individ_variation = DFlow([
NVPFlow((VAR_DIM * 2 + 1) * VAR_DIM,
name='nvp_{}'.format(i),
aux_vars=global_inf.output) for i in range(6)
],
init_sigma=0.01)
ind = individ_variation.output[0]
indivs[country] = ind
indiv_logdens.append(individ_variation.logdens)
indiv_priors.append(individ_variation_prior.logdens(ind))
model = VARmodel(data,
name='{}_model'.format(country),
var_dim=VAR_DIM,
mu=ind[tf.newaxis],
current_year=current_year)
models.append(model)
graph = tf.get_default_graph()
prior = tf.reduce_sum([model.priors for model in models
]) + tf.reduce_sum(indiv_priors) + tf.reduce_sum(
graph.get_collection('priors'))
logdensity = tf.reduce_sum([model.logdensities for model in models
]) + tf.reduce_sum(indiv_logdens) + tf.reduce_sum(
graph.get_collection('logdensities'))
kl = logdensity - prior
country_data = {c: d for c, d in zip(ccodes, datas)}
#BUILDING the model
current_year = tf.placeholder(tf.float32, shape=(), name='current_year')
tf.summary.scalar('current_year', current_year)
models = []
with tf.variable_scope(tf.get_variable_scope(),
dtype=floatX,
reuse=tf.AUTO_REUSE):
for country, data in country_data.items():
model = VARmodel(data,
name='{}_model'.format(country),
var_dim=VAR_DIM,
current_year=current_year)
models.append(model)
graph = tf.get_default_graph()
prior = tf.reduce_sum([model.priors for model in models])
logdensity = tf.reduce_sum([model.logdensities for model in models])
kl = logdensity - prior
kl /= 36 * 200 * 4
kls = tf.summary.scalar('KLd', kl)
summary = tf.summary.merge_all()