This change changes the type of left/right-arc collections from
vector[ArcC] to unordered_map[int, vector[Arc]], so that the arcs are
keyed by the head. This allows us to find all the left/right arcs for a
particular head in constant time in StateC::{L,R}.
Benchmarks with long docs (N is the number of text repetitions):
Before (using #10019):
N Time (s)
400 3.2
800 5.0
1600 9.5
3200 23.2
6400 66.8
12800 220.0
After (this commit):
N Time (s)
400 3.1
800 4.3
1600 6.7
3200 12.0
6400 22.0
12800 42.0
Related to #9858 and #10019.
* Speed up the StateC::L feature function
This function gets the n-th most-recent left-arc with a particular head.
Before this change, StateC::L would construct a vector of all left-arcs
with the given head and then pick the n-th most recent from that vector.
Since the number of left-arcs strongly correlates with the doc length
and the feature is constructed for every transition, this can make
transition-parsing quadratic.
With this change StateC::L:
- Searches left-arcs backwards.
- Stops early when the n-th matching transition is found.
- Does not construct a vector (reducing memory pressure).
This change doesn't avoid the linear search when the transition that is
queried does not occur in the left-arcs. Regardless, performance is
improved quite a bit with very long docs:
Before:
N Time
400 3.3
800 5.4
1600 11.6
3200 30.7
After:
N Time
400 3.2
800 5.0
1600 9.5
3200 23.2
We can probably do better with more tailored data structures, but I
first wanted to make a low-impact PR.
Found while investigating #9858.
* StateC::L: simplify loop
* Add test for #7035
* Update test for issue 7056
* Fix test
* Fix transitions method used in testing
* Fix state eol detection when rebuffer
* Clean up redundant fix
* 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
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.
* moving syntax folder to _parser_internals
* moving nn_parser and transition_system
* move nn_parser and transition_system out of internals folder
* moving nn_parser code into transition_system file
* rename transition_system to transition_parser
* moving parser_model and _state to ml
* move _state back to internals
* The Parser now inherits from Pipe!
* small code fixes
* removing unnecessary imports
* remove link_vectors_to_models
* transition_system to internals folder
* little bit more cleanup
* newlines
