def _setup_metrics(self):
self.metric_functions = {} # needed to shadow class variable
if self.decoder == 'generator' and self.decoder_obj.beam_width > 1:
# Generator Decoder w/ beam search
# beam search does not provide logits
self.metric_functions[LOSS] = SequenceLossMetric(from_logits=False)
else:
# Generator Decoder w/ no beam search and Tagger Decoder
self.metric_functions[LOSS] = SequenceLossMetric(from_logits=True)
self.metric_functions[TOKEN_ACCURACY] = TokenAccuracyMetric()
self.metric_functions[SEQUENCE_ACCURACY] = SequenceAccuracyMetric()
self.metric_functions[LAST_ACCURACY] = SequenceLastAccuracyMetric()
self.metric_functions[PERPLEXITY] = PerplexityMetric()
self.metric_functions[EDIT_DISTANCE] = EditDistanceMetric()
def _setup_loss(self):
if self.loss[TYPE] == 'softmax_cross_entropy':
self.train_loss_function = SequenceLoss()
elif self.loss[TYPE] == 'sampled_softmax_cross_entropy':
self.train_loss_function = SampledSoftmaxCrossEntropyLoss(
decoder_obj=self.decoder_obj,
num_classes=self.num_classes,
feature_loss=self.loss,
name='train_loss')
else:
raise ValueError("Loss type {} is not supported. Valid values are "
"'softmax_cross_entropy' or "
"'sampled_softmax_cross_entropy'".format(
self.loss[TYPE]))
self.eval_loss_function = SequenceLossMetric()