tf.nn.sampled_softmax_loss

Computes and returns the sampled softmax training loss.

tf.nn.sampled_softmax_loss(
    weights,
    biases,
    labels,
    inputs,
    num_sampled,
    num_classes,
    num_true=1,
    sampled_values=None,
    remove_accidental_hits=True,
    seed=None,
    name='sampled_softmax_loss'
)

This is a faster way to train a softmax classifier over a huge number of classes.

This operation is for training only. It is generally an underestimate of the full softmax loss.

A common use case is to use this method for training, and calculate the full softmax loss for evaluation or inference as in the following example:

if mode == "train":
  loss = tf.nn.sampled_softmax_loss(
      weights=weights,
      biases=biases,
      labels=labels,
      inputs=inputs,
      ...)
elif mode == "eval":
  logits = tf.matmul(inputs, tf.transpose(weights))
  logits = tf.nn.bias_add(logits, biases)
  labels_one_hot = tf.one_hot(labels, n_classes)
  loss = tf.nn.softmax_cross_entropy_with_logits(
      labels=labels_one_hot,
      logits=logits)

See our Candidate Sampling Algorithms Reference

Also see Section 3 of Jean et al., 2014 (pdf) for the math.

weightsA Tensor of shape [num_classes, dim], or a list of Tensor objects whose concatenation along dimension 0 has shape [num_classes, dim]. The (possibly-sharded) class embeddings.
biasesA Tensor of shape [num_classes]. The class biases.
labelsA Tensor of type int64 and shape [batch_size, num_true]. The target classes. Note that this format differs from the labels argument of nn.softmax_cross_entropy_with_logits.
inputsA Tensor of shape [batch_size, dim]. The forward activations of the input network.
num_sampledAn int. The number of classes to randomly sample per batch.
num_classesAn int. The number of possible classes.
num_trueAn int. The number of target classes per training example.
sampled_valuesa tuple of (sampled_candidates, true_expected_count, sampled_expected_count) returned by a *_candidate_sampler function. (if None, we default to log_uniform_candidate_sampler)
remove_accidental_hitsA bool. whether to remove "accidental hits" where a sampled class equals one of the target classes. Default is True.
seedrandom seed for candidate sampling. Default to None, which doesn't set the op-level random seed for candidate sampling.
nameA name for the operation (optional).

A batch_size 1-D tensor of per-example sampled softmax losses.