espnet2.asr.transducer.rnnt_multi_blank.utils.rnnt_helper.log_sum_exp

source

espnet2.asr.transducer.rnnt_multi_blank.utils.rnnt_helper.log_sum_exp(a: float, b: float)

Calculate the log of the sum of exponentials of two input values.

This function efficiently computes the logarithm of the sum of the exponentials of two floating-point numbers, a and b, while handling cases of negative infinity as defined in the global_constants module.

This implementation is designed for use in CUDA kernels, thus it uses the @cuda.jit decorator for Just-In-Time compilation. The function is also inlined for performance optimization.

• Parameters:
  • a (float) – The first input value.
  • b (float) – The second input value.
• Returns: The logarithm of the sum of exponentials of a and b.
• Return type: float

Examples

>>> result = log_sum_exp(1.0, 2.0)
>>> print(result)  # Output will be approximately 2.3133

>>> result = log_sum_exp(global_constants.FP32_NEG_INF, 3.0)
>>> print(result)  # Output will be 3.0, as -inf is ignored.

>>> result = log_sum_exp(4.0, global_constants.FP32_NEG_INF)
>>> print(result)  # Output will be 4.0, as -inf is ignored.

NOTE

The function assumes that inputs are valid floating-point numbers and uses constants defined in global_constants to handle edge cases effectively.