MaskedBinaryCrossEntropy

Bases: tf.keras.losses.Loss

Compute Binary Cross-Entropy with masking.

Parameters:

Name	Type	Description	Default
mask_value	int	value in y_true to be masked.	None
name	str	name of the loss function.	'masked_binary_crossentropy'
reduction	str	reduction method for the loss.	'none'

call(y_true, y_pred)

Calculate masked binary cross-entropy.

Define as:

$$H_p(q) = - \frac{1}{M} \sum_{i=1}^{N} m_i \cdot (y_i \cdot \log(p(y_i)) + (1-y_i) \cdot \log(1-p(y_i)))$$

where $y_i$ is the target, and $p(y_i)$ is the prediction $m_i$ is the mask, and $M$ is the number of unmasked units.

Warning

"$\cdot$" is element-wise multiplication.

Parameters:

Name	Type	Description	Default
y_true	tf.Tensor	target y with shape (batch_size, seq_len, feature)	required
y_pred	tf.Tensor	predicted y with shape (batch_size, seq_len, feature)	required

Returns:

Name	Type	Description
Loss	tf.Tensor	Loss values with shape (batch_size)