problems.py

This file stores the problems (a.k.a. objective function or optimizee) that you want LOIS to train or evaluate. Every problem should be in a format of function. Typically the function has two arguments:

• num_dims: The dimension of the problem.
• mode: A str indicates whether the problem is used during training or testing.

This file is called by the function get_config in util.py. You need to add this function in get_config as well following the template there.

An example of a problem is Rastrigin (square_cos):

def square_cos(num_dims=2, mode='train'):

    def build():
        """Builds loss graph."""

        batch_size=128
        stddev=0.01
        dtype=tf.float32
        if mode=='test':
        x = tf.get_variable(
            "x",
            shape=[batch_size, num_dims],
            dtype=dtype,
            initializer=tf.random_uniform_initializer(-3, 3))
        return ( tf.reduce_sum(x*x - 10*tf.math.cos(2*3.1415926*x), 1)+ 10*num_dims )


        # Trainable variable.
        x = tf.get_variable(
            "x",
            shape=[batch_size, num_dims],
            dtype=dtype,
            initializer=tf.random_uniform_initializer(-3, 3))

        # Non-trainable variables.
        w = tf.get_variable("w",
                            dtype=dtype,
                            initializer=indentity_init(batch_size, num_dims, stddev/num_dims),
                            trainable=False)

        y = tf.get_variable("y",
                            shape=[batch_size, num_dims],
                            dtype=dtype,
                            initializer=tf.random_normal_initializer(stddev=stddev/num_dims),
                            trainable=False)

        wcos = tf.get_variable("wcos",
                            shape=[batch_size, num_dims],
                            dtype=dtype,
                            initializer=tf.random_normal_initializer(mean=1.0, stddev=stddev/num_dims),
                            trainable=False)

        product = tf.squeeze(tf.matmul(w, tf.expand_dims(x, -1)))
        product2 = tf.reduce_sum(wcos*10*tf.math.cos(2*3.1415926*x), 1)


        return (tf.reduce_sum((product - y) ** 2, 1)) - tf.reduce_mean(product2) + 10*num_dims
    return build

Its caller in get_config function is:

elif  "square_cos" in problem_name:
    num_dims = int(problem_name.split('_')[-1])
    problem = problems.square_cos(batch_size=128, num_dims=num_dims, mode=mode)
    net_config = {"cw": {
        "net": "CoordinateWiseDeepLSTM",
        "net_options": {"layers": (20, 20)},
        "net_path": get_net_path("cw", path)
    }}
    net_assignments = None