View source on GitHub |
Samples a set of classes using the provided (fixed) base distribution.
tf.random.fixed_unigram_candidate_sampler(
true_classes,
num_true,
num_sampled,
unique,
range_max,
vocab_file='',
distortion=1.0,
num_reserved_ids=0,
num_shards=1,
shard=0,
unigrams=(),
seed=None,
name=None
)
This operation randomly samples a tensor of sampled classes
(sampled_candidates
) from the range of integers [0, range_max)
.
See the Candidate Sampling Algorithms Reference for a quick course on Candidate Sampling.
The elements of sampled_candidates
are drawn without replacement
(if unique=True
) or with replacement (if unique=False
) from
the base distribution.
The base distribution is read from a file or passed in as an in-memory array. There is also an option to skew the distribution by applying a distortion power to the weights.
In addition, this operation returns tensors true_expected_count
and sampled_expected_count
representing the number of times each
of the target classes (true_classes
) and the sampled
classes (sampled_candidates
) is expected to occur in an average
tensor of sampled classes. These values correspond to Q(y|x)
defined in the Candidate Sampling Algorithms
Reference.
If unique=True
, then these are post-rejection probabilities and we
compute them approximately.
Note that this function (and also other *_candidate_sampler
functions) only gives you the ingredients to implement the various
Candidate Sampling algorithms listed in the big table in the
Candidate Sampling Algorithms
Reference. You
still need to implement the algorithms yourself.
For example, according to that table, the phrase "negative samples"
may mean different things in different algorithms. For instance, in
NCE, "negative samples" means S_i
(which is just the sampled
classes) which may overlap with true classes, while in Sampled
Logistic, "negative samples" means S_i - T_i
which excludes the
true classes. The return value sampled_candidates
corresponds to
S_i
, not to any specific definition of "negative samples" in any
specific algorithm. It's your responsibility to pick an algorithm
and calculate the "negative samples" defined by that algorithm
(e.g. S_i - T_i
).
As another example, the true_classes
argument is for calculating
the true_expected_count
output (as a by-product of this function's
main calculation), which may be needed by some algorithms (according
to that table). It's not for excluding true classes in the return
value sampled_candidates
. Again that step is algorithm-specific
and should be carried out by you.