mirror of
https://github.com/explosion/spaCy.git
synced 2024-12-27 10:26:35 +03:00
bede11b67c
This patch does a few smallish things that tighten up the training workflow a little, and allow memory use during training to be reduced by letting the GoldCorpus stream data properly. Previously, the parser and entity recognizer read and saved labels as lists, with extra labels noted separately. Lists were used becaue ordering is very important, to ensure that the label-to-class mapping is stable. We now manage labels as nested dictionaries, first keyed by the action, and then keyed by the label. Values are frequencies. The trick is, how do we save new labels? We need to make sure we iterate over these in the same order they're added. Otherwise, we'll get different class IDs, and the model's predictions won't make sense. To allow stable sorting, we map the new labels to negative values. If we have two new labels, they'll be noted as having "frequency" -1 and -2. The next new label will then have "frequency" -3. When we sort by (frequency, label), we then get a stable sort. Storing frequencies then allows us to make the next nice improvement. Previously we had to iterate over the whole training set, to pre-process it for the deprojectivisation. This led to storing the whole training set in memory. This was most of the required memory during training. To prevent this, we now store the frequencies as we stream in the data, and deprojectivize as we go. Once we've built the frequencies, we can then apply a frequency cut-off when we decide how many classes to make. Finally, to allow proper data streaming, we also have to have some way of shuffling the iterator. This is awkward if the training files have multiple documents in them. To solve this, the GoldCorpus class now writes the training data to disk in msgpack files, one per document. We can then shuffle the data by shuffling the paths. This is a squash merge, as I made a lot of very small commits. Individual commit messages below. * Simplify label management for TransitionSystem and its subclasses * Fix serialization for new label handling format in parser * Simplify and improve GoldCorpus class. Reduce memory use, write to temp dir * Set actions in transition system * Require thinc 6.11.1.dev4 * Fix error in parser init * Add unicode declaration * Fix unicode declaration * Update textcat test * Try to get model training on less memory * Print json loc for now * Try rapidjson to reduce memory use * Remove rapidjson requirement * Try rapidjson for reduced mem usage * Handle None heads when projectivising * Stream json docs * Fix train script * Handle projectivity in GoldParse * Fix projectivity handling * Add minibatch_by_words util from ud_train * Minibatch by number of words in spacy.cli.train * Move minibatch_by_words util to spacy.util * Fix label handling * More hacking at label management in parser * Fix encoding in msgpack serialization in GoldParse * Adjust batch sizes in parser training * Fix minibatch_by_words * Add merge_subtokens function to pipeline.pyx * Register merge_subtokens factory * Restore use of msgpack tmp directory * Use minibatch-by-words in train * Handle retokenization in scorer * Change back-off approach for missing labels. Use 'dep' label * Update NER for new label management * Set NER tags for over-segmented words * Fix label alignment in gold * Fix label back-off for infrequent labels * Fix int type in labels dict key * Fix int type in labels dict key * Update feature definition for 8 feature set * Update ud-train script for new label stuff * Fix json streamer * Print the line number if conll eval fails * Update children and sentence boundaries after deprojectivisation * Export set_children_from_heads from doc.pxd * Render parses during UD training * Remove print statement * Require thinc 6.11.1.dev6. Try adding wheel as install_requires * Set different dev version, to flush pip cache * Update thinc version * Update GoldCorpus docs * Remove print statements * Fix formatting and links [ci skip]
204 lines
6.8 KiB
Cython
204 lines
6.8 KiB
Cython
# cython: infer_types=True
|
|
# coding: utf-8
|
|
from __future__ import unicode_literals
|
|
|
|
from cpython.ref cimport Py_INCREF
|
|
from cymem.cymem cimport Pool
|
|
from thinc.typedefs cimport weight_t
|
|
from collections import OrderedDict
|
|
import ujson
|
|
|
|
from ..structs cimport TokenC
|
|
from .stateclass cimport StateClass
|
|
from ..typedefs cimport attr_t
|
|
from ..compat import json_dumps
|
|
from .. import util
|
|
|
|
|
|
cdef weight_t MIN_SCORE = -90000
|
|
|
|
|
|
class OracleError(Exception):
|
|
pass
|
|
|
|
|
|
cdef void* _init_state(Pool mem, int length, void* tokens) except NULL:
|
|
cdef StateC* st = new StateC(<const TokenC*>tokens, length)
|
|
return <void*>st
|
|
|
|
|
|
cdef class TransitionSystem:
|
|
def __init__(self, StringStore string_table, labels_by_action):
|
|
self.mem = Pool()
|
|
self.strings = string_table
|
|
self.n_moves = 0
|
|
self._size = 100
|
|
|
|
self.c = <Transition*>self.mem.alloc(self._size, sizeof(Transition))
|
|
|
|
for action, label_strs in labels_by_action.items():
|
|
for label_str in label_strs:
|
|
self.add_action(int(action), label_str)
|
|
self.root_label = self.strings.add('ROOT')
|
|
self.init_beam_state = _init_state
|
|
|
|
def __reduce__(self):
|
|
labels_by_action = OrderedDict()
|
|
cdef Transition t
|
|
for trans in self.c[:self.n_moves]:
|
|
label_str = self.strings[trans.label]
|
|
labels_by_action.setdefault(trans.move, []).append(label_str)
|
|
return (self.__class__,
|
|
(self.strings, labels_by_action),
|
|
None, None)
|
|
|
|
def init_batch(self, docs):
|
|
cdef StateClass state
|
|
states = []
|
|
offset = 0
|
|
for doc in docs:
|
|
state = StateClass(doc, offset=offset)
|
|
self.initialize_state(state.c)
|
|
states.append(state)
|
|
offset += len(doc)
|
|
return states
|
|
|
|
def get_oracle_sequence(self, doc, GoldParse gold):
|
|
cdef Pool mem = Pool()
|
|
costs = <float*>mem.alloc(self.n_moves, sizeof(float))
|
|
is_valid = <int*>mem.alloc(self.n_moves, sizeof(int))
|
|
|
|
cdef StateClass state = StateClass(doc, offset=0)
|
|
self.initialize_state(state.c)
|
|
history = []
|
|
while not state.is_final():
|
|
self.set_costs(is_valid, costs, state, gold)
|
|
for i in range(self.n_moves):
|
|
if is_valid[i] and costs[i] <= 0:
|
|
action = self.c[i]
|
|
history.append(i)
|
|
action.do(state.c, action.label)
|
|
break
|
|
else:
|
|
print(gold.words)
|
|
print(gold.ner)
|
|
print(history)
|
|
raise ValueError("Could not find gold move")
|
|
return history
|
|
|
|
cdef int initialize_state(self, StateC* state) nogil:
|
|
pass
|
|
|
|
cdef int finalize_state(self, StateC* state) nogil:
|
|
pass
|
|
|
|
def finalize_doc(self, doc):
|
|
pass
|
|
|
|
def preprocess_gold(self, GoldParse gold):
|
|
raise NotImplementedError
|
|
|
|
def is_gold_parse(self, StateClass state, GoldParse gold):
|
|
raise NotImplementedError
|
|
|
|
cdef Transition lookup_transition(self, object name) except *:
|
|
raise NotImplementedError
|
|
|
|
cdef Transition init_transition(self, int clas, int move, attr_t label) except *:
|
|
raise NotImplementedError
|
|
|
|
def is_valid(self, StateClass stcls, move_name):
|
|
action = self.lookup_transition(move_name)
|
|
if action.move == 0:
|
|
return False
|
|
return action.is_valid(stcls.c, action.label)
|
|
|
|
cdef int set_valid(self, int* is_valid, const StateC* st) nogil:
|
|
cdef int i
|
|
for i in range(self.n_moves):
|
|
is_valid[i] = self.c[i].is_valid(st, self.c[i].label)
|
|
|
|
cdef int set_costs(self, int* is_valid, weight_t* costs,
|
|
StateClass stcls, GoldParse gold) except -1:
|
|
cdef int i
|
|
self.set_valid(is_valid, stcls.c)
|
|
cdef int n_gold = 0
|
|
for i in range(self.n_moves):
|
|
if is_valid[i]:
|
|
costs[i] = self.c[i].get_cost(stcls, &gold.c, self.c[i].label)
|
|
n_gold += costs[i] <= 0
|
|
else:
|
|
costs[i] = 9000
|
|
if n_gold <= 0:
|
|
print(gold.words)
|
|
print(gold.ner)
|
|
print([gold.c.ner[i].clas for i in range(gold.length)])
|
|
print([gold.c.ner[i].move for i in range(gold.length)])
|
|
print([gold.c.ner[i].label for i in range(gold.length)])
|
|
print("Self labels",
|
|
[self.c[i].label for i in range(self.n_moves)])
|
|
raise ValueError(
|
|
"Could not find a gold-standard action to supervise "
|
|
"the entity recognizer. The transition system has "
|
|
"%d actions." % (self.n_moves))
|
|
|
|
def get_class_name(self, int clas):
|
|
act = self.c[clas]
|
|
return self.move_name(act.move, act.label)
|
|
|
|
def add_action(self, int action, label_name):
|
|
cdef attr_t label_id
|
|
if not isinstance(label_name, int) and \
|
|
not isinstance(label_name, long):
|
|
label_id = self.strings.add(label_name)
|
|
else:
|
|
label_id = label_name
|
|
# Check we're not creating a move we already have, so that this is
|
|
# idempotent
|
|
for trans in self.c[:self.n_moves]:
|
|
if trans.move == action and trans.label == label_id:
|
|
return 0
|
|
if self.n_moves >= self._size:
|
|
self._size *= 2
|
|
self.c = <Transition*>self.mem.realloc(self.c, self._size * sizeof(self.c[0]))
|
|
self.c[self.n_moves] = self.init_transition(self.n_moves, action, label_id)
|
|
assert self.c[self.n_moves].label == label_id
|
|
self.n_moves += 1
|
|
return 1
|
|
|
|
def to_disk(self, path, **exclude):
|
|
with path.open('wb') as file_:
|
|
file_.write(self.to_bytes(**exclude))
|
|
|
|
def from_disk(self, path, **exclude):
|
|
with path.open('rb') as file_:
|
|
byte_data = file_.read()
|
|
self.from_bytes(byte_data, **exclude)
|
|
return self
|
|
|
|
def to_bytes(self, **exclude):
|
|
transitions = []
|
|
for trans in self.c[:self.n_moves]:
|
|
transitions.append({
|
|
'clas': trans.clas,
|
|
'move': trans.move,
|
|
'label': self.strings[trans.label],
|
|
'name': self.move_name(trans.move, trans.label)
|
|
})
|
|
serializers = {
|
|
'transitions': lambda: json_dumps(transitions),
|
|
'strings': lambda: self.strings.to_bytes()
|
|
}
|
|
return util.to_bytes(serializers, exclude)
|
|
|
|
def from_bytes(self, bytes_data, **exclude):
|
|
transitions = []
|
|
deserializers = {
|
|
'transitions': lambda b: transitions.extend(ujson.loads(b)),
|
|
'strings': lambda b: self.strings.from_bytes(b)
|
|
}
|
|
msg = util.from_bytes(bytes_data, deserializers, exclude)
|
|
for trans in transitions:
|
|
self.add_action(trans['move'], trans['label'])
|
|
return self
|