def mlstm()

in encoder.py [0:0]

• 40 lines of code
• 5 McCabe index (conditional complexity)

def mlstm(inputs, c, h, M, ndim, scope='lstm', wn=False):
    nin = inputs[0].get_shape()[1].value
    with tf.variable_scope(scope):
        wx = tf.get_variable("wx", [nin, ndim * 4], initializer=load_params)
        wh = tf.get_variable("wh", [ndim, ndim * 4], initializer=load_params)
        wmx = tf.get_variable("wmx", [nin, ndim], initializer=load_params)
        wmh = tf.get_variable("wmh", [ndim, ndim], initializer=load_params)
        b = tf.get_variable("b", [ndim * 4], initializer=load_params)
        if wn:
            gx = tf.get_variable("gx", [ndim * 4], initializer=load_params)
            gh = tf.get_variable("gh", [ndim * 4], initializer=load_params)
            gmx = tf.get_variable("gmx", [ndim], initializer=load_params)
            gmh = tf.get_variable("gmh", [ndim], initializer=load_params)

    if wn:
        wx = tf.nn.l2_normalize(wx, dim=0) * gx
        wh = tf.nn.l2_normalize(wh, dim=0) * gh
        wmx = tf.nn.l2_normalize(wmx, dim=0) * gmx
        wmh = tf.nn.l2_normalize(wmh, dim=0) * gmh

    cs = []
    for idx, x in enumerate(inputs):
        m = tf.matmul(x, wmx)*tf.matmul(h, wmh)
        z = tf.matmul(x, wx) + tf.matmul(m, wh) + b
        i, f, o, u = tf.split(z, 4, 1)
        i = tf.nn.sigmoid(i)
        f = tf.nn.sigmoid(f)
        o = tf.nn.sigmoid(o)
        u = tf.tanh(u)
        if M is not None:
            ct = f*c + i*u
            ht = o*tf.tanh(ct)
            m = M[:, idx, :]
            c = ct*m + c*(1-m)
            h = ht*m + h*(1-m)
        else:
            c = f*c + i*u
            h = o*tf.tanh(c)
        inputs[idx] = h
        cs.append(c)
    cs = tf.stack(cs)
    return inputs, cs, c, h