def doStep(w):
f_holdout, gw_holdout = func_holdout(w)
gw_holdout_norm = 0
gw_holdout_effective = ndict.clone(gw_holdout)
for i in w:
m1_holdout[i] += lambd * (gw_holdout[i] - m1_holdout[i])
m2_holdout[i] += lambd * (gw_holdout[i]**2 - m2_holdout[i])
gw_holdout_effective[i] /= np.sqrt(m2_holdout[i] + 1e-8)
gw_holdout_norm += (gw_holdout_effective[i]**2).sum()
gw_holdout_norm = np.sqrt(gw_holdout_norm)
f_tot = 0
gw_tot = ndict.cloneZeros(w)
alphas = []
for j in range(len(funcs)):
f, gw = funcs[j](w)
f_tot += f
gw_norm = 0
gw_effective = ndict.clone(gw)
for i in w:
# Update first and second moments
m1[j][i] += lambd * (gw[i] - m1[j][i])
m2[j][i] += lambd * (gw[i]**2 - m2[j][i])
gw_effective[i] /= np.sqrt(m2[j][i] + 1e-8)
gw_norm += (gw_effective[i]**2).sum()
gw_norm = np.sqrt(gw_norm)
# Compute dot product with holdout gradient
alpha = 0
for i in w:
alpha += (gw_effective[i] * gw_holdout_effective[i]).sum()
alpha /= gw_holdout_norm * gw_norm
alphas.append(alpha)
#alpha = (alpha > 0) * 1.0
for i in w:
# Accumulate gradient of subobjective
gw_tot[i] += alpha * gw[i] / np.sqrt(m2[j][i] + 1e-8)
#print 'alphas:', alphas
if batchi[0] > warmup:
for i in w:
w[i] += stepsize * gw_tot[i]
batchi[0] += 1
return f_tot
def sample_standard_auto(z, fgrad, n_burnin, n_samples):
hmc_dostep = hmc_step_autotune(n_steps=10, init_stepsize=1e-2, target=0.9)
def dostep(_z):
logpxz, _, _ = hmc_dostep(fgrad, _z)
return logpxz
# Burn-in phase
for i in range(n_burnin):
dostep(z)
# Sample
z_list = []
logpxz_list = []
for _ in range(n_samples):
logpxz = dostep(z)
#print 'logpxz:', logpxz
logpxz_list.append(logpxz.copy())
z_list.append(ndict.clone(z))
z, logpxz = combine_samples(z_list, logpxz_list)
return z, logpxz
def sample_standard_auto(z, fgrad, n_burnin, n_samples):
hmc_dostep = hmc_step_autotune(n_steps=10, init_stepsize=1e-2, target=0.9)
def dostep(_z):
logpxz, _, _ = hmc_dostep(fgrad, _z)
return logpxz
# Burn-in phase
for i in range(n_burnin):
dostep(z)
# Sample
z_list = []
logpxz_list = []
for _ in range(n_samples):
logpxz = dostep(z)
# print 'logpxz:', logpxz
logpxz_list.append(logpxz.copy())
z_list.append(ndict.clone(z))
z, logpxz = combine_samples(z_list, logpxz_list)
return z, logpxz
def hmc_step(fgrad, x0, _stepsize=1e-2, n_steps=20):
# === INITIALIZE
n_batch = x0.itervalues().next().shape[1]
stepsize = (np.random.uniform(size=(1, n_batch)) <
0.5).astype(float) * 2 - 1
stepsize *= _stepsize
if np.random.uniform() < 0.5:
stepsize *= -1
# Sample velocity
vnew = {}
for i in x0:
vnew[i] = np.random.normal(size=x0[i].shape)
# copy initial state
xnew = ndict.clone(x0)
v0 = ndict.clone(vnew)
# === LEAPFROG STEPS
# Compute velocity at time (t + stepsize/2)
# Compute position at time (t + stepsize)
logpxz0, g = fgrad(xnew)
for i in xnew:
vnew[i] += 0.5 * stepsize * g[i]
xnew[i] += stepsize * vnew[i]
# Perform leapfrog steps
for step in xrange(n_steps):
#print 'hmc_step:', step
logpxz, g = fgrad(xnew)
for i in xnew:
vnew[i] += stepsize * g[i]
xnew[i] += stepsize * vnew[i]
# Perform final half-step for velocity
logpxz1, g = fgrad(xnew)
for i in xnew:
vnew[i] += 0.5 * stepsize * g[i]
# === METROPOLIS-HASTINGS ACCEPT/REJECT
# Compute old and new Hamiltonians
k0 = 0
k1 = 0
for i in vnew:
k0 += (v0[i]**2).sum(axis=0, keepdims=True)
k1 += (vnew[i]**2).sum(axis=0, keepdims=True)
h0 = -logpxz0 + 0.5 * k0
h1 = -logpxz1 + 0.5 * k1
#print logpxz0, k0, logpxz1, k1
#print h0-h1
# Perform Metropolis-Hasting step
accept = np.exp(h0 - h1) >= np.random.uniform(size=h1.shape)
accept = accept.astype(float)
for i in xnew:
accept2 = np.dot(np.ones((xnew[i].shape[0], 1)), accept)
x0[i] = (1 - accept2) * x0[i] + accept2 * xnew[i]
# result: updated 'x0'
logpxz = (1 - accept) * logpxz0 + accept * logpxz1
return logpxz, accept
def hmc_step(fgrad, x0, _stepsize=1e-2, n_steps=20):
# === INITIALIZE
n_batch = x0.itervalues().next().shape[1]
stepsize = (np.random.uniform(size=(1, n_batch)) < 0.5).astype(float) * 2 - 1
stepsize *= _stepsize
if np.random.uniform() < 0.5:
stepsize *= -1
# Sample velocity
vnew = {}
for i in x0:
vnew[i] = np.random.normal(size=x0[i].shape)
# copy initial state
xnew = ndict.clone(x0)
v0 = ndict.clone(vnew)
# === LEAPFROG STEPS
# Compute velocity at time (t + stepsize/2)
# Compute position at time (t + stepsize)
logpxz0, g = fgrad(xnew)
for i in xnew:
vnew[i] += 0.5 * stepsize * g[i]
xnew[i] += stepsize * vnew[i]
# Perform leapfrog steps
for step in xrange(n_steps):
# print 'hmc_step:', step
logpxz, g = fgrad(xnew)
for i in xnew:
vnew[i] += stepsize * g[i]
xnew[i] += stepsize * vnew[i]
# Perform final half-step for velocity
logpxz1, g = fgrad(xnew)
for i in xnew:
vnew[i] += 0.5 * stepsize * g[i]
# === METROPOLIS-HASTINGS ACCEPT/REJECT
# Compute old and new Hamiltonians
k0 = 0
k1 = 0
for i in vnew:
k0 += (v0[i] ** 2).sum(axis=0, keepdims=True)
k1 += (vnew[i] ** 2).sum(axis=0, keepdims=True)
h0 = -logpxz0 + 0.5 * k0
h1 = -logpxz1 + 0.5 * k1
# print logpxz0, k0, logpxz1, k1
# print h0-h1
# Perform Metropolis-Hasting step
accept = np.exp(h0 - h1) >= np.random.uniform(size=h1.shape)
accept = accept.astype(float)
for i in xnew:
accept2 = np.dot(np.ones((xnew[i].shape[0], 1)), accept)
x0[i] = (1 - accept2) * x0[i] + accept2 * xnew[i]
# result: updated 'x0'
logpxz = (1 - accept) * logpxz0 + accept * logpxz1
return logpxz, accept
