NLLMulticlass

class ampligraph.latent_features.NLLMulticlass(loss_params={}, verbose=False)

Multiclass Negative Log-Likelihood loss.

Introduced in [aC15], this loss can be used when both the subject and objects are corrupted (to use it, pass corrupt_sides=['s,o'] in the embedding model parameters).

This loss was re-engineered in [KBK17] where only the object was corrupted to get improved performance (to use it in this way pass corrupt_sides ='o' in the embedding model parameters).

\[\mathcal{L(X)} = -\sum_{x_{e_1,e_2,r_k} \in X} log\,p(e_2|e_1,r_k) -\sum_{x_{e_1,e_2,r_k} \in X} log\,p(e_1|r_k, e_2)\]

Example

>>> import ampligraph.latent_features.loss_functions as lfs
>>> loss = lfs.NLLMulticlass({'reduction': 'mean'})
>>> isinstance(loss, lfs.NLLMulticlass)
True

>>> loss = lfs.get('multiclass_nll')
>>> isinstance(loss, lfs.NLLMulticlass)
True

Attributes

external_params
name

Methods

__init__([loss_params, verbose])

Initialize the loss.

__init__(loss_params={}, verbose=False)

Initialize the loss.

Parameters:

loss_params (dict) –

Dictionary of loss-specific hyperparams:

• ”reduction”: (str) - Specifies whether to “sum” or take the “mean” of loss per sample w.r.t. corruption (default: “sum”).