Criterion

Criterion is a Callable object which is the base class for all criterion objects in this toolkit. It serves the role of evaluating the model forward calculation results. Its output can be either a loss function used for training or an evaluation metric used for validation.

This base class has two abstract interface functions: criterion_init() for criterion initialization and __call__() for criterion forward calculation.

1. __call__() must be overridden if you want to make your own Criterion implementation.
2. criterion_init() is not mandatory to be overridden because some criteria can directly be applied to the input data without any initialization such as speechain.criterion.accuracy.Accuracy.

👆Back to the handbook page

Table of Contents

1. Criterion Library
2. API Document

Criterion Library

/speechain
    /criterion          
        /abs.py             # Abstract class of Criterion. Base of all Criterion implementations.
        /accuracy.py        # Criterion implementation of classification accuracy. Mainly used for ASR teacher-forcing accuracy.
        /bce_logits.py      # Criterion implementation of binary cross entropy. Mainly used for TTS stop flag.
        /cross_entropy.py   # Criterion implementation of cross entropy. Mainly used for ASR seq2seq loss function.
        /error_rate.py      # Criterion implementation of word and char error rate. Mainly used for ASR evaluation.
        /fbeta_score.py     # Criterion implementation of F_beta score. Mainly used for evaluation of TTS stop flag.
        /least_error.py     # Criterion implementation of mean square error and mean absolute error. Mainly used for TTS seq2seq loss function.

API Document

1. speechain.criterion.abs.Criterion.__init__
2. speechain.criterion.abs.Criterion.criterion_init
3. speechain.criterion.abs.Criterion.__call__

speechain.criterion.abs.Criterion.__init__(self, **criterion_conf)

• Description:
  This initialization function is shared by all Criterion subclasses. Currently, the shared logic only contains calling the initialization function of the parent class.
• Arguments:
  • **criterion_conf:
    The arguments used by criterion_init() for your customized Criterion initialization.

👆Back to the API list

speechain.criterion.abs.Criterion.criterion_init(self, **criterion_conf)

• Description:
  Abstract interface function for customized initialization of each Criterion subclass. This interface function is not mandatory to be overridden by your implementation.
• Arguments:
  • **criterion_conf:
    The arguments used for customized Criterion initialization. For more details, please refer to the docstring of your target Criterion subclass.

👆Back to the API list

speechain.criterion.abs.Criterion.__call__(self, **kwargs)

• Description:
  This abstract interface function receives the model forward calculation results and ground-truth labels. The output is a scalar which could be either trainable for parameter optimization or non-trainable for information recording.
  This interface function is mandatory to be overridden by your implementation.
• Arguments:
  • **kwargs:
    model forward calculation results and ground-truth labels.
    For more details, please refer to the docstring of __call__() of your target Criterion subclass.
• Return: A trainable or non-trainable scalar.
  For more details, please refer to the docstring of __call__() of your target Criterion subclass.

👆Back to the API list

👆Back to the table of contents