def srr_step_func(ss, zi_zmuv, p_masks, q_masks, si, mi_p, mi_q):
# transform the current belief state into an observation
si_as_x = self._from_si_to_x(si)
full_grad = T.log(1.0 + T.exp(ss)) * (self.x_out - si_as_x)
# get the masked belief state and gradient for primary policy
xi_for_p = (mi_p * self.x_out) + ((1.0 - mi_p) * si_as_x)
grad_for_p = mi_p * full_grad
# update the guide policy's revelation mask
new_to_q = (1.0 - mi_q) * q_masks
mip1_q = mi_q + new_to_q
# get the masked belief state and gradient for guide policy
# xi_for_q = (mip1_q * self.x_out) + ((1.0 - mip1_q) * si_as_x)
xi_for_q = xi_for_p
grad_for_q = mip1_q * full_grad
# get samples of next zi, according to the primary policy
zi_p_mean, zi_p_logvar = self.p_zi_given_xi.apply(
T.horizontal_stack(xi_for_p, grad_for_p), do_samples=False
)
zi_p = zi_p_mean + (T.exp(0.5 * zi_p_logvar) * zi_zmuv)
# get samples of next zi, according to the guide policy
zi_q_mean, zi_q_logvar = self.q_zi_given_xi.apply(
T.horizontal_stack(xi_for_q, grad_for_q), do_samples=False
)
zi_q = zi_q_mean + (T.exp(0.5 * zi_q_logvar) * zi_zmuv)
# make zi samples that can be switched between zi_p and zi_q
zi = (self.train_switch[0] * zi_q) + ((1.0 - self.train_switch[0]) * zi_p)
# compute relevant KLds for this step
kldi_q2p = gaussian_kld(zi_q_mean, zi_q_logvar, zi_p_mean, zi_p_logvar) # KL(q || p)
kldi_p2q = gaussian_kld(zi_p_mean, zi_p_logvar, zi_q_mean, zi_q_logvar) # KL(p || q)
kldi_p2g = gaussian_kld(zi_p_mean, zi_p_logvar, 0.0, 0.0) # KL(p || N(0, I))
# compute next si, given sampled zi (i.e. update the belief state)
hydra_out = self.p_sip1_given_zi.apply(zi)
si_step = hydra_out[0]
if self.step_type == "jump":
# jump steps always do a full swap of belief state
sip1 = si_step
else:
# additive steps adjust the belief state like an LSTM
write_gate = T.nnet.sigmoid(2.0 + hydra_out[1])
erase_gate = T.nnet.sigmoid(2.0 + hydra_out[2])
sip1 = (erase_gate * si) + (write_gate * si_step)
# update the primary policy's revelation mask
new_to_p = (1.0 - mi_p) * p_masks
mip1_p = mi_p + new_to_p
# compute NLL only for the newly revealed values
nlli = self._construct_nll_costs(sip1, self.x_out, new_to_p)
# each loop iteration produces the following values:
# sip1: belief state at end of current step
# mip1_p: revealed values mask to use in next step (primary)
# mip1_q: revealed values mask to use in next step (guide)
# nlli: NLL for values revealed at end of current step
# kldi_q2p: KL(q || p) for the current step
# kldi_p2q: KL(p || q) for the current step
# kldi_p2g: KL(p || N(0,I)) for the current step
return sip1, mip1_p, mip1_q, nlli, kldi_q2p, kldi_p2q, kldi_p2g
def imp_step_func(zi_zmuv, si):
si_as_x = self._si_as_x(si)
xi_unmasked = self.x_out
xi_masked = (self.x_mask * xi_unmasked) + \
((1.0 - self.x_mask) * si_as_x)
grad_unmasked = self.x_out - si_as_x
grad_masked = self.x_mask * grad_unmasked
# get samples of next zi, according to the global policy
zi_p_mean, zi_p_logvar = self.p_zi_given_xi.apply(xi_masked)
zi_p = zi_p_mean + (T.exp(0.5 * zi_p_logvar) * zi_zmuv)
# get samples of next zi, according to the guide policy
zi_q_mean, zi_q_logvar = self.q_zi_given_xi.apply(
T.concatenate([xi_masked, xi_unmasked], axis=1))
zi_q = zi_q_mean + (T.exp(0.5 * zi_q_logvar) * zi_zmuv)
# make zi samples that can be switched between zi_p and zi_q
zi = ((self.train_switch[0] * zi_q) + \
((1.0 - self.train_switch[0]) * zi_p))
# compute relevant KLds for this step
kldi_q2p = gaussian_kld(zi_q_mean, zi_q_logvar,
zi_p_mean, zi_p_logvar) # KL(q || p)
kldi_p2q = gaussian_kld(zi_p_mean, zi_p_logvar,
zi_q_mean, zi_q_logvar) # KL(p || q)
kldi_p2g = gaussian_kld(zi_p_mean, zi_p_logvar,
0.0, 0.0) # KL(p || global prior)
# compute the next si, given the sampled zi
hydra_out = self.p_sip1_given_zi.apply(zi)
si_step = hydra_out[0]
if (self.step_type == 'jump'):
# jump steps always completely overwrite the current guesses
sip1 = si_step
elif (self.step_type == 'add'):
# add steps just update the guesses additively
sip1 = si + si_step
elif (self.step_type == 'lstm'):
# LSTM-style updates with write and erase gates
write_gate = 1.1 * T.nnet.sigmoid(1.0 + hydra_out[1])
erase_gate = 1.1 * T.nnet.sigmoid(1.0 + hydra_out[2])
sip1 = (erase_gate * si) + (write_gate * si_step)
elif (self.step_type == 'layer'):
alpha_gate = T.nnet.sigmoid(hydra_out[1])
sip1 = (alpha_gate * si) + ((1.0 - alpha_gate) * si_step)
else:
assert False, "Unknown step type!"
# compute NLL for the current imputation
nlli = self._construct_nll_costs(sip1, self.x_out, self.x_mask)
return sip1, nlli, kldi_q2p, kldi_p2q, kldi_p2g
def imp_step_func(zi_zmuv, si):
si_as_x = self._si_as_x(si)
xi_unmasked = self.x_out
xi_masked = (self.x_mask * xi_unmasked) + \
((1.0 - self.x_mask) * si_as_x)
grad_unmasked = self.x_out - si_as_x
grad_masked = self.x_mask * grad_unmasked
# get samples of next zi, according to the global policy
zi_p_mean, zi_p_logvar = self.p_zi_given_xi.apply(xi_masked)
zi_p = zi_p_mean + (T.exp(0.5 * zi_p_logvar) * zi_zmuv)
# get samples of next zi, according to the guide policy
zi_q_mean, zi_q_logvar = self.q_zi_given_xi.apply(
T.concatenate([xi_masked, xi_unmasked], axis=1))
zi_q = zi_q_mean + (T.exp(0.5 * zi_q_logvar) * zi_zmuv)
# make zi samples that can be switched between zi_p and zi_q
zi = ((self.train_switch[0] * zi_q) + \
((1.0 - self.train_switch[0]) * zi_p))
# compute relevant KLds for this step
kldi_q2p = gaussian_kld(zi_q_mean, zi_q_logvar, zi_p_mean,
zi_p_logvar) # KL(q || p)
kldi_p2q = gaussian_kld(zi_p_mean, zi_p_logvar, zi_q_mean,
zi_q_logvar) # KL(p || q)
kldi_p2g = gaussian_kld(zi_p_mean, zi_p_logvar, 0.0,
0.0) # KL(p || global prior)
# compute the next si, given the sampled zi
hydra_out = self.p_sip1_given_zi.apply(zi)
si_step = hydra_out[0]
if (self.step_type == 'jump'):
# jump steps always completely overwrite the current guesses
sip1 = si_step
elif (self.step_type == 'add'):
# add steps just update the guesses additively
sip1 = si + si_step
elif (self.step_type == 'lstm'):
# LSTM-style updates with write and erase gates
write_gate = 1.1 * T.nnet.sigmoid(1.0 + hydra_out[1])
erase_gate = 1.1 * T.nnet.sigmoid(1.0 + hydra_out[2])
sip1 = (erase_gate * si) + (write_gate * si_step)
elif (self.step_type == 'layer'):
alpha_gate = T.nnet.sigmoid(hydra_out[1])
sip1 = (alpha_gate * si) + ((1.0 - alpha_gate) * si_step)
else:
assert False, "Unknown step type!"
# compute NLL for the current imputation
nlli = self._construct_nll_costs(sip1, self.x_out, self.x_mask)
return sip1, nlli, kldi_q2p, kldi_p2q, kldi_p2g
def _construct_compute_fe_terms(self):
"""
Construct theano function to compute the log-likelihood and posterior
KL-divergence terms for the variational free-energy.
"""
# setup some symbolic variables for theano to deal with
Xd = T.matrix()
Xc = T.zeros_like(Xd)
Xm = T.zeros_like(Xd)
# construct values to output
if self.x_type == 'bernoulli':
ll_term = log_prob_bernoulli(self.x, self.xg)
else:
ll_term = log_prob_gaussian2(self.x, self.xg, \
log_vars=self.bounded_logvar)
all_klds = gaussian_kld(self.q_z_given_x.output_mean, \
self.q_z_given_x.output_logvar, \
self.prior_mean, self.prior_logvar)
kld_term = T.sum(all_klds, axis=1)
# compile theano function for a one-sample free-energy estimate
fe_term_sample = theano.function(inputs=[Xd], \
outputs=[ll_term, kld_term], \
givens={self.Xd: Xd, self.Xc: Xc, self.Xm: Xm})
# construct a wrapper function for multi-sample free-energy estimate
def fe_term_estimator(X, sample_count):
ll_sum = np.zeros((X.shape[0],))
kld_sum = np.zeros((X.shape[0],))
for i in range(sample_count):
result = fe_term_sample(X)
ll_sum = ll_sum + result[0].ravel()
kld_sum = kld_sum + result[1].ravel()
mean_nll = -ll_sum / float(sample_count)
mean_kld = kld_sum / float(sample_count)
return [mean_nll, mean_kld]
return fe_term_estimator
def _construct_chain_kld_cost(self, cost_decay=0.1):
"""
Construct the posterior KL-d from prior part of cost to minimize.
This is for operation in "free chain" mode, where a seed point is used
to initialize a long(ish) running markov chain.
"""
assert((cost_decay >= 0.0) and (cost_decay <= 1.0))
obs_count = T.cast(self.Xd.shape[0], 'floatX')
kld_mean = self.IN.kld_mean[0]
kld_costs = []
step_weight = 1.0
step_weights = []
step_decay = cost_decay
for i in range(self.chain_len):
IN_i = self.IN_chain[i]
# basic variational term on KL divergence between post and prior
kld_i = gaussian_kld(IN_i.output_mean, IN_i.output_logvar, \
self.prior_mean, self.prior_logvar)
kld_i_costs = T.sum(kld_i, axis=1)
# sum and reweight the KLd cost for this step in the chain
c = T.mean(kld_i_costs)
kld_costs.append(step_weight * c)
step_weights.append(step_weight)
step_weight = step_weight * step_decay
kld_cost = sum(kld_costs) / sum(step_weights)
return kld_cost
def _construct_chain_kld_cost(self, cost_decay=0.1):
"""
Construct the posterior KL-d from prior part of cost to minimize.
This is for operation in "free chain" mode, where a seed point is used
to initialize a long(ish) running markov chain.
"""
assert ((cost_decay >= 0.0) and (cost_decay <= 1.0))
obs_count = T.cast(self.Xd.shape[0], 'floatX')
kld_mean = self.IN.kld_mean[0]
kld_costs = []
step_weight = 1.0
step_weights = []
step_decay = cost_decay
for i in range(self.chain_len):
IN_i = self.IN_chain[i]
# basic variational term on KL divergence between post and prior
kld_i = gaussian_kld(IN_i.output_mean, IN_i.output_logvar, \
self.prior_mean, self.prior_logvar)
kld_i_costs = T.sum(kld_i, axis=1)
# sum and reweight the KLd cost for this step in the chain
c = T.mean(kld_i_costs)
kld_costs.append(step_weight * c)
step_weights.append(step_weight)
step_weight = step_weight * step_decay
kld_cost = sum(kld_costs) / sum(step_weights)
return kld_cost
def _construct_compute_fe_terms(self):
"""
Construct theano function to compute the log-likelihood and posterior
KL-divergence terms for the variational free-energy.
"""
# construct values to output
if self.x_type == 'bernoulli':
ll_term = log_prob_bernoulli(self.x_in, self.xg)
else:
ll_term = log_prob_gaussian2(self.x_in, self.xg, \
log_vars=self.bounded_logvar)
all_klds = gaussian_kld(self.z_mean, self.z_logvar, \
self.prior_mean, self.prior_logvar)
kld_term = T.sum(all_klds, axis=1)
# compile theano function for a one-sample free-energy estimate
fe_term_sample = theano.function(inputs=[self.x_in], \
outputs=[ll_term, kld_term])
# construct a wrapper function for multi-sample free-energy estimate
def fe_term_estimator(X, sample_count):
X = to_fX(X)
ll_sum = np.zeros((X.shape[0], ))
kld_sum = np.zeros((X.shape[0], ))
for i in range(sample_count):
result = fe_term_sample(X)
ll_sum = ll_sum + result[0].ravel()
kld_sum = kld_sum + result[1].ravel()
mean_nll = -ll_sum / float(sample_count)
mean_kld = kld_sum / float(sample_count)
return [mean_nll, mean_kld]
return fe_term_estimator
def _construct_compute_fe_terms(self):
"""
Construct theano function to compute the log-likelihood and posterior
KL-divergence terms for the variational free-energy.
"""
# construct values to output
if self.x_type == 'bernoulli':
ll_term = log_prob_bernoulli(self.x_in, self.xg)
else:
ll_term = log_prob_gaussian2(self.x_in, self.xg, \
log_vars=self.bounded_logvar)
all_klds = gaussian_kld(self.z_mean, self.z_logvar, \
self.prior_mean, self.prior_logvar)
kld_term = T.sum(all_klds, axis=1)
# compile theano function for a one-sample free-energy estimate
fe_term_sample = theano.function(inputs=[self.x_in], \
outputs=[ll_term, kld_term])
# construct a wrapper function for multi-sample free-energy estimate
def fe_term_estimator(X, sample_count):
X = to_fX(X)
ll_sum = np.zeros((X.shape[0],))
kld_sum = np.zeros((X.shape[0],))
for i in range(sample_count):
result = fe_term_sample(X)
ll_sum = ll_sum + result[0].ravel()
kld_sum = kld_sum + result[1].ravel()
mean_nll = -ll_sum / float(sample_count)
mean_kld = kld_sum / float(sample_count)
return [mean_nll, mean_kld]
return fe_term_estimator
def imp_step_func(zi_zmuv, si):
si_as_x = self.obs_transform(si)
xi_masked = (self.x_mask * self.x_out) + \
((1.0 - self.x_mask) * si_as_x)
#grad_ll = self.x_out - xi_masked
# get samples of next zi, according to the global policy
zi_p_mean, zi_p_logvar = self.p_zi_given_xi.apply( \
xi_masked, do_samples=False)
zi_p = zi_p_mean + (T.exp(0.5 * zi_p_logvar) * zi_zmuv)
# get samples of next zi, according to the guide policy
zi_q_mean, zi_q_logvar = self.q_zi_given_x_xi.apply( \
T.horizontal_stack(xi_masked, self.x_out), \
do_samples=False)
zi_q = zi_q_mean + (T.exp(0.5 * zi_q_logvar) * zi_zmuv)
if self.use_osm_mode:
zi = zi_p
# compute relevant KLds for this step
kldi_q2p = gaussian_kld(zi_p_mean, zi_p_logvar, 0.0, 0.0)
kldi_p2q = gaussian_kld(zi_p_mean, zi_p_logvar, 0.0, 0.0)
else:
# make zi samples that can be switched between zi_p and zi_q
zi = ((self.train_switch[0] * zi_q) + \
((1.0 - self.train_switch[0]) * zi_p))
# compute relevant KLds for this step
kldi_q2p = gaussian_kld(zi_q_mean, zi_q_logvar, \
zi_p_mean, zi_p_logvar)
kldi_p2q = gaussian_kld(zi_p_mean, zi_p_logvar, \
zi_q_mean, zi_q_logvar)
# compute the next si, given the sampled zi
hydra_out = self.p_xip1_given_zi.apply(zi)
si_step = hydra_out[0]
if (self.step_type == 'jump'):
# jump steps always do a full swap (like standard VAE)
sip1 = si_step
else:
# additive steps adjust the current guesses incrementally
write_gate = T.nnet.sigmoid(2.0 + hydra_out[1])
erase_gate = T.nnet.sigmoid(2.0 + hydra_out[2])
# LSTM-style update
sip1 = (erase_gate * si) + (write_gate * si_step)
# normal update (this was used in workshop papers)
#sip1 = si + si_step
# compute NLL for the current imputation
nlli = self._construct_nll_costs(sip1, self.x_out, 0.0*self.x_mask)
return sip1, nlli, kldi_q2p, kldi_p2q
def imp_step_func(zi_zmuv, si):
si_as_x = self._from_si_to_x(si)
xi_unmasked = self.x_out
xi_masked = (self.x_mask * xi_unmasked) + \
((1.0 - self.x_mask) * si_as_x)
grad_unmasked = self.x_out - si_as_x
grad_masked = (self.x_mask * grad_unmasked) + \
((1.0 - self.x_mask) * self.grad_null)
# get samples of next zi, according to the global policy
zi_p_mean, zi_p_logvar = self.p_zi_given_xi.apply( \
T.horizontal_stack(xi_masked, grad_masked), \
do_samples=False)
zi_p = zi_p_mean + (T.exp(0.5 * zi_p_logvar) * zi_zmuv)
# get samples of next zi, according to the guide policy
zi_q_mean, zi_q_logvar = self.q_zi_given_xi.apply( \
T.horizontal_stack(xi_masked, grad_unmasked), \
do_samples=False)
zi_q = zi_q_mean + (T.exp(0.5 * zi_q_logvar) * zi_zmuv)
# make zi samples that can be switched between zi_p and zi_q
zi = ((self.train_switch[0] * zi_q) + \
((1.0 - self.train_switch[0]) * zi_p))
# compute relevant KLds for this step
kldi_q2p = gaussian_kld(zi_q_mean, zi_q_logvar, \
zi_p_mean, zi_p_logvar) # KL(q || p)
kldi_p2q = gaussian_kld(zi_p_mean, zi_p_logvar, \
zi_q_mean, zi_q_logvar) # KL(p || q)
kldi_p2g = gaussian_kld(zi_p_mean, zi_p_logvar, \
0.0, 0.0) # KL(p || global prior)
# compute the next si, given the sampled zi
hydra_out = self.p_sip1_given_zi.apply(zi)
si_step = hydra_out[0]
if (self.step_type == 'jump'):
# jump steps always completely overwrite the current guesses
sip1 = si_step
else:
# additive steps update the current guesses like an LSTM
write_gate = T.nnet.sigmoid(3.0 + hydra_out[1])
erase_gate = T.nnet.sigmoid(3.0 + hydra_out[2])
sip1 = (erase_gate * si) + (write_gate * si_step)
# compute NLL for the current imputation
nlli = self._construct_nll_costs(sip1, self.x_out, self.x_mask)
return sip1, nlli, kldi_q2p, kldi_p2q, kldi_p2g
def _construct_kld_costs(self):
"""
Construct the posterior KL-d from prior part of cost to minimize.
"""
# compute the KLds between posteriors and priors. we compute the KLd
# independently for each input and each latent variable dimension
kld_z = gaussian_kld(self.z_mean, self.z_logvar, \
self.prior_mean, self.prior_logvar)
kld_costs = T.sum(kld_z, axis=1, keepdims=True)
return kld_costs
def _construct_kld_costs(self):
"""
Construct the posterior KL-divergence part of cost to minimize.
"""
# construct a penalty that is L2-like near 0 and L1-like away from 0.
huber_pen = lambda x, d: \
((1.0 / (2.0 * d)) * ((T.abs_(x) < d) * (x**2.0))) + \
((T.abs_(x) >= d) * (T.abs_(x) - (d / 2.0)))
# construct KLd cost for the distributions over hi. the prior over
# hi is given by a distribution conditioned on si, which we estimate
# using self.p_hi_given_si[i]. the conditionals produced by each
# self.p_hi_given_si[i] will also be regularized towards a shared
# prior, e.g. a Gaussian with zero mean and unit variance.
kld_hi_conds = []
kld_hi_globs = []
for i in range(self.ir_steps):
kld_hi_cond = gaussian_kld( \
self.q_hi_given_x_si[i].output_mean, \
self.q_hi_given_x_si[i].output_logvar, \
self.p_hi_given_si[i].output_mean, \
self.p_hi_given_si[i].output_logvar)
kld_hi_glob = gaussian_kld( \
self.p_hi_given_si[i].output_mean, \
self.p_hi_given_si[i].output_logvar, \
0.0, 0.0)
kld_hi_cond_l1l2 = (self.l1l2_weight[0] * kld_hi_cond) + \
((1.0 - self.l1l2_weight[0]) * kld_hi_cond**2.0)
kld_hi_conds.append(T.sum(kld_hi_cond_l1l2, \
axis=1, keepdims=True))
kld_hi_globs.append(T.sum(kld_hi_glob**2.0, \
axis=1, keepdims=True))
# compute the batch-wise costs
kld_hi_cond = sum(kld_hi_conds)
kld_hi_glob = sum(kld_hi_globs)
# construct KLd cost for the distributions over z
kld_z_all = gaussian_kld(self.q_z_given_x.output_mean, \
self.q_z_given_x.output_logvar, \
0.0, 0.0)
kld_z_l1l2 = (self.l1l2_weight[0] * kld_z_all) + \
((1.0 - self.l1l2_weight[0]) * kld_z_all**2.0)
kld_z = T.sum(kld_z_l1l2, \
axis=1, keepdims=True)
return [kld_z, kld_hi_cond, kld_hi_glob]
def _construct_kld_costs(self):
"""
Construct the posterior KL-divergence part of cost to minimize.
"""
# construct a penalty that is L2-like near 0 and L1-like away from 0.
huber_pen = lambda x, d: \
((1.0 / (2.0 * d)) * ((T.abs_(x) < d) * (x**2.0))) + \
((T.abs_(x) >= d) * (T.abs_(x) - (d / 2.0)))
# construct KLd cost for the distributions over hi. the prior over
# hi is given by a distribution conditioned on si, which we estimate
# using self.p_hi_given_si[i]. the conditionals produced by each
# self.p_hi_given_si[i] will also be regularized towards a shared
# prior, e.g. a Gaussian with zero mean and unit variance.
kld_hi_conds = []
kld_hi_globs = []
for i in range(self.ir_steps):
kld_hi_cond = gaussian_kld( \
self.q_hi_given_x_si[i].output_mean, \
self.q_hi_given_x_si[i].output_logvar, \
self.p_hi_given_si[i].output_mean, \
self.p_hi_given_si[i].output_logvar)
kld_hi_glob = gaussian_kld( \
self.p_hi_given_si[i].output_mean, \
self.p_hi_given_si[i].output_logvar, \
0.0, 0.0)
kld_hi_cond_l1l2 = (self.l1l2_weight[0] * kld_hi_cond) + \
((1.0 - self.l1l2_weight[0]) * kld_hi_cond**2.0)
kld_hi_conds.append(T.sum(kld_hi_cond_l1l2, \
axis=1, keepdims=True))
kld_hi_globs.append(T.sum(kld_hi_glob**2.0, \
axis=1, keepdims=True))
# compute the batch-wise costs
kld_hi_cond = sum(kld_hi_conds)
kld_hi_glob = sum(kld_hi_globs)
# construct KLd cost for the distributions over z
kld_z_all = gaussian_kld(self.q_z_given_x.output_mean, \
self.q_z_given_x.output_logvar, \
0.0, 0.0)
kld_z_l1l2 = (self.l1l2_weight[0] * kld_z_all) + \
((1.0 - self.l1l2_weight[0]) * kld_z_all**2.0)
kld_z = T.sum(kld_z_l1l2, \
axis=1, keepdims=True)
return [kld_z, kld_hi_cond, kld_hi_glob]
def _construct_kld_costs(self):
"""
Construct the posterior KL-d from prior part of cost to minimize.
"""
# compute the KLds between posteriors and priors. we compute the KLd
# independently for each input and each latent variable dimension
kld_z = gaussian_kld(self.z_mean, self.z_logvar, \
self.prior_mean, self.prior_logvar)
kld_costs = T.sum(kld_z, axis=1, keepdims=True)
return kld_costs
def _construct_compute_post_klds(self):
"""
Construct theano function to compute the info about the variational
approximate posteriors for some inputs.
"""
# setup some symbolic variables for theano to deal with
all_klds = gaussian_kld(self.z_mean, self.z_logvar, \
self.prior_mean, self.prior_logvar)
# compile theano function for a one-sample free-energy estimate
kld_func = theano.function(inputs=[self.x_in], outputs=all_klds)
return kld_func
def _construct_compute_post_klds(self):
"""
Construct theano function to compute the info about the variational
approximate posteriors for some inputs.
"""
# setup some symbolic variables for theano to deal with
all_klds = gaussian_kld(self.z_mean, self.z_logvar, \
self.prior_mean, self.prior_logvar)
# compile theano function for a one-sample free-energy estimate
kld_func = theano.function(inputs=[self.x_in], outputs=all_klds)
return kld_func
def _construct_compute_post_klds(self):
"""
Construct theano function to compute the info about the variational
approximate posteriors for some inputs.
"""
# setup some symbolic variables for theano to deal with
x = T.matrix()
# construct symbolic expressions for the desired KLds
cond_klds = []
glob_klds = []
for i in range(self.ir_steps):
kld_hi_cond = gaussian_kld(self.q_hi_given_x_si[i].output_mean, \
self.q_hi_given_x_si[i].output_logvar, \
self.p_hi_given_si[i].output_mean, \
self.p_hi_given_si[i].output_logvar)
kld_hi_glob = gaussian_kld(self.p_hi_given_si[i].output_mean, \
self.p_hi_given_si[i].output_logvar, 0.0, 0.0)
cond_klds.append(kld_hi_cond)
glob_klds.append(kld_hi_glob)
# gather conditional and global klds for all IR steps
all_klds = cond_klds + glob_klds
# gather kld for the initialization step
kld_z_all = gaussian_kld(self.q_z_given_x.output_mean, \
self.q_z_given_x.output_logvar, \
0.0, 0.0)
all_klds.append(kld_z_all)
# compile theano function for a one-sample free-energy estimate
kld_func = theano.function(inputs=[x], outputs=all_klds, \
givens={ self.x: x })
def post_kld_computer(X):
f_all_klds = kld_func(X)
f_kld_z = f_all_klds[-1]
f_kld_hi_cond = np.zeros(f_all_klds[0].shape)
f_kld_hi_glob = np.zeros(f_all_klds[0].shape)
for j in range(self.ir_steps):
f_kld_hi_cond += f_all_klds[j]
f_kld_hi_glob += f_all_klds[j + self.ir_steps]
return [f_kld_z, f_kld_hi_cond, f_kld_hi_glob]
return post_kld_computer
def _construct_kld_cost(self):
"""
Compute (analytically) the KL divergence between each approximate
posterior encoded by self.mu/self.sigma and the isotropic Gaussian
distribution with mean 0 and standard deviation self.prior_sigma.
"""
prior_mu = 0.0
prior_logvar = np.log(self.prior_sigma**2.0)
post_klds = gaussian_kld(self.output_mean, self.output_logvar, \
prior_mu, prior_logvar)
kld_cost = T.sum(post_klds, axis=1, keepdims=True)
return kld_cost
def _construct_kld_costs(self):
"""
Construct the posterior KL-d from prior part of cost to minimize.
"""
# compute the KLds between posteriors and priors. we compute the KLd
# independently for each input and each latent variable dimension
kld_z = gaussian_kld(self.z_mean, self.z_logvar, \
self.prior_mean, self.prior_logvar)
# compute the batch-wise L1 and L2 penalties on per-dim KLds
kld_l1_costs = T.sum(kld_z, axis=1, keepdims=True)
kld_l2_costs = (kld_l1_costs - self.kld_z_mean[0])**2.0
return [kld_l1_costs, kld_l2_costs]
def _construct_compute_post_klds(self):
"""
Construct theano function to compute the info about the variational
approximate posteriors for some inputs.
"""
# setup some symbolic variables for theano to deal with
x = T.matrix()
# construct symbolic expressions for the desired KLds
cond_klds = []
glob_klds = []
for i in range(self.ir_steps):
kld_hi_cond = gaussian_kld(self.q_hi_given_x_si[i].output_mean, \
self.q_hi_given_x_si[i].output_logvar, \
self.p_hi_given_si[i].output_mean, \
self.p_hi_given_si[i].output_logvar)
kld_hi_glob = gaussian_kld(self.p_hi_given_si[i].output_mean, \
self.p_hi_given_si[i].output_logvar, 0.0, 0.0)
cond_klds.append(kld_hi_cond)
glob_klds.append(kld_hi_glob)
# gather conditional and global klds for all IR steps
all_klds = cond_klds + glob_klds
# gather kld for the initialization step
kld_z_all = gaussian_kld(self.q_z_given_x.output_mean, \
self.q_z_given_x.output_logvar, \
0.0, 0.0)
all_klds.append(kld_z_all)
# compile theano function for a one-sample free-energy estimate
kld_func = theano.function(inputs=[x], outputs=all_klds, \
givens={ self.x: x })
def post_kld_computer(X):
f_all_klds = kld_func(X)
f_kld_z = f_all_klds[-1]
f_kld_hi_cond = np.zeros(f_all_klds[0].shape)
f_kld_hi_glob = np.zeros(f_all_klds[0].shape)
for j in range(self.ir_steps):
f_kld_hi_cond += f_all_klds[j]
f_kld_hi_glob += f_all_klds[j + self.ir_steps]
return [f_kld_z, f_kld_hi_cond, f_kld_hi_glob]
return post_kld_computer
def _construct_kld_costs(self, p=1.0):
"""
Construct the posterior KL-divergence part of cost to minimize.
"""
kld_hi_q2ps = []
kld_hi_p2qs = []
for i in range(self.ir_steps):
kld_hi_q2p = self.kldi_q2p[i]
kld_hi_p2q = self.kldi_p2q[i]
kld_hi_q2ps.append(T.sum(kld_hi_q2p**p, \
axis=1, keepdims=True))
kld_hi_p2qs.append(T.sum(kld_hi_p2q**p, \
axis=1, keepdims=True))
# compute the batch-wise costs
kld_hi_q2p = sum(kld_hi_q2ps)
kld_hi_p2q = sum(kld_hi_p2qs)
# construct KLd cost for the distributions over z
kld_z_q2ps = gaussian_kld(self.q_z_mean, self.q_z_logvar, \
self.p_z_mean, self.p_z_logvar)
kld_z_p2qs = gaussian_kld(self.p_z_mean, self.p_z_logvar, \
self.q_z_mean, self.q_z_logvar)
kld_z_q2p = T.sum(kld_z_q2ps**p, axis=1, keepdims=True)
kld_z_p2q = T.sum(kld_z_p2qs**p, axis=1, keepdims=True)
return [kld_z_q2p, kld_z_p2q, kld_hi_q2p, kld_hi_p2q]
def ir_step_func(hi_zmuv, sim1):
# get variables used throughout this refinement step
sim1_obs = self.obs_transform(sim1) # transform state -> obs
grad_ll = self.x_out - sim1_obs
# get samples of next hi, conditioned on current si
hi_p_mean, hi_p_logvar = self.p_hi_given_si.apply( \
sim1_obs, do_samples=False)
# now we build the model for variational hi given si
hi_q_mean, hi_q_logvar = self.q_hi_given_x_si.apply( \
T.horizontal_stack(grad_ll, sim1_obs), \
do_samples=False)
hi_q = (T.exp(0.5 * hi_q_logvar) * hi_zmuv) + hi_q_mean
hi_p = (T.exp(0.5 * hi_p_logvar) * hi_zmuv) + hi_p_mean
# make hi samples that can be switched between hi_p and hi_q
hi = ( ((self.train_switch[0] * hi_q) + \
((1.0 - self.train_switch[0]) * hi_p)) )
# p_sip1_given_si_hi is conditioned on si and hi.
ig_vals, fg_vals, in_vals = self.p_sip1_given_si_hi.apply(hi)
# get the transformed values (for an LSTM style update)
i_gate = 1.0 * T.nnet.sigmoid(ig_vals + 2.0)
f_gate = 1.0 * T.nnet.sigmoid(fg_vals + 2.0)
# perform an LSTM-like update of the state sim1 -> si
si = (in_vals * i_gate) + (sim1 * f_gate)
# compute generator NLL for this step
nlli = self.log_prob_func(self.x_out, self.obs_transform(si))
# compute relevant KLds for this step
kldi_q2p = gaussian_kld(hi_q_mean, hi_q_logvar, \
hi_p_mean, hi_p_logvar)
kldi_p2q = gaussian_kld(hi_p_mean, hi_p_logvar, \
hi_q_mean, hi_q_logvar)
return si, nlli, kldi_q2p, kldi_p2q
def _construct_compute_post_klds(self):
"""
Construct theano function to compute the info about the variational
approximate posteriors for some inputs.
"""
# setup some symbolic variables for theano to deal with
Xd = T.matrix()
Xc = T.zeros_like(Xd)
Xm = T.zeros_like(Xd)
all_klds = gaussian_kld(self.q_z_given_x.output_mean, \
self.q_z_given_x.output_logvar, \
self.prior_mean, self.prior_logvar)
# compile theano function for a one-sample free-energy estimate
kld_func = theano.function(inputs=[Xd], outputs=all_klds, \
givens={self.Xd: Xd, self.Xc: Xc, self.Xm: Xm})
return kld_func
def _construct_compute_post_klds(self):
"""
Construct theano function to compute the info about the variational
approximate posteriors for some inputs.
"""
# setup some symbolic variables for theano to deal with
Xd = T.matrix()
Xc = T.zeros_like(Xd)
Xm = T.zeros_like(Xd)
all_klds = gaussian_kld(self.q_z_given_x.output_mean, \
self.q_z_given_x.output_logvar, \
self.prior_mean, self.prior_logvar)
# compile theano function for a one-sample free-energy estimate
kld_func = theano.function(inputs=[Xd], outputs=all_klds, \
givens={self.Xd: Xd, self.Xc: Xc, self.Xm: Xm})
return kld_func
def _construct_kld_costs(self):
"""
Construct the posterior KL-d from prior part of cost to minimize.
"""
# compute the KLds between posteriors and priors. we compute the KLd
# independently for each input and each latent variable dimension
kld_z = gaussian_kld(self.q_z_given_x.output_mean, \
self.q_z_given_x.output_logvar, \
self.prior_mean, self.prior_logvar)
# compute the batch-wise L1 and L2 penalties on per-dim KLds
kld_l1_costs = T.sum(kld_z, axis=1, keepdims=True)
derp1 = T.mean(kld_l1_costs)
derp2 = kld_l1_costs > derp1
batch_kld_mean = theano.gradient.disconnected_grad(derp1)
mask = theano.gradient.disconnected_grad(derp2)
kld_l2_costs = T.sum(((kld_l1_costs-batch_kld_mean)**2.0 * mask), \
axis=1, keepdims=True)
return [kld_l1_costs, kld_l2_costs]
def _construct_kld_costs(self):
"""
Construct the posterior KL-d from prior part of cost to minimize.
"""
# compute the KLds between posteriors and priors. we compute the KLd
# independently for each input and each latent variable dimension
kld_z = gaussian_kld(self.q_z_given_x.output_mean, \
self.q_z_given_x.output_logvar, \
self.prior_mean, self.prior_logvar)
# compute the batch-wise L1 and L2 penalties on per-dim KLds
kld_l1_costs = T.sum(kld_z, axis=1, keepdims=True)
derp1 = T.mean(kld_l1_costs)
derp2 = kld_l1_costs > derp1
batch_kld_mean = theano.gradient.disconnected_grad(derp1)
mask = theano.gradient.disconnected_grad(derp2)
kld_l2_costs = T.sum(((kld_l1_costs-batch_kld_mean)**2.0 * mask), \
axis=1, keepdims=True)
return [kld_l1_costs, kld_l2_costs]
def _construct_compute_post_stats(self):
"""
Construct theano function to compute some stats describing the latent
posteriors inferred by this model.
"""
# construct the stats to compute
obs_count = T.cast(self.Xd.shape[0], 'floatX')
prior_mean = 0.0
prior_logvar = 0.0
all_klds = gaussian_kld(self.IN.output_mean, self.IN.output_logvar, \
prior_mean, prior_logvar)
obs_klds = T.sum(all_klds, axis=1)
dim_klds = T.sum(all_klds, axis=0) / obs_count
dim_vars = T.sum(self.IN.output_mean**2.0, axis=0) / obs_count
# make a theano function to compute them
outputs = [all_klds, obs_klds, dim_klds, dim_vars]
func = theano.function(inputs=[self.Xd, self.Xc, self.Xm], \
outputs=outputs)
return func
def chain_step_func(zi_zmuv, xim1):
# get mean and logvar of z samples for this step
zi_mean, zi_logvar = self.IN.apply(xim1, do_samples=False)
# transform ZMUV samples to get desired samples
zi = (T.exp(0.5 * zi_logvar) * zi_zmuv) + zi_mean
# get the next generated xi (pre-transformation)
outputs = self.GN.apply(zi)
xti = outputs[-1]
# apply the observation "mean" transform
xgi = self.xt_transform(xti)
# compute NLL for this step
if self.chain_type == 'walkout':
x_true = self.x_d
else:
x_true = xim1
nlli = self._log_prob(x_true, xgi).flatten()
kldi = T.sum(gaussian_kld(zi_mean, zi_logvar, \
self.p_z_mean, self.p_z_logvar), axis=1)
return xgi, nlli, kldi
def _construct_compute_fe_terms(self):
"""
Construct theano function to compute the log-likelihood and posterior
KL-divergence terms for the variational free-energy.
"""
# setup some symbolic variables for theano to deal with
Xd = T.matrix()
Xc = T.zeros_like(Xd)
Xm = T.zeros_like(Xd)
# construct values to output
if self.x_type == 'bernoulli':
ll_term = log_prob_bernoulli(self.x, self.xg)
else:
ll_term = log_prob_gaussian2(self.x, self.xg, \
log_vars=self.bounded_logvar)
all_klds = gaussian_kld(self.q_z_given_x.output_mean, \
self.q_z_given_x.output_logvar, \
self.prior_mean, self.prior_logvar)
kld_term = T.sum(all_klds, axis=1)
# compile theano function for a one-sample free-energy estimate
fe_term_sample = theano.function(inputs=[Xd], \
outputs=[ll_term, kld_term], \
givens={self.Xd: Xd, self.Xc: Xc, self.Xm: Xm})
# construct a wrapper function for multi-sample free-energy estimate
def fe_term_estimator(X, sample_count):
ll_sum = np.zeros((X.shape[0], ))
kld_sum = np.zeros((X.shape[0], ))
for i in range(sample_count):
result = fe_term_sample(X)
ll_sum = ll_sum + result[0].ravel()
kld_sum = kld_sum + result[1].ravel()
mean_nll = -ll_sum / float(sample_count)
mean_kld = kld_sum / float(sample_count)
return [mean_nll, mean_kld]
return fe_term_estimator
def __init__(self, rng=None, \
x_in=None, x_out=None, \
p_s0_given_z=None, \
p_hi_given_si=None, \
p_sip1_given_si_hi=None, \
q_z_given_x=None, \
q_hi_given_x_si=None, \
obs_dim=None, \
z_dim=None, h_dim=None, \
ir_steps=4, params=None, \
shared_param_dicts=None):
# setup a rng for this GIPair
self.rng = RandStream(rng.randint(100000))
# grab the user-provided parameters
self.params = params
self.x_type = self.params['x_type']
assert((self.x_type == 'bernoulli') or (self.x_type == 'gaussian'))
if 'obs_transform' in self.params:
assert((self.params['obs_transform'] == 'sigmoid') or \
(self.params['obs_transform'] == 'none'))
if self.params['obs_transform'] == 'sigmoid':
self.obs_transform = lambda x: T.nnet.sigmoid(20.0 * T.tanh(0.05 * x))
else:
self.obs_transform = lambda x: x
else:
self.obs_transform = lambda x: T.nnet.sigmoid(20.0 * T.tanh(0.05 * x))
if self.x_type == 'bernoulli':
self.obs_transform = lambda x: T.nnet.sigmoid(20.0 * T.tanh(0.05 * x))
self.shared_param_dicts = shared_param_dicts
# record the dimensions of various spaces relevant to this model
self.obs_dim = obs_dim
self.z_dim = z_dim
self.h_dim = h_dim
self.ir_steps = ir_steps
# grab handles to the relevant InfNets
self.q_z_given_x = q_z_given_x
self.q_hi_given_x_si = q_hi_given_x_si
self.p_s0_given_z = p_s0_given_z
self.p_hi_given_si = p_hi_given_si
self.p_sip1_given_si_hi = p_sip1_given_si_hi
# record the symbolic variables that will provide inputs to the
# computation graph created to describe this MultiStageModel
self.x_in = x_in
self.x_out = x_out
self.hi_zmuv = T.tensor3() # for ZMUV Gaussian samples to use in scan
# setup switching variable for changing between sampling/training
zero_ary = to_fX( np.zeros((1,)) )
self.train_switch = theano.shared(value=zero_ary, name='msm_train_switch')
self.set_train_switch(1.0)
# setup a variable for controlling dropout noise
self.drop_rate = theano.shared(value=zero_ary, name='msm_drop_rate')
self.set_drop_rate(0.0)
# this weight balances l1 vs. l2 penalty on posterior KLds
self.lam_kld_l1l2 = theano.shared(value=zero_ary, name='msm_lam_kld_l1l2')
self.set_lam_kld_l1l2(1.0)
if self.shared_param_dicts is None:
# initialize "optimizable" parameters specific to this MSM
init_vec = to_fX( np.zeros((self.z_dim,)) )
self.p_z_mean = theano.shared(value=init_vec, name='msm_p_z_mean')
self.p_z_logvar = theano.shared(value=init_vec, name='msm_p_z_logvar')
init_vec = to_fX( np.zeros((self.obs_dim,)) )
self.obs_logvar = theano.shared(value=zero_ary, name='msm_obs_logvar')
self.bounded_logvar = 8.0 * T.tanh((1.0/8.0) * self.obs_logvar)
self.shared_param_dicts = {}
self.shared_param_dicts['p_z_mean'] = self.p_z_mean
self.shared_param_dicts['p_z_logvar'] = self.p_z_logvar
self.shared_param_dicts['obs_logvar'] = self.obs_logvar
else:
self.p_z_mean = self.shared_param_dicts['p_z_mean']
self.p_z_logvar = self.shared_param_dicts['p_z_logvar']
self.obs_logvar = self.shared_param_dicts['obs_logvar']
self.bounded_logvar = 8.0 * T.tanh((1.0/8.0) * self.obs_logvar)
# setup a function for computing reconstruction log likelihood
if self.x_type == 'bernoulli':
self.log_prob_func = lambda xo, xh: \
(-1.0 * log_prob_bernoulli(xo, xh))
else:
self.log_prob_func = lambda xo, xh: \
(-1.0 * log_prob_gaussian2(xo, xh, \
log_vars=self.bounded_logvar))
# get a drop mask that drops things with probability p
drop_scale = 1. / (1. - self.drop_rate[0])
drop_rnd = self.rng.uniform(size=self.x_out.shape, \
low=0.0, high=1.0, dtype=theano.config.floatX)
drop_mask = drop_scale * (drop_rnd > self.drop_rate[0])
#############################
# Setup self.z and self.s0. #
#############################
print("Building MSM step 0...")
drop_x = drop_mask * self.x_in
self.q_z_mean, self.q_z_logvar, self.z = \
self.q_z_given_x.apply(drop_x, do_samples=True)
# get initial observation state
self.s0, _ = self.p_s0_given_z.apply(self.z, do_samples=False)
# gather KLd and NLL for the initialization step
self.init_klds = gaussian_kld(self.q_z_mean, self.q_z_logvar, \
self.p_z_mean, self.p_z_logvar)
self.init_nlls = -1.0 * \
self.log_prob_func(self.x_out, self.obs_transform(self.s0))
##################################################
# Setup the iterative generation loop using scan #
##################################################
def ir_step_func(hi_zmuv, sim1):
# get variables used throughout this refinement step
sim1_obs = self.obs_transform(sim1) # transform state -> obs
grad_ll = self.x_out - sim1_obs
# get samples of next hi, conditioned on current si
hi_p_mean, hi_p_logvar = self.p_hi_given_si.apply( \
sim1_obs, do_samples=False)
# now we build the model for variational hi given si
hi_q_mean, hi_q_logvar = self.q_hi_given_x_si.apply( \
T.horizontal_stack(grad_ll, sim1_obs), \
do_samples=False)
hi_q = (T.exp(0.5 * hi_q_logvar) * hi_zmuv) + hi_q_mean
hi_p = (T.exp(0.5 * hi_p_logvar) * hi_zmuv) + hi_p_mean
# make hi samples that can be switched between hi_p and hi_q
hi = ( ((self.train_switch[0] * hi_q) + \
((1.0 - self.train_switch[0]) * hi_p)) )
# p_sip1_given_si_hi is conditioned on si and hi.
ig_vals, fg_vals, in_vals = self.p_sip1_given_si_hi.apply(hi)
# get the transformed values (for an LSTM style update)
i_gate = 1.0 * T.nnet.sigmoid(ig_vals + 2.0)
f_gate = 1.0 * T.nnet.sigmoid(fg_vals + 2.0)
# perform an LSTM-like update of the state sim1 -> si
si = (in_vals * i_gate) + (sim1 * f_gate)
# compute generator NLL for this step
nlli = self.log_prob_func(self.x_out, self.obs_transform(si))
# compute relevant KLds for this step
kldi_q2p = gaussian_kld(hi_q_mean, hi_q_logvar, \
hi_p_mean, hi_p_logvar)
kldi_p2q = gaussian_kld(hi_p_mean, hi_p_logvar, \
hi_q_mean, hi_q_logvar)
return si, nlli, kldi_q2p, kldi_p2q
init_values = [self.s0, None, None, None]
self.scan_results, self.scan_updates = theano.scan(ir_step_func, \
outputs_info=init_values, sequences=self.hi_zmuv)
self.si = self.scan_results[0]
self.nlli = self.scan_results[1]
self.kldi_q2p = self.scan_results[2]
self.kldi_p2q = self.scan_results[3]
######################################################################
# ALL SYMBOLIC VARS NEEDED FOR THE OBJECTIVE SHOULD NOW BE AVAILABLE #
######################################################################
# shared var learning rate for generator and inferencer
zero_ary = to_fX( np.zeros((1,)) )
self.lr_1 = theano.shared(value=zero_ary, name='msm_lr_1')
self.lr_2 = theano.shared(value=zero_ary, name='msm_lr_2')
# shared var momentum parameters for generator and inferencer
self.mom_1 = theano.shared(value=zero_ary, name='msm_mom_1')
self.mom_2 = theano.shared(value=zero_ary, name='msm_mom_2')
# init parameters for controlling learning dynamics
self.set_sgd_params()
# init shared var for weighting nll of data given posterior sample
self.lam_nll = theano.shared(value=zero_ary, name='msm_lam_nll')
self.set_lam_nll(lam_nll=1.0)
# init shared var for weighting prior kld against reconstruction
self.lam_kld_z = theano.shared(value=zero_ary, name='msm_lam_kld_z')
self.lam_kld_q2p = theano.shared(value=zero_ary, name='msm_lam_kld_q2p')
self.lam_kld_p2q = theano.shared(value=zero_ary, name='msm_lam_kld_p2q')
self.set_lam_kld(lam_kld_z=1.0, lam_kld_q2p=0.7, lam_kld_p2q=0.3)
# init shared var for controlling l2 regularization on params
self.lam_l2w = theano.shared(value=zero_ary, name='msm_lam_l2w')
self.set_lam_l2w(1e-5)
# Grab all of the "optimizable" parameters in "group 1"
self.q_params = []
self.q_params.extend(self.q_z_given_x.mlp_params)
self.q_params.extend(self.q_hi_given_x_si.mlp_params)
# Grab all of the "optimizable" parameters in "group 2"
self.p_params = [self.p_z_mean, self.p_z_logvar]
self.p_params.extend(self.p_hi_given_si.mlp_params)
self.p_params.extend(self.p_sip1_given_si_hi.mlp_params)
self.p_params.extend(self.p_s0_given_z.mlp_params)
# Make a joint list of parameters group 1/2
self.joint_params = self.q_params + self.p_params
#################################
# CONSTRUCT THE KLD-BASED COSTS #
#################################
self.kld_z_q2p, self.kld_z_p2q, self.kld_hi_q2p, self.kld_hi_p2q = \
self._construct_kld_costs(p=1.0)
self.kld_z = (self.lam_kld_q2p[0] * self.kld_z_q2p) + \
(self.lam_kld_p2q[0] * self.kld_z_p2q)
self.kld_hi = (self.lam_kld_q2p[0] * self.kld_hi_q2p) + \
(self.lam_kld_p2q[0] * self.kld_hi_p2q)
self.kld_costs = (self.lam_kld_z[0] * self.kld_z) + self.kld_hi
# now do l2 KLd costs
self.kl2_z_q2p, self.kl2_z_p2q, self.kl2_hi_q2p, self.kl2_hi_p2q = \
self._construct_kld_costs(p=2.0)
self.kl2_z = (self.lam_kld_q2p[0] * self.kl2_z_q2p) + \
(self.lam_kld_p2q[0] * self.kl2_z_p2q)
self.kl2_hi = (self.lam_kld_q2p[0] * self.kl2_hi_q2p) + \
(self.lam_kld_p2q[0] * self.kl2_hi_p2q)
self.kl2_costs = (self.lam_kld_z[0] * self.kl2_z) + self.kl2_hi
# compute joint l1/l2 KLd cost
self.kld_l1l2_costs = (self.lam_kld_l1l2[0] * self.kld_costs) + \
((1.0 - self.lam_kld_l1l2[0]) * self.kl2_costs)
# compute "mean" (rather than per-input) costs
self.kld_cost = T.mean(self.kld_costs)
self.kl2_cost = T.mean(self.kl2_costs)
self.kld_l1l2_cost = T.mean(self.kld_l1l2_costs)
#################################
# CONSTRUCT THE NLL-BASED COSTS #
#################################
self.nll_costs = self.nlli[-1]
self.nll_cost = self.lam_nll[0] * T.mean(self.nll_costs)
########################################
# CONSTRUCT THE REST OF THE JOINT COST #
########################################
param_reg_cost = self._construct_reg_costs()
self.reg_cost = self.lam_l2w[0] * param_reg_cost
self.joint_cost = self.nll_cost + self.kld_l1l2_cost + \
self.reg_cost
##############################
# CONSTRUCT A PER-INPUT COST #
##############################
self.obs_costs = self.nll_costs + self.kld_l1l2_costs
# Get the gradient of the joint cost for all optimizable parameters
print("Computing gradients of self.joint_cost...")
self.joint_grads = OrderedDict()
grad_list = T.grad(self.joint_cost, self.joint_params)
for i, p in enumerate(self.joint_params):
self.joint_grads[p] = grad_list[i]
# Construct the updates for the generator and inferencer networks
self.q_updates = get_adam_updates(params=self.q_params, \
grads=self.joint_grads, alpha=self.lr_1, \
beta1=self.mom_1, beta2=self.mom_2, \
mom2_init=1e-3, smoothing=1e-5, max_grad_norm=10.0)
self.p_updates = get_adam_updates(params=self.p_params, \
grads=self.joint_grads, alpha=self.lr_2, \
beta1=self.mom_1, beta2=self.mom_2, \
mom2_init=1e-3, smoothing=1e-5, max_grad_norm=10.0)
self.joint_updates = OrderedDict()
for k in self.q_updates:
self.joint_updates[k] = self.q_updates[k]
for k in self.p_updates:
self.joint_updates[k] = self.p_updates[k]
# add scan updates, which seem to be required
for k in self.scan_updates:
self.joint_updates[k] = self.scan_updates[k]
# Construct a function for jointly training the generator/inferencer
print("Compiling cost computer...")
self.compute_raw_klds = self._construct_raw_klds()
print("Compiling training function...")
self.train_joint = self._construct_train_joint()
print("Compiling free-energy sampler...")
self.compute_fe_terms = self._construct_compute_fe_terms()
print("Compiling open-loop model sampler...")
self.sample_from_prior = self._construct_sample_from_prior()
print("Compiling data-guided model sampler...")
self.sample_from_input = self._construct_sample_from_input()
return
def __init__(self, rng=None,
x_in=None, x_out=None,
p_h_given_z=None,
p_x_given_h=None,
q_z_given_x=None,
q_h_given_z_x=None,
x_dim=None,
z_dim=None,
h_dim=None,
h_det_dim=None,
params=None,
shared_param_dicts=None):
# setup a rng for this GIPair
self.rng = RandStream(rng.randint(100000))
# grab the user-provided parameters
self.params = params
self.x_type = self.params['x_type']
assert((self.x_type == 'bernoulli') or (self.x_type == 'gaussian'))
if 'obs_transform' in self.params:
assert((self.params['obs_transform'] == 'sigmoid') or \
(self.params['obs_transform'] == 'none'))
if self.params['obs_transform'] == 'sigmoid':
self.obs_transform = lambda x: T.nnet.sigmoid(x)
else:
self.obs_transform = lambda x: x
else:
self.obs_transform = lambda x: T.nnet.sigmoid(x)
if self.x_type == 'bernoulli':
self.obs_transform = lambda x: T.nnet.sigmoid(x)
self.shared_param_dicts = shared_param_dicts
# record the dimensions of various spaces relevant to this model
self.x_dim = x_dim
self.z_dim = z_dim
self.h_dim = h_dim
self.h_det_dim = h_det_dim
# grab handles to the relevant HydraNets
self.q_z_given_x = q_z_given_x
self.q_h_given_z_x = q_h_given_z_x
self.p_h_given_z = p_h_given_z
self.p_x_given_h = p_x_given_h
# record the symbolic variables that will provide inputs to the
# computation graph created to describe this MultiStageModel
self.x_in = x_in
self.x_out = x_out
# setup switching variable for changing between sampling/training
zero_ary = to_fX( np.zeros((1,)) )
self.train_switch = theano.shared(value=zero_ary, name='tsm_train_switch')
self.set_train_switch(1.0)
if self.shared_param_dicts is None:
# initialize "optimizable" parameters specific to this MSM
init_vec = to_fX( np.zeros((1,self.z_dim)) )
self.p_z_mean = theano.shared(value=init_vec, name='tsm_p_z_mean')
self.p_z_logvar = theano.shared(value=init_vec, name='tsm_p_z_logvar')
self.obs_logvar = theano.shared(value=zero_ary, name='tsm_obs_logvar')
self.bounded_logvar = 8.0 * T.tanh((1.0/8.0) * self.obs_logvar)
self.shared_param_dicts = {}
self.shared_param_dicts['p_z_mean'] = self.p_z_mean
self.shared_param_dicts['p_z_logvar'] = self.p_z_logvar
self.shared_param_dicts['obs_logvar'] = self.obs_logvar
else:
self.p_z_mean = self.shared_param_dicts['p_z_mean']
self.p_z_logvar = self.shared_param_dicts['p_z_logvar']
self.obs_logvar = self.shared_param_dicts['obs_logvar']
self.bounded_logvar = 8.0 * T.tanh((1.0/8.0) * self.obs_logvar)
##############################################
# Setup the TwoStageModels main computation. #
##############################################
print("Building TSM...")
# samples of "hidden" latent state (from both p and q)
z_q_mean, z_q_logvar = self.q_z_given_x.apply(self.x_in)
z_q = reparametrize(z_q_mean, z_q_logvar, rng=self.rng)
z_p_mean = self.p_z_mean.repeat(z_q.shape[0], axis=0)
z_p_logvar = self.p_z_logvar.repeat(z_q.shape[0], axis=0)
z_p = reparametrize(z_p_mean, z_p_logvar, rng=self.rng)
self.z = (self.train_switch[0] * z_q) + \
((1.0 - self.train_switch[0]) * z_p)
# compute relevant KLds for this step
self.kld_z_q2p = gaussian_kld(z_q_mean, z_q_logvar,
z_p_mean, z_p_logvar)
self.kld_z_p2q = gaussian_kld(z_p_mean, z_p_logvar,
z_q_mean, z_q_logvar)
# samples of "hidden" latent state (from both p and q)
h_p_mean, h_p_logvar = self.p_h_given_z.apply(self.z)
h_p = reparametrize(h_p_mean, h_p_logvar, rng=self.rng)
h_q_mean, h_q_logvar = self.q_h_given_z_x.apply(
T.concatenate([h_p_mean, self.x_out], axis=1))
h_q = reparametrize(h_q_mean, h_q_logvar, rng=self.rng)
# compute "stochastic" and "deterministic" parts of latent state
h_sto = (self.train_switch[0] * h_q) + \
((1.0 - self.train_switch[0]) * h_p)
h_det = h_p_mean
if self.h_det_dim is None:
# don't pass forward any deterministic state
self.h = h_sto
else:
# pass forward some deterministic state
self.h = T.concatenate([h_det[:,:self.h_det_dim],
h_sto[:,self.h_det_dim:]], axis=1)
# compute relevant KLds for this step
self.kld_h_q2p = gaussian_kld(h_q_mean, h_q_logvar,
h_p_mean, h_p_logvar)
self.kld_h_p2q = gaussian_kld(h_p_mean, h_p_logvar,
h_q_mean, h_q_logvar)
# p_x_given_h generates an observation x conditioned on the "hidden"
# latent variables h.
self.x_gen, _ = self.p_x_given_h.apply(self.h)
######################################################################
# ALL SYMBOLIC VARS NEEDED FOR THE OBJECTIVE SHOULD NOW BE AVAILABLE #
######################################################################
# shared var learning rate for generator and inferencer
zero_ary = to_fX( np.zeros((1,)) )
self.lr = theano.shared(value=zero_ary, name='tsm_lr')
# shared var momentum parameters for generator and inferencer
self.mom_1 = theano.shared(value=zero_ary, name='tsm_mom_1')
self.mom_2 = theano.shared(value=zero_ary, name='tsm_mom_2')
# init parameters for controlling learning dynamics
self.set_sgd_params()
# init shared var for weighting nll of data given posterior sample
self.lam_nll = theano.shared(value=zero_ary, name='tsm_lam_nll')
self.set_lam_nll(lam_nll=1.0)
# init shared var for weighting prior kld against reconstruction
self.lam_kld_q2p = theano.shared(value=zero_ary, name='tsm_lam_kld_q2p')
self.lam_kld_p2q = theano.shared(value=zero_ary, name='tsm_lam_kld_p2q')
self.set_lam_kld(lam_kld_q2p=1.0, lam_kld_p2q=0.0)
# init shared var for controlling l2 regularization on params
self.lam_l2w = theano.shared(value=zero_ary, name='tsm_lam_l2w')
self.set_lam_l2w(1e-5)
# get optimizable parameters belonging to the TwoStageModel
self_params = [self.obs_logvar] #+ [self.p_z_mean, self.p_z_logvar]
# get optimizable parameters belonging to the underlying networks
child_params = []
child_params.extend(self.q_z_given_x.mlp_params)
child_params.extend(self.q_h_given_z_x.mlp_params)
child_params.extend(self.p_h_given_z.mlp_params)
child_params.extend(self.p_x_given_h.mlp_params)
# make a joint list of all optimizable parameters
self.joint_params = self_params + child_params
#################################
# CONSTRUCT THE KLD-BASED COSTS #
#################################
self.kld_z = (self.lam_kld_q2p[0] * self.kld_z_q2p) + \
(self.lam_kld_p2q[0] * self.kld_z_p2q)
self.kld_h = (self.lam_kld_q2p[0] * self.kld_h_q2p) + \
(self.lam_kld_p2q[0] * self.kld_h_p2q)
self.kld_costs = T.sum(self.kld_z, axis=1) + \
T.sum(self.kld_h, axis=1)
# compute "mean" (rather than per-input) costs
self.kld_cost = T.mean(self.kld_costs)
#################################
# CONSTRUCT THE NLL-BASED COSTS #
#################################
self.nll_costs = self._construct_nll_costs(self.x_out)
self.nll_cost = self.lam_nll[0] * T.mean(self.nll_costs)
########################################
# CONSTRUCT THE REST OF THE JOINT COST #
########################################
param_reg_cost = self._construct_reg_costs()
self.reg_cost = self.lam_l2w[0] * param_reg_cost
self.joint_cost = self.nll_cost + self.kld_cost + self.reg_cost
##############################
# CONSTRUCT A PER-INPUT COST #
##############################
self.obs_costs = self.nll_costs + self.kld_costs
# get the gradient of the joint cost for all optimizable parameters
print("Computing gradients of self.joint_cost...")
self.joint_grads = OrderedDict()
grad_list = T.grad(self.joint_cost, self.joint_params)
for i, p in enumerate(self.joint_params):
self.joint_grads[p] = grad_list[i]
# construct the updates for the generator and inferencer networks
all_updates = get_adam_updates(params=self.joint_params,
grads=self.joint_grads, alpha=self.lr,
beta1=self.mom_1, beta2=self.mom_2,
mom2_init=1e-3, smoothing=1e-4, max_grad_norm=5.0)
self.joint_updates = OrderedDict()
for k in all_updates:
self.joint_updates[k] = all_updates[k]
# Construct a function for jointly training the generator/inferencer
print("Compiling training function...")
self.train_joint = self._construct_train_joint()
print("Compiling free-energy sampler...")
self.compute_fe_terms = self._construct_compute_fe_terms()
print("Compiling open-loop model sampler...")
self.sample_from_prior = self._construct_sample_from_prior()
return
def __init__(self, rng=None,
x_in=None, x_mask=None, x_out=None, \
p_h_given_x=None, \
p_s0_given_h=None, \
p_zi_given_xi=None, \
p_sip1_given_zi=None, \
p_x_given_si=None, \
q_h_given_x=None, \
q_zi_given_xi=None, \
params=None, \
shared_param_dicts=None):
# setup a rng for this GIPair
self.rng = RandStream(rng.randint(100000))
# grab the user-provided parameters
self.params = params
self.x_dim = self.params['x_dim']
self.h_dim = self.params['h_dim']
self.z_dim = self.params['z_dim']
self.s_dim = self.params['s_dim']
self.use_p_x_given_si = self.params['use_p_x_given_si']
self.imp_steps = self.params['imp_steps']
self.step_type = self.params['step_type']
self.x_type = self.params['x_type']
if self.use_p_x_given_si:
print("Constructing hypotheses via p_x_given_si...")
else:
print("Constructing hypotheses directly in x-space...")
assert(self.s_dim == self.x_dim)
assert((self.x_type == 'bernoulli') or (self.x_type == 'gaussian'))
if 'obs_transform' in self.params:
assert((self.params['obs_transform'] == 'sigmoid') or \
(self.params['obs_transform'] == 'none'))
if self.params['obs_transform'] == 'sigmoid':
self.obs_transform = lambda x: T.nnet.sigmoid(x)
else:
self.obs_transform = lambda x: x
else:
self.obs_transform = lambda x: T.nnet.sigmoid(x)
if self.x_type == 'bernoulli':
self.obs_transform = lambda x: T.nnet.sigmoid(x)
self.shared_param_dicts = shared_param_dicts
assert((self.step_type == 'add') or (self.step_type == 'jump'))
# grab handles to the relevant InfNets
self.p_h_given_x = p_h_given_x
self.p_s0_given_h = p_s0_given_h
self.p_zi_given_xi = p_zi_given_xi
self.p_sip1_given_zi = p_sip1_given_zi
self.p_x_given_si = p_x_given_si
self.q_h_given_x = q_h_given_x
self.q_zi_given_xi = q_zi_given_xi
# record the symbolic variables that will provide inputs to the
# computation graph created to describe this MultiStageModel
self.x_in = x_in
self.x_out = x_out
self.x_mask = x_mask
self.zi_zmuv = T.tensor3()
# setup switching variable for changing between sampling/training
zero_ary = to_fX( np.zeros((1,)) )
self.train_switch = theano.shared(value=zero_ary, name='gpsi_train_switch')
self.set_train_switch(1.0)
if self.shared_param_dicts is None:
# initialize parameters "owned" by this model
init_ary = to_fX( np.zeros((self.x_dim,)) )
self.s_null = theano.shared(value=init_ary, name='gpis_sn')
self.grad_null = theano.shared(value=init_ary, name='gpsi_gn')
self.obs_logvar = theano.shared(value=zero_ary, name='gpsi_obs_logvar')
self.bounded_logvar = 8.0 * T.tanh((1.0/8.0) * self.obs_logvar[0])
self.shared_param_dicts = {}
self.shared_param_dicts['s_null'] = self.s_null
self.shared_param_dicts['grad_null'] = self.grad_null
self.shared_param_dicts['obs_logvar'] = self.obs_logvar
self.x_null = self._from_si_to_x(self.s_null)
else:
# grab the parameters required by this model from a given dict
self.s_null = self.shared_param_dicts['s_null']
self.grad_null = self.shared_param_dicts['grad_null']
self.obs_logvar = self.shared_param_dicts['obs_logvar']
self.bounded_logvar = 8.0 * T.tanh((1.0/8.0) * self.obs_logvar[0])
self.x_null = self._from_si_to_x(self.s_null)
##############################################
# Compute results of the initialization step #
##############################################
self.x_init = (self.x_mask * self.x_in) + \
((1.0 - self.x_mask) * self.x_null)
# sample from primary and guide conditionals over h
h_p_mean, h_p_logvar, h_p = \
self.p_h_given_x.apply(self.x_init, do_samples=True)
h_q_mean, h_q_logvar, h_q = \
self.q_h_given_x.apply(self.x_in, do_samples=True)
# make h samples that can be switched between h_p and h_q
self.h = ((self.train_switch[0] * h_q) + \
((1.0 - self.train_switch[0]) * h_p))
# get the emitted initial state s0 (sampled via either p or q)
hydra_out = self.p_s0_given_h.apply(self.h)
self.s0 = hydra_out[0]
# compute NLL reconstruction cost for the initialization step
self.nll0 = self._construct_nll_costs(self.s0, self.x_out, self.x_mask)
# compute KLds for the initialization step
self.kldh_q2p = gaussian_kld(h_q_mean, h_q_logvar, \
h_p_mean, h_p_logvar) # KL(q || p)
self.kldh_p2q = gaussian_kld(h_p_mean, h_p_logvar, \
h_q_mean, h_q_logvar) # KL(p || q)
self.kldh_p2g = gaussian_kld(h_p_mean, h_p_logvar, \
0.0, 0.0) # KL(p || global prior)
##################################################
# Setup the iterative imputation loop using scan #
##################################################
self.ones_mask = T.ones_like(self.x_mask)
def imp_step_func(zi_zmuv, si):
si_as_x = self._from_si_to_x(si)
xi_unmasked = self.x_out
xi_masked = (self.x_mask * xi_unmasked) + \
((1.0 - self.x_mask) * si_as_x)
grad_unmasked = self.x_out - si_as_x
grad_masked = (self.x_mask * grad_unmasked) + \
((1.0 - self.x_mask) * self.grad_null)
# get samples of next zi, according to the global policy
zi_p_mean, zi_p_logvar = self.p_zi_given_xi.apply( \
T.horizontal_stack(xi_masked, grad_masked), \
do_samples=False)
zi_p = zi_p_mean + (T.exp(0.5 * zi_p_logvar) * zi_zmuv)
# get samples of next zi, according to the guide policy
zi_q_mean, zi_q_logvar = self.q_zi_given_xi.apply( \
T.horizontal_stack(xi_masked, grad_unmasked), \
do_samples=False)
zi_q = zi_q_mean + (T.exp(0.5 * zi_q_logvar) * zi_zmuv)
# make zi samples that can be switched between zi_p and zi_q
zi = ((self.train_switch[0] * zi_q) + \
((1.0 - self.train_switch[0]) * zi_p))
# compute relevant KLds for this step
kldi_q2p = gaussian_kld(zi_q_mean, zi_q_logvar, \
zi_p_mean, zi_p_logvar) # KL(q || p)
kldi_p2q = gaussian_kld(zi_p_mean, zi_p_logvar, \
zi_q_mean, zi_q_logvar) # KL(p || q)
kldi_p2g = gaussian_kld(zi_p_mean, zi_p_logvar, \
0.0, 0.0) # KL(p || global prior)
# compute the next si, given the sampled zi
hydra_out = self.p_sip1_given_zi.apply(zi)
si_step = hydra_out[0]
if (self.step_type == 'jump'):
# jump steps always completely overwrite the current guesses
sip1 = si_step
else:
# additive steps update the current guesses like an LSTM
write_gate = T.nnet.sigmoid(3.0 + hydra_out[1])
erase_gate = T.nnet.sigmoid(3.0 + hydra_out[2])
sip1 = (erase_gate * si) + (write_gate * si_step)
# compute NLL for the current imputation
nlli = self._construct_nll_costs(sip1, self.x_out, self.x_mask)
return sip1, nlli, kldi_q2p, kldi_p2q, kldi_p2g
# apply scan op for the sequential imputation loop
init_vals = [self.s0, None, None, None, None]
self.scan_results, self.scan_updates = theano.scan(imp_step_func, \
outputs_info=init_vals, sequences=self.zi_zmuv)
self.si = self.scan_results[0]
self.nlli = self.scan_results[1]
self.kldi_q2p = self.scan_results[2]
self.kldi_p2q = self.scan_results[3]
self.kldi_p2g = self.scan_results[4]
######################################################################
# ALL SYMBOLIC VARS NEEDED FOR THE OBJECTIVE SHOULD NOW BE AVAILABLE #
######################################################################
# shared var learning rate for generator and inferencer
zero_ary = to_fX( np.zeros((1,)) )
self.lr = theano.shared(value=zero_ary, name='gpsi_lr')
# shared var momentum parameters for generator and inferencer
self.mom_1 = theano.shared(value=zero_ary, name='gpsi_mom_1')
self.mom_2 = theano.shared(value=zero_ary, name='gpsi_mom_2')
# init parameters for controlling learning dynamics
self.set_sgd_params()
# init shared var for weighting nll of data given posterior sample
self.lam_nll = theano.shared(value=zero_ary, name='gpsi_lam_nll')
self.set_lam_nll(lam_nll=1.0)
# init shared var for weighting prior kld against reconstruction
self.lam_kld_p = theano.shared(value=zero_ary, name='gpsi_lam_kld_p')
self.lam_kld_q = theano.shared(value=zero_ary, name='gpsi_lam_kld_q')
self.lam_kld_g = theano.shared(value=zero_ary, name='gpsi_lam_kld_g')
self.lam_kld_s = theano.shared(value=zero_ary, name='gpsi_lam_kld_s')
self.set_lam_kld(lam_kld_p=0.0, lam_kld_q=1.0, lam_kld_g=0.0, lam_kld_s=0.0)
# init shared var for controlling l2 regularization on params
self.lam_l2w = theano.shared(value=zero_ary, name='msm_lam_l2w')
self.set_lam_l2w(1e-5)
# Grab all of the "optimizable" parameters in the model
self.joint_params = [self.s_null, self.grad_null, self.obs_logvar]
self.joint_params.extend(self.p_zi_given_xi.mlp_params)
self.joint_params.extend(self.p_sip1_given_zi.mlp_params)
self.joint_params.extend(self.p_x_given_si.mlp_params)
self.joint_params.extend(self.q_zi_given_xi.mlp_params)
#################################
# CONSTRUCT THE KLD-BASED COSTS #
#################################
self.kld_p, self.kld_q, self.kld_g, self.kld_s = \
self._construct_kld_costs(p=1.0)
self.kld_costs = (self.lam_kld_p[0] * self.kld_p) + \
(self.lam_kld_q[0] * self.kld_q) + \
(self.lam_kld_g[0] * self.kld_g) + \
(self.lam_kld_s[0] * self.kld_s)
self.kld_cost = T.mean(self.kld_costs)
#################################
# CONSTRUCT THE NLL-BASED COSTS #
#################################
self.nll_costs = self.nlli[-1]
self.nll_cost = self.lam_nll[0] * T.mean(self.nll_costs)
self.nll_bounds = self.nll_costs.ravel() + self.kld_q.ravel()
self.nll_bound = T.mean(self.nll_bounds)
########################################
# CONSTRUCT THE REST OF THE JOINT COST #
########################################
param_reg_cost = self._construct_reg_costs()
self.reg_cost = self.lam_l2w[0] * param_reg_cost
self.joint_cost = self.nll_cost + self.kld_cost + self.reg_cost
##############################
# CONSTRUCT A PER-TRIAL COST #
##############################
self.obs_costs = self.nll_costs + self.kld_costs
# Get the gradient of the joint cost for all optimizable parameters
print("Computing gradients of self.joint_cost...")
self.joint_grads = OrderedDict()
grad_list = T.grad(self.joint_cost, self.joint_params)
for i, p in enumerate(self.joint_params):
self.joint_grads[p] = grad_list[i]
# Construct the updates for the generator and inferencer networks
self.joint_updates = get_adam_updates(params=self.joint_params, \
grads=self.joint_grads, alpha=self.lr, \
beta1=self.mom_1, beta2=self.mom_2, \
mom2_init=1e-3, smoothing=1e-5, max_grad_norm=10.0)
for k, v in self.scan_updates.items():
self.joint_updates[k] = v
# Construct a function for jointly training the generator/inferencer
print("Compiling training function...")
self.train_joint = self._construct_train_joint()
print("Compiling free-energy sampler...")
self.compute_fe_terms = self._construct_compute_fe_terms()
print("Compiling best step cost computer...")
self.compute_per_step_cost = self._construct_compute_per_step_cost()
print("Compiling data-guided imputer sampler...")
self.sample_imputer = self._construct_sample_imputer()
# make easy access points for some interesting parameters
#self.gen_inf_weights = self.p_zi_given_xi.shared_layers[0].W
return
def __init__(self, rng=None, \
x_in=None, x_out=None, \
p_h_given_z=None, \
p_x_given_h=None, \
q_z_given_x=None, \
q_h_given_z_x=None, \
x_dim=None, \
z_dim=None, \
h_dim=None, \
params=None, \
shared_param_dicts=None):
# setup a rng for this GIPair
self.rng = RandStream(rng.randint(100000))
# grab the user-provided parameters
self.params = params
self.x_type = self.params['x_type']
assert((self.x_type == 'bernoulli') or (self.x_type == 'gaussian'))
if 'obs_transform' in self.params:
assert((self.params['obs_transform'] == 'sigmoid') or \
(self.params['obs_transform'] == 'none'))
if self.params['obs_transform'] == 'sigmoid':
self.obs_transform = lambda x: T.nnet.sigmoid(x)
else:
self.obs_transform = lambda x: x
else:
self.obs_transform = lambda x: T.nnet.sigmoid(x)
if self.x_type == 'bernoulli':
self.obs_transform = lambda x: T.nnet.sigmoid(x)
self.shared_param_dicts = shared_param_dicts
# record the dimensions of various spaces relevant to this model
self.x_dim = x_dim
self.z_dim = z_dim
self.h_dim = h_dim
# grab handles to the relevant InfNets
self.q_z_given_x = q_z_given_x
self.q_h_given_z_x = q_h_given_z_x
self.p_h_given_z = p_h_given_z
self.p_x_given_h = p_x_given_h
# record the symbolic variables that will provide inputs to the
# computation graph created to describe this MultiStageModel
self.x_in = x_in
self.x_out = x_out
# setup switching variable for changing between sampling/training
zero_ary = to_fX( np.zeros((1,)) )
self.train_switch = theano.shared(value=zero_ary, name='tsm_train_switch')
self.set_train_switch(1.0)
if self.shared_param_dicts is None:
# initialize "optimizable" parameters specific to this MSM
init_vec = to_fX( np.zeros((1,self.z_dim)) )
self.p_z_mean = theano.shared(value=init_vec, name='tsm_p_z_mean')
self.p_z_logvar = theano.shared(value=init_vec, name='tsm_p_z_logvar')
self.obs_logvar = theano.shared(value=zero_ary, name='tsm_obs_logvar')
self.bounded_logvar = 8.0 * T.tanh((1.0/8.0) * self.obs_logvar)
self.shared_param_dicts = {}
self.shared_param_dicts['p_z_mean'] = self.p_z_mean
self.shared_param_dicts['p_z_logvar'] = self.p_z_logvar
self.shared_param_dicts['obs_logvar'] = self.obs_logvar
else:
self.p_z_mean = self.shared_param_dicts['p_z_mean']
self.p_z_logvar = self.shared_param_dicts['p_z_logvar']
self.obs_logvar = self.shared_param_dicts['obs_logvar']
self.bounded_logvar = 8.0 * T.tanh((1.0/8.0) * self.obs_logvar)
##############################################
# Setup the TwoStageModels main computation. #
##############################################
print("Building TSM...")
# samples of "hidden" latent state (from both p and q)
z_q_mean, z_q_logvar, z_q = \
self.q_z_given_x.apply(self.x_in, do_samples=True)
z_p_mean = self.p_z_mean.repeat(z_q.shape[0], axis=0)
z_p_logvar = self.p_z_logvar.repeat(z_q.shape[0], axis=0)
zmuv = self.rng.normal(size=z_q.shape, avg=0.0, std=1.0, \
dtype=theano.config.floatX)
z_p = (T.exp(0.5*z_p_logvar) * zmuv) + z_p_mean
self.z = (self.train_switch[0] * z_q) + \
((1.0 - self.train_switch[0]) * z_p)
# compute relevant KLds for this step
self.kld_z_q2p = gaussian_kld(z_q_mean, z_q_logvar, \
z_p_mean, z_p_logvar)
self.kld_z_p2q = gaussian_kld(z_p_mean, z_p_logvar, \
z_q_mean, z_q_logvar)
# samples of "hidden" latent state (from both p and q)
h_p_mean, h_p_logvar, h_p = self.p_h_given_z.apply(self.z)
h_q_mean, h_q_logvar, h_q = self.q_h_given_z_x.apply( \
T.horizontal_stack(h_p_mean, h_p_logvar, self.x_out))
self.h = (self.train_switch[0] * h_q) + \
((1.0 - self.train_switch[0]) * h_p)
# compute relevant KLds for this step
self.kld_h_q2p = gaussian_kld(h_q_mean, h_q_logvar, \
h_p_mean, h_p_logvar)
self.kld_h_p2q = gaussian_kld(h_p_mean, h_p_logvar, \
h_q_mean, h_q_logvar)
# p_x_given_h generates an observation x conditioned on the "hidden"
# latent variables h.
self.x_gen, _ = self.p_x_given_h.apply(self.h, do_samples=False)
######################################################################
# ALL SYMBOLIC VARS NEEDED FOR THE OBJECTIVE SHOULD NOW BE AVAILABLE #
######################################################################
# shared var learning rate for generator and inferencer
zero_ary = to_fX( np.zeros((1,)) )
self.lr = theano.shared(value=zero_ary, name='tsm_lr')
# shared var momentum parameters for generator and inferencer
self.mom_1 = theano.shared(value=zero_ary, name='tsm_mom_1')
self.mom_2 = theano.shared(value=zero_ary, name='tsm_mom_2')
# init parameters for controlling learning dynamics
self.set_sgd_params()
# init shared var for weighting nll of data given posterior sample
self.lam_nll = theano.shared(value=zero_ary, name='tsm_lam_nll')
self.set_lam_nll(lam_nll=1.0)
# init shared var for weighting prior kld against reconstruction
self.lam_kld_q2p = theano.shared(value=zero_ary, name='tsm_lam_kld_q2p')
self.lam_kld_p2q = theano.shared(value=zero_ary, name='tsm_lam_kld_p2q')
self.set_lam_kld(lam_kld_q2p=1.0, lam_kld_p2q=0.0)
# init shared var for controlling l2 regularization on params
self.lam_l2w = theano.shared(value=zero_ary, name='tsm_lam_l2w')
self.set_lam_l2w(1e-5)
# get optimizable parameters belonging to the TwoStageModel
self_params = [self.obs_logvar] #+ [self.p_z_mean, self.p_z_logvar]
# get optimizable parameters belonging to the underlying networks
child_params = []
child_params.extend(self.q_z_given_x.mlp_params)
child_params.extend(self.q_h_given_z_x.mlp_params)
child_params.extend(self.p_h_given_z.mlp_params)
child_params.extend(self.p_x_given_h.mlp_params)
# make a joint list of all optimizable parameters
self.joint_params = self_params + child_params
#################################
# CONSTRUCT THE KLD-BASED COSTS #
#################################
self.kld_z = (self.lam_kld_q2p[0] * self.kld_z_q2p) + \
(self.lam_kld_p2q[0] * self.kld_z_p2q)
self.kld_h = (self.lam_kld_q2p[0] * self.kld_h_q2p) + \
(self.lam_kld_p2q[0] * self.kld_h_p2q)
self.kld_costs = T.sum(self.kld_z, axis=1) + \
T.sum(self.kld_h, axis=1)
# compute "mean" (rather than per-input) costs
self.kld_cost = T.mean(self.kld_costs)
#################################
# CONSTRUCT THE NLL-BASED COSTS #
#################################
self.nll_costs = self._construct_nll_costs(self.x_out)
self.nll_cost = self.lam_nll[0] * T.mean(self.nll_costs)
########################################
# CONSTRUCT THE REST OF THE JOINT COST #
########################################
param_reg_cost = self._construct_reg_costs()
self.reg_cost = self.lam_l2w[0] * param_reg_cost
self.joint_cost = self.nll_cost + self.kld_cost + self.reg_cost
##############################
# CONSTRUCT A PER-INPUT COST #
##############################
self.obs_costs = self.nll_costs + self.kld_costs
# get the gradient of the joint cost for all optimizable parameters
print("Computing gradients of self.joint_cost...")
self.joint_grads = OrderedDict()
grad_list = T.grad(self.joint_cost, self.joint_params)
for i, p in enumerate(self.joint_params):
self.joint_grads[p] = grad_list[i]
# construct the updates for the generator and inferencer networks
all_updates = get_adam_updates(params=self.joint_params, \
grads=self.joint_grads, alpha=self.lr, \
beta1=self.mom_1, beta2=self.mom_2, \
mom2_init=1e-3, smoothing=1e-4, max_grad_norm=5.0)
self.joint_updates = OrderedDict()
for k in all_updates:
self.joint_updates[k] = all_updates[k]
# Construct a function for jointly training the generator/inferencer
print("Compiling training function...")
self.train_joint = self._construct_train_joint()
print("Compiling free-energy sampler...")
self.compute_fe_terms = self._construct_compute_fe_terms()
print("Compiling open-loop model sampler...")
self.sample_from_prior = self._construct_sample_from_prior()
return
def __init__(self, rng=None, \
x_in=None, y_in=None, \
q_z_given_x=None, \
class_count=None, \
z_dim=None, \
use_samples=None):
# setup a rng for this GIPair
self.rng = RandStream(rng.randint(100000))
# record the dimensions of various spaces relevant to this model
self.class_count = class_count
self.z_dim = z_dim
self.shared_dim = q_z_given_x.shared_layers[-1].out_dim
self.use_samples = use_samples
# grab handles to the relevant InfNets
self.q_z_given_x = q_z_given_x
# record the symbolic variables that will provide inputs to the
# computation graph created to describe this MultiStageModel
self.x_in = x_in
self.y_in = y_in
# setup switching variable for changing between sampling/training
zero_ary = to_fX( np.zeros((1,)) )
# setup a variable for controlling dropout noise
self.drop_rate = theano.shared(value=zero_ary, name='cm_drop_rate')
self.set_drop_rate(0.0)
# initialize classification layer parameters
init_mat = to_fX(0.01 * npr.randn(self.shared_dim, self.class_count))
init_vec = to_fX( np.zeros((self.class_count,)) )
self.W_class = theano.shared(value=init_mat, name='cm_W_class')
self.b_class = theano.shared(value=init_vec, name='cm_b_class')
# initialize "optimizable" parameters specific to this CM
init_vec = to_fX( np.zeros((self.z_dim,)) )
self.p_z_mean = theano.shared(value=init_vec, name='cm_p_z_mean')
self.p_z_logvar = theano.shared(value=init_vec, name='cm_p_z_logvar')
#################
# Setup self.z. #
#################
self.q_z_mean, self.q_z_logvar, self.q_z_samples = \
self.q_z_given_x.apply(self.x_in, do_samples=True)
self.q_z_samples = self.q_z_given_x.apply_shared(self.x_in)
# get a drop mask that drops things with probability p
drop_scale = 1. / (1. - self.drop_rate[0])
drop_rnd = self.rng.uniform(size=self.q_z_samples.shape, \
low=0.0, high=1.0, dtype=theano.config.floatX)
drop_mask = drop_scale * (drop_rnd > self.drop_rate[0])
# get a droppy version of either z mean or z samples
# if self.use_samples:
# self.z = self.q_z_samples * drop_mask
# else:
# self.z = self.q_z_mean * drop_mask
self.z = self.q_z_samples * drop_mask
# compute class predictions
self.y_out = T.dot(self.z, self.W_class) + self.b_class
# compute KLds for training via variational free-energy
self.kld_z_q2ps = gaussian_kld(self.q_z_mean, self.q_z_logvar, \
self.p_z_mean, self.p_z_logvar)
self.kld_z_p2qs = gaussian_kld(self.p_z_mean, self.p_z_logvar, \
self.q_z_mean, self.q_z_logvar)
######################################################################
# ALL SYMBOLIC VARS NEEDED FOR THE OBJECTIVE SHOULD NOW BE AVAILABLE #
######################################################################
# shared var learning rate for generator and inferencer
zero_ary = to_fX( np.zeros((1,)) )
self.lr_1 = theano.shared(value=zero_ary, name='cm_lr_1')
self.lr_2 = theano.shared(value=zero_ary, name='cm_lr_2')
# shared var momentum parameters for generator and inferencer
self.mom_1 = theano.shared(value=zero_ary, name='cm_mom_1')
self.mom_2 = theano.shared(value=zero_ary, name='cm_mom_2')
# init parameters for controlling learning dynamics
self.set_sgd_params()
# init shared var for weighting nll of data given posterior sample
self.lam_nll = theano.shared(value=zero_ary, name='cm_lam_nll')
self.set_lam_nll(lam_nll=1.0)
# init shared var for weighting prior kld against reconstruction
self.lam_kld_q2p = theano.shared(value=zero_ary, name='cm_lam_kld_q2p')
self.lam_kld_p2q = theano.shared(value=zero_ary, name='cm_lam_kld_p2q')
self.set_lam_kld(lam_kld_q2p=0.9, lam_kld_p2q=0.0)
# init shared var for controlling l2 regularization on params
self.lam_l2w = theano.shared(value=zero_ary, name='cm_lam_l2w')
self.set_lam_l2w(1e-5)
# Grab all of the "optimizable" parameters
self.joint_params = [self.p_z_mean, self.p_z_logvar, \
self.W_class, self.b_class]
self.joint_params.extend(self.q_z_given_x.mlp_params)
#################################
# CONSTRUCT THE NLL-BASED COSTS #
#################################
self.nll_costs = self.lam_nll[0] * self._construct_nll_costs(self.y_in)
self.nll_cost = T.mean(self.nll_costs)
#################################
# CONSTRUCT THE KLD-BASED COSTS #
#################################
self.kld_z_q2p, self.kld_z_p2q = self._construct_kld_costs(p=1.0)
self.kld_costs = (self.lam_kld_q2p[0] * self.kld_z_q2p) + \
(self.lam_kld_p2q[0] * self.kld_z_p2q)
self.kld_cost = T.mean(self.kld_costs)
##################################
# CONSTRUCT THE FINAL JOINT COST #
##################################
param_reg_cost = self._construct_reg_costs()
self.reg_cost = self.lam_l2w[0] * param_reg_cost
self.joint_cost = self.nll_cost + self.kld_cost + self.reg_cost
##############################
# CONSTRUCT A PER-INPUT COST #
##############################
self.obs_costs = self.nll_costs + self.kld_costs
# Get the gradient of the joint cost for all optimizable parameters
print("Computing gradients of self.joint_cost...")
self.joint_grads = OrderedDict()
grad_list = T.grad(self.joint_cost, self.joint_params)
for i, p in enumerate(self.joint_params):
self.joint_grads[p] = grad_list[i]
# Construct the updates for the model parameters
self.joint_updates = get_adam_updates(params=self.joint_params, \
grads=self.joint_grads, alpha=self.lr_1, \
beta1=self.mom_1, beta2=self.mom_2, \
mom2_init=1e-3, smoothing=1e-4, max_grad_norm=10.0)
# Construct a function for jointly training the generator/inferencer
print("Compiling training function...")
self.train_joint = self._construct_train_joint()
print("Compiling class error estimator...")
self.class_error = self._construct_class_error()
print("Compiling free-energy sampler...")
self.compute_fe_terms = self._construct_compute_fe_terms()
# make easy access points for some interesting parameters
self.inf_weights = self.q_z_given_x.shared_layers[0].W
return
def __init__(self, rng=None, \
x_in=None, x_out=None, \
p_s_given_z=None, \
p_h_given_s=None, \
p_x_given_s_h=None, \
q_z_given_x=None, \
q_h_given_x_s=None, \
x_dim=None, \
z_dim=None, \
s_dim=None, \
h_dim=None, \
params=None, \
shared_param_dicts=None):
# setup a rng for this GIPair
self.rng = RandStream(rng.randint(100000))
# grab the user-provided parameters
self.params = params
self.x_type = self.params['x_type']
assert((self.x_type == 'bernoulli') or (self.x_type == 'gaussian'))
if 'obs_transform' in self.params:
assert((self.params['obs_transform'] == 'sigmoid') or \
(self.params['obs_transform'] == 'none'))
if self.params['obs_transform'] == 'sigmoid':
self.obs_transform = lambda x: T.nnet.sigmoid(x)
else:
self.obs_transform = lambda x: x
else:
self.obs_transform = lambda x: T.nnet.sigmoid(x)
if self.x_type == 'bernoulli':
self.obs_transform = lambda x: T.nnet.sigmoid(x)
self.shared_param_dicts = shared_param_dicts
# record the dimensions of various spaces relevant to this model
self.x_dim = x_dim
self.z_dim = z_dim
self.s_dim = s_dim
self.h_dim = h_dim
# grab handles to the relevant InfNets
self.q_z_given_x = q_z_given_x
self.q_h_given_x_s = q_h_given_x_s
self.p_s_given_z = p_s_given_z
self.p_h_given_s = p_h_given_s
self.p_x_given_s_h = p_x_given_s_h
# record the symbolic variables that will provide inputs to the
# computation graph created to describe this MultiStageModel
self.x_in = x_in
self.x_out = x_out
self.batch_reps = T.lscalar()
# setup switching variable for changing between sampling/training
zero_ary = to_fX( np.zeros((1,)) )
self.train_switch = theano.shared(value=zero_ary, name='msm_train_switch')
self.set_train_switch(1.0)
# setup a variable for controlling dropout noise
self.drop_rate = theano.shared(value=zero_ary, name='msm_drop_rate')
self.set_drop_rate(0.0)
# this weight balances l1 vs. l2 penalty on posterior KLds
self.lam_kld_l1l2 = theano.shared(value=zero_ary, name='msm_lam_kld_l1l2')
self.set_lam_kld_l1l2(1.0)
if self.shared_param_dicts is None:
# initialize "optimizable" parameters specific to this MSM
init_vec = to_fX( np.zeros((self.z_dim,)) )
self.p_z_mean = theano.shared(value=init_vec, name='msm_p_z_mean')
self.p_z_logvar = theano.shared(value=init_vec, name='msm_p_z_logvar')
init_vec = to_fX( np.zeros((self.x_dim,)) )
self.obs_logvar = theano.shared(value=zero_ary, name='msm_obs_logvar')
self.bounded_logvar = 8.0 * T.tanh((1.0/8.0) * self.obs_logvar)
self.shared_param_dicts = {}
self.shared_param_dicts['p_z_mean'] = self.p_z_mean
self.shared_param_dicts['p_z_logvar'] = self.p_z_logvar
self.shared_param_dicts['obs_logvar'] = self.obs_logvar
else:
self.p_z_mean = self.shared_param_dicts['p_z_mean']
self.p_z_logvar = self.shared_param_dicts['p_z_logvar']
self.obs_logvar = self.shared_param_dicts['obs_logvar']
self.bounded_logvar = 8.0 * T.tanh((1.0/8.0) * self.obs_logvar)
# get a drop mask that drops things with probability p
drop_scale = 1. / (1. - self.drop_rate[0])
drop_rnd = self.rng.uniform(size=self.x_out.shape, \
low=0.0, high=1.0, dtype=theano.config.floatX)
drop_mask = drop_scale * (drop_rnd > self.drop_rate[0])
##############################################
# Setup the TwoStageModels main computation. #
##############################################
print("Building TSM...")
# samples of "first" latent state
drop_x = drop_mask * self.x_in
z_q_mean, z_q_logvar, self.z = \
self.q_z_given_x.apply(drop_x, do_samples=True)
# compute relevant KLds for this step
self.kld_z_q2ps = gaussian_kld(z_q_mean, z_q_logvar, \
self.p_z_mean, self.p_z_logvar)
self.kld_z_p2qs = gaussian_kld(self.p_z_mean, self.p_z_logvar, \
z_q_mean, z_q_logvar)
# transform "first" latent state into "second" latent state
self.s, _ = self.p_s_given_z.apply(self.z, do_samples=False)
# get samples of h, conditioned on current s
h_p_mean, h_p_logvar, h_p = self.p_h_given_s.apply( \
self.s, do_samples=True)
# get variational samples of h, given s and x_out
h_q_mean, h_q_logvar, h_q = self.q_h_given_x_s.apply( \
T.horizontal_stack(self.x_out, self.s), \
do_samples=True)
# make h samples that can be switched between h_p and h_q
self.h = (self.train_switch[0] * h_q) + \
((1.0 - self.train_switch[0]) * h_p)
# compute relevant KLds for this step
self.kld_h_q2ps = gaussian_kld(h_q_mean, h_q_logvar, \
h_p_mean, h_p_logvar)
self.kld_h_p2qs = gaussian_kld(h_p_mean, h_p_logvar, \
h_q_mean, h_q_logvar)
# p_x_given_s_h is conditioned on s and h.
self.x_gen, _ = self.p_x_given_s_h.apply( \
T.horizontal_stack(self.s, self.h), \
do_samples=False)
######################################################################
# ALL SYMBOLIC VARS NEEDED FOR THE OBJECTIVE SHOULD NOW BE AVAILABLE #
######################################################################
# shared var learning rate for generator and inferencer
zero_ary = to_fX( np.zeros((1,)) )
self.lr_1 = theano.shared(value=zero_ary, name='msm_lr_1')
self.lr_2 = theano.shared(value=zero_ary, name='msm_lr_2')
# shared var momentum parameters for generator and inferencer
self.mom_1 = theano.shared(value=zero_ary, name='msm_mom_1')
self.mom_2 = theano.shared(value=zero_ary, name='msm_mom_2')
# init parameters for controlling learning dynamics
self.set_sgd_params()
# init shared var for weighting nll of data given posterior sample
self.lam_nll = theano.shared(value=zero_ary, name='msm_lam_nll')
self.set_lam_nll(lam_nll=1.0)
# init shared var for weighting prior kld against reconstruction
self.lam_kld_z = theano.shared(value=zero_ary, name='msm_lam_kld_z')
self.lam_kld_q2p = theano.shared(value=zero_ary, name='msm_lam_kld_q2p')
self.lam_kld_p2q = theano.shared(value=zero_ary, name='msm_lam_kld_p2q')
self.set_lam_kld(lam_kld_z=1.0, lam_kld_q2p=0.7, lam_kld_p2q=0.3)
# init shared var for controlling l2 regularization on params
self.lam_l2w = theano.shared(value=zero_ary, name='msm_lam_l2w')
self.set_lam_l2w(1e-5)
# Grab all of the "optimizable" parameters in "group 1"
self.group_1_params = []
self.group_1_params.extend(self.q_z_given_x.mlp_params)
self.group_1_params.extend(self.q_h_given_x_s.mlp_params)
# Grab all of the "optimizable" parameters in "group 2"
self.group_2_params = [self.p_z_mean, self.p_z_logvar]
self.group_2_params.extend(self.p_s_given_z.mlp_params)
self.group_2_params.extend(self.p_h_given_s.mlp_params)
self.group_2_params.extend(self.p_x_given_s_h.mlp_params)
# Make a joint list of parameters group 1/2
self.joint_params = self.group_1_params + self.group_2_params
#################################
# CONSTRUCT THE KLD-BASED COSTS #
#################################
self.kld_z_q2p, self.kld_z_p2q, self.kld_h_q2p, self.kld_h_p2q = \
self._construct_kld_costs(p=1.0)
self.kld_z = (self.lam_kld_q2p[0] * self.kld_z_q2p) + \
(self.lam_kld_p2q[0] * self.kld_z_p2q)
self.kld_h = (self.lam_kld_q2p[0] * self.kld_h_q2p) + \
(self.lam_kld_p2q[0] * self.kld_h_p2q)
self.kld_costs = (self.lam_kld_z[0] * self.kld_z) + self.kld_h
# now do l2 KLd costs
self.kl2_z_q2p, self.kl2_z_p2q, self.kl2_h_q2p, self.kl2_h_p2q = \
self._construct_kld_costs(p=2.0)
self.kl2_z = (self.lam_kld_q2p[0] * self.kl2_z_q2p) + \
(self.lam_kld_p2q[0] * self.kl2_z_p2q)
self.kl2_h = (self.lam_kld_q2p[0] * self.kl2_h_q2p) + \
(self.lam_kld_p2q[0] * self.kl2_h_p2q)
self.kl2_costs = (self.lam_kld_z[0] * self.kl2_z) + self.kl2_h
# compute joint l1/l2 KLd cost
self.kld_l1l2_costs = (self.lam_kld_l1l2[0] * self.kld_costs) + \
((1.0 - self.lam_kld_l1l2[0]) * self.kl2_costs)
# compute "mean" (rather than per-input) costs
self.kld_cost = T.mean(self.kld_costs)
self.kl2_cost = T.mean(self.kl2_costs)
self.kld_l1l2_cost = T.mean(self.kld_l1l2_costs)
#################################
# CONSTRUCT THE NLL-BASED COSTS #
#################################
self.nll_costs = self._construct_nll_costs(self.x_out)
self.nll_cost = self.lam_nll[0] * T.mean(self.nll_costs)
########################################
# CONSTRUCT THE REST OF THE JOINT COST #
########################################
param_reg_cost = self._construct_reg_costs()
self.reg_cost = self.lam_l2w[0] * param_reg_cost
self.joint_cost = self.nll_cost + self.kld_l1l2_cost + self.reg_cost
##############################
# CONSTRUCT A PER-INPUT COST #
##############################
self.obs_costs = self.nll_costs + self.kld_l1l2_costs
# Get the gradient of the joint cost for all optimizable parameters
print("Computing gradients of self.joint_cost...")
self.joint_grads = OrderedDict()
grad_list = T.grad(self.joint_cost, self.joint_params)
for i, p in enumerate(self.joint_params):
self.joint_grads[p] = grad_list[i]
# Construct the updates for the generator and inferencer networks
self.group_1_updates = get_adam_updates(params=self.group_1_params, \
grads=self.joint_grads, alpha=self.lr_1, \
beta1=self.mom_1, beta2=self.mom_2, \
mom2_init=1e-3, smoothing=1e-5, max_grad_norm=10.0)
self.group_2_updates = get_adam_updates(params=self.group_2_params, \
grads=self.joint_grads, alpha=self.lr_2, \
beta1=self.mom_1, beta2=self.mom_2, \
mom2_init=1e-3, smoothing=1e-5, max_grad_norm=10.0)
self.joint_updates = OrderedDict()
for k in self.group_1_updates:
self.joint_updates[k] = self.group_1_updates[k]
for k in self.group_2_updates:
self.joint_updates[k] = self.group_2_updates[k]
# Construct a function for jointly training the generator/inferencer
print("Compiling training function...")
self.train_joint = self._construct_train_joint()
print("Compiling free-energy sampler...")
self.compute_fe_terms = self._construct_compute_fe_terms()
print("Compiling open-loop model sampler...")
self.sample_from_prior = self._construct_sample_from_prior()
print("Compiling data-guided model sampler...")
self.sample_from_input = self._construct_sample_from_input()
# make easy access points for some interesting parameters
self.gen_gen_weights = self.p_x_given_s_h.mu_layers[-1].W
return
