def __init__(self, layer_sizes, input_tensor, learning_rate=.001, hidden_tensor=None, n_fantasy_states=100, persistent_contrastive_divergence=True, constrastive_divergence_level=1):
self.layer_sizes = layer_sizes
self.machines = []
current_input = input_tensor
current_input_size = int(input_tensor.get_shape()[-1])
for i, size in enumerate(layer_sizes):
rbm = RBM(current_input_size, size, current_input, learning_rate=learning_rate,
n_fantasy_states=n_fantasy_states, name='dbn_{}'.format(i), persistent_contrastive_divergence=persistent_contrastive_divergence, contrastive_divergence_level=constrastive_divergence_level)
current_input_size = size
current_input = rbm.hidden_state_plus
self.machines.append(rbm)
# def __init__(self, n_features, n_hidden, visible_placeholder=None, contrastive_divergence_level=1, learning_rate=1e-3,
# persistent_contrastive_divergence=True, n_fantasy_states=100, name='rbm'):
if hidden_tensor is None:
self.hidden_tensor = tf.placeholder(tf.float32, shape=[None, layer_sizes[-1]])
else:
self.hidden_tensor = hidden_tensor
current_hidden = self.hidden_tensor
self.hidden_tensor = self.hidden_tensor
for i in range(len(layer_sizes)):
# dirty hack to connect back to front for sampling
# TODO: Make this better
index = len(layer_sizes) - 1 - i
rbm = self.machines[index]
with tf.name_scope(rbm.name):
visible_prop = tf.nn.sigmoid(tf.matmul(current_hidden, rbm.w, transpose_b=True) + rbm.b_v)
rbm.visible_from_hidden = RBM._sample_from(visible_prop, 'visible_from_hidden')
current_hidden = rbm.visible_from_hidden
