KGE.constraint

Functions

Lp_regularization(X, p, axis)	Standard Lp-regularization
clip_constraint(X, p, value, axis)	Clip embeddings
normalized_embeddings(X, p, value, axis)	Normalized embeddings
soft_constraint(X, p, value, axis)	Soft constraint

KGE.constraint.Lp_regularization(X, p, axis)

Standard Lp-regularization

The standard Lp-regularization:

..math ::

regularization ~ term =

sum left| X right|_p^p

Parameters

• X (tf.Tensor) – Tensor to be regularized

• p (int) – p-norm

• axis (int or tuple) – along what axis

Returns

constraint term

Return type

tf.Tensor

KGE.constraint.clip_constraint(X, p, value, axis)

Clip embeddings

If X’s p-norm exceeds value, clip the value that let p-norm of X equals value.

Parameters

• X (tf.Tensor) – Tensor to be constraint

• p (int) – p-norm

• value (float) – restrict value

• axis (int or tuple) – along what axis

Returns

constraint tensor with same shape as X

Return type

tf.Tensor

KGE.constraint.normalized_embeddings(X, p, value, axis)

Normalized embeddings

Normalized X into p-norm equals value.

Parameters

• X (tf.Tensor) – Tensor to be normalized

• p (int) – p-norm

• value (float) – restrict value

• axis (int or tuple) – along what axis

Returns

normalized tensor with same shape as X

Return type

tf.Tensor

KGE.constraint.soft_constraint(X, p, value, axis)

Soft constraint

Soft constraint that described in TransH:

\[regularization ~ term = \sum \left[ \left\| \textbf{X} \right\|_p^2 - value \right]_+\]

where \([x]_+ = max(0,x)\)

Parameters

• X (tf.Tensor) – Tensor to be constraint

• p (int) – p-norm

• value (float) – restrict value

• axis (int or tuple) – along what axis

Returns

regularization term

Return type

tf.Tensor