Vocabulary(method='indexing', use_pad=True, use_unk=True, special_tokens=None, do_counting=True, pad_value=None, unk_value=None)¶
This class will store “Elements” that are added, assign “Ids” to them and return “Representations” if queried. These three are the main concepts in this class.
Element: Any hash-able instance that the user want to store.
Id: Each element will have an unique Id, which is an integer.
Representation: according to the configuration, the representation for an element could be an integer (in this case, would be “Id”), or an one-hot vector (in this case, would be a list of integer).
The class adopts the special elements from Texar-Pytorch, which are:
<PAD>: which will be mapped into Id of 0 or -1 and have different representation according to different setting.
<UNK>: if added into the vocabulary, will be the default element if the queried element is not found.
Note that these two special tokens are necessary for the system in certain cases and thus must present in the vocabulary. The behavior of these special tokens are pre-defined based on different settings. To get around the default behavior (for example, if you have a pre-defined vocabulary with different setups), you can instruct the class to not adding these tokens automatically, and use the
Here is a table on how our Vocabulary class behavior under different settings. Element0 means the first element that is added to the vocabulary. Elements added later will be element1, element2 and so on. They will follow the same behavior as element0. For readability, they are not listed in the table.
custom (handle and implemented by the user)
-1 -> <PAD> 0->element0 (be careful)
method (str) – The method to represent element in vocabulary, currently supporting “indexing” and “one-hot”.
use_pad (bool) – Whether to add <PAD> element to the vocabulary on creation. It will be added to the vocabulary first, but the id of it depends on the specific settings.
use_unk (bool) – Whether to add <UNK> element to the vocabulary on creation. Elements that are not found in vocabulary will be directed to <UNK> element. It will be added right after the <PAD> element if provided.
special_tokens (List[str]) – Additional special tokens to be added, they will be added at the beginning of vocabulary (but right after the <UNK> token) one by one.
do_counting (bool) – Whether the vocabulary class will count the elements.
pad_value – A customized value/representation to be used for padding, for example, following the PyTorch convention you may want to use -100. This value is only needed when use_pad is True. Default is None, where the value of padding is determined by the system.
unk_value – A customized value/representation to be used for unknown value (unk). This value is only needed when use_unk is True. Default is None, where the value of UNK is determined by the system.
Get the counts of the vocabulary element.
e – The element to get counts for. It can be the element id or the element’s raw type.
The count of the element.
mark_special_element(element_id, element_name, representation=None)¶
Mark a particular (but already existed) index in the vocabulary to be a special required element (i.e PAD or UNK).
element_id (int) – The id to be set for the special element.
element_name (str) – The name of this element to be set, it can be one of PAD, UNK.
representation – The representation/value that this element should be assigned. Default is None, then its representation will be computed from the internal indexing.
Check whether the element is a special token.
add_special_element(element, element_id=None, representation=None, special_token_name=None)¶
This function will add special elements to the vocabulary, such as UNK, PAD, BOS, CLS symbols. Some special tokens will not be filtered by any VocabFilter. Some special tokens has their unique behavior in the system.
Note: most of the time, you don’t have to call this method yourself, but should let the init function to handle that.
element (str) – The surface form of this special element.
element_id (Optional[int]) – The to be used for this special token. If not provided, the vocabulary will use the next id internally. If the provided id is occupied, a ValueError will be thrown. The id can be any integer, including negative ones.
representation – The representation you want to assign to this special token. If None, the representation may be computed based on the index (which depends on the vocabulary setting).
special_token_name (Optional[str]) – An internal name of this special token. This only matters for the base special tokens: <PAD> or <UNK>, and the name should be “PAD” and “UNK” respectively. Any other name here is considered invalid, and a ValueError will be thrown if provided.
add_element(element, representation=None, count=1)¶
This function will add a regular element to the vocabulary.
element (Hashable) – The element to be added.
representation – The vocabulary representation of this element will use this value. For example, you may want to use -100 for ignored tokens for PyTorch skipped tokens. Note that the class do not check whether this representation is used by another element, so the caller have to manage the behavior itself.
count (int) – the count to be incremented for this element, default is 1 (i.e. consider it appear once on every add). This value will have effect only if do_counting is True.
The internal id of the element.
This function will map id to element.
This function will map element to representation.
element (Hashable) – The queried element. It can be either the same type as the element, or string (for the special tokens).
- The corresponding representation
of the element. Check the behavior of this function under different setting in the documentation.
- Return type
KeyError – If element is not found and vocabulary does not use <UNK> element.
Create a dictionary from the vocabulary storing all the known elements.
- Returns: The vocabulary as a Dict from ElementType to the
representation of the element (could be Integer or One-hot vector, depending on the settings of this class).
This function checks whether an element is added to vocabulary.
element (Hashable) – The queried element.
Whether element is found.
- Return type
This function will loop over the (element, id) pair inside this class.
Iterables of (element, id) pair.
- Return type
This function will get the representation of the PAD element for the vocabulary. The representation depends on the settings of this class, it can be an integer or a list of int (e.g. a vector).
This function will create a new vocabulary object, which is based on the current vocabulary, but filter out elements that appear fewer times than the min_count value. Calling this function will cause a full iteration over the vocabulary, thus normally, it should be called after collecting all the vocabulary in the dataset.
vocab_filter – The filter used to filter the vocabulary.
A new vocabulary after filtering.