mirror of
https://github.com/SheffieldML/GPy.git
synced 2026-05-30 14:35:15 +02:00
weave speed up for tie framework
This commit is contained in:
Zhenwen Dai
parent
7326d19848
commit
1142b17790
2 changed files with 224 additions and 45 deletions
|
|
@ -69,6 +69,7 @@ class Tie(Parameterized):
|
|||
self.label_buf = None
|
||||
self.buf_idx = None
|
||||
self._untie_ = None
|
||||
self._label_to_idx = None
|
||||
|
||||
def getTiedParamList(self):
|
||||
if self.tied_param is None:
|
||||
|
|
@ -85,18 +86,44 @@ class Tie(Parameterized):
|
|||
"""
|
||||
assert self.tied_param is not None
|
||||
read = np.zeros((self.tied_param.size,),dtype=np.uint8)
|
||||
def _sync_val_p(p, tieparam, read):
|
||||
if p.tie is not None:
|
||||
labels = np.unique(p.tie)
|
||||
labels = labels[labels>0]
|
||||
for l in labels:
|
||||
if toTiedParam and read[tieparam.tie==l] == 0:
|
||||
val = p[p.tie==l][0]
|
||||
tieparam[tieparam.tie==l] = val
|
||||
read[tieparam.tie==l] = 1
|
||||
else:
|
||||
val = tieparam[tieparam.tie==l]
|
||||
p[p.tie==l] = val
|
||||
try:
|
||||
from scipy import weave
|
||||
def _sync_val_p(p, tieparam, read):
|
||||
if p.tie is not None:
|
||||
# ltoi = self._label_to_idx
|
||||
ltoi = np.zeros((self.tied_param.size+1,),dtype=np.int32)
|
||||
ltoi[self.tied_param.tie] = range(self.tied_param.size)
|
||||
totie = 1 if toTiedParam else 0
|
||||
p_idx = p._raveled_index()
|
||||
p_tie = p.tie.flatten()
|
||||
p_org = p._original_.flat
|
||||
p_size = p.size
|
||||
code = """
|
||||
for(int i=0;i<p_size;i++) {
|
||||
if(p_tie[i]>0) {
|
||||
if(totie==1 && read[ltoi[p_tie[i]]]==0) {
|
||||
tieparam[ltoi[p_tie[i]]] = p_org[int(p_idx[i])];
|
||||
read[ltoi[p_tie[i]]] = 1;
|
||||
} else {
|
||||
p_org[int(p_idx[i])] = tieparam[ltoi[p_tie[i]]];
|
||||
}
|
||||
}
|
||||
}
|
||||
"""
|
||||
weave.inline(code, arg_names=['p_org','p_idx','tieparam','p_tie','p_size','ltoi','read','totie'])
|
||||
except:
|
||||
def _sync_val_p(p, tieparam, read):
|
||||
if p.tie is not None:
|
||||
labels = np.unique(p.tie)
|
||||
labels = labels[labels>0]
|
||||
for l in labels:
|
||||
if toTiedParam and read[tieparam.tie==l] == 0:
|
||||
val = p[p.tie==l][0]
|
||||
tieparam[tieparam.tie==l] = val
|
||||
read[tieparam.tie==l] = 1
|
||||
else:
|
||||
val = tieparam[tieparam.tie==l]
|
||||
p[p.tie==l] = val
|
||||
for p in plist:
|
||||
self._traverse_param(_sync_val_p, (p,self.tied_param,read), [])
|
||||
|
||||
|
|
@ -108,38 +135,151 @@ class Tie(Parameterized):
|
|||
"""
|
||||
assert self.tied_param is not None
|
||||
read = np.zeros((self.tied_param.size,),dtype=np.uint8)
|
||||
cons = [c[0] if len(c)>0 else None for c in self.tied_param.constraints.properties_for(range(self.tied_param.size))]
|
||||
def _sync_constraints_p(p, tieparam, read, cons):
|
||||
from scipy import weave
|
||||
def _sync_constraint_totie(p, tieparam, read):
|
||||
if p.tie is not None:
|
||||
p = p._original_
|
||||
if p is tieparam:
|
||||
return
|
||||
labels = np.unique(p.tie)
|
||||
labels = labels[labels>0]
|
||||
for l in labels:
|
||||
idx = np.where(tieparam.tie==l)[0][0]
|
||||
conslist = p.constraints.properties_to_index_dict(np.where(p.tie.flat==l)[0])
|
||||
if toTiedParam and read[tieparam.tie==l] == 0:
|
||||
if len(conslist)==0:
|
||||
if cons[idx] is not None:
|
||||
tieparam[idx:idx+1].unconstrain()
|
||||
else:
|
||||
c = conslist.keys()[0]
|
||||
if len(conslist)>1 or len(conslist[c])!= (p.tie==l).sum():
|
||||
p[p.tie==l].constrain(c)
|
||||
if c != cons[idx]:
|
||||
tieparam[idx:idx+1].constrain(c)
|
||||
cons[idx] = c
|
||||
read[tieparam.tie==l] = 1
|
||||
else:
|
||||
if cons[idx] is None:
|
||||
if len(conslist)>0:
|
||||
p[p.tie==l].unconstrain()
|
||||
else:
|
||||
if len(conslist)!=1 or conslist.keys()[0]!=cons[idx] or len(conslist[cons[idx]])!= (p.tie==l).sum():
|
||||
p[p.tie==l].constrain(cons[idx])
|
||||
# ltoi = self._label_to_idx
|
||||
ltoi = np.zeros((self.tied_param.size+1,),dtype=np.int32)
|
||||
ltoi[self.tied_param.tie] = range(self.tied_param.size)
|
||||
p_tie = p.tie.flatten()
|
||||
p_size = p.size
|
||||
cons_tie = self.tied_param.constraints
|
||||
cons_p = p.constraints
|
||||
for c in cons_p.properties():
|
||||
idx_p = cons_p[c]
|
||||
idx_p_size = idx_p.size
|
||||
idx_tie = cons_tie[c]
|
||||
flag = np.zeros((self.tied_param.size,),dtype=np.uint8)
|
||||
flag[idx_tie] = -1
|
||||
flag[read==1] = -1
|
||||
code1 = """
|
||||
for(int i=0;i<idx_p_size;i++) {
|
||||
int label = int(p_tie[int(idx_p[i])]);
|
||||
if(label>0) {
|
||||
int idx_tie=ltoi[label];
|
||||
if(flag[idx_tie]==0) {flag[idx_tie]=1;}
|
||||
}
|
||||
}
|
||||
"""
|
||||
weave.inline(code1, arg_names=['idx_p_size','idx_p','p_tie','ltoi','flag'])
|
||||
if (flag==1).sum()>0:
|
||||
tieparam[flag==1].constrain(c)
|
||||
read[flag==1] = 1
|
||||
flag = np.zeros((self.tied_param.size,),dtype=np.uint8)
|
||||
flag[:]= -1
|
||||
uncons_p = np.ones((p.size,),dtype=np.uint8)
|
||||
for c,ind in cons_tie.iteritems():
|
||||
flag[ind] = 0
|
||||
flag[read==1] = -1
|
||||
for c,ind in cons_p.iteritems():
|
||||
uncons_p[ind] = 0
|
||||
code2 = """
|
||||
for(int i=0;i<p_size;i++) {
|
||||
if(uncons_p[i]==1 && p_tie[i]>0){
|
||||
int idx_tie=ltoi[p_tie[i]];
|
||||
if(flag[idx_tie]==0) {flag[idx_tie]=1;}
|
||||
}
|
||||
}
|
||||
"""
|
||||
weave.inline(code2, arg_names=['p_size','uncons_p','p_tie','ltoi','flag'])
|
||||
if (flag==1).sum()>0:
|
||||
tieparam[flag==1].unconstrain()
|
||||
read[flag==1] = 1
|
||||
|
||||
def _sync_constraint_toparam(p, tieparam):
|
||||
if p.tie is not None:
|
||||
p = p._original_
|
||||
if p is tieparam:
|
||||
return
|
||||
# ltoi = self._label_to_idx
|
||||
ltoi = np.zeros((self.tied_param.size+1,),dtype=np.int32)
|
||||
ltoi[self.tied_param.tie] = range(self.tied_param.size)
|
||||
p_tie = p.tie.flatten()
|
||||
p_size = p.size
|
||||
cons_tie = self.tied_param.constraints
|
||||
cons_p = p.constraints
|
||||
for c in cons_tie.properties():
|
||||
idx_p = cons_p[c]
|
||||
idx_tie = cons_tie[c]
|
||||
flag_tie = np.zeros((tieparam.size,),dtype=np.uint8)
|
||||
flag_tie[idx_tie]=1
|
||||
flag = np.zeros((p.size,),dtype=np.uint8)
|
||||
flag[idx_p] = -1
|
||||
|
||||
code1 = """
|
||||
for(int i=0;i<p_size;i++) {
|
||||
if(flag[i]==0 && p_tie[i]>0 && flag_tie[ltoi[p_tie[i]]]==1) {flag[i]=1;}
|
||||
}
|
||||
"""
|
||||
weave.inline(code1, arg_names=['p_size','p_tie','ltoi','flag','flag_tie'])
|
||||
if (flag==1).sum()>0:
|
||||
flag = flag.reshape(*p.shape)
|
||||
p[flag==1].constrain(c)
|
||||
uncons_tie = np.ones((self.tied_param.size,),dtype=np.uint8)
|
||||
flag = np.zeros((p.size,),dtype=np.uint8)
|
||||
flag[:]= -1
|
||||
for c,ind in cons_tie.iteritems():
|
||||
uncons_tie[ind] = 0
|
||||
for c,ind in cons_p.iteritems():
|
||||
flag[ind] = 0
|
||||
code2 = """
|
||||
for(int i=0;i<p_size;i++) {
|
||||
if(flag[i]==0 && p_tie[i]>0 && uncons_tie[ltoi[p_tie[i]]]==1){ flag[i]=1;}
|
||||
}
|
||||
"""
|
||||
weave.inline(code2, arg_names=['p_size','uncons_tie','p_tie','ltoi','flag'])
|
||||
if (flag==1).sum()>0:
|
||||
flag = flag.reshape(*p.shape)
|
||||
p[flag==1].unconstrain()
|
||||
if toTiedParam:
|
||||
for p in plist:
|
||||
self._traverse_param(_sync_constraint_totie, (p,self.tied_param,read), [])
|
||||
for p in plist:
|
||||
self._traverse_param(_sync_constraints_p, (p,self.tied_param,read, cons), [])
|
||||
self._traverse_param(_sync_constraint_toparam, (p,self.tied_param), [])
|
||||
|
||||
# def _sync_constraints(self, plist, toTiedParam=True):
|
||||
# """
|
||||
# Ensure the consistency of the constraints of tied parameters.
|
||||
# if toTieParam is true, the constraints of tied parameters will be synchronized among themselves and to the *TiedParam*,
|
||||
# otherwise all the tied parameters will be synchronized according to the constraints in *TiedParam*.
|
||||
# """
|
||||
# assert self.tied_param is not None
|
||||
# read = np.zeros((self.tied_param.size,),dtype=np.uint8)
|
||||
# cons = [c[0] if len(c)>0 else None for c in self.tied_param.constraints.properties_for(range(self.tied_param.size))]
|
||||
# def _sync_constraints_p(p, tieparam, read, cons):
|
||||
# if p.tie is not None:
|
||||
# p = p._original_
|
||||
# if p is tieparam:
|
||||
# return
|
||||
# labels = np.unique(p.tie)
|
||||
# labels = labels[labels>0]
|
||||
# for l in labels:
|
||||
# idx = np.where(tieparam.tie==l)[0][0]
|
||||
# conslist = p.constraints.properties_to_index_dict(np.where(p.tie.flat==l)[0])
|
||||
# if toTiedParam and read[tieparam.tie==l] == 0:
|
||||
# if len(conslist)==0:
|
||||
# if cons[idx] is not None:
|
||||
# tieparam[idx:idx+1].unconstrain()
|
||||
# else:
|
||||
# c = conslist.keys()[0]
|
||||
# if len(conslist)>1 or len(conslist[c])!= (p.tie==l).sum():
|
||||
# p[p.tie==l].constrain(c)
|
||||
# if c != cons[idx]:
|
||||
# tieparam[idx:idx+1].constrain(c)
|
||||
# cons[idx] = c
|
||||
# read[tieparam.tie==l] = 1
|
||||
# else:
|
||||
# if cons[idx] is None:
|
||||
# if len(conslist)>0:
|
||||
# p[p.tie==l].unconstrain()
|
||||
# else:
|
||||
# if len(conslist)!=1 or conslist.keys()[0]!=cons[idx] or len(conslist[cons[idx]])!= (p.tie==l).sum():
|
||||
# p[p.tie==l].constrain(cons[idx])
|
||||
# for p in plist:
|
||||
# self._traverse_param(_sync_constraints_p, (p,self.tied_param,read, cons), [])
|
||||
|
||||
|
||||
@staticmethod
|
||||
|
|
@ -449,14 +589,53 @@ class Tie(Parameterized):
|
|||
|
||||
def collate_gradient(self):
|
||||
if self.tied_param is not None:
|
||||
self.tied_param.gradient = 0.
|
||||
[np.put(self.tied_param.gradient, i, self._highest_parent_.gradient[self.label_buf==self.label_buf[self.buf_idx[i]]].sum())
|
||||
for i in xrange(self.tied_param.size)]
|
||||
try:
|
||||
from scipy import weave
|
||||
pa_grad = self._highest_parent_.gradient
|
||||
label_buf = self.label_buf
|
||||
buf_idx = self.buf_idx
|
||||
tied_grad = self.tied_param.gradient
|
||||
t_size = self.tied_param.size
|
||||
p_size = self._highest_parent_.gradient.size
|
||||
code="""
|
||||
for(int i=0;i<t_size;i++) {
|
||||
int l = label_buf[buf_idx[i]];
|
||||
tied_grad[i] = 0;
|
||||
double g=0;
|
||||
for(int j=0;j<p_size;j++) {
|
||||
if(label_buf[j]==l) {g += pa_grad[j];}
|
||||
}
|
||||
tied_grad[i] = g;
|
||||
}
|
||||
"""
|
||||
weave.inline(code, arg_names=['pa_grad','label_buf','buf_idx','tied_grad','t_size','p_size'])
|
||||
except:
|
||||
self.tied_param.gradient = 0.
|
||||
[np.put(self.tied_param.gradient, i, self._highest_parent_.gradient[self.label_buf==self.label_buf[self.buf_idx[i]]].sum())
|
||||
for i in xrange(self.tied_param.size)]
|
||||
|
||||
def propagate_val(self):
|
||||
if self.tied_param is not None:
|
||||
for i in xrange(self.tied_param.size):
|
||||
self._highest_parent_.param_array[self.label_buf==self.label_buf[self.buf_idx[i]]] = self.tied_param[i]
|
||||
try:
|
||||
from scipy import weave
|
||||
param_array = self._highest_parent_.param_array
|
||||
label_buf = self.label_buf
|
||||
buf_idx = self.buf_idx
|
||||
tied_param = self.tied_param
|
||||
t_size = self.tied_param.size
|
||||
p_size = param_array.size
|
||||
code="""
|
||||
for(int i=0;i<t_size;i++) {
|
||||
int l = label_buf[buf_idx[i]];
|
||||
for(int j=0;j<p_size;j++) {
|
||||
if(label_buf[j]==l) {param_array[j] = tied_param[i];}
|
||||
}
|
||||
}
|
||||
"""
|
||||
weave.inline(code, arg_names=['param_array','label_buf','buf_idx','tied_param','t_size','p_size'])
|
||||
except:
|
||||
for i in xrange(self.tied_param.size):
|
||||
self._highest_parent_.param_array[self.label_buf==self.label_buf[self.buf_idx[i]]] = self.tied_param[i]
|
||||
self._PROPAGATE_VAL_ = True
|
||||
|
||||
#=========================================
|
||||
|
|
|
|||
|
|
@ -84,7 +84,7 @@ class SparseGP_MPI(SparseGP):
|
|||
if self.mpi_comm != None:
|
||||
if self._IN_OPTIMIZATION_ and self.mpi_comm.rank==0:
|
||||
self.mpi_comm.Bcast(np.int32(1),root=0)
|
||||
self.mpi_comm.Bcast(p, root=0)
|
||||
self.mpi_comm.Bcast(p, root=0)
|
||||
SparseGP.optimizer_array.fset(self,p)
|
||||
|
||||
def optimize(self, optimizer=None, start=None, **kwargs):
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue