Skip to content

Latest commit

 

History

History
92 lines (60 loc) · 3.01 KB

CrfLayer.md

File metadata and controls

92 lines (60 loc) · 3.01 KB
Raw

CCrfLayer Class

This class implements a layer that trains and calculates transition probabilities in a conditional random field (CRF).

Settings

The number of classes in the CRF

void SetNumberOfClasses( int numberOfClasses );

Empty class

void SetPaddingClass( int paddingClass );

Sets the number of the empty class used to fill the sequence end.

Variational dropout

void SetDropoutRate( float newDropoutRate );

O_BestPrevClass output computation

void SetBestPrevClassEnabled( bool enabled );

Enables computation of the first (O_BestPrevClass) output during training. This setting allows you to connect the CBestSequenceLayer and get predictions for every batch during training.

By default disabled during training. During inference the first output is always computed, ignoring this setting.

Trainable parameters

The layer trains the transition probabilities, but they are not available to the user.

Inputs

The layer may have one or two inputs.

The first input accepts a blob with the object sequences, of the dimensions:

  • BatchLength is the sequence length
  • BatchWidth is the number of sequences in the set
  • ListSize is equal to 1
  • Height * Width * Depth * Channels is the object size

[Optional] The second input accepts a blob with int data that contains the correct class sequences. This input is required for training. The blob dimensions are:

  • BatchLength is equal to the first input BatchLength
  • BatchWidth is equal to the first input BatchWidth
  • the other dimensions are equal to 1

Outputs

The layer may have two to three outputs.

The first output is a blob with int data that contains optimal class sequences. The blob dimensions are:

  • BatchLength is equal to the inputs' BatchLength
  • BatchWidth is equal to the inputs' BatchWidth
  • Channels is equal to GetNumberOfClasses()
  • the other dimensions are equal to 1

The second output is a blob with float data that contains non-normalized logarithm of optimal class sequences probabilities. It has the same dimensions as the first output.

[Optional] The third output will be there only if the layer has two inputs. It contains a blob with non-normalized logarithm of the correct class sequences probabilities. The blob dimensions are:

  • BatchLength is equal to the inputs' BatchLength
  • BatchWidth is equal to the inputs' BatchWidth
  • the other dimensions are equal to 1

The CBestSequenceLayer class extracts the best sequence using the first two outputs.