explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2025-08-06 21:30:22 +03:00
Code Issues Packages Projects Releases Wiki Activity

Replace references to thinc.backends.linalg with CBlas

Browse Source 
Backport of:
https://github.com/explosion/spaCy/pull/11292

Co-authored-by: Madeesh Kannan <shadeMe@users.noreply.github.com>
This commit is contained in:
Daniël de Kok 2022-12-02 15:52:09 +01:00
parent 1573477c75
commit 1d705cc683
1 changed files with 17 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
spacy/ml/parser_model.pyx
View File

@ -3,7 +3,6 @@ cimport numpy as np
from libc.math cimport exp from libc.math cimport exp
from libc.string cimport memset, memcpy from libc.string cimport memset, memcpy
from libc.stdlib cimport calloc, free, realloc from libc.stdlib cimport calloc, free, realloc
from thinc.backends.linalg cimport Vec, VecVec
from thinc.backends.cblas cimport saxpy, sgemm from thinc.backends.cblas cimport saxpy, sgemm
import numpy import numpy
@ -102,11 +101,10 @@ cdef void predict_states(CBlas cblas, ActivationsC* A, StateC** states,
sum_state_features(cblas, A.unmaxed, sum_state_features(cblas, A.unmaxed,
W.feat_weights, A.token_ids, n.states, n.feats, n.hiddens * n.pieces) W.feat_weights, A.token_ids, n.states, n.feats, n.hiddens * n.pieces)
for i in range(n.states): for i in range(n.states):
VecVec.add_i(&A.unmaxed[i*n.hiddens*n.pieces], saxpy(cblas)(n.hiddens * n.pieces, 1., W.feat_bias, 1, &A.unmaxed[i*n.hiddens*n.pieces], 1)
W.feat_bias, 1., n.hiddens * n.pieces)
for j in range(n.hiddens): for j in range(n.hiddens):
index = i * n.hiddens * n.pieces + j * n.pieces index = i * n.hiddens * n.pieces + j * n.pieces
which = Vec.arg_max(&A.unmaxed[index], n.pieces) which = _arg_max(&A.unmaxed[index], n.pieces)
A.hiddens[i*n.hiddens + j] = A.unmaxed[index + which] A.hiddens[i*n.hiddens + j] = A.unmaxed[index + which]
memset(A.scores, 0, n.states * n.classes * sizeof(float)) memset(A.scores, 0, n.states * n.classes * sizeof(float))
if W.hidden_weights == NULL: if W.hidden_weights == NULL:
@ -119,8 +117,7 @@ cdef void predict_states(CBlas cblas, ActivationsC* A, StateC** states,
0.0, A.scores, n.classes) 0.0, A.scores, n.classes)
# Add bias # Add bias
for i in range(n.states): for i in range(n.states):
VecVec.add_i(&A.scores[i*n.classes], saxpy(cblas)(n.classes, 1., W.hidden_bias, 1, &A.scores[i*n.classes], 1)
W.hidden_bias, 1., n.classes)
# Set unseen classes to minimum value # Set unseen classes to minimum value
i = 0 i = 0
min_ = A.scores[0] min_ = A.scores[0]
@ -158,7 +155,8 @@ cdef void cpu_log_loss(float* d_scores,
"""Do multi-label log loss""" """Do multi-label log loss"""
cdef double max_, gmax, Z, gZ cdef double max_, gmax, Z, gZ
best = arg_max_if_gold(scores, costs, is_valid, O) best = arg_max_if_gold(scores, costs, is_valid, O)
guess = Vec.arg_max(scores, O) guess = _arg_max(scores, O)
if best == -1 or guess == -1: if best == -1 or guess == -1:
# These shouldn't happen, but if they do, we want to make sure we don't # These shouldn't happen, but if they do, we want to make sure we don't
# cause an OOB access. # cause an OOB access.
@ -488,3 +486,15 @@ cdef class precompute_hiddens:
return d_best.reshape((d_best.shape + (1,))) return d_best.reshape((d_best.shape + (1,)))
return state_vector, backprop_relu return state_vector, backprop_relu
cdef inline int _arg_max(const float* scores, const int n_classes) nogil:
if n_classes == 2:
return 0 if scores[0] > scores[1] else 1
cdef int i
cdef int best = 0
cdef float mode = scores[0]
for i in range(1, n_classes):
if scores[i] > mode:
mode = scores[i]
best = i
return best