tft.vocabulary( x, top_k=None, frequency_threshold=None, vocab_filename=None, store_frequency=False, weights=None, labels=None, use_adjusted_mutual_info=False, min_diff_from_avg=None, coverage_top_k=None, coverage_frequency_threshold=None, key_fn=None, fingerprint_shuffle=False, name=None )
Computes the unique values of a
Tensor over the whole dataset.
Computes The unique values taken by
x, which can be a
SparseTensor of any size. The unique values will be aggregated over all
x and all instances.
In case one of the tokens contains the '\n' or '\r' characters or is empty it will be discarded since we are currently writing the vocabularies as text files. This behavior will likely be fixed/improved in the future.
If an integer
Tensor is provided, its semantic type should be categorical
not a continuous/numeric, since computing a vocabulary over a continuous
feature is not appropriate.
The unique values are sorted by decreasing frequency and then reverse lexicographical order (e.g. [('a', 5), ('c', 3), ('b', 3)]).
For large datasets it is highly recommended to either set frequency_threshold or top_k to control the size of the output, and also the run time of this operation.
When labels are provided, we filter the vocabulary based on the relationship between the token's presence in a record and the label for that record, using (possibly adjusted) Mutual Information. Note: If labels are provided, the x input must be a unique set of per record, as the semantics of the mutual information calculation depend on a multi-hot representation of the input. Having unique input tokens per row is advisable but not required for a frequency-based vocabulary.
WARNING: The following is experimental and is still being actively worked on.
key_fn if you would like to generate a vocabulary with coverage over
A "coverage vocabulary" is the union of two vocabulary "arms". The "standard arm" of the vocabulary is equivalent to the one generated by the same function call with no coverage arguments. Adding coverage only appends additional entries to the end of the standard vocabulary.
The "coverage arm" of the vocabulary is determined by taking the
coverage_top_k most frequent unique terms per key. A term's key is obtained
key_fn to the term. Use
coverage_frequency_threshold to lower
bound the frequency of entries in the coverage arm of the vocabulary.
Note this is currently implemented for the case where the key is contained within each vocabulary entry (b/117796748).
x: A categorical/discrete input
SparseTensorwith dtype tf.string or tf.int[8|16|32|64]. The inputs should generally be unique per row (i.e. a bag of words/ngrams representation).
top_k: Limit the generated vocabulary to the first
top_kelements. If set to None, the full vocabulary is generated.
frequency_threshold: Limit the generated vocabulary only to elements whose absolute frequency is >= to the supplied threshold. If set to None, the full vocabulary is generated. Absolute frequency means the number of occurrences of the element in the dataset, as opposed to the proportion of instances that contain that element.
vocab_filename: The file name for the vocabulary file. If none, the "uniques" scope name in the context of this graph will be used as the file name. If not None, should be unique within a given preprocessing function. NOTE To make your pipelines resilient to implementation details please set
vocab_filenamewhen you are using the vocab_filename on a downstream component.
store_frequency: If True, frequency of the words is stored in the vocabulary file. In the case labels are provided, the mutual information is stored in the file instead. Each line in the file will be of the form 'frequency word'. NOTE: if this is True then the computed vocabulary cannot be used with
tft.apply_vocabularydirectly, since frequencies are added to the beginning of each row of the vocabulary, which the mapper will not ignore.
weights: (Optional) Weights
Tensorfor the vocabulary. It must have the same shape as x.
labels: (Optional) Labels dense
Tensorfor the vocabulary. If provided, the vocabulary is calculated based on mutual information with the label, rather than frequency. The labels must have the same batch dimension as x. If x is sparse, labels should be a 1D tensor reflecting row-wise labels. If x is dense, labels can either be a 1D tensor of row-wise labels, or a dense tensor of the identical shape as x (i.e. element-wise labels). Labels should be a discrete integerized tensor (If the label is numeric, it should first be bucketized; If the label is a string, an integer vocabulary should first be applied). Note:
SparseTensorlabels are not yet supported (b/134931826).
use_adjusted_mutual_info: If true, and labels are provided, calculate vocabulary using adjusted rather than raw mutual information.
min_diff_from_avg: MI (or AMI) of a feature x label will be adjusted to zero whenever the difference between count and the expected (average) count is lower than min_diff_from_average. This can be thought of as a regularizing parameter that pushes small MI/AMI values to zero. If None, a default parameter will be selected based on the size of the dataset (see calculate_recommended_min_diff_from_avg).
coverage_top_k: (Optional), (Experimental) The minimum number of elements per key to be included in the vocabulary.
coverage_frequency_threshold: (Optional), (Experimental) Limit the coverage arm of the vocabulary only to elements whose absolute frequency is >= this threshold for a given key.
key_fn: (Optional), (Experimental) A fn that takes in a single entry of
xand returns the corresponding key for coverage calculation. If this is
None, no coverage arm is added to the vocabulary.
fingerprint_shuffle: (Optional), (Experimental) Whether to sort the vocabularies by fingerprint instead of counts. This is useful for load balancing on the training parameter servers. Shuffle only happens while writing the files, so all the filters above (top_k, frequency_threshold, etc) will still take effect.
name: (Optional) A name for this operation.
The path name for the vocabulary file containing the unique values of
frequency_thresholdis negative. If
coverage_frequency_thresholdis negative. If either
coverage_frequency_thresholdis specified and
key_fnis not. If
key_fnis specified and neither