def __init__(self, logger, learning_rate, input_dim, z_dim, ae_h_dims, *args, **kwargs):
self.scope = "AE"
self.logger = logger
self.learning_rate = learning_rate
self.input_dim = input_dim
self.z_dim = z_dim
self.enc_layer_dims = [input_dim, *ae_h_dims, z_dim]
self.dec_layer_dims = [z_dim, *list(reversed(ae_h_dims)), input_dim] #todo : just reverse enc_layer_dims
self.logger.info("[*] Building AE model")
with tf.variable_scope(self.scope):
self.input = tf.placeholder(tf.float32, [None, self.input_dim])
enc = Encoder(self.enc_layer_dims)
dec = Decoder(self.dec_layer_dims)
z_layer = enc.encode(self.input)
# todo : how to handle output?
_, _, self.output = dec.decode(z_layer)
# todo: refactoring get theta method --> get solver?
enc_theta = enc.get_theta()
dec_theta = dec.get_theta()
self.theta = [*enc_theta, *dec_theta]
#l2_loss = enc.get_l2_loss() +
self.recon_loss = tf.reduce_mean(tf.square(self.input-self.output))
self.solver = tf.train.AdamOptimizer(self.learning_rate).minimize(self.recon_loss, var_list=theta)
def test_encoder_output_shape(self):
with self.test_session():
input_dim = 100
output_dim = 10
batch_size = 512
layer_dims = [input_dim, 50, output_dim]
X = tf.constant(np.random.random(batch_size * input_dim),
shape=[batch_size, input_dim],
dtype=tf.float32)
enc = Encoder(layer_dims)
self.assertShapeEqual(
np.reshape(np.zeros(batch_size * output_dim),
[batch_size, output_dim]), enc.encode(X))