AdaptiveLogSoftmaxWithLoss(in_features, n_classes, cutoffs, div_value=4.0, head_bias=False)¶
Efficient softmax approximation as described in Efficient softmax approximation for GPUs by Edouard Grave, Armand Joulin, Moustapha Cissé, David Grangier, and Hervé Jégou.
Adaptive softmax is an approximate strategy for training models with large output spaces. It is most effective when the label distribution is highly imbalanced, for example in natural language modelling, where the word frequency distribution approximately follows the Zipf’s law.
Adaptive softmax partitions the labels into several clusters, according to their frequency. These clusters may contain different number of targets each. Additionally, clusters containing less frequent labels assign lower dimensional embeddings to those labels, which speeds up the computation. For each minibatch, only clusters for which at least one target is present are evaluated.
The idea is that the clusters which are accessed frequently (like the first one, containing most frequent labels), should also be cheap to compute – that is, contain a small number of assigned labels.
We highly recommend taking a look at the original paper for more details.
cutoffsshould be an ordered Sequence of integers sorted in the increasing order. It controls number of clusters and the partitioning of targets into clusters. For example setting
cutoffs = [10, 100, 1000]means that first 10 targets will be assigned to the ‘head’ of the adaptive softmax, targets 11, 12, …, 100 will be assigned to the first cluster, and targets 101, 102, …, 1000 will be assigned to the second cluster, while targets 1001, 1002, …, n_classes - 1 will be assigned to the last, third cluster.
div_valueis used to compute the size of each additional cluster, which is given as , where is the cluster index (with clusters for less frequent words having larger indices, and indices starting from ).
head_biasif set to True, adds a bias term to the ‘head’ of the adaptive softmax. See paper for details. Set to False in the official implementation.
Labels passed as inputs to this module should be sorted according to their frequency. This means that the most frequent label should be represented by the index 0, and the least frequent label should be represented by the index n_classes - 1.
This module returns a
lossfields. See further documentation for details.
To compute log-probabilities for all classes, the
log_probmethod can be used.
in_features (int) – Number of features in the input tensor
n_classes (int) – Number of classes in the dataset
cutoffs (Sequence) – Cutoffs used to assign targets to their buckets
div_value (float, optional) – value used as an exponent to compute sizes of the clusters. Default: 4.0
head_bias (bool, optional) – If
True, adds a bias term to the ‘head’ of the adaptive softmax. Default:
output is a Tensor of size
Ncontaining computed target log probabilities for each example
loss is a Scalar representing the computed negative log likelihood loss
- Return type
target: where each value satisfies
Computes log probabilities for all
input (Tensor) – a minibatch of examples
log-probabilities of for each class in range , where is a parameter passed to