def _weighted_loss(y_true, y_pred, sample_weight):
def _standardize(losses, sample_weight):
if isinstance(sample_weight, np.ndarray):
if (sample_weight.shape[-1] != y_true.shape[-1]) and (
sample_weight.ndim < y_true.ndim):
sample_weight = np.expand_dims(sample_weight, -1)
if losses.ndim < sample_weight.ndim:
losses = np.expand_dims(losses, -1)
return losses, sample_weight
yt, yp = _maybe_cast_to_tensor(y_true, y_pred)
losses = K.eval(getattr(keras_losses, name)(yt, yp))
# negative is standard in TF2, but was positive in TF1
if name == 'cosine_similarity' and (not TF_2 and TF_KERAS):
losses = -losses
losses, sample_weight = _standardize(losses, sample_weight)
return np.mean(losses * sample_weight)
def _fn(y_true, y_pred):
yt, yp = _maybe_cast_to_tensor(y_true, y_pred)
# sample_weight makes no sense for keras `metrics`
return K.eval(getattr(keras_metrics, name)(yt, yp))