def gen_proposal(self, ancestor = None):
assert()
#this needs to be updated
assert(ancestor is not None)
if np.linalg.norm(ancestor.other["gr"]) < 10**-10:
#we are close to a local maximum
f = ancestor.other["lrate"]
prop_dist = self.jump_dist
else:
#we are at a distance to a local maximum
#step in direction of gradient.
(f, theta_1, fval_1, back_off_tmp) = find_step_size(ancestor.sample, ancestor.other["lrate"], ancestor.lpost, ancestor.other["conj"], func = self.lpost)
self.back_off_count += back_off_tmp
step_mean = f * 0.5 * ancestor.other["conj"]
cov = ideal_covar(step_mean, main_var_scale = 1, other_var = 0.5) # , fix_main_var=1
prop_dist = mvnorm(step_mean, cov)
step = prop_dist.rvs()
samp = ancestor.sample + prop_dist.rvs()
(lp, gr) = self.lpost_and_grad(samp)
momentum = max(0, np.float(gr.T.dot(gr - ancestor.other["gr"]) / ancestor.other["gr"].T.dot(ancestor.other["gr"])))
conj_dir_1 = gr + momentum * ancestor.other["conj"]
lprop = prop_dist.logpdf(step)
rval = PmcSample(ancestor = ancestor,
sample = samp,
lpost = lp,
lprop = lprop,
prop_obj = self,
lweight = lp - prop_dist.logpdf(step),
other = {"lrate":f, "gr":gr, "conj":conj_dir_1})
return rval
def gen_proposal(self, ancestor = None):
assert()
#this needs to be updated
assert(ancestor is not None)
if np.linalg.norm(ancestor.other["gr"]) < 10**-10:
#we are close to a local maximum
f = ancestor.other["lrate"]
prop_dist = self.jump_dist
else:
#we are at a distance to a local maximum
#step in direction of gradient.
(f, theta_1, fval_1, back_off_tmp) = find_step_size(ancestor.sample, ancestor.other["lrate"], ancestor.lpost, ancestor.other["conj"], func = self.lpost)
self.back_off_count += back_off_tmp
step_mean = f * 0.5 * ancestor.other["conj"]
cov = ideal_covar(step_mean, main_var_scale = 1, other_var = 0.5) # , fix_main_var=1
prop_dist = mvnorm(step_mean, cov)
step = prop_dist.rvs()
samp = ancestor.sample + prop_dist.rvs()
(lp, gr) = self.lpost_and_grad(samp)
momentum = max(0, np.float(gr.T.dot(gr - ancestor.other["gr"]) / ancestor.other["gr"].T.dot(ancestor.other["gr"])))
conj_dir_1 = gr + momentum * ancestor.other["conj"]
lprop = prop_dist.logpdf(step)
rval = PmcSample(ancestor = ancestor,
sample = samp,
lpost = lp,
lprop = lprop,
prop_obj = self,
lweight = lp - prop_dist.logpdf(step),
other = {"lrate":f, "gr":gr, "conj":conj_dir_1})
return rval
def gen_proposal(self, ancestor=None):
self.prepare_ancestor(ancestor)
f = ancestor.other["lrate"]
if True:
(f, theta_1, lpost_1, grad_1,
foo) = find_step_size(ancestor.sample,
f,
ancestor.lpost,
ancestor.other["gr"],
func_and_grad=self.lpost_and_grad)
sc_gr = f * 0.5 * ancestor.other["gr"]
cov = ideal_covar(sc_gr,
fix_main_var=self.main_var,
other_var=self.other_var) # , fix_main_var=1
step_dist = mvnorm(sc_gr, cov)
#prop_dist = mvnorm(ancestor.sample + sc_gr, self.cov)
#print(ancestor.lpost, lpost_1)
else:
step_dist = mvnorm(np.zeros(ancestor.sample.size),
np.eye(ancestor.sample.size) * self.main_var)
(new_samp,
step) = gen_sample_prototype(ancestor,
self,
step_dist=step_dist,
lpost_and_grad_func=self.lpost_and_grad)
new_samp.other["lrate"] = f
return new_samp
def gen_proposal(self, ancestor = None):
assert(ancestor is not None)
assert(ancestor.sample is not None)
rval = []
if ancestor.lpost is None:
if "gr" not in ancestor.other:
(ancestor.lpost, ancestor.other["gr"]) = self.lpost_and_grad(ancestor.sample)
else:
ancestor.lpost = self.lpost_and_grad(ancestor.sample)[0]
elif "gr" not in ancestor.other:
ancestor.other["old"] = False
ancestor.other["gr"] = self.lpost_and_grad(ancestor.sample)[1]
assert(ancestor.other["gr"].size == ancestor.sample.size)
if "lrate" in ancestor.other:
f = ancestor.other["lrate"]
else:
f = self.lrate
if np.linalg.norm(ancestor.other["gr"]) < 10**-10:
#we are close to a local maximum
print("jumping")
prop_dist = self.jump_dist
else:
#we are at a distance to a local maximum
#step in direction of gradient.
assert(ancestor.other["gr"].size == ancestor.sample.size)
(f, theta_1, lpost_1, grad_1, back_off_tmp) = find_step_size(ancestor.sample, f, ancestor.lpost, ancestor.other["gr"], func_and_grad = self.lpost_and_grad)
self.back_off_count += len(back_off_tmp)
if False and ancestor.lprop is not None:
for (f, samp, lp, gr) in back_off_tmp:
rval.append(PmcSample(ancestor = ancestor,
sample = samp,
lpost = lp,
lprop = ancestor.lprop,
lweight = lp - ancestor.lprop,
prop_obj = ancestor.prop_obj,
other = {"lrate":f, "gr":gr, "old":True}))
mean_step = f * self.prop_mean_on_line * ancestor.other["gr"]
prop_mean = ancestor.sample + mean_step
cov = ideal_covar(mean_step, main_var_scale = self.main_var_scale, other_var = self.other_var, fix_main_var = self.fix_main_var) # , fix_main_var=1
prop_dist = mvnorm(prop_mean, cov)
#print(ancestor.lpost, lpost_1)
samp = prop_dist.rvs()
(lp, gr) = self.lpost_and_grad(samp)
print(ancestor.lpost, self.lpost_and_grad(prop_mean)[0])
lprop = prop_dist.logpdf(samp)
assert(ancestor.other["gr"].size == ancestor.sample.size)
assert(gr.size == samp.size)
rval.append(PmcSample(ancestor = ancestor,
sample = samp,
lpost = lp,
lprop = lprop,
lweight = lp - lprop,
prop_obj = self,
other = {"lrate":f, "gr":gr, "old":True}))
return rval
def gen_proposal(self, ancestor = None):
assert(ancestor is not None)
rval = []
old = True
if "gr" not in ancestor.other:
old = False
ancestor.other["old"] = False
ancestor.other["gr"] = self.lpost_and_grad(ancestor.sample)[1]
assert(ancestor.other["gr"].size == ancestor.sample.size)
if "lrate" in ancestor.other:
f = ancestor.other["lrate"]
else:
f = self.lrate
if np.linalg.norm(ancestor.other["gr"]) < 10**-10:
#we are close to a local maximum
print("jumping")
prop_dist = self.jump_dist
else:
#we are at a distance to a local maximum
#step in direction of gradient.
assert(ancestor.other["gr"].size == ancestor.sample.size)
(f, theta_1, lpost_1, grad_1, back_off_tmp) = find_step_size(ancestor.sample, f, ancestor.lpost, ancestor.other["gr"], func_and_grad = self.lpost_and_grad)
self.back_off_count += len(back_off_tmp)
if False and ancestor.lprop is not None:
for (f, samp, lp, gr) in back_off_tmp:
rval.append(PmcSample(ancestor = ancestor,
sample = samp,
lpost = lp,
lprop = ancestor.lprop,
lweight = lp - ancestor.lprop,
prop_obj = ancestor.prop_obj,
other = {"lrate":f, "gr":gr, "old":True}))
step_mean = f * self.prop_mean_on_line * ancestor.other["gr"]
cov = ideal_covar(step_mean, main_var_scale = self.main_var_scale, other_var = self.other_var, fix_main_var = self.fix_main_var) # , fix_main_var=1
prop_dist = mvnorm(step_mean, cov)
step = prop_dist.rvs()
samp = ancestor.sample + step
(lp, gr) = self.lpost_and_grad(samp)
lprop = prop_dist.logpdf(step)
assert(ancestor.other["gr"].size == ancestor.sample.size)
assert(gr.size == samp.size)
rval.append(PmcSample(ancestor = ancestor,
sample = samp,
lpost = lp,
lprop = lprop,
lweight = lp - lprop,
prop_obj = self,
other = {"lrate":f, "gr":gr, "old":True}))
return rval
def gen_proposal(self, ancestor = None):
self.prepare_ancestor(ancestor)
f = ancestor.other["lrate"]
if True:
(f, theta_1, lpost_1,
grad_1, foo) = find_step_size(ancestor.sample, f, ancestor.lpost,
ancestor.other["gr"],
func_and_grad = self.lpost_and_grad)
sc_gr = f * 0.5* ancestor.other["gr"]
cov = ideal_covar(sc_gr, fix_main_var = self.main_var, other_var = self.other_var) # , fix_main_var=1
step_dist = mvnorm(sc_gr, cov)
#prop_dist = mvnorm(ancestor.sample + sc_gr, self.cov)
#print(ancestor.lpost, lpost_1)
else:
step_dist = mvnorm(np.zeros(ancestor.sample.size), np.eye(ancestor.sample.size)*self.main_var)
(new_samp, step) = gen_sample_prototype(ancestor, self,
step_dist = step_dist,
lpost_and_grad_func = self.lpost_and_grad)
new_samp.other["lrate"] = f
return new_samp