def setup_losses_and_metrics(self):
def selectL(loss):
if type(loss) is dict:
loss = {k: selectLoss(v) for k, v in loss.items()}
else:
loss = selectLoss(loss)
return loss
self.loss_metrics = selectL(self.config[a.loss_metric])
self.loss_weights = self.config.get("loss_weights")
self.class_weights = self.config.get("class_weights")
if self.loss_weights is None and type(self.loss_metrics) is dict:
self.loss_weights = [1.0 for _ in range(len(self.loss_metrics))]
if type(self.config[a.metrics]) is list:
self.metrics_name = [
selectMetric(m) for m in self.config[a.metrics]
]
else:
def selectM(metric):
if type(metric) is list:
return [selectMetric(m_i) for m_i in metric]
else:
return selectMetric(metric)
self.metrics_name = {
n: selectM(m)
for n, m in self.config[a.metrics].items()
}
def apply_metric(metric_name, y_true, y_pred) -> float:
"""Perform the computation of provided metric.
:meta private:
Args:
metric_name (str|callable): If ``str`` then it needs to be a metric available in ``deephyper.nas.metrics``.
y_true (array): Array of true predictions.
y_pred (array): Array of predicted predictions
Returns:
float: a scalar value of the computed metric.
"""
metric_func = selectMetric(metric_name)
metric = tf.reduce_mean(
metric_func(
tf.convert_to_tensor(y_true, dtype=np.float32),
tf.convert_to_tensor(y_pred, dtype=np.float32),
)).numpy()
return metric
def selectM(metric):
if type(metric) is list:
return [selectMetric(m_i) for m_i in metric]
else:
return selectMetric(metric)