Python networks.discriminator() Examples

def _get_optimizer(gen_lr, dis_lr):
  """Returns generator optimizer and discriminator optimizer.

  Args:
    gen_lr: A scalar float `Tensor` or a Python number.  The Generator learning
        rate.
    dis_lr: A scalar float `Tensor` or a Python number.  The Discriminator
        learning rate.

  Returns:
    A tuple of generator optimizer and discriminator optimizer.
  """
  # beta1 follows
  # https://github.com/junyanz/CycleGAN/blob/master/options.lua
  gen_opt = tf.train.AdamOptimizer(gen_lr, beta1=0.5, use_locking=True)
  dis_opt = tf.train.AdamOptimizer(dis_lr, beta1=0.5, use_locking=True)
  return gen_opt, dis_opt 

def _define_model(images_x, images_y):
  """Defines a CycleGAN model that maps between images_x and images_y.

  Args:
    images_x: A 4D float `Tensor` of NHWC format.  Images in set X.
    images_y: A 4D float `Tensor` of NHWC format.  Images in set Y.

  Returns:
    A `CycleGANModel` namedtuple.
  """
  cyclegan_model = tfgan.cyclegan_model(
      generator_fn=networks.generator,
      discriminator_fn=networks.discriminator,
      data_x=images_x,
      data_y=images_y)

  # Add summaries for generated images.
  tfgan.eval.add_cyclegan_image_summaries(cyclegan_model)

  return cyclegan_model 

def define_train_ops(gan_model, gan_loss, **kwargs):
  """Defines progressive GAN train ops.

  Args:
    gan_model: A `GANModel` namedtuple.
    gan_loss: A `GANLoss` namedtuple.
    **kwargs: A dictionary of
        'adam_beta1': A float of Adam optimizer beta1.
        'adam_beta2': A float of Adam optimizer beta2.
        'generator_learning_rate': A float of generator learning rate.
        'discriminator_learning_rate': A float of discriminator learning rate.

  Returns:
    A tuple of `GANTrainOps` namedtuple and a list variables tracking the state
    of optimizers.
  """
  with tf.variable_scope('progressive_gan_train_ops') as var_scope:
    beta1, beta2 = kwargs['adam_beta1'], kwargs['adam_beta2']
    gen_opt = tf.train.AdamOptimizer(kwargs['generator_learning_rate'], beta1,
                                     beta2)
    dis_opt = tf.train.AdamOptimizer(kwargs['discriminator_learning_rate'],
                                     beta1, beta2)
    gan_train_ops = tfgan.gan_train_ops(gan_model, gan_loss, gen_opt, dis_opt)
  return gan_train_ops, tf.get_collection(
      tf.GraphKeys.GLOBAL_VARIABLES, scope=var_scope.name) 

def define_train_ops(gan_model, gan_loss, **kwargs):
  """Defines progressive GAN train ops.

  Args:
    gan_model: A `GANModel` namedtuple.
    gan_loss: A `GANLoss` namedtuple.
    **kwargs: A dictionary of
        'adam_beta1': A float of Adam optimizer beta1.
        'adam_beta2': A float of Adam optimizer beta2.
        'generator_learning_rate': A float of generator learning rate.
        'discriminator_learning_rate': A float of discriminator learning rate.

  Returns:
    A tuple of `GANTrainOps` namedtuple and a list variables tracking the state
    of optimizers.
  """
  with tf.variable_scope('progressive_gan_train_ops') as var_scope:
    beta1, beta2 = kwargs['adam_beta1'], kwargs['adam_beta2']
    gen_opt = tf.train.AdamOptimizer(kwargs['generator_learning_rate'], beta1,
                                     beta2)
    dis_opt = tf.train.AdamOptimizer(kwargs['discriminator_learning_rate'],
                                     beta1, beta2)
    gan_train_ops = tfgan.gan_train_ops(gan_model, gan_loss, gen_opt, dis_opt)
  return gan_train_ops, tf.get_collection(
      tf.GraphKeys.GLOBAL_VARIABLES, scope=var_scope.name) 

def _define_model(images_x, images_y):
  """Defines a CycleGAN model that maps between images_x and images_y.

  Args:
    images_x: A 4D float `Tensor` of NHWC format.  Images in set X.
    images_y: A 4D float `Tensor` of NHWC format.  Images in set Y.

  Returns:
    A `CycleGANModel` namedtuple.
  """
  cyclegan_model = tfgan.cyclegan_model(
      generator_fn=networks.generator,
      discriminator_fn=networks.discriminator,
      data_x=images_x,
      data_y=images_y)

  # Add summaries for generated images.
  tfgan.eval.add_image_comparison_summaries(
      cyclegan_model, num_comparisons=3, display_diffs=False)
  tfgan.eval.add_gan_model_image_summaries(
      cyclegan_model, grid_size=int(np.sqrt(FLAGS.batch_size)))

  return cyclegan_model 

def _get_optimizer(gen_lr, dis_lr):
  """Returns generator optimizer and discriminator optimizer.

  Args:
    gen_lr: A scalar float `Tensor` or a Python number.  The Generator learning
        rate.
    dis_lr: A scalar float `Tensor` or a Python number.  The Discriminator
        learning rate.

  Returns:
    A tuple of generator optimizer and discriminator optimizer.
  """
  # beta1 follows
  # https://github.com/junyanz/CycleGAN/blob/master/options.lua
  gen_opt = tf.train.AdamOptimizer(gen_lr, beta1=0.5, use_locking=True)
  dis_opt = tf.train.AdamOptimizer(dis_lr, beta1=0.5, use_locking=True)
  return gen_opt, dis_opt 

def _get_optimizer(gen_lr, dis_lr):
  """Returns generator optimizer and discriminator optimizer.

  Args:
    gen_lr: A scalar float `Tensor` or a Python number.  The Generator learning
        rate.
    dis_lr: A scalar float `Tensor` or a Python number.  The Discriminator
        learning rate.

  Returns:
    A tuple of generator optimizer and discriminator optimizer.
  """
  # beta1 follows
  # https://github.com/junyanz/CycleGAN/blob/master/options.lua
  gen_opt = tf.train.AdamOptimizer(gen_lr, beta1=0.5, use_locking=True)
  dis_opt = tf.train.AdamOptimizer(dis_lr, beta1=0.5, use_locking=True)
  return gen_opt, dis_opt 

def _define_model(images_x, images_y):
  """Defines a CycleGAN model that maps between images_x and images_y.

  Args:
    images_x: A 4D float `Tensor` of NHWC format.  Images in set X.
    images_y: A 4D float `Tensor` of NHWC format.  Images in set Y.

  Returns:
    A `CycleGANModel` namedtuple.
  """
  cyclegan_model = tfgan.cyclegan_model(
      generator_fn=networks.generator,
      discriminator_fn=networks.discriminator,
      data_x=images_x,
      data_y=images_y)

  # Add summaries for generated images.
  tfgan.eval.add_cyclegan_image_summaries(cyclegan_model)

  return cyclegan_model 

def _lr(gen_lr_base, dis_lr_base):
  """Return the generator and discriminator learning rates."""
  gen_lr = tf.train.exponential_decay(
      learning_rate=gen_lr_base,
      global_step=tf.train.get_or_create_global_step(),
      decay_steps=100000,
      decay_rate=0.8,
      staircase=True,)
  dis_lr = dis_lr_base

  return gen_lr, dis_lr 

def test_discriminator_run(self):
    img_batch = tf.zeros([3, 70, 70, 3])
    disc_output = networks.discriminator(img_batch)
    with self.test_session() as sess:
      sess.run(tf.global_variables_initializer())
      sess.run(disc_output) 

def define_loss(gan_model, **kwargs):
  """Defines progressive GAN losses.

  The generator and discriminator both use wasserstein loss. In addition,
  a small penalty term is added to the discriminator loss to prevent it getting
  too large.

  Args:
    gan_model: A `GANModel` namedtuple.
    **kwargs: A dictionary of
        'gradient_penalty_weight': A float of gradient norm target for
            wasserstein loss.
        'gradient_penalty_target': A float of gradient penalty weight for
            wasserstein loss.
        'real_score_penalty_weight': A float of Additional penalty to keep
            the scores from drifting too far from zero.

  Returns:
    A `GANLoss` namedtuple.
  """
  gan_loss = tfgan.gan_loss(
      gan_model,
      generator_loss_fn=tfgan.losses.wasserstein_generator_loss,
      discriminator_loss_fn=tfgan.losses.wasserstein_discriminator_loss,
      gradient_penalty_weight=kwargs['gradient_penalty_weight'],
      gradient_penalty_target=kwargs['gradient_penalty_target'],
      gradient_penalty_epsilon=0.0)

  real_score_penalty = tf.reduce_mean(
      tf.square(gan_model.discriminator_real_outputs))
  tf.summary.scalar('real_score_penalty', real_score_penalty)

  return gan_loss._replace(
      discriminator_loss=(
          gan_loss.discriminator_loss +
          kwargs['real_score_penalty_weight'] * real_score_penalty)) 

def test_discriminator_run(self):
    img_batch = tf.zeros([3, 70, 70, 3])
    disc_output = networks.discriminator(img_batch)
    with self.test_session() as sess:
      sess.run(tf.global_variables_initializer())
      sess.run(disc_output) 

def test_discriminator_invalid_input(self):
    wrong_dim_input = tf.zeros([5, 32, 32])
    with self.assertRaisesRegexp(ValueError, 'Shape must be rank 4'):
      networks.discriminator(wrong_dim_input)

    not_fully_defined = tf.placeholder(tf.float32, [3, None, 32, 3])
    with self.assertRaisesRegexp(ValueError, 'Shape .* is not fully defined'):
      networks.compression_model(not_fully_defined) 

def _optimizer(gen_lr, dis_lr):
  # First is generator optimizer, second is discriminator.
  adam_kwargs = {
      'epsilon': 1e-8,
      'beta1': 0.5,
  }
  return (tf.train.AdamOptimizer(gen_lr, **adam_kwargs),
          tf.train.AdamOptimizer(dis_lr, **adam_kwargs)) 

def _lr(gen_lr_base, dis_lr_base):
  """Return the generator and discriminator learning rates."""
  gen_lr_kwargs = {
      'decay_steps': 60000,
      'decay_rate': 0.9,
      'staircase': True,
  }
  gen_lr = tf.train.exponential_decay(
      learning_rate=gen_lr_base,
      global_step=tf.train.get_or_create_global_step(),
      **gen_lr_kwargs)
  dis_lr = dis_lr_base

  return gen_lr, dis_lr 

def _get_gan_model(generator_inputs, generated_data, real_data,
                   generator_scope):
  """Manually construct and return a GANModel tuple."""
  generator_vars = tf.contrib.framework.get_trainable_variables(generator_scope)

  discriminator_fn = networks.discriminator
  with tf.variable_scope('discriminator') as dis_scope:
    discriminator_gen_outputs = discriminator_fn(generated_data)
  with tf.variable_scope(dis_scope, reuse=True):
    discriminator_real_outputs = discriminator_fn(real_data)
  discriminator_vars = tf.contrib.framework.get_trainable_variables(
      dis_scope)

  # Manually construct GANModel tuple.
  gan_model = tfgan.GANModel(
      generator_inputs=generator_inputs,
      generated_data=generated_data,
      generator_variables=generator_vars,
      generator_scope=generator_scope,
      generator_fn=None,  # not necessary
      real_data=real_data,
      discriminator_real_outputs=discriminator_real_outputs,
      discriminator_gen_outputs=discriminator_gen_outputs,
      discriminator_variables=discriminator_vars,
      discriminator_scope=dis_scope,
      discriminator_fn=discriminator_fn)

  return gan_model 

def train():
    real_img = tf.placeholder(tf.float32, [None, H, W, 3])
    label = tf.placeholder(tf.int32, [None])
    z = tf.placeholder(tf.float32, [None, 100])
    one_hot_label = tf.one_hot(label, NUMS_CLASS)
    labeled_z = tf.concat([z, one_hot_label], axis=1)
    G = generator("generator")
    D = discriminator("discriminator")
    fake_img = G(labeled_z)
    class_fake_logits, adv_fake_logits = D(fake_img, NUMS_CLASS)
    class_real_logits, adv_real_logits = D(real_img, NUMS_CLASS)
    loss_d_real = -tf.reduce_mean(tf.log(adv_real_logits + EPSILON))
    loss_d_fake = -tf.reduce_mean(tf.log(1 - adv_fake_logits + EPSILON))
    loss_cls_real = -tf.reduce_mean(tf.log(tf.reduce_sum(class_real_logits * one_hot_label, axis=1) + EPSILON))
    loss_cls_fake = -tf.reduce_mean(tf.log(tf.reduce_sum(class_fake_logits * one_hot_label, axis=1) + EPSILON))
    D_loss = loss_d_real + loss_d_fake + loss_cls_real
    G_loss =  -tf.reduce_mean(tf.log(adv_fake_logits + EPSILON)) + loss_cls_fake

    D_opt = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(D_loss, var_list=D.var_list())
    G_opt = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(G_loss, var_list=G.var_list())
    sess = tf.Session()
    sess.run(tf.global_variables_initializer())
    saver = tf.train.Saver()
    data, labels = read_face_data("./dataset/face_woman_man.mat")
    for i in range(50000):
        s = time.time()
        for j in range(1):
            BATCH, LABELS, Z = get_batch_face(data, labels, BATCHSIZE)
            BATCH = BATCH / 127.5 - 1.0
            sess.run(D_opt, feed_dict={real_img: BATCH, label: LABELS, z: Z})
        sess.run(G_opt, feed_dict={real_img: BATCH, label: LABELS, z: Z})
        e = time.time()
        if i % 100 == 0:
            [D_LOSS, G_LOSS, FAKE_IMG] = sess.run([D_loss, G_loss, fake_img], feed_dict={real_img: BATCH, label: LABELS, z: Z})
            Image.fromarray(np.uint8((FAKE_IMG[0, :, :, :] + 1) * 127.5)).save("./results/" + str(i) +"_" + str(int(LABELS[0])) + ".jpg")
            print("Iteration: %d, D_loss: %f, G_loss: %f, update_time: %f"%(i, D_LOSS, G_LOSS, e-s))
        if i % 500 == 0:
            saver.save(sess, "./save_para/model.ckpt")
    pass 

def test_discriminator(self):
    batch_size = 5
    image = tf.random_uniform([batch_size, 32, 32, 3], -1, 1)
    dis_output = networks.discriminator(image, None)
    with self.test_session(use_gpu=True) as sess:
      sess.run(tf.global_variables_initializer())
      dis_output_np = dis_output.eval()

    self.assertAllEqual([batch_size, 1], dis_output_np.shape) 

def test_discriminator_run(self):
    img_batch = tf.zeros([3, 70, 70, 3])
    disc_output = networks.discriminator(img_batch)
    with self.test_session() as sess:
      sess.run(tf.global_variables_initializer())
      sess.run(disc_output) 

def test_discriminator_graph(self):
    # Check graph construction for a number of image size/depths and batch
    # sizes.
    for batch_size, patch_size in zip([3, 6], [70, 128]):
      tf.reset_default_graph()
      img = tf.ones([batch_size, patch_size, patch_size, 3])
      disc_output = networks.discriminator(img)

      self.assertEqual(2, disc_output.shape.ndims)
      self.assertEqual(batch_size, disc_output.shape.as_list()[0]) 

def _lr(gen_lr_base, dis_lr_base):
  """Return the generator and discriminator learning rates."""
  gen_lr = tf.train.exponential_decay(
      learning_rate=gen_lr_base,
      global_step=tf.train.get_or_create_global_step(),
      decay_steps=100000,
      decay_rate=0.8,
      staircase=True,)
  dis_lr = dis_lr_base

  return gen_lr, dis_lr 

def define_loss(gan_model, **kwargs):
  """Defines progressive GAN losses.

  The generator and discriminator both use wasserstein loss. In addition,
  a small penalty term is added to the discriminator loss to prevent it getting
  too large.

  Args:
    gan_model: A `GANModel` namedtuple.
    **kwargs: A dictionary of
        'gradient_penalty_weight': A float of gradient norm target for
            wasserstein loss.
        'gradient_penalty_target': A float of gradient penalty weight for
            wasserstein loss.
        'real_score_penalty_weight': A float of Additional penalty to keep
            the scores from drifting too far from zero.

  Returns:
    A `GANLoss` namedtuple.
  """
  gan_loss = tfgan.gan_loss(
      gan_model,
      generator_loss_fn=tfgan.losses.wasserstein_generator_loss,
      discriminator_loss_fn=tfgan.losses.wasserstein_discriminator_loss,
      gradient_penalty_weight=kwargs['gradient_penalty_weight'],
      gradient_penalty_target=kwargs['gradient_penalty_target'],
      gradient_penalty_epsilon=0.0)

  real_score_penalty = tf.reduce_mean(
      tf.square(gan_model.discriminator_real_outputs))
  tf.summary.scalar('real_score_penalty', real_score_penalty)

  return gan_loss._replace(
      discriminator_loss=(
          gan_loss.discriminator_loss +
          kwargs['real_score_penalty_weight'] * real_score_penalty)) 

def test_discriminator_run(self):
    img_batch = tf.zeros([3, 70, 70, 3])
    disc_output = networks.discriminator(img_batch)
    with self.test_session() as sess:
      sess.run(tf.global_variables_initializer())
      sess.run(disc_output) 

def test_discriminator_invalid_input(self):
    wrong_dim_input = tf.zeros([5, 32, 32])
    with self.assertRaisesRegexp(ValueError, 'Shape must be rank 4'):
      networks.discriminator(wrong_dim_input)

    not_fully_defined = tf.placeholder(tf.float32, [3, None, 32, 3])
    with self.assertRaisesRegexp(ValueError, 'Shape .* is not fully defined'):
      networks.compression_model(not_fully_defined) 

def _optimizer(gen_lr, dis_lr):
  # First is generator optimizer, second is discriminator.
  adam_kwargs = {
      'epsilon': 1e-8,
      'beta1': 0.5,
  }
  return (tf.train.AdamOptimizer(gen_lr, **adam_kwargs),
          tf.train.AdamOptimizer(dis_lr, **adam_kwargs)) 

def _lr(gen_lr_base, dis_lr_base):
  """Return the generator and discriminator learning rates."""
  gen_lr_kwargs = {
      'decay_steps': 60000,
      'decay_rate': 0.9,
      'staircase': True,
  }
  gen_lr = tf.train.exponential_decay(
      learning_rate=gen_lr_base,
      global_step=tf.train.get_or_create_global_step(),
      **gen_lr_kwargs)
  dis_lr = dis_lr_base

  return gen_lr, dis_lr 

def _get_gan_model(generator_inputs, generated_data, real_data,
                   generator_scope):
  """Manually construct and return a GANModel tuple."""
  generator_vars = tf.contrib.framework.get_trainable_variables(generator_scope)

  discriminator_fn = networks.discriminator
  with tf.variable_scope('discriminator') as dis_scope:
    discriminator_gen_outputs = discriminator_fn(generated_data)
  with tf.variable_scope(dis_scope, reuse=True):
    discriminator_real_outputs = discriminator_fn(real_data)
  discriminator_vars = tf.contrib.framework.get_trainable_variables(
      dis_scope)

  # Manually construct GANModel tuple.
  gan_model = tfgan.GANModel(
      generator_inputs=generator_inputs,
      generated_data=generated_data,
      generator_variables=generator_vars,
      generator_scope=generator_scope,
      generator_fn=None,  # not necessary
      real_data=real_data,
      discriminator_real_outputs=discriminator_real_outputs,
      discriminator_gen_outputs=discriminator_gen_outputs,
      discriminator_variables=discriminator_vars,
      discriminator_scope=dis_scope,
      discriminator_fn=discriminator_fn)

  return gan_model 

def test_discriminator_invalid_input(self):
    wrong_dim_input = tf.zeros([5, 32, 32])
    with self.assertRaisesRegexp(ValueError, 'Shape must be rank 4'):
      networks.discriminator(wrong_dim_input)

    not_fully_defined = tf.placeholder(tf.float32, [3, None, 32, 3])
    with self.assertRaisesRegexp(ValueError, 'Shape .* is not fully defined'):
      networks.compression_model(not_fully_defined) 

def test_discriminator_run(self):
    img_batch = tf.zeros([3, 70, 70, 3])
    disc_output = networks.discriminator(img_batch)
    with self.test_session() as sess:
      sess.run(tf.global_variables_initializer())
      sess.run(disc_output) 

def test_discriminator_graph(self):
    # Check graph construction for a number of image size/depths and batch
    # sizes.
    for batch_size, patch_size in zip([3, 6], [70, 128]):
      tf.reset_default_graph()
      img = tf.ones([batch_size, patch_size, patch_size, 3])
      disc_output = networks.discriminator(img)

      self.assertEqual(2, disc_output.shape.ndims)
      self.assertEqual(batch_size, disc_output.shape[0])