_context);
> @@ -221,23 +232,35 @@ DNNModel *ff_dnn_load_model_native(const char *model_filename)
> 
> static void convolve(const float *input, float *output, const ConvolutionalParams *conv_params, int width, int height)
> {
> -    int y, x, n_filter, ch, kernel_y, kernel_x;
>     int radius = conv_params->kernel_size >> 1;
>     int src_linesize = width * conv_params->input_num;
>     int filter_linesize = conv_params->kernel_size * conv_params->input_num;
>     int filter_size = conv_params->kernel_size * filter_linesize;
> +    int pad_size = (conv_params->padding_method == VALID) ? (conv_params->kernel_size - 1) / 2 : 0;
> 
> -    for (y = 0; y < height; ++y){
> -        for (x = 0; x < width; ++x){
> -            for (n_filter = 0; n_filter < conv_params->output_num; ++n_filter){
> +    for (int y = pad_size; y < height - pad_size; ++y){
> +        for (int x = pad_size; x < width - pad_size; ++x){
> +            for (int n_filter = 0; n_filter < conv_params->output_num; ++n_filter){
>                 output[n_filter] = conv_params->biases[n_filter];
> -                for (ch = 0; ch < conv_params->input_num; ++ch){
> -                    for (kernel_y = 0; kernel_y < conv_params->kernel_size; ++kernel_y){
> -                        for (kernel_x = 0; kernel_x < conv_params->kernel_size; ++kernel_x){
> -                            output[n_filter] += input[CLAMP_TO_EDGE(y + kernel_y - radius, height) * src_linesize +
> -                                                      CLAMP_TO_EDGE(x + kernel_x - radius, width) * conv_params->input_num + ch] *
> -                                                conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
> -                                                                    kernel_x * conv_params->input_num + ch];
> +
> +                for (int ch = 0; ch < conv_params->input_num; ++ch){
> +                    for (int kernel_y = 0; kernel_y < conv_params->kernel_size; ++kernel_y){
> +                        for (int kernel_x = 0; kernel_x < conv_params->kernel_size; ++kernel_x){
> +                            float input_pel;
> +                            if(conv_params->padding_method == SAME_CLAMP_TO_EDGE){
> +                                int y_pos = CLAMP_TO_EDGE(y + kernel_y - radius, height);
> +                                int x_pos = CLAMP_TO_EDGE(x + kernel_x - radius, width);
> +                                input_pel = input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
> +                            }else{
attention all the code style above section. : } else { and  for (…) {
> +                                int y_pos = y + kernel_y - radius;
> +                                int x_pos = x + kernel_x - radius;
> +                                input_pel = (x_pos < 0 || x_pos >= width || y_pos < 0 || y_pos >= height) ? 0.0 : 
> +                                                   input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
> +                            }
> +
> +
> +                            output[n_filter] += input_pel * conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
> +                                                                                kernel_x * conv_params->input_num + ch];
>                         }
>                     }
>                 }
> @@ -308,6 +331,11 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model)
>             conv_params = (ConvolutionalParams *)network->layers[layer].params;
>             convolve(network->layers[layer - 1].output, network->layers[layer].output, conv_params, cur_width, cur_height);
>             cur_channels = conv_params->output_num;
> +            if(conv_params->padding_method == VALID){
> +                int pad_size = conv_params->kernel_size - 1;
> +                cur_height -= pad_size;
> +                cur_width -= pad_size;
> +            }
>             break;
>         case DEPTH_TO_SPACE:
>             depth_to_space_params = (DepthToSpaceParams *)network->layers[layer].params;
> diff --git a/libavfilter/dnn_backend_native.h b/libavfilter/dnn_backend_native.h
> index 51d4cac955..a609e09754 100644
> --- a/libavfilter/dnn_backend_native.h
> +++ b/libavfilter/dnn_backend_native.h
> @@ -34,6 +34,8 @@ typedef enum {INPUT, CONV, DEPTH_TO_SPACE} DNNLayerType;
> 
> typedef enum {RELU, TANH, SIGMOID} DNNActivationFunc;
> 
> +typedef enum {VALID, SAME, SAME_CLAMP_TO_EDGE} DNNPaddingFunc;
> +
> typedef struct Layer{
>     DNNLayerType type;
>     float *output;
> @@ -43,6 +45,7 @@ typedef struct Layer{
> typedef struct ConvolutionalParams{
>     int32_t input_num, output_num, kernel_size;
>     DNNActivationFunc activation;
> +    DNNPaddingFunc padding_method;
>     float *kernel;
>     float *biases;
> } ConvolutionalParams;
> -- 
> 2.17.1
> 
> 
> 
> 
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
Guo, Yejun May 9, 2019, 1:52 a.m. UTC | #2 
> -----Original Message-----

> From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf Of

> xwmeng@pku.edu.cn

> Sent: Wednesday, May 08, 2019 5:34 PM

> To: ffmpeg-devel@ffmpeg.org

> Subject: [FFmpeg-devel] [PATCH] libavfilter: Add multiple padding methods in

> FFmpeg dnn native mode.

> 

> 

> 

> 

> This patch is for the support of derain filter project in GSoC. It adds supports for

> the following operations:


it is a general patch, not special for derain, so I think we don't need mention derain here,
just explain it in general.

> 

> 

> 

> 

>  (1) Conv padding method: "SAME", "VALID" and "SAME_CLAMP_TO_EDGE"

> 

> 

> 

> 

> These operations are all needed in derain filter. As we discussed before, the

> "SAME_CLAMP_TO_EDGE" method is the same as dnn native padding method

> in the current implementation. And the sr model generation code should be

> So I sent a PR

> (https://github.com/HighVoltageRocknRoll/sr/pull/4)to the original sr

> repo(https://github.com/HighVoltageRocknRoll/sr).

> 

> 

> 

> From c0724bb304a6f4c3ca935cccda5b810e5c4eceb1 Mon Sep 17 00:00:00

> 2001

> From: Xuewei Meng <xwmeng@pku.edu.cn>

> Date: Wed, 8 May 2019 17:32:30 +0800

> Subject: [PATCH] Add multiple padding method in dnn native

> 

> Signed-off-by: Xuewei Meng <xwmeng@pku.edu.cn>

> ---

>  libavfilter/dnn_backend_native.c | 52 ++++++++++++++++++++++++--------

>  libavfilter/dnn_backend_native.h |  3 ++

>  2 files changed, 43 insertions(+), 12 deletions(-)

> 

> diff --git a/libavfilter/dnn_backend_native.c b/libavfilter/dnn_backend_native.c

> index 70d857f5f2..b7c0508d91 100644

> --- a/libavfilter/dnn_backend_native.c

> +++ b/libavfilter/dnn_backend_native.c

> @@ -59,6 +59,12 @@ static DNNReturnType set_input_output_native(void

> *model, DNNData *input, DNNDat

>                  return DNN_ERROR;

>              }

>              cur_channels = conv_params->output_num;

> +

> +            if(conv_params->padding_method == VALID){

> +                int pad_size = conv_params->kernel_size - 1;

> +                cur_height -= pad_size;

> +                cur_width -= pad_size;

> +            }

>              break;

>          case DEPTH_TO_SPACE:

>              depth_to_space_params = (DepthToSpaceParams

> *)network->layers[layer].params;

> @@ -75,6 +81,10 @@ static DNNReturnType set_input_output_native(void

> *model, DNNData *input, DNNDat

>          if (network->layers[layer].output){

>              av_freep(&network->layers[layer].output);

>          }

> +

> +        if(cur_height <= 0 || cur_width <= 0)

> +            return DNN_ERROR;

> +

>          network->layers[layer].output = av_malloc(cur_height * cur_width *

> cur_channels * sizeof(float));

>          if (!network->layers[layer].output){

>              return DNN_ERROR;

> @@ -157,13 +167,14 @@ DNNModel *ff_dnn_load_model_native(const char

> *model_filename)

>                  ff_dnn_free_model_native(&model);

>                  return NULL;

>              }

> +            conv_params->padding_method =

> (int32_t)avio_rl32(model_file_context);

>              conv_params->activation =

> (int32_t)avio_rl32(model_file_context);

>              conv_params->input_num =

> (int32_t)avio_rl32(model_file_context);

>              conv_params->output_num =

> (int32_t)avio_rl32(model_file_context);

>              conv_params->kernel_size =

> (int32_t)avio_rl32(model_file_context);

>              kernel_size = conv_params->input_num *

> conv_params->output_num *

>                            conv_params->kernel_size *

> conv_params->kernel_size;

> -            dnn_size += 16 + (kernel_size + conv_params->output_num <<

> 2);

> +            dnn_size += 20 + (kernel_size + conv_params->output_num <<

> 2);

>              if (dnn_size > file_size || conv_params->input_num <= 0 ||

>                  conv_params->output_num <= 0 ||

> conv_params->kernel_size <= 0){

>                  avio_closep(&model_file_context);

> @@ -221,23 +232,35 @@ DNNModel *ff_dnn_load_model_native(const char

> *model_filename)

> 

>  static void convolve(const float *input, float *output, const

> ConvolutionalParams *conv_params, int width, int height)

>  {

> -    int y, x, n_filter, ch, kernel_y, kernel_x;

>      int radius = conv_params->kernel_size >> 1;

>      int src_linesize = width * conv_params->input_num;

>      int filter_linesize = conv_params->kernel_size *

> conv_params->input_num;

>      int filter_size = conv_params->kernel_size * filter_linesize;

> +    int pad_size = (conv_params->padding_method == VALID) ?

> (conv_params->kernel_size - 1) / 2 : 0;

> 

> -    for (y = 0; y < height; ++y){

> -        for (x = 0; x < width; ++x){

> -            for (n_filter = 0; n_filter < conv_params->output_num;

> ++n_filter){

> +    for (int y = pad_size; y < height - pad_size; ++y){

> +        for (int x = pad_size; x < width - pad_size; ++x){

> +            for (int n_filter = 0; n_filter < conv_params->output_num;

> ++n_filter){

>                  output[n_filter] = conv_params->biases[n_filter];

> -                for (ch = 0; ch < conv_params->input_num; ++ch){

> -                    for (kernel_y = 0; kernel_y <

> conv_params->kernel_size; ++kernel_y){

> -                        for (kernel_x = 0; kernel_x <

> conv_params->kernel_size; ++kernel_x){

> -                            output[n_filter] +=

> input[CLAMP_TO_EDGE(y + kernel_y - radius, height) * src_linesize +

> -

> CLAMP_TO_EDGE(x + kernel_x - radius, width) * conv_params->input_num + ch]

> *

> -

> conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +

> -

> kernel_x * conv_params->input_num + ch];

> +

> +                for (int ch = 0; ch < conv_params->input_num; ++ch){

> +                    for (int kernel_y = 0; kernel_y <

> conv_params->kernel_size; ++kernel_y){

> +                        for (int kernel_x = 0; kernel_x <

> conv_params->kernel_size; ++kernel_x){

> +                            float input_pel;

> +                            if(conv_params->padding_method ==

> SAME_CLAMP_TO_EDGE){

> +                                int y_pos = CLAMP_TO_EDGE(y +

> kernel_y - radius, height);

> +                                int x_pos = CLAMP_TO_EDGE(x +

> kernel_x - radius, width);

> +                                input_pel = input[y_pos * src_linesize

> + x_pos * conv_params->input_num + ch];

> +                            }else{

> +                                int y_pos = y + kernel_y - radius;

> +                                int x_pos = x + kernel_x - radius;

> +                                input_pel = (x_pos < 0 || x_pos >=

> width || y_pos < 0 || y_pos >= height) ? 0.0 :

> +                                                   input[y_pos *

> src_linesize + x_pos * conv_params->input_num + ch];

> +                            }

> +

> +

> +                            output[n_filter] += input_pel *

> conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +

> +

> kernel_x * conv_params->input_num + ch];

>                          }

>                      }

>                  }

> @@ -308,6 +331,11 @@ DNNReturnType ff_dnn_execute_model_native(const

> DNNModel *model)

>              conv_params = (ConvolutionalParams

> *)network->layers[layer].params;

>              convolve(network->layers[layer - 1].output,

> network->layers[layer].output, conv_params, cur_width, cur_height);

>              cur_channels = conv_params->output_num;

> +            if(conv_params->padding_method == VALID){

> +                int pad_size = conv_params->kernel_size - 1;

> +                cur_height -= pad_size;

> +                cur_width -= pad_size;

> +            }

>              break;

>          case DEPTH_TO_SPACE:

>              depth_to_space_params = (DepthToSpaceParams

> *)network->layers[layer].params;

> diff --git a/libavfilter/dnn_backend_native.h b/libavfilter/dnn_backend_native.h

> index 51d4cac955..a609e09754 100644

> --- a/libavfilter/dnn_backend_native.h

> +++ b/libavfilter/dnn_backend_native.h

> @@ -34,6 +34,8 @@ typedef enum {INPUT, CONV, DEPTH_TO_SPACE}

> DNNLayerType;

> 

>  typedef enum {RELU, TANH, SIGMOID} DNNActivationFunc;

> 

> +typedef enum {VALID, SAME, SAME_CLAMP_TO_EDGE} DNNPaddingFunc;


DNNPaddingFunc might have conflict with the pad layer, how about rename to DNNConvPaddingParam

> +

>  typedef struct Layer{

>      DNNLayerType type;

>      float *output;

> @@ -43,6 +45,7 @@ typedef struct Layer{

>  typedef struct ConvolutionalParams{

>      int32_t input_num, output_num, kernel_size;

>      DNNActivationFunc activation;

> +    DNNPaddingFunc padding_method;

>      float *kernel;

>      float *biases;

>  } ConvolutionalParams;

> --

> 2.17.1

> 

> 

> 

> 

> _______________________________________________

> ffmpeg-devel mailing list

> ffmpeg-devel@ffmpeg.org

> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

> 

> To unsubscribe, visit link above, or email

> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
xwmeng@pku.edu.cn May 9, 2019, 1 p.m. UTC | #3 
> -----原始邮件-----
> 发件人: "Guo, Yejun" <yejun.guo@intel.com>
> 发送时间: 2019-05-09 09:52:46 (星期四)
> 收件人: "FFmpeg development discussions and patches" <ffmpeg-devel@ffmpeg.org>
> 抄送: 
> 主题: Re: [FFmpeg-devel] [PATCH] libavfilter: Add multiple padding methods in FFmpeg dnn native mode.
> 
> 
> 
> > -----Original Message-----
> > From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf Of
> > xwmeng@pku.edu.cn
> > Sent: Wednesday, May 08, 2019 5:34 PM
> > To: ffmpeg-devel@ffmpeg.org
> > Subject: [FFmpeg-devel] [PATCH] libavfilter: Add multiple padding methods in
> > FFmpeg dnn native mode.
> > 
> > 
> > 
> > 
> > This patch is for the support of derain filter project in GSoC. It adds supports for
> > the following operations:
> 
> it is a general patch, not special for derain, so I think we don't need mention derain here,
> just explain it in general.
> 
> > 
> > 
> > 
> > 
> >  (1) Conv padding method: "SAME", "VALID" and "SAME_CLAMP_TO_EDGE"
> > 
> > 
> > 
> > 
> > These operations are all needed in derain filter. As we discussed before, the
> > "SAME_CLAMP_TO_EDGE" method is the same as dnn native padding method
> > in the current implementation. And the sr model generation code should be
> > So I sent a PR
> > (https://github.com/HighVoltageRocknRoll/sr/pull/4)to the original sr
> > repo(https://github.com/HighVoltageRocknRoll/sr).
> > 
> > 
> > 
> > From c0724bb304a6f4c3ca935cccda5b810e5c4eceb1 Mon Sep 17 00:00:00
> > 2001
> > From: Xuewei Meng <xwmeng@pku.edu.cn>
> > Date: Wed, 8 May 2019 17:32:30 +0800
> > Subject: [PATCH] Add multiple padding method in dnn native
> > 
> > Signed-off-by: Xuewei Meng <xwmeng@pku.edu.cn>
> > ---
> >  libavfilter/dnn_backend_native.c | 52 ++++++++++++++++++++++++--------
> >  libavfilter/dnn_backend_native.h |  3 ++
> >  2 files changed, 43 insertions(+), 12 deletions(-)
> > 
> > diff --git a/libavfilter/dnn_backend_native.c b/libavfilter/dnn_backend_native.c
> > index 70d857f5f2..b7c0508d91 100644
> > --- a/libavfilter/dnn_backend_native.c
> > +++ b/libavfilter/dnn_backend_native.c
> > @@ -59,6 +59,12 @@ static DNNReturnType set_input_output_native(void
> > *model, DNNData *input, DNNDat
> >                  return DNN_ERROR;
> >              }
> >              cur_channels = conv_params->output_num;
> > +
> > +            if(conv_params->padding_method == VALID){
> > +                int pad_size = conv_params->kernel_size - 1;
> > +                cur_height -= pad_size;
> > +                cur_width -= pad_size;
> > +            }
> >              break;
> >          case DEPTH_TO_SPACE:
> >              depth_to_space_params = (DepthToSpaceParams
> > *)network->layers[layer].params;
> > @@ -75,6 +81,10 @@ static DNNReturnType set_input_output_native(void
> > *model, DNNData *input, DNNDat
> >          if (network->layers[layer].output){
> >              av_freep(&network->layers[layer].output);
> >          }
> > +
> > +        if(cur_height <= 0 || cur_width <= 0)
> > +            return DNN_ERROR;
> > +
> >          network->layers[layer].output = av_malloc(cur_height * cur_width *
> > cur_channels * sizeof(float));
> >          if (!network->layers[layer].output){
> >              return DNN_ERROR;
> > @@ -157,13 +167,14 @@ DNNModel *ff_dnn_load_model_native(const char
> > *model_filename)
> >                  ff_dnn_free_model_native(&model);
> >                  return NULL;
> >              }
> > +            conv_params->padding_method =
> > (int32_t)avio_rl32(model_file_context);
> >              conv_params->activation =
> > (int32_t)avio_rl32(model_file_context);
> >              conv_params->input_num =
> > (int32_t)avio_rl32(model_file_context);
> >              conv_params->output_num =
> > (int32_t)avio_rl32(model_file_context);
> >              conv_params->kernel_size =
> > (int32_t)avio_rl32(model_file_context);
> >              kernel_size = conv_params->input_num *
> > conv_params->output_num *
> >                            conv_params->kernel_size *
> > conv_params->kernel_size;
> > -            dnn_size += 16 + (kernel_size + conv_params->output_num <<
> > 2);
> > +            dnn_size += 20 + (kernel_size + conv_params->output_num <<
> > 2);
> >              if (dnn_size > file_size || conv_params->input_num <= 0 ||
> >                  conv_params->output_num <= 0 ||
> > conv_params->kernel_size <= 0){
> >                  avio_closep(&model_file_context);
> > @@ -221,23 +232,35 @@ DNNModel *ff_dnn_load_model_native(const char
> > *model_filename)
> > 
> >  static void convolve(const float *input, float *output, const
> > ConvolutionalParams *conv_params, int width, int height)
> >  {
> > -    int y, x, n_filter, ch, kernel_y, kernel_x;
> >      int radius = conv_params->kernel_size >> 1;
> >      int src_linesize = width * conv_params->input_num;
> >      int filter_linesize = conv_params->kernel_size *
> > conv_params->input_num;
> >      int filter_size = conv_params->kernel_size * filter_linesize;
> > +    int pad_size = (conv_params->padding_method == VALID) ?
> > (conv_params->kernel_size - 1) / 2 : 0;
> > 
> > -    for (y = 0; y < height; ++y){
> > -        for (x = 0; x < width; ++x){
> > -            for (n_filter = 0; n_filter < conv_params->output_num;
> > ++n_filter){
> > +    for (int y = pad_size; y < height - pad_size; ++y){
> > +        for (int x = pad_size; x < width - pad_size; ++x){
> > +            for (int n_filter = 0; n_filter < conv_params->output_num;
> > ++n_filter){
> >                  output[n_filter] = conv_params->biases[n_filter];
> > -                for (ch = 0; ch < conv_params->input_num; ++ch){
> > -                    for (kernel_y = 0; kernel_y <
> > conv_params->kernel_size; ++kernel_y){
> > -                        for (kernel_x = 0; kernel_x <
> > conv_params->kernel_size; ++kernel_x){
> > -                            output[n_filter] +=
> > input[CLAMP_TO_EDGE(y + kernel_y - radius, height) * src_linesize +
> > -
> > CLAMP_TO_EDGE(x + kernel_x - radius, width) * conv_params->input_num + ch]
> > *
> > -
> > conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
> > -
> > kernel_x * conv_params->input_num + ch];
> > +
> > +                for (int ch = 0; ch < conv_params->input_num; ++ch){
> > +                    for (int kernel_y = 0; kernel_y <
> > conv_params->kernel_size; ++kernel_y){
> > +                        for (int kernel_x = 0; kernel_x <
> > conv_params->kernel_size; ++kernel_x){
> > +                            float input_pel;
> > +                            if(conv_params->padding_method ==
> > SAME_CLAMP_TO_EDGE){
> > +                                int y_pos = CLAMP_TO_EDGE(y +
> > kernel_y - radius, height);
> > +                                int x_pos = CLAMP_TO_EDGE(x +
> > kernel_x - radius, width);
> > +                                input_pel = input[y_pos * src_linesize
> > + x_pos * conv_params->input_num + ch];
> > +                            }else{
> > +                                int y_pos = y + kernel_y - radius;
> > +                                int x_pos = x + kernel_x - radius;
> > +                                input_pel = (x_pos < 0 || x_pos >=
> > width || y_pos < 0 || y_pos >= height) ? 0.0 :
> > +                                                   input[y_pos *
> > src_linesize + x_pos * conv_params->input_num + ch];
> > +                            }
> > +
> > +
> > +                            output[n_filter] += input_pel *
> > conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
> > +
> > kernel_x * conv_params->input_num + ch];
> >                          }
> >                      }
> >                  }
> > @@ -308,6 +331,11 @@ DNNReturnType ff_dnn_execute_model_native(const
> > DNNModel *model)
> >              conv_params = (ConvolutionalParams
> > *)network->layers[layer].params;
> >              convolve(network->layers[layer - 1].output,
> > network->layers[layer].output, conv_params, cur_width, cur_height);
> >              cur_channels = conv_params->output_num;
> > +            if(conv_params->padding_method == VALID){
> > +                int pad_size = conv_params->kernel_size - 1;
> > +                cur_height -= pad_size;
> > +                cur_width -= pad_size;
> > +            }
> >              break;
> >          case DEPTH_TO_SPACE:
> >              depth_to_space_params = (DepthToSpaceParams
> > *)network->layers[layer].params;
> > diff --git a/libavfilter/dnn_backend_native.h b/libavfilter/dnn_backend_native.h
> > index 51d4cac955..a609e09754 100644
> > --- a/libavfilter/dnn_backend_native.h
> > +++ b/libavfilter/dnn_backend_native.h
> > @@ -34,6 +34,8 @@ typedef enum {INPUT, CONV, DEPTH_TO_SPACE}
> > DNNLayerType;
> > 
> >  typedef enum {RELU, TANH, SIGMOID} DNNActivationFunc;
> > 
> > +typedef enum {VALID, SAME, SAME_CLAMP_TO_EDGE} DNNPaddingFunc;
> 
> DNNPaddingFunc might have conflict with the pad layer, how about rename to DNNConvPaddingParam

Yeah, DNNConvPaddingParam is better

> 
> > +
> >  typedef struct Layer{
> >      DNNLayerType type;
> >      float *output;
> > @@ -43,6 +45,7 @@ typedef struct Layer{
> >  typedef struct ConvolutionalParams{
> >      int32_t input_num, output_num, kernel_size;
> >      DNNActivationFunc activation;
> > +    DNNPaddingFunc padding_method;
> >      float *kernel;
> >      float *biases;
> >  } ConvolutionalParams;
> > --
> > 2.17.1
> > 
> > 
> > 
> > 
> > _______________________________________________
> > ffmpeg-devel mailing list
> > ffmpeg-devel@ffmpeg.org
> > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> > 
> > To unsubscribe, visit link above, or email
> > ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
diff mbox

Patch

diff --git a/libavfilter/dnn_backend_native.c b/libavfilter/dnn_backend_native.c
index 70d857f5f2..b7c0508d91 100644
--- a/libavfilter/dnn_backend_native.c
+++ b/libavfilter/dnn_backend_native.c
@@ -59,6 +59,12 @@  static DNNReturnType set_input_output_native(void *model, DNNData *input, DNNDat
                 return DNN_ERROR;
             }
             cur_channels = conv_params->output_num;
+
+            if(conv_params->padding_method == VALID){
+                int pad_size = conv_params->kernel_size - 1;
+                cur_height -= pad_size;
+                cur_width -= pad_size;
+            }
             break;
         case DEPTH_TO_SPACE:
             depth_to_space_params = (DepthToSpaceParams *)network->layers[layer].params;
@@ -75,6 +81,10 @@  static DNNReturnType set_input_output_native(void *model, DNNData *input, DNNDat
         if (network->layers[layer].output){
             av_freep(&network->layers[layer].output);
         }
+
+        if(cur_height <= 0 || cur_width <= 0)
+            return DNN_ERROR; 
+        
         network->layers[layer].output = av_malloc(cur_height * cur_width * cur_channels * sizeof(float));
         if (!network->layers[layer].output){
             return DNN_ERROR;
@@ -157,13 +167,14 @@  DNNModel *ff_dnn_load_model_native(const char *model_filename)
                 ff_dnn_free_model_native(&model);
                 return NULL;
             }
+            conv_params->padding_method = (int32_t)avio_rl32(model_file_context);
             conv_params->activation = (int32_t)avio_rl32(model_file_context);
             conv_params->input_num = (int32_t)avio_rl32(model_file_context);
             conv_params->output_num = (int32_t)avio_rl32(model_file_context);
             conv_params->kernel_size = (int32_t)avio_rl32(model_file_context);
             kernel_size = conv_params->input_num * conv_params->output_num *
                           conv_params->kernel_size * conv_params->kernel_size;
-            dnn_size += 16 + (kernel_size + conv_params->output_num << 2);
+            dnn_size += 20 + (kernel_size + conv_params->output_num << 2);
             if (dnn_size > file_size || conv_params->input_num <= 0 ||
                 conv_params->output_num <= 0 || conv_params->kernel_size <= 0){
                 avio_closep(&model_file_context);
@@ -221,23 +232,35 @@  DNNModel *ff_dnn_load_model_native(const char *model_filename)
 
 static void convolve(const float *input, float *output, const ConvolutionalParams *conv_params, int width, int height)
 {
-    int y, x, n_filter, ch, kernel_y, kernel_x;
     int radius = conv_params->kernel_size >> 1;
     int src_linesize = width * conv_params->input_num;
     int filter_linesize = conv_params->kernel_size * conv_params->input_num;
     int filter_size = conv_params->kernel_size * filter_linesize;
+    int pad_size = (conv_params->padding_method == VALID) ? (conv_params->kernel_size - 1) / 2 : 0;
 
-    for (y = 0; y < height; ++y){
-        for (x = 0; x < width; ++x){
-            for (n_filter = 0; n_filter < conv_params->output_num; ++n_filter){
+    for (int y = pad_size; y < height - pad_size; ++y){
+        for (int x = pad_size; x < width - pad_size; ++x){
+            for (int n_filter = 0; n_filter < conv_params->output_num; ++n_filter){
                 output[n_filter] = conv_params->biases[n_filter];
-                for (ch = 0; ch < conv_params->input_num; ++ch){
-                    for (kernel_y = 0; kernel_y < conv_params->kernel_size; ++kernel_y){
-                        for (kernel_x = 0; kernel_x < conv_params->kernel_size; ++kernel_x){
-                            output[n_filter] += input[CLAMP_TO_EDGE(y + kernel_y - radius, height) * src_linesize +
-                                                      CLAMP_TO_EDGE(x + kernel_x - radius, width) * conv_params->input_num + ch] *
-                                                conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
-                                                                    kernel_x * conv_params->input_num + ch];
+
+                for (int ch = 0; ch < conv_params->input_num; ++ch){
+                    for (int kernel_y = 0; kernel_y < conv_params->kernel_size; ++kernel_y){
+                        for (int kernel_x = 0; kernel_x < conv_params->kernel_size; ++kernel_x){
+                            float input_pel;
+                            if(conv_params->padding_method == SAME_CLAMP_TO_EDGE){
+                                int y_pos = CLAMP_TO_EDGE(y + kernel_y - radius, height);
+                                int x_pos = CLAMP_TO_EDGE(x + kernel_x - radius, width);
+                                input_pel = input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
+                            }else{
+                                int y_pos = y + kernel_y - radius;
+                                int x_pos = x + kernel_x - radius;
+                                input_pel = (x_pos < 0 || x_pos >= width || y_pos < 0 || y_pos >= height) ? 0.0 : 
+                                                   input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
+                            }
+
+
+                            output[n_filter] += input_pel * conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
+                                                                                kernel_x * conv_params->input_num + ch];
                         }
                     }
                 }
@@ -308,6 +331,11 @@  DNNReturnType ff_dnn_execute_model_native(const DNNModel *model)
             conv_params = (ConvolutionalParams *)network->layers[layer].params;
             convolve(network->layers[layer - 1].output, network->layers[layer].output, conv_params, cur_width, cur_height);
             cur_channels = conv_params->output_num;
+            if(conv_params->padding_method == VALID){
+                int pad_size = conv_params->kernel_size - 1;
+                cur_height -= pad_size;
+                cur_width -= pad_size;
+            }
             break;
         case DEPTH_TO_SPACE:
             depth_to_space_params = (DepthToSpaceParams *)network->layers[layer].params;
diff --git a/libavfilter/dnn_backend_native.h b/libavfilter/dnn_backend_native.h
index 51d4cac955..a609e09754 100644
--- a/libavfilter/dnn_backend_native.h
+++ b/libavfilter/dnn_backend_native.h
@@ -34,6 +34,8 @@  typedef enum {INPUT, CONV, DEPTH_TO_SPACE} DNNLayerType;
 
 typedef enum {RELU, TANH, SIGMOID} DNNActivationFunc;
 
+typedef enum {VALID, SAME, SAME_CLAMP_TO_EDGE} DNNPaddingFunc;
+
 typedef struct Layer{
     DNNLayerType type;
     float *output;
@@ -43,6 +45,7 @@  typedef struct Layer{
 typedef struct ConvolutionalParams{
     int32_t input_num, output_num, kernel_size;
     DNNActivationFunc activation;
+    DNNPaddingFunc padding_method;
     float *kernel;
     float *biases;
 } ConvolutionalParams;