def mask_loss(loss, mask_value=MASK_VALUE):
"""Generates a new loss function that ignores values where `y_true == mask_value`.
# Arguments
loss: str; name of the keras loss function from `keras.losses`
mask_value: int; which values should be masked
# Returns
function; Masked version of the `loss`
# Example
```python
categorical_crossentropy_masked = mask_loss("categorical_crossentropy")
```
"""
loss_fn = kloss.deserialize(loss)
def masked_loss_fn(y_true, y_pred):
# currently not suppoerd with NA's:
# - there is no K.is_nan impolementation in keras.backend
# - https://github.com/fchollet/keras/issues/1628
mask = K.cast(K.not_equal(y_true, mask_value), K.floatx())
# we divide by the mean to correct for the number of done loss evaluations
return loss_fn(y_true * mask, y_pred * mask) / K.mean(mask)
masked_loss_fn.__name__ = loss + "_masked"
return masked_loss_fn
def test_serializing_loss_class():
orig_loss_class = MSE_MAE_loss(0.3)
with custom_object_scope({'MSE_MAE_loss': MSE_MAE_loss}):
serialized = losses.serialize(orig_loss_class)
with custom_object_scope({'MSE_MAE_loss': MSE_MAE_loss}):
deserialized = losses.deserialize(serialized)
assert isinstance(deserialized, MSE_MAE_loss)
assert deserialized.mse_fraction == 0.3
def test_serializing_loss_class():
orig_loss_class = MSE_MAE_loss(0.3)
with custom_object_scope({'MSE_MAE_loss': MSE_MAE_loss}):
serialized = losses.serialize(orig_loss_class)
with custom_object_scope({'MSE_MAE_loss': MSE_MAE_loss}):
deserialized = losses.deserialize(serialized)
assert isinstance(deserialized, MSE_MAE_loss)
assert deserialized.mse_fraction == 0.3
