def get_loss_train(self):
"""Calculates (if needed) cross-entropy losses for the training set."""
if self.loss_train is None:
if self.logits_train is not None:
self.loss_train = utils.log_loss_from_logits(
self.labels_train, self.logits_train)
else:
self.loss_train = utils.log_loss(self.labels_train,
self.probs_train)
return self.loss_train
def calculate_losses(model, data, labels):
"""Calculate losses of model prediction on data, provided true labels.
Args:
model: model to make prediction
data: samples
labels: true labels of samples (integer valued)
Returns:
preds: probability vector of each sample
loss: cross entropy loss of each sample
"""
pred = model.predict(data)
loss = log_loss(labels, pred)
return pred, loss
def get_loss_test(self):
"""Calculates (if needed) cross-entropy losses for the test set.
Returns:
Loss (or None if neither the loss nor the labels are present).
"""
if self.loss_test is None:
if self.labels_test is None:
return None
if self.logits_test is not None:
self.loss_test = utils.log_loss_from_logits(
self.labels_test, self.logits_test)
else:
self.loss_test = utils.log_loss(self.labels_test,
self.probs_test)
return self.loss_test
def calculate_losses(estimator, input_fn, labels):
"""Get predictions and losses for samples.
The assumptions are 1) the loss is cross-entropy loss, and 2) user have
specified prediction mode to return predictions, e.g.,
when mode == tf.estimator.ModeKeys.PREDICT, the model function returns
tf.estimator.EstimatorSpec(mode=mode, predictions=tf.nn.softmax(logits)).
Args:
estimator: model to make prediction
input_fn: input function to be used in estimator.predict
labels: true labels of samples
Returns:
preds: probability vector of each sample
loss: cross entropy loss of each sample
"""
pred = np.array(list(estimator.predict(input_fn=input_fn)))
loss = log_loss(labels, pred)
return pred, loss
