def __init__(self,
dim_z2h,
dim_h2v,
activation_fn=tf.nn.relu,
temp=.1,
hard=False,
train_pi=True):
self.prior_h = NADE(dim_z2h[1], dim_z2h[0], temp)
self.dim_h2v = dim_h2v
self.w_h2v, self.b_h2v = [], []
if activation_fn == tf.nn.relu:
const = 1.0
else:
const = 4.0
assert len(self.dim_h2v) > 1
for d1, d2, i in zip(dim_h2v[:-1], dim_h2v[1:],
xrange(1, len(self.dim_h2v))):
self.w_h2v.append(
tfe.Variable(tf_xavier_init(d1, d2, const=const),
name='sbn.w_h2v.' + str(i)))
self.b_h2v.append(
tfe.Variable(tf.zeros([d2]),
dtype=tf.float32,
name='sbn.b_h2v.' + str(i)))
self.activation_fn = activation_fn
self.temp = temp
self.hard = hard
def build_model(dataset, trainingparams, hyperparams, hidden_size):
print '\n### Initializing NADE ... ',
start_time = t.time()
model = NADE(dataset,
learning_rate=trainingparams['learning_rate'],
decrease_constant=trainingparams['decrease_constant'],
hidden_size=hidden_size,
random_seed=hyperparams['random_seed'],
batch_size=trainingparams['batch_size'],
hidden_activation=activation_functions[hyperparams['hidden_activation']],
momentum=trainingparams['momentum'],
dropout_rate=trainingparams['dropout_rate'],
weights_initialization=hyperparams['weights_initialization'],
tied=hyperparams['tied'])
print utils.get_done_text(start_time), "###"
# printParams(model)
return model
def __init__(self,
dim_zh2,
dims_h2h1,
dims_h1v,
activation_fn=tf.nn.relu,
temp=.1,
hard=False,
train_pi=True):
self.prior_h2 = NADE(dim_zh2[1], dim_zh2[0], temp)
self.dec_h2h1 = DEC(dims_h2h1,
name='h2h1',
activation_fn=activation_fn,
temp=temp,
hard=hard)
self.dec_h1v = DEC(dims_h1v,
name='h1v',
activation_fn=activation_fn,
temp=temp,
hard=hard)
def __init__(self, num_dims, num_hidden, temp=.1, hard=False):
self.nade = NADE(num_dims, num_hidden, temperature=temp, hard=hard)
