* Switch converters to generator functions
To reduce the memory usage when converting large corpora, refactor the
convert methods to be generator functions.
* Update tests
* Get basic beam tests working
* Get basic beam tests working
* Compile _beam_utils
* Remove prints
* Test beam density
* Beam parser seems to train
* Draft beam NER
* Upd beam
* Add hypothesis as dev dependency
* Implement missing is-gold-parse method
* Implement early update
* Fix state hashing
* Fix test
* Fix test
* Default to non-beam in parser constructor
* Improve oracle for beam
* Start refactoring beam
* Update test
* Refactor beam
* Update nn
* Refactor beam and weight by cost
* Update ner beam settings
* Update test
* Add __init__.pxd
* Upd test
* Fix test
* Upd test
* Fix test
* Remove ring buffer history from StateC
* WIP change arc-eager transitions
* Add state tests
* Support ternary sent start values
* Fix arc eager
* Fix NER
* Pass oracle cut size for beam
* Fix ner test
* Fix beam
* Improve StateC.clone
* Improve StateClass.borrow
* Work directly with StateC, not StateClass
* Remove print statements
* Fix state copy
* Improve state class
* Refactor parser oracles
* Fix arc eager oracle
* Fix arc eager oracle
* Use a vector to implement the stack
* Refactor state data structure
* Fix alignment of sent start
* Add get_aligned_sent_starts method
* Add test for ae oracle when bad sentence starts
* Fix sentence segment handling
* Avoid Reduce that inserts illegal sentence
* Update preset SBD test
* Fix test
* Remove prints
* Fix sent starts in Example
* Improve python API of StateClass
* Tweak comments and debug output of arc eager
* Upd test
* Fix state test
* Fix state test
* add test for multi-label textcat reproducibility
* remove positive_label
* fix lengths dtype
* fix comments
* remove comment that we should not have forgotten :-)
Remove the non-working `--use-chars` option from the train CLI. The
implementation of the option across component types and the CLI settings
could be fixed, but the `CharacterEmbed` model does not work on GPU in
v2 so it's better to remove it.
* define new architectures for the pretraining objective
* add loss function as attr of the omdel
* cleanup
* cleanup
* shorten name
* fix typo
* remove unused error
Preserve `token.spacy` corresponding to the span end token in the
original doc rather than adjusting for the current offset.
* If not modifying in place, this checks in the original document
(`doc.c` rather than `tokens`).
* If modifying in place, the document has not been modified past the
current span start position so the value at the current span end
position is valid.
* When checking for token alignments, check not only that the tokens are
identical but that the character positions are both at the start of a
token.
It's possible for the tokens to be identical even though the two
tokens aren't aligned one-to-one in a case like `["a'", "''"]` vs.
`["a", "''", "'"]`, where the middle tokens are identical but should not
be aligned on the token level at character position 2 since it's the
start of one token but the middle of another.
* Use the lowercased version of the token texts to create the
character-to-token alignment because lowercasing can change the string
length (e.g., for `İ`, see the not-a-bug bug report:
https://bugs.python.org/issue34723)
* Only set NORM on Token in retokenizer
Instead of setting `NORM` on both the token and lexeme, set `NORM` only
on the token.
The retokenizer tries to set all possible attributes with
`Token/Lexeme.set_struct_attr` so that it doesn't have to enumerate
which attributes are available for each. `NORM` is the only attribute
that's stored on both and for most cases it doesn't make sense to set
the global norms based on a individual retokenization. For lexeme-only
attributes like `IS_STOP` there's no way to avoid the global side
effects, but I think that `NORM` would be better only on the token.
* Fix test
