Skip to content

Logs

VAE.logs

Collection of log functions for model training.

VAE.logs.ActiveUnits 

ActiveUnits(z_mean, z_log_var, kl_threshold=0.1)

Number of active units.

Return metric for number of active units that are above the threshold specified in kl_threshold.

Parameters:

  • z_mean (Tensor) –

    Output node of encoder, specifying z_mean.

  • z_log_var (Tensor) –

    Output node of encoder, specifying z_log_var.

  • kl_threshold (float, default: 0.1 ) –

    Lower bound for the KL divergence above which a latent dimension is assumed to be active.

Returns:

  • callable –

    Metrics that returns the number of active units.

Source code in VAE/logs.py 
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
def ActiveUnits(z_mean: tf.Tensor, z_log_var: tf.Tensor, kl_threshold: float = 0.1) -> callable:
    """Number of active units.

    Return metric for number of active units that are above the threshold specified in `kl_threshold`.

    Parameters:
        z_mean:
            Output node of encoder, specifying `z_mean`.
        z_log_var:
            Output node of encoder, specifying `z_log_var`.
        kl_threshold:
            Lower bound for the KL divergence above which a latent dimension is assumed to be active.

    Returns:
        Metrics that returns the number of active units.

    """
    def active_units(y_true, y_pred):
        """Prepare function that takes y_true and y_pred as input"""
        kl_loss = 1 + z_log_var - tf.square(z_mean) - tf.exp(z_log_var)
        kl_loss *= -0.5
        kl_loss = tf.reduce_mean(kl_loss, axis=0)
        is_active = tf.greater(kl_loss, kl_threshold)
        is_active = tf.cast(is_active, kl_loss.dtype)

        return tf.reduce_sum(is_active)

    return active_units

VAE.logs.Beta 

Beta(beta)

Beta value.

Returns metrics for beta value, which can be added to the model metrics for logging purposes.

Parameters:

  • beta (Tensor) –

    Beta multiplier of the KL divergence.

Returns:

  • callable –

    Metric function that reduces the beta tensor to a scalar.

Source code in VAE/logs.py 
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
def Beta(beta: tf.Tensor) -> callable:
    """Beta value.

    Returns metrics for beta value, which can be added to the model metrics for logging purposes.

    Parameters:
        beta:
            Beta multiplier of the KL divergence.

    Returns:
        Metric function that reduces the beta tensor to a scalar.
    """
    def beta_value(y_true, y_pred):
        return tf.reduce_mean(beta)

    return beta_value