explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2025-02-02 12:44:10 +03:00
Code Issues Packages Projects Releases Wiki Activity

Track which indices are being used

Browse Source
This commit is contained in:
Matthew Honnibal 2017-03-24 00:23:32 +01:00
parent 755d7d486c
commit abb209f631
2 changed files with 20 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
spacy/_ml.pxd
View File

@ -1,9 +1,10 @@
from thinc.linear.features cimport ConjunctionExtracter
from thinc.typedefs cimport atom_t, weight_t
from thinc.structs cimport FeatureC
from libc.stdint cimport uint32_t
from libc.stdint cimport uint32_t, uint64_t
cimport numpy as np
from cymem.cymem cimport Pool
from libcpp.vector cimport vector
cdef class LinearModel:
@ -14,6 +15,7 @@ cdef class LinearModel:
cdef Pool mem
cdef weight_t* W
cdef weight_t* d_W
cdef vector[uint64_t]* _indices
cdef void hinge_lossC(self, weight_t* d_scores,
const weight_t* scores, const weight_t* costs) nogil

20
spacy/_ml.pyx
View File

@ -15,6 +15,9 @@ cimport numpy as np
import numpy
np.import_array()
from thinc.neural.optimizers import Adam
from thinc.neural.ops import NumpyOps
cdef class LinearModel:
def __init__(self, int nr_class, templates, weight_t learn_rate=0.001,
@ -28,6 +31,10 @@ cdef class LinearModel:
sizeof(weight_t))
self.d_W = <weight_t*>self.mem.alloc(self.nr_weight * self.nr_class,
sizeof(weight_t))
self._indices = new vector[uint64_t]()
def __dealloc__(self):
del self._indices
cdef void hinge_lossC(self, weight_t* d_scores,
const weight_t* scores, const weight_t* costs) nogil:
@ -129,12 +136,19 @@ cdef class LinearModel:
# Sort them, to improve memory access pattern
libcpp.algorithm.sort(indices.begin(), indices.end())
for idx in indices:
W = &self.W[idx * nr_class]
d_W = &self.d_W[idx * nr_class]
for clas in range(nr_class):
if d_scores[clas] < 0:
W[clas] -= self.learn_rate * max(-10., d_scores[clas])
d_W[clas] += max(-10., d_scores[clas])
else:
W[clas] -= self.learn_rate * min(10., d_scores[clas])
d_W[clas] += min(10., d_scores[clas])
def finish_update(self, optimizer):
cdef np.npy_intp[1] shape
shape[0] = self.nr_weight * self.nr_class
W_arr = np.PyArray_SimpleNewFromData(1, shape, np.NPY_FLOAT, self.W)
dW_arr = np.PyArray_SimpleNewFromData(1, shape, np.NPY_FLOAT, self.d_W)
optimizer(W_arr, dW_arr, key=1)
@property
def nr_active_feat(self):