def logtildegamma(G, Z, X, supp, w, G_mu, G_Sigma, eta):
prob = gv.sigmoid(w + gv.logit(Z))
llh = np.sum(supp * X * clog(prob)) + np.sum(supp *
(1 - X) * clog(1 - prob))
prior = logpdf(G, G_mu, G_Sigma)
#print llh, prior
return llh + prior
def fun(w0):
return gv.sigmoid(w0 +
gv.logit(Zs_ok[..., f])).mean() - Xbar[l0,
l1, f]
if 0:
l0 = l1 = 4
f = 10
def fun(w0):
return gv.sigmoid(w0 +
gv.logit(Zs_ok[..., f])).mean() - Xbar[l0,
l1, f]
#import IPython
#IPython.embed()
w[l0, l1, f] = find_zero(fun, l=-6, u=6, depth=15)
LZ = gv.logit(Zs_ok.clip(EPS, 1 - EPS))
if 0 and loop == 0:
import IPython
IPython.embed()
BINS = 50
LZ_counts = np.zeros((F, BINS))
LZ_values = np.zeros((F, BINS))
for f in xrange(F):
ret = np.histogram(LZ[..., f], BINS)
LZ_counts[f] = ret[0]
LZ_values[f] = (ret[1][1:] + ret[1][:-1]) / 2
print LZ_counts[0]
print LZ_values[0]
def clogit(x):
return gv.logit(gv.bclip(x, 0.025))
def logtildegamma(G, Z, X, supp, w, G_mu, G_Sigma, eta):
prob = gv.sigmoid(w + gv.logit(Z))
llh = np.sum(supp * X * clog(prob)) + np.sum(supp * (1 - X) * clog(1 - prob))
prior = logpdf(G, G_mu, G_Sigma)
#print llh, prior
return llh + prior
def fun(w0):
return gv.sigmoid(w0 + gv.logit(Zs_ok[...,f])).mean() - Xbar[l0,l1,f]
w[l0,l1,f] = find_zero(fun, l=-6, u=6, depth=15)
if 0:
l0 = l1 = 4
f = 10
def fun(w0):
return gv.sigmoid(w0 + gv.logit(Zs_ok[...,f])).mean() - Xbar[l0,l1,f]
#import IPython
#IPython.embed()
w[l0,l1,f] = find_zero(fun, l=-6, u=6, depth=15)
LZ = gv.logit(Zs_ok.clip(EPS, 1-EPS))
if 0 and loop == 0:
import IPython
IPython.embed()
BINS = 50
LZ_counts = np.zeros((F, BINS))
LZ_values = np.zeros((F, BINS))
for f in xrange(F):
ret = np.histogram(LZ[...,f], BINS)
LZ_counts[f] = ret[0]
LZ_values[f] = (ret[1][1:] + ret[1][:-1])/2
print LZ_counts[0]
