apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-05 01:32:40 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

parallelizing backprop of cholesky

Browse source
This commit is contained in:
James Hensman 2015-06-08 15:25:16 +01:00
parent 59771c8956
commit 04c14a9b4c
5 changed files with 14239 additions and 123 deletions
Download patch file Download diff file Expand all files Collapse all files

28
GPy/util/choleskies_cython.pyx
View file

@ -5,6 +5,7 @@
# Copyright James Hensman and Alan Saul 2015
import numpy as np
from cython.parallel import prange, parallel
cimport numpy as np
def flat_to_triang(np.ndarray[double, ndim=2] flat, int M):
@ -57,3 +58,30 @@ def backprop_gradient(np.ndarray[double, ndim=2] dL, np.ndarray[double, ndim=2]
dL_dK[k, k] -= L[j, k] * dL_dK[j, k]
dL_dK[k, k] /= (2. * L[k, k])
return dL_dK
def backprop_gradient_par(double[:,:] dL, double[:,:] L):
cdef double[:,:] dL_dK = np.tril(dL).copy()
cdef int N = L.shape[0]
cdef int k, j, i
for k in range(N - 1, -1, -1):
with nogil, parallel():
for i in prange(k + 1, N):
for j in range(k+1, i+1):
dL_dK[i, k] -= dL_dK[i, j] * L[j, k]
for j in range(i, N):
dL_dK[i, k] -= dL_dK[j, i] * L[j, k]
for j in range(k + 1, N):
dL_dK[j, k] /= L[k, k]
dL_dK[k, k] -= L[j, k] * dL_dK[j, k]
dL_dK[k, k] /= (2. * L[k, k])
return dL_dK
cdef extern from "cholesky_backprop.h":
void chol_backprop(int N, double* dL, double* L)
def backprop_gradient_par_c(np.ndarray[double, ndim=2] dL, np.ndarray[double, ndim=2] L):
cdef np.ndarray[double, ndim=2] dL_dK = np.tril(dL) # makes a copy, c-contig
cdef int N = L.shape[0]
chol_backprop(N, <double*> dL_dK.data, <double*> L.data)
return dL_dK