espnet2.asr.transducer.rnnt_multi_blank.utils.cpu_utils.cpu_rnnt.CpuRNNT_index

source

class espnet2.asr.transducer.rnnt_multi_blank.utils.cpu_utils.cpu_rnnt.CpuRNNT_index(U: int, maxU: int, minibatch: int, alphabet_size: int, batch_first: bool)

Bases: object

Class to compute the CPU-based index for RNNT (Recurrent Neural Network Transducer) operations, mimicking pointer indexing as done in CUDA kernels.

This class facilitates the mapping of target samples to flattened tensors, allowing for efficient access to indices based on the current target sample, the maximum padded target sample length, and the vocabulary size.

U

Length of the current target sample (without padding).

• Type: int

maxU

Maximum length of the padded target samples.

• Type: int

minibatch

Minibatch index.

• Type: int

alphabet_size

Size of the vocabulary including RNNT blank (V + 1).

• Type: int

batch_first

Flag determining if batch index is first or third.

• Type: bool

• Parameters:

  • U (int) – Length of the current target sample (without padding).
  • maxU (int) – Max length of the padded target samples.
  • minibatch (int) – Minibatch index.
  • alphabet_size (int) – Size of the vocabulary including RNNT blank (V + 1).
  • batch_first (bool) – Flag to indicate if batch dimension is first (True) or third (False).

__call__(t

int, u: int, v: Optional[int] = None) -> int: Returns the computed index for the given parameters.

Examples

>>> index = CpuRNNT_index(U=10, maxU=15, minibatch=2, alphabet_size=30,
...                        batch_first=True)
>>> idx = index(t=3, u=5)
>>> print(idx)
53  # Example output based on given parameters

NOTE

The index computation is designed to emulate the memory access patterns of CUDA kernels on the CPU, which is essential for performance in RNNT training and inference.

A placeholder Index computation class that emits the resolved index in a

flattened tensor, mimicing pointer indexing in CUDA kernels on the CPU.

• Parameters:
  • U – Length of the current target sample (without padding).
  • maxU – Max Length of the padded target samples.
  • minibatch – Minibatch index
  • alphabet_size – Size of the vocabulary including RNNT blank - V+1.
  • batch_first – Bool flag determining if batch index is first or third.