CGruLayer Class

• CGruLayer Class
  • Settings
  • Trainable parameters
    • Gate weights matrix
    • Gate free terms
    • Output weights matrix
    • Output free terms
  • Inputs
  • Outputs

This class implements a GRU layer that works with a set of vector sequences.

The result is a vector sequence of the same length, with each vector the length of GetHiddenSize().

Settings

Trainable parameters

Gate weights matrix

CPtr<CDnnBlob> GetGateWeightsData() const;

The gate weights are put into a two-dimensional matrix represented by a blob of the dimensions:

• BatchLength * BatchWidth * ListSize is equal to 2 * GetHiddenSize()
• Height * Width * Depth * Channels is equal to the same dimension of the input plus GetHiddenSize().

Along the BatchLength * BatchWidth * ListSize axis the gate weights are sorted in the following order:

G_Update = 0, // Update gate
G_Reset,      // Reset gate

Along the Height * Width * Depth * Channels axis the weights are put into following order:

• from the start to the input Height * Width * Depth * Channels the weights for the input vectors
• the rest GetHiddenSize() coordinates correspond to the weights for the previous step result.

Gate free terms

CPtr<CDnnBlob> GetGateFreeTermData() const;

The free terms for the gates are represented by a blob of 2 * GetHiddenSize() total size.

Output weights matrix

CPtr<CDnnBlob> GetMainWeightsData() const;

The output weights are put into a two-dimensional matrix represented by a blob of the dimensions:

• BatchLength * BatchWidth * ListSize is equal to GetHiddenSize()
• Height * Width * Depth * Channels is equal to the same dimension fo the input plus GetHiddenSize()

Output free terms

The free terms for the output are represented by a blob of GetHiddenSize() total size.

Inputs

The layer has 1 to 2 inputs:

1. The set of vector sequences.
2. [Optional] The initial previous step result that should be used on the first step. If you do not connect this input all zeros will be used.

Output

The single output returns a blob of the dimensions:

• BatchLength, BatchWidth, and ListSize equal the corresponding input dimensions
• Height, Width, and Depth are equal to 1
• Channels equals GetHiddenSize()