explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2024-11-10 19:57:17 +03:00
Code Issues Packages Projects Releases Wiki Activity

* Prepare for new models to be plugged in by using Example class

Browse Source
This commit is contained in:
Matthew Honnibal 2015-06-26 13:51:39 +02:00
parent 75aeccc064
commit 9282a8e72c
5 changed files with 56 additions and 46 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
setup.py
View File

@ -151,7 +151,8 @@ MOD_NAMES = ['spacy.parts_of_speech', 'spacy.strings',
'spacy.lexeme', 'spacy.vocab', 'spacy.tokens', 'spacy.spans',
'spacy.morphology',
'spacy.syntax.stateclass',
'spacy._ml', 'spacy.tokenizer', 'spacy.en.attrs',
'spacy._ml', 'spacy._theano',
'spacy.tokenizer', 'spacy.en.attrs',
'spacy.en.pos', 'spacy.syntax.parser',
'spacy.syntax.transition_system',
'spacy.syntax.arc_eager',

5
spacy/_ml.pxd
View File

@ -14,9 +14,14 @@ from .tokens cimport Tokens
cdef int arg_max(const weight_t* scores, const int n_classes) nogil
cdef int arg_max_if_true(const weight_t* scores, const int* is_valid, int n_classes) nogil
cdef int arg_max_if_zero(const weight_t* scores, const int* costs, int n_classes) nogil
cdef class Model:
cdef int n_classes
cdef int n_feats
cdef const weight_t* score(self, atom_t* context) except NULL
cdef int set_scores(self, weight_t* scores, atom_t* context) except -1

18
spacy/_ml.pyx
View File

@ -24,7 +24,7 @@ cdef int arg_max(const weight_t* scores, const int n_classes) nogil:
return best
cdef int arg_max_if_true(const weight_t* scores, const bint* is_valid,
cdef int arg_max_if_true(const weight_t* scores, const int* is_valid,
const int n_classes) nogil:
cdef int i
cdef int best = 0
@ -54,21 +54,25 @@ cdef class Model:
model_loc = path.join(model_loc, 'model')
self.n_classes = n_classes
self._extractor = Extractor(templates)
self.n_feats = self._extractor.n_templ
self._model = LinearModel(n_classes, self._extractor.n_templ)
self.model_loc = model_loc
if self.model_loc and path.exists(self.model_loc):
self._model.load(self.model_loc, freq_thresh=0)
def predict(self, Example eg):
self.set_scores(eg.scores, eg.atoms)
eg.guess = arg_max_if_true(eg.scores, eg.is_valid, self.n_classes)
self.set_scores(<weight_t*>eg.scores.data, <atom_t*>eg.atoms.data)
eg.guess = arg_max_if_true(<weight_t*>eg.scores.data, <int*>eg.is_valid.data,
self.n_classes)
def train(self, Example eg):
self.set_scores(eg.scores, eg.atoms)
eg.guess = arg_max_if_true(eg.scores, eg.is_valid, self.n_classes)
eg.best = arg_max_if_zero(eg.scores, eg.costs, self.n_classes)
self.set_scores(<weight_t*>eg.scores.data, <atom_t*>eg.atoms.data)
eg.guess = arg_max_if_true(<weight_t*>eg.scores.data,
<int*>eg.is_valid.data, self.n_classes)
eg.best = arg_max_if_zero(<weight_t*>eg.scores.data, <int*>eg.costs.data,
self.n_classes)
eg.cost = eg.costs[eg.guess]
self.update(eg.atoms, eg.guess, eg.best, eg.cost)
self.update(<atom_t*>eg.atoms.data, eg.guess, eg.best, eg.cost)
cdef const weight_t* score(self, atom_t* context) except NULL:
cdef int n_feats

64
spacy/_theano.pyx
View File

@ -1,44 +1,44 @@
from thinc.example cimport Example
from thinc.api cimport Example
from thinc.typedefs cimport weight_t
from ._ml cimport arg_max_if_true
from ._ml cimport arg_max_if_zero
import numpy
from os import path
cdef class TheanoModel(Model):
def __init__(self, n_classes, input_layer, train_func, predict_func, model_loc=None):
def __init__(self, n_classes, input_spec, train_func, predict_func, model_loc=None):
if model_loc is not None and path.isdir(model_loc):
model_loc = path.join(model_loc, 'model')
self.n_classes = n_classes
tables = []
lengths = []
for window_size, n_dims, vocab_size in input_structure:
tables.append(EmbeddingTable(n_dims, vocab_size, initializer))
lengths.append(window_size)
self.input_layer = InputLayer(lengths, tables)
self.eta = 0.001
self.mu = 0.9
self.t = 1
initializer = lambda: 0.2 * numpy.random.uniform(-1.0, 1.0)
self.input_layer = InputLayer(input_spec, initializer)
self.train_func = train_func
self.predict_func = predict_func
self.n_classes = n_classes
self.n_feats = len(self.input_layer)
self.model_loc = model_loc
if self.model_loc and path.exists(self.model_loc):
self._model.load(self.model_loc, freq_thresh=0)
def train(self, Instance eg):
pass
def predict(self, Instance eg):
cdef const weight_t* score(self, atom_t* context) except NULL:
self.set_scores(self._scores, context)
return self._scores
cdef int set_scores(self, weight_t* scores, atom_t* context) except -1:
# TODO f(context) --> Values
self._input_layer.fill(self._x, self._values, use_avg=False)
theano_scores = self._predict(self._x)
def predict(self, Example eg):
self.input_layer.fill(eg.embeddings, eg.atoms)
theano_scores = self.predict_func(eg.embeddings)
cdef int i
for i in range(self.n_classes):
output[i] = theano_scores[i]
eg.scores[i] = theano_scores[i]
eg.guess = arg_max_if_true(<weight_t*>eg.scores.data, <int*>eg.is_valid.data,
self.n_classes)
cdef int update(self, atom_t* context, class_t guess, class_t gold, int cost) except -1:
# TODO f(context) --> Values
self._input_layer.fill(self._x, self._values, use_avg=False)
def train(self, Example eg):
self.predict(eg)
update, t, eta, mu = self.train_func(eg.embeddings, eg.scores, eg.costs)
self.input_layer.update(eg.atoms, update, self.t, self.eta, self.mu)
eg.best = arg_max_if_zero(<weight_t*>eg.scores.data, <int*>eg.costs.data,
self.n_classes)
eg.cost = eg.costs[eg.guess]
self.t += 1

12
spacy/syntax/parser.pyx
View File

@ -68,11 +68,11 @@ cdef class Parser:
cdef StateClass stcls = StateClass.init(tokens.data, tokens.length)
self.moves.initialize_state(stcls)
cdef Example eg = Example(self.model.n_classes, CONTEXT_SIZE)
cdef Example eg = Example(self.model.n_classes, CONTEXT_SIZE, self.model.n_feats)
while not stcls.is_final():
eg.wipe()
fill_context(eg.atoms, stcls)
self.moves.set_valid(eg.is_valid, stcls)
fill_context(<atom_t*>eg.atoms.data, stcls)
self.moves.set_valid(<bint*>eg.is_valid.data, stcls)
self.model.predict(eg)
@ -84,12 +84,12 @@ cdef class Parser:
self.moves.preprocess_gold(gold)
cdef StateClass stcls = StateClass.init(tokens.data, tokens.length)
self.moves.initialize_state(stcls)
cdef Example eg = Example(self.model.n_classes, CONTEXT_SIZE)
cdef Example eg = Example(self.model.n_classes, CONTEXT_SIZE, self.model.n_feats)
cdef int cost = 0
while not stcls.is_final():
eg.wipe()
fill_context(eg.atoms, stcls)
self.moves.set_costs(eg.is_valid, eg.costs, stcls, gold)
fill_context(<atom_t*>eg.atoms.data, stcls)
self.moves.set_costs(<bint*>eg.is_valid.data, <int*>eg.costs.data, stcls, gold)
self.model.train(eg)