Similar to how vectors are handled, move the vocab lookups to be loaded
at the start of training rather than when the vocab is initialized,
since the vocab doesn't have access to the full config when it's
created.
The option moves from `nlp.load_vocab_data` to `training.lookups`.
Typically these tables will come from `spacy-lookups-data`, but any
`Lookups` object can be provided.
The loading from `spacy-lookups-data` is now strict, so configs for each
language should specify the exact tables required. This also makes it
easier to control whether the larger clusters and probs tables are
included.
To load `lexeme_norm` from `spacy-lookups-data`:
```
[training.lookups]
@misc = "spacy.LoadLookupsData.v1"
lang = ${nlp.lang}
tables = ["lexeme_norm"]
```
In order to make it easier to construct `Doc` objects as training data,
modify how missing and blocked entity tokens are set to prioritize
setting `O` and missing entity tokens for training purposes over setting
blocked entity tokens.
* `Doc.ents` setter sets tokens outside entity spans to `O` regardless
of the current state of each token
* For `Doc.ents`, setting a span with a missing label sets the `ent_iob`
to missing instead of blocked
* `Doc.block_ents(spans)` marks spans as hard `O` for use with the
`EntityRecognizer`
* Refactor Docs.is_ flags
* Add derived `Doc.has_annotation` method
* `Doc.has_annotation(attr)` returns `True` for partial annotation
* `Doc.has_annotation(attr, require_complete=True)` returns `True` for
complete annotation
* Add deprecation warnings to `is_tagged`, `is_parsed`, `is_sentenced`
and `is_nered`
* Add `Doc._get_array_attrs()`, which returns a full list of `Doc` attrs
for use with `Doc.to_array`, `Doc.to_bytes` and `Doc.from_docs`. The
list is the `DocBin` attributes list plus `SPACY` and `LENGTH`.
Notes on `Doc.has_annotation`:
* `HEAD` is converted to `DEP` because heads don't have an unset state
* Accept `IS_SENT_START` as a synonym of `SENT_START`
Additional changes:
* Add `NORM`, `ENT_ID` and `SENT_START` to default attributes for
`DocBin`
* In `Doc.from_array()` the presence of `DEP` causes `HEAD` to override
`SENT_START`
* In `Doc.from_array()` using `attrs` other than
`Doc._get_array_attrs()` (i.e., a user's custom list rather than our
default internal list) with both `HEAD` and `SENT_START` shows a warning
that `HEAD` will override `SENT_START`
* `set_children_from_heads` does not require dependency labels to set
sentence boundaries and sets `sent_start` for all non-sentence starts to
`-1`
* Fix call to set_children_form_heads
Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
* Clean up spacy.tokens
* Update `set_children_from_heads`:
* Don't check `dep` when setting lr_* or sentence starts
* Set all non-sentence starts to `False`
* Use `set_children_from_heads` in `Token.head` setter
* Reduce similar/duplicate code (admittedly adds a bit of overhead)
* Update sentence starts consistently
* Remove unused `Doc.set_parse`
* Minor changes:
* Declare cython variables (to avoid cython warnings)
* Clean up imports
* Modify set_children_from_heads to set token range
Modify `set_children_from_heads` so that it adjust tokens within a
specified range rather then the whole document.
Modify the `Token.head` setter to adjust only the tokens affected by the
new head assignment.
For languages without provided models and with lemmatizer rules in
`spacy-lookups-data`, make the rule-based lemmatizer the default:
Bengali, Persian, Norwegian, Swedish
Modify `Token.morph` property so that `Token.c.morph` can be reset back
to an internal value of `0`. Allow setting `Token.morph` from a hash as
long as the morph string is already in the `StringStore`, setting it
indirectly through `Token.morph_` so that the value is added to the
morphology. If the hash is not in the `StringStore`, raise an error.
* raise error if no valid Example objects were found during initialization
* fix max_length parameter
* remove commit from other branch
Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
* ensure Language passes on valid examples for initialization
* fix tagger model initialization
* check for valid get_examples across components
* assume labels were added before begin_training
* fix senter initialization
* fix morphologizer initialization
* use methods to check arguments
* test textcat init, requires thinc>=8.0.0a31
* fix tok2vec init
* fix entity linker init
* use islice
* fix simple NER
* cleanup debug model
* fix assert statements
* fix tests
* throw error when adding a label if the output layer can't be resized anymore
* fix test
* add failing test for simple_ner
* UX improvements
* morphologizer UX
* assume begin_training gets a representative set and processes the labels
* remove assumptions for output of untrained NER model
* restore test for original purpose
This was the only registry that expected the registered objects to be dictionaries instead of functions that return something. We can still support plain dicts but we should also support functions for consistency
The Language.use_params method was failing if you passed in None, which
meant we had to use awkward conditionals for the parameter averaging.
This solves the problem.
Add official support for the `DependencyMatcher`. Redesign the pattern
specification. Fix and extend operator implementations. Update API docs
and add usage docs.
Patterns
--------
Refactor pattern structure to:
```
{
"LEFT_ID": str,
"REL_OP": str,
"RIGHT_ID": str,
"RIGHT_ATTRS": dict,
}
```
The first node contains only `RIGHT_ID` and `RIGHT_ATTRS` and all
subsequent nodes contain all four keys.
New operators
-------------
Because of the way patterns are constructed from left to right, it's
helpful to have `follows` operators along with `precedes` operators. Add
operators for simple precedes / follows alongside immediate precedes /
follows.
* `.*`: precedes
* `;`: immediately follows
* `;*`: follows
Operator fixes
--------------
* `<` and `<<` do not include the node itself
* Fix reversed order for all operators involving linear precedence (`.`,
all sibling operators)
* Linear precedence operators do not match nodes outside the same parse
Additional fixes
----------------
* Use v3 Matcher API
* Support `get` and `remove`
* Support pickling
Follow-ups to the parser efficiency fix.
* Avoid introducing new counter for number of pushes
* Base cut on number of transitions, keeping it more even
* Reintroduce the randomization we had in v2.
The parser training makes use of a trick for long documents, where we
use the oracle to cut up the document into sections, so that we can have
batch items in the middle of a document. For instance, if we have one
document of 600 words, we might make 6 states, starting at words 0, 100,
200, 300, 400 and 500.
The problem is for v3, I screwed this up and didn't stop parsing! So
instead of a batch of [100, 100, 100, 100, 100, 100], we'd have a batch
of [600, 500, 400, 300, 200, 100]. Oops.
The implementation here could probably be improved, it's annoying to
have this extra variable in the state. But this'll do.
This makes the v3 parser training 5-10 times faster, depending on document
lengths. This problem wasn't in v2.
A long time ago we went to some trouble to try to clean up "unused"
strings, to avoid the `StringStore` growing in long-running processes.
This never really worked reliably, and I think it was a really wrong
approach. It's much better to let the user reload the `nlp` object as
necessary, now that the string encoding is stable (in v1, the string IDs
were sequential integers, making reloading the NLP object really
annoying.)
The extra book-keeping does make some performance difference, and the
feature is unsed, so it's past time we killed it.
* Prevent Tagger model init with 0 labels
Raise an error before trying to initialize a tagger model with 0 labels.
* Add dummy tagger label for test
* Remove tagless tagger model initializiation
* Fix error number after merge
* Add dummy tagger label to test
* Fix formatting
Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
* rename to spacy-transformers.TransformerListener
* add some more tok2vec tests
* use select_pipes
* fix docs - annotation setter was not changed in the end
Sort the returned matches by rule order (the `match_id`) so that the
rules are applied in the order they were added. This is necessary, for
instance, if the `AttributeRuler` is used for the tag map and later
rules require POS tags.
Serialize `AttributeRuler.patterns` instead of the individual lists to
simplify the serialized and so that patterns are reloaded exactly as
they were originally provided (preserving `_attrs_unnormed`).
