auPR metric

from sklearn.metrics import mean_squared_error

from sklearn.metrics import mean_absolute_error

from sklearn.metrics import precision_score

from sklearn.metrics import precision_recall_curve

from sklearn.metrics import auc

from scipy.stats import pearsonr

  return pearsonr(y, y_pred)[0]**2

def auPR_score(y, y_pred):

  """Compute area under precision-recall curve"""

  assert y_pred.shape == y.shape

  n_classes = y_pred.shape[1]

  scores = []

  for i in range(n_classes):

    precision, recall, _ = precision_recall_curve(y[:, i], y_pred[:, i])

    scores.append(auc(recall, precision))

  return np.mean(scores)

def rms_score(y_true, y_pred):

  """Computes RMS error."""

  return np.sqrt(mean_squared_error(y_true, y_pred))

      if self.metric.__name__ in [

          "roc_auc_score", "matthews_corrcoef", "recall_score",

          "accuracy_score", "kappa_score", "precision_score",

          "balanced_accuracy_score"

          "balanced_accuracy_score", "auPR_score"

      ]:

        mode = "classification"

      elif self.metric.__name__ in [

      # TODO(rbharath): This has been a major source of bugs. Is there a more

      # robust characterization of which metrics require class-probs and which

      # don't?

      if "roc_auc_score" in self.name:

      if "roc_auc_score" in self.name or "auPR_score" in self.name:

        y_true = to_one_hot(y_true).astype(int)

        y_pred = np.reshape(y_pred, (n_samples, n_classes))

      else:

    # Fit trained model

    model.fit(dataset, nb_epoch=20)

    model.save()

    # Eval model on train

    scores = model.evaluate(dataset, [regression_metric])

    # Fit trained model

    model.fit(dataset, nb_epoch=50)

    model.save()

    # Eval model on train

    scores = model.evaluate(dataset, [regression_metric])

    # Fit trained model

    model.fit(dataset, nb_epoch=20)

    model.save()

    # Eval model on train

    scores = model.evaluate(dataset, [classification_metric])

    # Fit trained model

    model.fit(dataset, nb_epoch=40)

    model.save()

    # Eval model on train

    scores = model.evaluate(dataset, [regression_metric])