How do I mask a loss function in Keras with the TensorFlow backend?

If you're not using masks as in Yu-Yang's answer, you can try this.

If you have your target data Y with length and padded with the mask value, you can:

import keras.backend as K
def custom_loss(yTrue,yPred):

    #find which values in yTrue (target) are the mask value
    isMask = K.equal(yTrue, maskValue) #true for all mask values

    #since y is shaped as (batch, length, features), we need all features to be mask values
    isMask = K.all(isMask, axis=-1) #the entire output vector must be true
        #this second line is only necessary if the output features are more than 1

    #transform to float (0 or 1) and invert
    isMask = K.cast(isMask, dtype=K.floatx())
    isMask = 1 - isMask #now mask values are zero, and others are 1

    #multiply this by the inputs:
       #maybe you might need K.expand_dims(isMask) to add the extra dimension removed by K.all
     yTrue = yTrue * isMask   
     yPred = yPred * isMask

     return someLossFunction(yTrue,yPred)

---

If you have padding only for the input data, or if Y has no length, you can have your own mask outside the function:

masks = [
   [1,1,1,1,1,1,0,0,0],
   [1,1,1,1,0,0,0,0,0],
   [1,1,1,1,1,1,1,1,0]
]
 #shape (samples, length). If it fails, make it (samples, length, 1). 

import keras.backend as K

masks = K.constant(masks)

Since masks depend on your input data, you can use your mask value to know where to put zeros, such as:

masks = np.array((X_train == maskValue).all(), dtype='float64')    
masks = 1 - masks

#here too, if you have a problem with dimensions in the multiplications below
#expand masks dimensions by adding a last dimension = 1.

And make your function taking masks from outside of it (you must recreate the loss function if you change the input data):

def customLoss(yTrue,yPred):

    yTrue = masks*yTrue
    yPred = masks*yPred

    return someLossFunction(yTrue,yPred)

---

Does anyone know if keras automatically masks the loss function?? Since it provides a Masking layer and says nothing about the outputs, maybe it does it automatically?

If there's a mask in your model, it'll be propagated layer-by-layer and eventually applied to the loss. So if you're padding and masking the sequences in a correct way, the loss on the padding placeholders would be ignored.

Some Details:

It's a bit involved to explain the whole process, so I'll just break it down to several steps:

1. In compile(), the mask is collected by calling compute_mask() and applied to the loss(es) (irrelevant lines are ignored for clarity).

weighted_losses = [_weighted_masked_objective(fn) for fn in loss_functions]

# Prepare output masks.
masks = self.compute_mask(self.inputs, mask=None)
if masks is None:
    masks = [None for _ in self.outputs]
if not isinstance(masks, list):
    masks = [masks]

# Compute total loss.
total_loss = None
with K.name_scope('loss'):
    for i in range(len(self.outputs)):
        y_true = self.targets[i]
        y_pred = self.outputs[i]
        weighted_loss = weighted_losses[i]
        sample_weight = sample_weights[i]
        mask = masks[i]
        with K.name_scope(self.output_names[i] + '_loss'):
            output_loss = weighted_loss(y_true, y_pred,
                                        sample_weight, mask)

2. Inside Model.compute_mask(), run_internal_graph() is called.
3. Inside run_internal_graph(), the masks in the model is propagated layer-by-layer from the model's inputs to outputs by calling Layer.compute_mask() for each layer iteratively.

So if you're using a Masking layer in your model, you shouldn't worry about the loss on the padding placeholders. The loss on those entries will be masked out as you've probably already seen inside _weighted_masked_objective().

A Small Example:

max_sentence_length = 5
character_number = 2

input_tensor = Input(shape=(max_sentence_length, character_number))
masked_input = Masking(mask_value=0)(input_tensor)
output = LSTM(3, return_sequences=True)(masked_input)
model = Model(input_tensor, output)
model.compile(loss='mae', optimizer='adam')

X = np.array([[[0, 0], [0, 0], [1, 0], [0, 1], [0, 1]],
              [[0, 0], [0, 1], [1, 0], [0, 1], [0, 1]]])
y_true = np.ones((2, max_sentence_length, 3))
y_pred = model.predict(X)
print(y_pred)
[[[ 0.          0.          0.        ]
  [ 0.          0.          0.        ]
  [-0.11980877  0.05803877  0.07880752]
  [-0.00429189  0.13382857  0.19167568]
  [ 0.06817091  0.19093043  0.26219055]]

 [[ 0.          0.          0.        ]
  [ 0.0651961   0.10283815  0.12413475]
  [-0.04420842  0.137494    0.13727818]
  [ 0.04479844  0.17440712  0.24715884]
  [ 0.11117355  0.21645413  0.30220413]]]

# See if the loss computed by model.evaluate() is equal to the masked loss
unmasked_loss = np.abs(1 - y_pred).mean()
masked_loss = np.abs(1 - y_pred[y_pred != 0]).mean()

print(model.evaluate(X, y_true))
0.881977558136

print(masked_loss)
0.881978

print(unmasked_loss)
0.917384

As can be seen from this example, the loss on the masked part (the zeroes in y_pred) is ignored, and the output of model.evaluate() is equal to masked_loss.

---

EDIT:

If there's a recurrent layer with return_sequences=False, the mask stop propagates (i.e., the returned mask is None). In RNN.compute_mask():

def compute_mask(self, inputs, mask):
    if isinstance(mask, list):
        mask = mask[0]
    output_mask = mask if self.return_sequences else None
    if self.return_state:
        state_mask = [None for _ in self.states]
        return [output_mask] + state_mask
    else:
        return output_mask

In your case, if I understand correctly, you want a mask that's based on y_true, and whenever the value of y_true is [0, 0, 1] (the one-hot encoding of "#") you want the loss to be masked. If so, you need to mask the loss values in a somewhat similar way to Daniel's answer.

The main difference is the final average. The average should be taken over the number of unmasked values, which is just K.sum(mask). And also, y_true can be compared to the one-hot encoded vector [0, 0, 1] directly.

def get_loss(mask_value):
    mask_value = K.variable(mask_value)
    def masked_categorical_crossentropy(y_true, y_pred):
        # find out which timesteps in `y_true` are not the padding character '#'
        mask = K.all(K.equal(y_true, mask_value), axis=-1)
        mask = 1 - K.cast(mask, K.floatx())

        # multiply categorical_crossentropy with the mask
        loss = K.categorical_crossentropy(y_true, y_pred) * mask

        # take average w.r.t. the number of unmasked entries
        return K.sum(loss) / K.sum(mask)
    return masked_categorical_crossentropy

masked_categorical_crossentropy = get_loss(np.array([0, 0, 1]))
model = Model(input_tensor, output)
model.compile(loss=masked_categorical_crossentropy, optimizer='adam')

The output of the above code then shows that the loss is computed only on the unmasked values:

model.evaluate: 1.08339476585
tf unmasked_loss: 1.08989
tf masked_loss: 1.08339

The value is different from yours because I've changed the axis argument in tf.reverse from [0,1] to [1].