Skip to content

TransposedConvLayer.md

Latest commit

 

History

History
118 lines (83 loc) · 3.72 KB

TransposedConvLayer.md

File metadata and controls

118 lines (83 loc) · 3.72 KB

CTransposedConvLayer Class

This class implements a layer that performs transposed convolution (sometimes also called deconvolution or up-convolution) on a set of two-dimensional multi-channel images. Padding and dilated convolution are supported.

Settings

Filters size

void SetFilterHeight( int filterHeight );
void SetFilterWidth( int filterWidth );
void SetFilterCount( int filterCount );

Sets the filters' size and number.

Convolution stride

void SetStrideHeight( int strideHeight );
void SetStrideWidth( int strideWidth );

Sets the convolution stride. By default, the stride is 1.

Padding

void SetPaddingHeight( int paddingHeight );
void SetPaddingWidth( int paddingWidth );

Sets the width and height of padding that should be removed from the convolution result. For example, if SetPaddingWidth( 1 );, two columns - one on the right and one on the left - will be cut off of the resulting image. By default these values are set to 0.

Dilated convolution

void SetDilationHeight( int dilationHeight );
void SetDilationWidth( int dilationWidth );

Sets the vertical and horizontal step values for dilated convolution. Dilated convolution applies the filter not to the consecutive pixels of the original image but to pixels with the gaps between.

By default, these values are equal to 1: no dilation, consecutive pixels are used.

Using the free terms

void SetZeroFreeTerm(bool isZeroFreeTerm);

Specifies if the free terms should be used. If you set this value to true, the free terms vector will be set to all zeros and won't be trained. By default, this value is set to false.

Trainable parameters

Filters

CPtr<CDnnBlob> GetFilterData() const;

The filters are represented by a blob of the following dimensions:

  • BatchLength is equal to 1
  • BatchWidth is equal to the inputs' Channels * Depth
  • ListSize is equal to 1
  • Height is equal to GetFilterHeight()
  • Width is equal to GetFilterWidth()
  • Depth is equal to 1
  • Channels is equal to GetFilterCount()

Free terms

CPtr<CDnnBlob> GetFreeTermData() const;

The free terms are represented by a blob of the total size equal to the number of filters used (GetFilterCount()).

Inputs

Each input accepts a blob with several images. The dimensions of all inputs should be the same:

  • BatchLength * BatchWidth * ListSize - the number of images in the set.
  • Height - the images' height.
  • Width - the images' width.
  • Depth * Channels - the number of channels the image format uses.

Outputs

For each input the layer has one output. It contains a blob with the result of convolution. The output blob dimensions are:

  • BatchLength is equal to the input BatchLength.
  • BatchWidth is equal to the input BatchWidth.
  • ListSize is equal to the input ListSize.
  • Height can be calculated from the input Height as StrideHeight * (Height - 1) + (FilterHeight - 1) * DilationHeight + 1 - 2 * PaddingHeight.
  • Width can be calculated from the input Width as StrideWidth * (Width - 1) + (FilterWidth - 1) * DilationWidth + 1 - 2 * PaddingWidth.
  • Depth is equal to 1.
  • Channels is equal to GetFilterCount().