AccumulativeLookupLayer.md

CAccumulativeLookupLayer Class

• CAccumulativeLookupLayer Class
  • Settings
  • Trainable parameters
    • The representation table
  • Inputs
  • Outputs

This class implements a layer that trains fixed-length vector representations for the values of a discrete feature.

See Word2Vec, GloVe, etc.

This layer can work only with one feature; when several values of the feature are passed, the sum of the corresponding vectors is returned.

Settings

// Size of the representation table
struct CLookupDimension {
    int VectorCount; // the number of vectors
    int VectorSize; // the vector length
};

void SetDimension( const CLookupDimension& newDimension );

Sets the size of the vector table.

Trainable parameters

The representation table

CPtr<CDnnBlob> GetEmbeddings() const;

Gets the table with the trained vectors. The blob storing the table has the following dimensions:

• BatchLength is equal to GetDimension().VectorCount;
• BatchWidth is equal to GetDimension().VectorSize.

Inputs

The single input accepts a blob with int data that contains the feature values, of the dimensions:

• BatchLength * BatchWidth * ListSize is the number of different values the feature can take
• Height * Width * Depth * Channels is the number of values in the set

Outputs

The single output contains a blob with the sum of vector representations of the given feature values. The 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 is equal to 1
• Width is equal to 1
• Depth is equal to 1
• Channels is equal to GetDimension().VectorSize.