with tf.device("/gpu:0"):
    x = tf.placeholder(tf.float32, [None, feature_dim], name="input")
    y_ = tf.placeholder(tf.float32, [None, output_dim],name="output")
    phase = tf.placeholder(tf.bool, name='phase')

    weights = {
        'hidden1': tf.Variable(tf.random_normal([feature_dim, hidden_num_units1], stddev=1 / np.sqrt(feature_dim)), name="w1"),
        'hidden2': tf.Variable(
            tf.random_normal([hidden_num_units1, hidden_num_units2], stddev=1 / np.sqrt(hidden_num_units1)), name="w2"),
        'output': tf.Variable(
            tf.random_normal([hidden_num_units2, output_num_units], stddev=1 / np.sqrt(hidden_num_units2)), name="w3")
    }

    biases = {
        'hidden1': tf.Variable(tf.random_normal([hidden_num_units1], stddev=1 / np.sqrt(hidden_num_units1)), name="b1"),
        'hidden2': tf.Variable(tf.random_normal([hidden_num_units2], stddev=1 / np.sqrt(hidden_num_units2)), name="b2"),
        'output': tf.Variable(tf.random_normal([output_num_units], stddev=1 / np.sqrt(output_num_units)), name="b3")
    }

def network1(data):
    with tf.name_scope(name="layer_1"):
        h1 = tf.add(tf.matmul(data, weights['hidden1']), biases['hidden1'])
        h1_BN = tf.contrib.layers.batch_norm(h1,
                                          center=True, scale=True,
                                          is_training=phase,
                                          scope='hidden_layer1_bn')
        h1_relu = tf.nn.relu(h1_BN)
    with tf.name_scope(name="layer_2"):
        h2 = tf.add(tf.matmul(h1_relu, weights['hidden2']), biases['hidden2'])
        h2_BN = tf.contrib.layers.batch_norm(h2,
                                          center=True, scale=True,
                                          is_training=phase,
                                          scope='hidden_layer2_bn')
        h2_relu = tf.nn.relu(h2_BN)

    with tf.name_scope(name="output_layer"):
        output_layer = tf.add(tf.matmul(h2_relu, weights['output']), biases['output'])
        output__layer_BN = tf.contrib.layers.batch_norm(output_layer,
                                          center=True, scale=True,
                                          is_training=phase,
                                          scope='output_bn')
        output = tf.sigmoid(output__layer_BN, name="f" )
    return output

def network1(data):
    weights = {
        'hidden1': tf.Variable(tf.random_normal([feature_dim, hidden_num_units1], stddev=1 / np.sqrt(feature_dim)), name="w1"),
        'hidden2': tf.Variable(
            tf.random_normal([hidden_num_units1, hidden_num_units2], stddev=1 / np.sqrt(hidden_num_units1)), name="w2"),
        'output': tf.Variable(
            tf.random_normal([hidden_num_units2, output_num_units], stddev=1 / np.sqrt(hidden_num_units2)), name="w3")
    }

    biases = {
        'hidden1': tf.Variable(tf.random_normal([hidden_num_units1], stddev=1 / np.sqrt(hidden_num_units1)), name="b1"),
        'hidden2': tf.Variable(tf.random_normal([hidden_num_units2], stddev=1 / np.sqrt(hidden_num_units2)), name="b2"),
        'output': tf.Variable(tf.random_normal([output_num_units], stddev=1 / np.sqrt(output_num_units)), name="b3")
    }

    with tf.name_scope(name = "layer_1"):
        hidden_layer1 = tf.add(tf.matmul(data, weights['hidden1']), biases['hidden1'])
        hidden_layer1 = tf.nn.relu(hidden_layer1)
    with tf.name_scope(name = "layer_1_dropout"):
        drop_out1 = tf.nn.dropout(hidden_layer1, keep_prob)  # DROP-OUT here
    with tf.name_scope(name = "layer_2"):
        hidden_layer2 = tf.add(tf.matmul(drop_out1, weights['hidden2']), biases['hidden2'])
        hidden_layer2 = tf.nn.relu(hidden_layer2)
    with tf.name_scope(name = "layer_2_dropout"):
        drop_out2 = tf.nn.dropout(hidden_layer2, keep_prob)  # DROP-OUT here
    with tf.name_scope(name = "output_layer"):
        output = tf.add(tf.matmul(drop_out2, weights['output']), biases['output'])
        output = tf.nn.sigmoid(output,  name="f")
    return output