CTCGreedyDecoderSeqLen#

Versioned name: CTCGreedyDecoderSeqLen-6

Category: Sequence processing

Short description: CTCGreedyDecoderSeqLen performs greedy decoding of the logits provided as the first input. The sequence lengths are provided as the second input.

Detailed description:

This operation is similar to the TensorFlow CTCGreedyDecoder.

The operation CTCGreedyDecoderSeqLen implements best path decoding. Decoding is done in two steps:

  1. Concatenate the most probable classes per time-step which yields the best path.

  2. Remove duplicate consecutive elements if the attribute merge_repeated is true and then remove all blank elements.

Sequences in the batch can have different length. The lengths of sequences are coded in the second input integer tensor sequence_length.

The main difference between CTCGreedyDecoder and CTCGreedyDecoderSeqLen is in the second input. CTCGreedyDecoder uses 2D input floating-point tensor with sequence masks for each sequence in the batch while CTCGreedyDecoderSeqLen uses 1D integer tensor with sequence lengths.

Attributes

  • merge_repeated

    • Description: merge_repeated is a flag for merging repeated labels during the CTC calculation. If the value is false the sequence ABB*B*B (where ‘*’ is the blank class) will look like ABBBB. But if the value is true, the sequence will be ABBB.

    • Range of values: true or false

    • Type: boolean

    • Default value: true

    • Required: no

  • classes_index_type

    • Description: the type of output tensor with classes indices

    • Range of values: “i64” or “i32”

    • Type: string

    • Default value: “i32”

    • Required: no

  • sequence_length_type

    • Description: the type of output tensor with sequence length

    • Range of values: “i64” or “i32”

    • Type: string

    • Default value: “i32”

    • Required: no

Inputs

  • 1: data - input tensor of type T_F of shape [N, T, C] with a batch of sequences. Where T is the maximum sequence length, N is the batch size and C is the number of classes. Required.

  • 2: sequence_length - input tensor of type T_I of shape [N] with sequence lengths. The values of sequence length must be less or equal to T. Required.

  • 3: blank_index - scalar or 1D tensor with 1 element of type T_I. Specifies the class index to use for the blank class. Regardless of the value of merge_repeated attribute, if the output index for a given batch and time step corresponds to the blank_index, no new element is emitted. Default value is C-1. Optional.

Output

  • 1: Output tensor of type T_IND1 shape [N, T] and containing the decoded classes. All elements that do not code sequence classes are filled with -1.

  • 2: Output tensor of type T_IND2 shape [N] and containing length of decoded class sequence for each batch.

Types

  • T_F: any supported floating-point type.

  • T_I: int32 or int64.

  • T_IND1: int32 or int64 and depends on classes_index_type attribute.

  • T_IND2: int32 or int64 and depends on sequence_length_type attribute.

Example

<layer ... type="CTCGreedyDecoderSeqLen" version="opset6">
    <data merge_repeated="true" classes_index_type="i64" sequence_length_type="i64"/>
    <input>
        <port id="0">
            <dim>8</dim>
            <dim>20</dim>
            <dim>128</dim>
        </port>
        <port id="1">
            <dim>8</dim>
        </port>
        <port id="2"/>  <!-- blank_index = 120 -->
    </input>
    <output>
        <port id="0" precision="I64">
            <dim>8</dim>
            <dim>20</dim>
        </port>
        <port id="1" precision="I64">
            <dim>8</dim>
        </port>
    </output>
</layer>