[FFmpeg-devel] [PATCH V3 1/2] dnn/native: add native support for avg_pool

[Fu, Ting]

> -----Original Message-----
> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of Guo,
> Yejun
> Sent: Monday, August 3, 2020 10:20 AM
> To: FFmpeg development discussions and patches <ffmpeg-devel at ffmpeg.org>
> Subject: Re: [FFmpeg-devel] [PATCH V3 1/2] dnn/native: add native support for
> avg_pool
> 
> 
> 
> > -----Original Message-----
> > From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of Ting
> > Fu
> > Sent: 2020年7月30日 18:03
> > To: ffmpeg-devel at ffmpeg.org
> > Subject: [FFmpeg-devel] [PATCH V3 1/2] dnn/native: add native support
> > for avg_pool
> >
> > Not support pooling strides in channel dimension now.
> > It can be tested with the model generated with below python script:
> >
> > import tensorflow as tf
> > import numpy as np
> > import imageio
> >
> > in_img = imageio.imread('input_odd.jpg') in_img =
> > in_img.astype(np.float32)/255.0 in_data = in_img[np.newaxis, :]
> >
> > x = tf.placeholder(tf.float32, shape=[1, None, None, 3],
> > name='dnn_in') x_pool = tf.nn.avg_pool(x, ksize=[1,2,2,1],
> > strides=[1,2,2,1], padding='SAME')
> 
> I don't see input/output channel is set in this function parameter, why we need
> them in struct AvgPoolParams?

Hi Yejun,

The in/out_channel in struct AvgPoolParams are for assert() the channel of input image is equal to model params.
If read image with  rgba then the native will not process correctly, so I think this part is necessary.

> 
[...]
> > +    if (avgpool_params->padding_method == SAME) {
> > +        height_end = height;
> > +        width_end = width;
> > +        height_radius = avgpool_params->kernel_size - ((height - 1) %
> > kernel_strides + 1);
> > +        width_radius = avgpool_params->kernel_size - ((width - 1) %
> > kernel_strides + 1);
> > +        height_radius = height_radius < 0 ? 0 : height_radius >> 1;
> > +        width_radius = width_radius < 0 ? 0 : width_radius >> 1;
> > +        output_height = ceil(height / (kernel_strides * 1.0));
> > +        output_width = ceil(width / (kernel_strides * 1.0));
> > +    } else {
> []
> add an assert here, since avg_pool only accepts 'VALID' or 'SAME', while
> padding_method has three enum.
> avassert0(padding_method==VALID)
> 

Sure, will add in patch V4.

> > +        height_end = height - avgpool_params->kernel_size + 1;
[...]
> > diff --git a/libavfilter/dnn/dnn_backend_native_layer_avgpool.h
> > b/libavfilter/dnn/dnn_backend_native_layer_avgpool.h
> > new file mode 100644
> > index 0000000000..0b37a8f64b
> > --- /dev/null
> > +++ b/libavfilter/dnn/dnn_backend_native_layer_avgpool.h
> > @@ -0,0 +1,35 @@
> > +/*
> > + * Copyright (c) 2018 Sergey Lavrushkin
> []
> remove name here

Sorry, will modify in next patch version.

Thank you for your review.
Ting FU
> 
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email ffmpeg-devel-request at ffmpeg.org
> with subject "unsubscribe".