Simple gradient descent using mxnet

Question

I'm trying to use MXNet's gradient descent optimizers to minimize a function. The equivalent example in Tensorflow would be:

import tensorflow as tf

x = tf.Variable(2, name='x', dtype=tf.float32)
log_x = tf.log(x)
log_x_squared = tf.square(log_x)

optimizer = tf.train.GradientDescentOptimizer(0.5)
train = optimizer.minimize(log_x_squared)

init = tf.initialize_all_variables()

def optimize():
  with tf.Session() as session:
    session.run(init)
    print("starting at", "x:", session.run(x), "log(x)^2:", session.run(log_x_squared))
    for step in range(10):  
      session.run(train)
      print("step", step, "x:", session.run(x), "log(x)^2:", session.run(log_x_squared))

I am not sure how to accomplish the same in MXNet. The optimizer API documentation does not appear to have an equivalent method. Here's what I've tried so far. The main confusion has been around the need to pass training data:

import mxnet as mx

x = mx.sym.Variable('data')
log_x = mx.sym.log(x)
log_x_squared = mx.sym.square(log_x)

mod = mx.mod.Module(log_x_squared)  # Create a module where the loss function
                                    # is the one we want to optimize
mod.bind(data_shapes=[('data', (1,1))])  # ?? not sure if this is correct - we
                                         # are saying our input is a scalar
mod.init_params()
mod.init_optimizer()  # SGD is default

mod.fit()  # ?? must pass data_iter to fit

It seems like the x variable should be somehow fed back in as the data_iter but I don't know how to accomplish this.

Update: thanks to kevinthesun for their excellent answer! Here is a working minimization routine built on top of a single hidden-layer neural net:

import mxnet as mx
import numpy as np


def minimize(objective_function,
             initial_params,
             max_iters=1000,
             optimizer='sgd',
             optimizer_params=(('learning_rate', 0.1),),
             tol=1e-8):

    class InitialParam(mx.init.Initializer):

        def __init__(self, vals):
            super(InitialParam, self).__init__()
            self._vals = vals

        def _init_weight(self, _, arr):
            arr[:] = self._vals.asnumpy()[:, np.newaxis]


    x = mx.sym.Variable('data')
    params_len = initial_params.shape[0]
    fc = mx.sym.FullyConnected(data=x, name='fc1',
                               num_hidden=params_len,
                               no_bias=True)

    # Passing the FullyConnected layer into the objective function
    # is difficult to manipulate. If the fully connected layer represents
    # [x, y] for optimizing a 2 dimensional function f(x, y) it is easier
    # to work with x, and y. So we split the fully connected layer into a
    # number of symbols for each parameter:
    param_syms = []
    for i in range(params_len):
        ps = mx.sym.slice(fc, begin=(0, i), end=(1, i + 1))
        param_syms.append(ps)

    # The loss function for the network is our objective function.
    loss = mx.sym.MakeLoss(objective_function(param_syms))
    mod = mx.mod.Module(loss)

    mod.bind(data_shapes=[('data', (1,))])
    mod.init_params(InitialParam(initial_params))
    mod.init_optimizer(optimizer=optimizer,
                       optimizer_params=optimizer_params)

    (o_name, o_shape), = mod.output_shapes

    i = 0
    params = initial_params
    old_val = np.full(o_shape, np.nan)
    while i < max_iters:
        mod.forward_backward(mx.io.DataBatch(
            data=[mx.nd.ones((1,))])) 
        mod.update()
        params = mod.get_params()[0]['fc1_weight']
        val = mod.get_outputs()[0].asnumpy()
        if np.allclose(old_val, val, atol=tol):
            print 'Function value: {}'.format(val)
            print 'Iterations: {}'.format(i)
            return params

        old_val = val
        i += 1

    return params

and using it:

def my_func(x):
    return (x[0] + 1) ** 2

p = minimize(my_func, mx.nd.array([1.0]))
p.asnumpy()

>>> array([[-0.99999988]], dtype=float32)

and another:

def my_func(x):
    return (x[0] + 1) ** 2 + (x[1] - 2) ** 2 + (x[2] + 3) ** 2

p = minimize(my_func, mx.nd.array([1.0, 1.5, 2.0]))
p.asnumpy()

>>> array([[-0.99996436],
           [ 1.99999106],
           [-2.99991083]], dtype=float32)

Answer 1

Currently it is not as easy as tensorflow to optimize a simple function using MXNet, due to the lack of support in frontend.

First you need a loss Function as the last layer of your network. Here it is log_x_squared. Use MakeLoss to create a loss function.

Second is the input and weights. Since currently in MXNet Variable is not counted as trainable weight, you need to set x as weight. Here is a workaround: Set a 'fake' input variable which is always to be 1. After it add a fullyconnected layer with 1 hidden unit and no bias. This gives us "1 * x". Now our x is a weight.

Third if you would like to optimize multiple times on single data sample, module.fit might not be the best choice. After initializing optimizer. You just need to call module.forward_backward() and module.update() multiple times. For forward_backward function you need to pass a databatch, which is a simpler interface comparing to dataiter. Here we just need to pass a constant ndarray of 1 every time.

Actually we construct a computation graph of log(1 * x) ^ 2 and x becomes a weight instead of variable.

Anyway, we should consider providing a similar interface of tensorflow to optimize variable.

Hope this is useful info!

Answer 2

mxnet is different from tensorflow. You need to read this tutorial