def __init__(self, input_dim, output_dim, training=True):
self.training = training
nb = NetworkBuilder()
with tf.name_scope("Input"):
self.input = tf.placeholder(tf.float32,
shape=[None, input_dim, input_dim, 1],
name="input")
with tf.name_scope("Output"):
self.output = tf.placeholder(tf.float32,
shape=[None, output_dim],
name="output")
with tf.name_scope("ImageModel"):
model = self.input
model = nb.add_batch_normalization(model, self.training)
model = nb.add_conv_layer(model,
output_size=64,
feature_size=(4, 4),
padding='SAME',
activation=tf.nn.relu)
model = nb.add_max_pooling_layer(model)
model = nb.add_dropout(model, 0.1, self.training)
model = nb.add_conv_layer(model,
64,
feature_size=(4, 4),
activation=tf.nn.relu,
padding='VALID')
model = nb.add_max_pooling_layer(model)
model = nb.add_dropout(model, 0.3, self.training)
model = nb.flatten(model)
model = nb.add_dense_layer(model, 256, tf.nn.relu)
model = nb.add_dropout(model, 0.5, self.training)
model = nb.add_dense_layer(model, 64, tf.nn.relu)
model = nb.add_batch_normalization(model, self.training)
self.logits = nb.add_dense_layer(model,
output_dim,
activation=tf.nn.softmax)
