Skip to content

Latest commit

 

History

History
82 lines (55 loc) · 2.8 KB

InterpolationLayer.md

File metadata and controls

82 lines (55 loc) · 2.8 KB
Raw

CInterpolationLayer Class

The class implements a layer that changes the sizes of some axes of the blob and fills the new elements with interpolated values.

At this moment only linear interpolation is supported.

Settings

Dimension sizes

void SetRule( TBlobDim dim, const CRule& rule );

Sets the rule to be used when calculating the size of dim blob dimension.

It's recommended to use one of these functions to create the CRule:

  • CInterpolationLayer::CRule::Resize( int newSize ) to set the dim of the output to newSize
  • CInterpolationLayer::CRule::Scale( float scale ) to multiply the dim by scale (relative to the input blob).

By default layer doesn't change any of the dimensions.

Coordinate system

void SetCoords( TInterpolationCoords newCoords );

Sets the way layer calculates the coordinates of the new elements xNew relative to the elements of the input xOld.

Supported values:

  • TInterpolationCoords::HalfPixel - xOld = ( xNew + 0.5 ) / scale - 0.5
  • TInterpolationCoords::PytorchHalfPixel - xOld = ( newSize > 1 ) ? ( xNew + 0.5 ) / scale - 0.5 : 0
  • TInterpolationCoords::AlignCorners - xOld = xNew * ( oldSize - 1 ) / ( newSize - 1 )
  • TInterpolationCoords::Asymmetric - xOld = xNew / scale where:
  • oldSize and newSize are the sizes of the dimension before and after the transformation
  • scale - coefficient used to multiply input dimension (newSize / oldSize in case of fixed output size)

By default TInterpolationCoords::Asymmetric is used.

Coordinate rounding

void SetRound( TInterpolationRound newRound );

Sets how the coordinates xOld are rounded. If used then every element of the output will be one of the elements of the input (no interpolated values).

Supported values:

  • TInterpolationRound::None - don't round, calculate interpolated values
  • TInterpolationRound::RoundPreferFloor - less or equal to the half values are rounded down
  • TInterpolationRound::RoundPreferCeil - less than half values are rounded down
  • TInterpolationRound::Floor - zeroing non-integer part
  • TInterpolationRound::Ceil - if non-integer part is present, it's rounded up.

By default rounding is not used (TInterpolationRound::None).

Trainable parameters

There are no trainable parameters for this layer.

Inputs

The single input accepts a blob of any size.

Outputs

The single output contains a blob where each dim size is calculated based on GetRule( dim ).