`latte.metrics.keras.disentanglement`

Module Contents

Classes

`MutualInformationGap`	Calculate Mutual Information Gap (MIG) between latent vectors and attributes.
`DependencyAwareMutualInformationGap`	Calculate Dependency-Aware Mutual Information Gap (DMIG) between latent vectors and attributes
`DependencyAwareLatentInformationGap`	Calculate Dependency-Aware Latent Information Gap (DLIG) between latent vectors and attributes
`DependencyBlindMutualInformationGap`	Calculate Dependency-Blind Mutual Information Gap (XMIG) between latent vectors and attributes
`SeparateAttributePredictability`	Calculate Separate Attribute Predictability (SAP) between latent vectors and attributes
`Modularity`	Calculate Modularity between latent vectors and attributes

Attributes

`MIG`	alias for `MutualInformationGap`
`DMIG`	alias for `DependencyAwareMutualInformationGap`
`DLIG`	alias for `DependencyAwareLatentInformationGap`
`XMIG`	alias for `DependencyBlindMutualInformationGap`
`SAP`	alias for `SeparateAttributePredictability`

class MutualInformationGap(reg_dim=None, discrete=False, fill_reg_dim=False)

Bases: latte.metrics.keras.wrapper.KerasMetricWrapper

Calculate Mutual Information Gap (MIG) between latent vectors and attributes.

Mutual Information Gap measures the degree of disentanglement. For each attribute, MIG is calculated by difference in the mutual informations between that of the attribute and its most informative latent dimension, and that of the attribute and its second-most informative latent dimension. Mathematically, MIG is given by

\[\operatorname{MIG}(a_i, \mathbf{z}) = \dfrac{\mathcal{I}(a_i, z_j)-\mathcal{I}(a_i, z_k)}{\mathcal{H}(a_i)},\]

where \(j=\operatorname{arg}\max_n \mathcal{I}(a_i, z_n)\), \(k=\operatorname{arg}\max_{n≠j} \mathcal{I}(a_i, z_n)\), \(\mathcal{I}(\cdot,\cdot)\) is mutual information, and \(\mathcal{H}(\cdot)\) is entropy.

If reg_dim is specified, \(j\) is instead overwritten to reg_dim[i], while \(k=\operatorname{arg}\max_{n≠j} \mathcal{I}(a_i, z_n)\) as usual.

MIG is best applied for independent attributes.

Parameters:

reg_dim (Optional[List], optional) – regularized dimensions, by default None Attribute a[:, i] is regularized by z[:, reg_dim[i]]. If reg_dim is provided, the first mutual information is always taken between the regularized dimension and the attribute, and MIG may be negative.
discrete (bool, optional) – Whether the attributes are discrete, by default False
fill_reg_dim (bool, optional) – Whether to automatically fill reg_dim with range(n_attributes), by default False. If fill_reg_dim is True, the reg_dim behavior is the same as the dependency-aware family. This option is mainly used for compatibility with the dependency-aware family in a bundle.

See also

bundles.DependencyAwareMutualInformationBundle: Dependency-Aware Mutual Information Bundle
DependencyAwareMutualInformationGap: Dependency-Aware Mutual Information Gap
DependencyBlindMutualInformationGap: Dependency-Blind Mutual Information Gap
DependencyAwareLatentInformationGap: Dependency-Aware Latent Information Gap

References

[1]

Chen, X. Li, R. Grosse, and D. Duvenaud, “Isolating sources of disentanglement in variational autoencoders”, in Proceedings of the 32nd International Conference on Neural Information Processing Systems, 2018.

update_state(z, a)

Update metric states. This function converts the tensors to numpy arrays then append the latent vectors and attributes to the internal state lists.

Parameters:

z (tf.Tensor, (n_samples, n_features)) – a batch of latent vectors
a (tf.Tensor, (n_samples, n_attributes) or (n_samples,)) – a batch of attribute(s)

result()

Compute metric values from the current state. The latent vectors and attributes in the internal states are concatenated along the sample dimension and passed to the metric function to obtain the metric values.

Returns:: MIG for each attribute
Return type:: tf.Tensor, (n_attributes,)

class DependencyAwareMutualInformationGap(reg_dim=None, discrete=False)

Bases: latte.metrics.keras.wrapper.KerasMetricWrapper

Calculate Dependency-Aware Mutual Information Gap (DMIG) between latent vectors and attributes

Dependency-Aware Mutual Information Gap (DMIG) is a dependency-aware version of MIG that accounts for attribute interdependence observed in real-world data. Mathematically, DMIG is given by

\[\operatorname{DMIG}(a_i, \mathbf{z}) = \dfrac{\mathcal{I}(a_i, z_j)-\mathcal{I}(a_i, z_k)}{\mathcal{H}(a_i|a_l)},\]

where \(j=\operatorname{arg}\max_n \mathcal{I}(a_i, z_n)\), \(k=\operatorname{arg}\max_{n≠j} \mathcal{I}(a_i, z_n)\), \(\mathcal{I}(\cdot,\cdot)\) is mutual information, \(\mathcal{H}(\cdot|\cdot)\) is conditional entropy, and \(a_l\) is the attribute regularized by \(z_k\). If \(z_k\) is not regularizing any attribute, DMIG reduces to the usual MIG. DMIG compensates for the reduced maximum possible value of the numerator due to attribute interdependence.

If reg_dim is specified, \(j\) is instead overwritten to reg_dim[i], while \(k=\operatorname{arg}\max_{n≠j} \mathcal{I}(a_i, z_n)\) as usual.

Parameters:

reg_dim (Optional[List], optional) – regularized dimensions, by default None Attribute a[:, i] is regularized by z[:, reg_dim[i]]. If None, a[:, i] is assumed to be regularized by z[:, i].
discrete (bool, optional) – Whether the attributes are discrete, by default False

See also

bundles.DependencyAwareMutualInformationBundle: Dependency-Aware Mutual Information Bundle
MutualInformationGap: Mutual Information Gap
DependencyBlindMutualInformationGap: Dependency-Blind Mutual Information Gap
DependencyAwareLatentInformationGap: Dependency-Aware Latent Information Gap

References

[1]

1. Watcharasupat and A. Lerch, “Evaluation of Latent Space Disentanglement in the Presence of Interdependent Attributes”, in Extended Abstracts of the Late-Breaking Demo Session of the 22nd International Society for Music Information Retrieval Conference, 2021.

[2]

1. Watcharasupat, “Controllable Music: Supervised Learning of Disentangled Representations for Music Generation”, 2021.

update_state(z, a)

Update metric states. This function converts the tensors to numpy arrays then append the latent vectors and attributes to the internal state lists.

Parameters:

z (tf.Tensor, (n_samples, n_features)) – a batch of latent vectors
a (tf.Tensor, (n_samples, n_attributes) or (n_samples,)) – a batch of attribute(s)

result()

Compute metric values from the current state. The latent vectors and attributes in the internal states are concatenated along the sample dimension and passed to the metric function to obtain the metric values.

Returns:: DMIG for each attribute
Return type:: tf.Tensor, (n_attributes,)

class DependencyAwareLatentInformationGap(reg_dim=None, discrete=False)

Bases: latte.metrics.keras.wrapper.KerasMetricWrapper

Calculate Dependency-Aware Latent Information Gap (DLIG) between latent vectors and attributes

Dependency-aware Latent Information Gap (DLIG) is a latent-centric counterpart to DMIG. DLIG evaluates disentanglement of a set of semantic attributes \(\{a_i\}\) with respect to a latent dimension \(z_d\) such that

\[\operatorname{DLIG}(\{a_i\}, z_d) = \dfrac{\mathcal{I}(a_j, z_d)-\mathcal{I}(a_k, z_d)}{\mathcal{H}(a_j|a_k)},\]

where \(j=\operatorname{arg}\max_i \mathcal{I}(a_i, z_d)\), \(k=\operatorname{arg}\max_{i≠j} \mathcal{I}(a_i, z_d)\), \(\mathcal{I}(\cdot,\cdot)\) is mutual information, and \(\mathcal{H}(\cdot|\cdot)\) is conditional entropy.

If reg_dim is specified, \(j\) is instead overwritten to reg_dim[i], while \(k=\operatorname{arg}\max_{i≠j} \mathcal{I}(a_i, z_d)\) as usual.

Parameters:

reg_dim (Optional[List], optional) – regularized dimensions, by default None Attribute a[:, i] is regularized by z[:, reg_dim[i]]. If None, a[:, i] is assumed to be regularized by z[:, i].
discrete (bool, optional) – Whether the attributes are discrete, by default False

See also

bundles.DependencyAwareMutualInformationBundle: Dependency-Aware Mutual Information Bundle
MutualInformationGap: Mutual Information Gap
DependencyBlindMutualInformationGap: Dependency-Blind Mutual Information Gap
DependencyAwareMutualInformationGap: Dependency-Aware Mutual Information Gap

References

[1]

1. Watcharasupat, “Controllable Music: Supervised Learning of Disentangled Representations for Music Generation”, 2021.

update_state(z, a)

Update metric states. This function converts the tensors to numpy arrays then append the latent vectors and attributes to the internal state lists.

Parameters:

z (tf.Tensor, (n_samples, n_features)) – a batch of latent vectors
a (tf.Tensor, (n_samples, n_attributes) or (n_samples,)) – a batch of attribute(s)

result()

Compute metric values from the current state. The latent vectors and attributes in the internal states are concatenated along the sample dimension and passed to the metric function to obtain the metric values.

Returns:: DLIG for each attribute-regularizing latent dimension
Return type:: tf.Tensor, (n_attributes,)

class DependencyBlindMutualInformationGap(reg_dim=None, discrete=False)

Bases: latte.metrics.keras.wrapper.KerasMetricWrapper

Calculate Dependency-Blind Mutual Information Gap (XMIG) between latent vectors and attributes

Dependency-blind Mutual Information Gap (XMIG) is a complementary metric to MIG and DMIG that measures the gap in mutual information with the subtrahend restricted to dimensions which do not regularize any attribute. XMIG is given by

\[\operatorname{XMIG}(a_i, \mathbf{z}) = \dfrac{\mathcal{I}(a_i, z_j)-\mathcal{I}(a_i, z_k)}{\mathcal{H}(a_i)},\]

where \(j=\operatorname{arg}\max_d \mathcal{I}(a_i, z_d)\), \(k=\operatorname{arg}\max_{d∉\mathcal{D}} \mathcal{I}(a_i, z_d)\), \(\mathcal{I}(\cdot,\cdot)\) is mutual information, \(\mathcal{H}(\cdot)\) is entropy, and \(\mathcal{D}\) is a set of latent indices which do not regularize any attribute. XMIG allows monitoring of latent disentanglement exclusively against attribute-unregularized latent dimensions.

If reg_dim is specified, \(j\) is instead overwritten to reg_dim[i], while \(k=\operatorname{arg}\max_{d∉\mathcal{D}} \mathcal{I}(a_i, z_d)\) as usual.

Parameters:

reg_dim (Optional[List], optional) – regularized dimensions, by default None Attribute a[:, i] is regularized by z[:, reg_dim[i]]. If None, a[:, i] is assumed to be regularized by z[:, i].
discrete (bool, optional) – Whether the attributes are discrete, by default False

See also

bundles.DependencyAwareMutualInformationBundle: Dependency-Aware Mutual Information Bundle
MutualInformationGap: Mutual Information Gap
DependencyAwareMutualInformationGap: Dependency-Aware Mutual Information Gap
DependencyAwareLatentInformationGap: Dependency-Aware Latent Information Gap

References

[1]

1. Watcharasupat, “Controllable Music: Supervised Learning of Disentangled Representations for Music Generation”, 2021.

update_state(z, a)

Update metric states. This function converts the tensors to numpy arrays then append the latent vectors and attributes to the internal state lists.

Parameters:

z (tf.Tensor, (n_samples, n_features)) – a batch of latent vectors
a (tf.Tensor, (n_samples, n_attributes) or (n_samples,)) – a batch of attribute(s)

result()

Compute metric values from the current state. The latent vectors and attributes in the internal states are concatenated along the sample dimension and passed to the metric function to obtain the metric values.

Returns:: XMIG for each attribute
Return type:: tf.Tensor, (n_attributes,)

class SeparateAttributePredictability(reg_dim=None, discrete=False, l2_reg=1.0, thresh=1e-12)

Bases: latte.metrics.keras.wrapper.KerasMetricWrapper

Calculate Separate Attribute Predictability (SAP) between latent vectors and attributes

Separate Attribute Predictability (SAP) is similar in nature to MIG but, instead of mutual information, uses the coefficient of determination for continuous attributes and classification accuracy for discrete attributes to measure the extent of relationship between a latent dimension and an attribute. SAP is given by

\[\operatorname{SAP}(a_i, \mathbf{z}) = \mathcal{S}(a_i, z_j)-\mathcal{S}(a_i, z_k),\]

where \(j=\operatorname{arg}\max_d \mathcal{S}(a_i, z_d)\), \(k=\operatorname{arg}\max_{d≠j} \mathcal{S}(a_i, z_d)\), and \(\mathcal{S}(\cdot,\cdot)\) is either the coefficient of determination or classification accuracy.

If reg_dim is specified, \(j\) is instead overwritten to reg_dim[i], while \(k=\operatorname{arg}\max_{d≠j} \mathcal{S}(a_i, z_d)\) as usual.

Parameters:

reg_dim (Optional[List], optional) – regularized dimensions, by default None Attribute a[:, i] is regularized by z[:, reg_dim[i]]. If None, a[:, i] is assumed to be regularized by z[:, i].
discrete (bool, optional) – Whether the attributes are discrete, by default False
l2_reg (float, optional) – regularization parameter for linear classifier, by default 1.0. Ignored if discrete is False. See sklearn.svm.LinearSVC for more details.
thresh (float, optional) – threshold for latent vector variance, by default 1e-12. Latent dimensions with variance below thresh will have SAP contribution zeroed. Ignored if discrete is True.

See also

sklearn.svm.LinearSVC: Linear SVC

References

[1]

Kumar, P. Sattigeri, and A. Balakrishnan, “Variational inference of disentangled latent concepts from unlabeled observations”, in Proceedings of the 6th International Conference on Learning Representations, 2018.

update_state(z, a)

Update metric states. This function converts the tensors to numpy arrays then append the latent vectors and attributes to the internal state lists.

Parameters:

z (tf.Tensor, (n_samples, n_features)) – a batch of latent vectors
a (tf.Tensor, (n_samples, n_attributes) or (n_samples,)) – a batch of attribute(s)

result()

Compute metric values from the current state. The latent vectors and attributes in the internal states are concatenated along the sample dimension and passed to the metric function to obtain the metric values.

Returns:: SAP for each attribute
Return type:: tf.Tensor, (n_attributes,)

class Modularity(reg_dim=None, discrete=False, thresh=1e-12)

Bases: latte.metrics.keras.wrapper.KerasMetricWrapper

Calculate Modularity between latent vectors and attributes

Modularity is a letent-centric measure of disentanglement based on mutual information. Modularity measures the degree in which a latent dimension contains information about only one attribute, and is given by

\[\operatorname{Modularity}(\{a_i\}, z_d) = 1-\dfrac{\sum_{i≠j}(\mathcal{I}(a_i, z_d)/\mathcal{I}(a_j, z_d))^2}{|{a_i}| -1},\]

where \(j=\operatorname{arg}\max_i \mathcal{I}(a_i, z_d)\), and \(\mathcal{I}(\cdot,\cdot)\) is mutual information.

reg_dim is currently ignored in Modularity.

Parameters:

reg_dim (Optional[List], optional) – regularized dimensions, by default None. Attribute a[:, i] is regularized by z[:, reg_dim[i]]. If None, a[:, i] is assumed to be regularized by z[:, i].
discrete (bool, optional) – Whether the attributes are discrete, by default False
thresh (float, optional) – threshold for mutual information, by default 1e-12. Latent-attribute pair with variance below thresh will have modularity contribution zeroed.

References

[1]

Ridgeway and M. C. Mozer, “Learning deep disentangled embeddings with the F-statistic loss,” in Proceedings of the 32nd International Conference on Neural Information Processing Systems, 2018, pp. 185–194.

update_state(z, a)

Update metric states. This function converts the tensors to numpy arrays then append the latent vectors and attributes to the internal state lists.

Parameters:

z (tf.Tensor, (n_samples, n_features)) – a batch of latent vectors
a (tf.Tensor, (n_samples, n_attributes) or (n_samples,)) – a batch of attribute(s)

result()

Compute metric values from the current state. The latent vectors and attributes in the internal states are concatenated along the sample dimension and passed to the metric function to obtain the metric values.

Returns:: Modularity for each attribute-regularizing latent dimension
Return type:: tf.Tensor, (n_attributes,)

MIG: alias for MutualInformationGap

DMIG: alias for DependencyAwareMutualInformationGap

DLIG: alias for DependencyAwareLatentInformationGap

XMIG: alias for DependencyBlindMutualInformationGap

SAP: alias for SeparateAttributePredictability

latte.metrics.keras.disentanglement

Module Contents

Classes

Attributes

`latte.metrics.keras.disentanglement`