vae_ssl.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

from __future__ import absolute_import
from __future__ import print_function
from __future__ import division
import os
import time

import tensorflow as tf
from six.moves import range
import numpy as np
import zhusuan as zs

from examples import conf
from examples.utils import dataset


@zs.meta_bayesian_net(scope="gen", reuse_variables=True)
def build_gen(n, x_dim, n_class, z_dim, n_particles):
    bn = zs.BayesianNet()
    z_mean = tf.zeros([n, z_dim])
    z = bn.normal("z", z_mean, std=1., group_ndims=1, n_samples=n_particles)
    h_from_z = tf.layers.dense(z, 500)
    y_logits = tf.zeros([n, n_class])
    y = bn.onehot_categorical("y", y_logits)
    h_from_y = tf.layers.dense(tf.cast(y, tf.float32), 500)
    h = tf.nn.relu(h_from_z + h_from_y)
    h = tf.layers.dense(h, 500, activation=tf.nn.relu)
    x_logits = tf.layers.dense(h, x_dim)
    bn.bernoulli("x", x_logits, group_ndims=1)
    return bn


@zs.reuse_variables(scope="variational")
def qz_xy(x, y, z_dim, n_particles):
    bn = zs.BayesianNet()
    h = tf.layers.dense(tf.cast(tf.concat([x, y], -1), tf.float32), 500,
                        activation=tf.nn.relu)
    h = tf.layers.dense(h, 500, activation=tf.nn.relu)
    z_mean = tf.layers.dense(h, z_dim)
    z_logstd = tf.layers.dense(h, z_dim)
    bn.normal("z", z_mean, logstd=z_logstd, group_ndims=1,
              n_samples=n_particles)
    return bn


@zs.reuse_variables("classifier")
def qy_x(x, n_class):
    h = tf.layers.dense(tf.cast(x, tf.float32), 500, activation=tf.nn.relu)
    h = tf.layers.dense(h, 500, activation=tf.nn.relu)
    y_logits = tf.layers.dense(h, n_class)
    return y_logits


def main():
    tf.set_random_seed(1234)
    np.random.seed(1234)

    # Load MNIST
    data_path = os.path.join(conf.data_dir, "mnist.pkl.gz")
    x_labeled, t_labeled, x_unlabeled, x_test, t_test = \
        dataset.load_mnist_semi_supervised(data_path, one_hot=True)
    x_test = np.random.binomial(1, x_test, size=x_test.shape)
    n_labeled, x_dim = x_labeled.shape
    n_class = 10

    # Define model parameters
    z_dim = 100

    # Define training/evaluation parameters
    lb_samples = 10
    beta = 1200.
    epochs = 3000
    batch_size = 100
    test_batch_size = 100
    iters = x_unlabeled.shape[0] // batch_size
    test_iters = x_test.shape[0] // test_batch_size
    test_freq = 10

    # Build the computation graph
    n = tf.placeholder(tf.int32, shape=[], name="n")
    n_particles = tf.placeholder(tf.int32, shape=[], name="n_particles")
    model = build_gen(n, x_dim, n_class, z_dim, n_particles)

    # Labeled
    x_labeled_ph = tf.placeholder(tf.float32, shape=[None, x_dim], name="x_l")
    x_labeled = tf.cast(
        tf.less(tf.random_uniform(tf.shape(x_labeled_ph)), x_labeled_ph),
        tf.int32)
    y_labeled_ph = tf.placeholder(tf.int32, shape=[None, n_class], name="y_l")
    variational = qz_xy(x_labeled, y_labeled_ph, z_dim, n_particles)

    labeled_lower_bound = tf.reduce_mean(
        zs.variational.elbo(model,
                            observed={"x": x_labeled, "y": y_labeled_ph},
                            variational=variational,
                            axis=0))

    # Unlabeled
    # TODO: n not match.

    x_unlabeled_ph = tf.placeholder(tf.float32, shape=[None, x_dim],
                                    name="x_u")
    x_unlabeled = tf.cast(
        tf.less(tf.random_uniform(tf.shape(x_unlabeled_ph)), x_unlabeled_ph),
        tf.int32)
    y_diag = tf.eye(n_class, dtype=tf.int32)
    y_u = tf.reshape(tf.tile(y_diag[None, ...], [n, 1, 1]), [-1, n_class])
    x_u = tf.reshape(tf.tile(x_unlabeled[:, None, ...], [1, n_class, 1]),
                     [-1, x_dim])
    variational = qz_xy(x_u, y_u, z_dim, n_particles)
    lb_z = zs.variational.elbo(model,
                               observed={"x": x_u, "y": y_u},
                               variational=variational,
                               axis=0)
    # sum over y
    lb_z = tf.reshape(lb_z, [-1, n_class])
    qy_logits_u = qy_x(x_unlabeled_ph, n_class)
    qy_u = tf.nn.softmax(qy_logits_u) + 1e-8
    qy_u /= tf.reduce_sum(qy_u, 1, keepdims=True)
    log_qy_u = tf.log(qy_u)
    unlabeled_lower_bound = tf.reduce_mean(
        tf.reduce_sum(qy_u * (lb_z - log_qy_u), 1))

    # Build classifier
    qy_logits_l = qy_x(x_labeled, n_class)
    qy_l = tf.nn.softmax(qy_logits_l)
    pred_y = tf.argmax(qy_l, 1)
    acc = tf.reduce_sum(
        tf.cast(tf.equal(pred_y, tf.argmax(y_labeled_ph, 1)), tf.float32) /
        tf.cast(tf.shape(x_labeled)[0], tf.float32))
    onehot_cat = zs.distributions.OnehotCategorical(qy_logits_l)
    log_qy_x = onehot_cat.log_prob(y_labeled_ph)
    classifier_cost = -beta * tf.reduce_mean(log_qy_x)

    # Gather gradients
    cost = -(labeled_lower_bound + unlabeled_lower_bound -
             classifier_cost) / 2.
    optimizer = tf.train.AdamOptimizer(learning_rate=3e-4)
    grads = optimizer.compute_gradients(cost)
    infer_op = optimizer.apply_gradients(grads)

    # Run the inference
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        for epoch in range(1, epochs + 1):
            time_epoch = -time.time()
            np.random.shuffle(x_unlabeled)
            lbs_labeled, lbs_unlabeled, train_accs = [], [], []

            for t in range(iters):
                labeled_indices = np.random.randint(0, n_labeled,
                                                    size=batch_size)
                x_labeled_batch = x_labeled[labeled_indices]
                y_labeled_batch = t_labeled[labeled_indices]
                x_unlabeled_batch = x_unlabeled[t * batch_size:
                                                (t + 1) * batch_size]
                _, lb_labeled, lb_unlabeled, train_acc = sess.run(
                    [infer_op, labeled_lower_bound, unlabeled_lower_bound,
                     acc],
                    feed_dict={x_labeled_ph: x_labeled_batch,
                               y_labeled_ph: y_labeled_batch,
                               x_unlabeled_ph: x_unlabeled_batch,
                               n_particles: lb_samples,
                               n: batch_size})
                lbs_labeled.append(lb_labeled)
                lbs_unlabeled.append(lb_unlabeled)
                train_accs.append(train_acc)

            time_epoch += time.time()
            print('Epoch {} ({:.1f}s), Lower bound: labeled = {}, '
                  'unlabeled = {} Accuracy: {:.2f}%'.
                  format(epoch, time_epoch, np.mean(lbs_labeled),
                         np.mean(lbs_unlabeled), np.mean(train_accs) * 100.))

            if epoch % test_freq == 0:
                time_test = -time.time()
                test_lls_labeled, test_lls_unlabeled, test_accs = [], [], []

                for t in range(test_iters):
                    test_x_batch = x_test[
                        t * test_batch_size: (t + 1) * test_batch_size]
                    test_y_batch = t_test[
                        t * test_batch_size: (t + 1) * test_batch_size]
                    test_ll_labeled, test_ll_unlabeled, test_acc = sess.run(
                        [labeled_lower_bound, unlabeled_lower_bound, acc],
                        feed_dict={x_labeled: test_x_batch,
                                   y_labeled_ph: test_y_batch,
                                   x_unlabeled: test_x_batch,
                                   n_particles: lb_samples,
                                   n: test_batch_size})
                    test_lls_labeled.append(test_ll_labeled)
                    test_lls_unlabeled.append(test_ll_unlabeled)
                    test_accs.append(test_acc)

                time_test += time.time()
                print('>>> TEST ({:.1f}s)'.format(time_test))
                print('>> Test lower bound: labeled = {}, unlabeled = {}'.
                      format(np.mean(test_lls_labeled),
                             np.mean(test_lls_unlabeled)))
                print('>> Test accuracy: {:.2f}%'.format(
                    100. * np.mean(test_accs)))


if __name__ == "__main__":
    main()