Python tensorflow.reduce_mean() Examples

def test_adam(self):
        with self.test_session() as sess:
            w = tf.get_variable(
                "w",
                shape=[3],
                initializer=tf.constant_initializer([0.1, -0.2, -0.1]))
            x = tf.constant([0.4, 0.2, -0.5])
            loss = tf.reduce_mean(tf.square(x - w))
            tvars = tf.trainable_variables()
            grads = tf.gradients(loss, tvars)
            global_step = tf.train.get_or_create_global_step()
            optimizer = optimization.AdamWeightDecayOptimizer(learning_rate=0.2)
            train_op = optimizer.apply_gradients(zip(grads, tvars), global_step)
            init_op = tf.group(tf.global_variables_initializer(),
                               tf.local_variables_initializer())
            sess.run(init_op)
            for _ in range(100):
                sess.run(train_op)
            w_np = sess.run(w)
            self.assertAllClose(w_np.flat, [0.4, 0.2, -0.5], rtol=1e-2, atol=1e-2) 

def minibatch_stddev_layer(x, group_size=4):
    with tf.variable_scope('MinibatchStddev'):
        group_size = tf.minimum(group_size, tf.shape(x)[0])     # Minibatch must be divisible by (or smaller than) group_size.
        s = x.shape                                             # [NCHW]  Input shape.
        y = tf.reshape(x, [group_size, -1, s[1], s[2], s[3]])   # [GMCHW] Split minibatch into M groups of size G.
        y = tf.cast(y, tf.float32)                              # [GMCHW] Cast to FP32.
        y -= tf.reduce_mean(y, axis=0, keep_dims=True)           # [GMCHW] Subtract mean over group.
        y = tf.reduce_mean(tf.square(y), axis=0)                # [MCHW]  Calc variance over group.
        y = tf.sqrt(y + 1e-8)                                   # [MCHW]  Calc stddev over group.
        y = tf.reduce_mean(y, axis=[1,2,3], keep_dims=True)      # [M111]  Take average over fmaps and pixels.
        y = tf.cast(y, x.dtype)                                 # [M111]  Cast back to original data type.
        y = tf.tile(y, [group_size, 1, s[2], s[3]])             # [N1HW]  Replicate over group and pixels.
        return tf.concat([x, y], axis=1)                        # [NCHW]  Append as new fmap.

#----------------------------------------------------------------------------
# Generator network used in the paper. 

def maximum_mean_discrepancy(x, y, kernel=utils.gaussian_kernel_matrix):
  r"""Computes the Maximum Mean Discrepancy (MMD) of two samples: x and y.

  Maximum Mean Discrepancy (MMD) is a distance-measure between the samples of
  the distributions of x and y. Here we use the kernel two sample estimate
  using the empirical mean of the two distributions.

  MMD^2(P, Q) = || \E{\phi(x)} - \E{\phi(y)} ||^2
              = \E{ K(x, x) } + \E{ K(y, y) } - 2 \E{ K(x, y) },

  where K = <\phi(x), \phi(y)>,
    is the desired kernel function, in this case a radial basis kernel.

  Args:
      x: a tensor of shape [num_samples, num_features]
      y: a tensor of shape [num_samples, num_features]
      kernel: a function which computes the kernel in MMD. Defaults to the
              GaussianKernelMatrix.

  Returns:
      a scalar denoting the squared maximum mean discrepancy loss.
  """
  with tf.name_scope('MaximumMeanDiscrepancy'):
    # \E{ K(x, x) } + \E{ K(y, y) } - 2 \E{ K(x, y) }
    cost = tf.reduce_mean(kernel(x, x))
    cost += tf.reduce_mean(kernel(y, y))
    cost -= 2 * tf.reduce_mean(kernel(x, y))

    # We do not allow the loss to become negative.
    cost = tf.where(cost > 0, cost, 0, name='value')
  return cost 

def _build_input(self):
        self.tails = tf.placeholder(tf.int32, [None])
        self.heads = tf.placeholder(tf.int32, [None])
        self.targets = tf.one_hot(indices=self.heads, depth=self.num_entity)
            
        if not self.query_is_language:
            self.queries = tf.placeholder(tf.int32, [None, self.num_step])
            self.query_embedding_params = tf.Variable(self._random_uniform_unit(
                                                          self.num_query + 1, # <END> token 
                                                          self.query_embed_size), 
                                                      dtype=tf.float32)
        
            rnn_inputs = tf.nn.embedding_lookup(self.query_embedding_params, 
                                                self.queries)
        else:
            self.queries = tf.placeholder(tf.int32, [None, self.num_step, self.num_word])
            self.vocab_embedding_params = tf.Variable(self._random_uniform_unit(
                                                          self.num_vocab + 1, # <END> token
                                                          self.vocab_embed_size),
                                                      dtype=tf.float32)
            embedded_query = tf.nn.embedding_lookup(self.vocab_embedding_params, 
                                                    self.queries)
            rnn_inputs = tf.reduce_mean(embedded_query, axis=2)

        return rnn_inputs 

def build_pgd_attack(self, eps):
        victim_embeddings = tf.constant(self.victim_embeddings, dtype=tf.float32)

        def one_step_attack(image, grad):
            """
            core components of this attack are:
            (a) PGD adversarial attack (https://arxiv.org/pdf/1706.06083.pdf)
            (b) momentum (https://arxiv.org/pdf/1710.06081.pdf)
            (c) input diversity (https://arxiv.org/pdf/1803.06978.pdf)
            """
            orig_image = image
            image = self.structure(image)
            image = (image - 127.5) / 128.0
            image = image + tf.random_uniform(tf.shape(image), minval=-1e-2, maxval=1e-2)
            prelogits, _ = self.network.inference(image, 1.0, False, bottleneck_layer_size=512)
            embeddings = tf.nn.l2_normalize(prelogits, 1, 1e-10, name='embeddings')

            embeddings = tf.reshape(embeddings[0], [512, 1])
            objective = tf.reduce_mean(tf.matmul(victim_embeddings, embeddings))  # to be maximized

            noise, = tf.gradients(objective, orig_image)

            noise = noise / tf.reduce_mean(tf.abs(noise), [1, 2, 3], keep_dims=True)
            noise = 0.9 * grad + noise

            adv = tf.clip_by_value(orig_image + tf.sign(noise) * 1.0, lower_bound, upper_bound)
            return adv, noise

        input = tf.to_float(self.image_batch)
        lower_bound = tf.clip_by_value(input - eps, 0, 255.)
        upper_bound = tf.clip_by_value(input + eps, 0, 255.)

        with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
            adv, _ = tf.while_loop(
                lambda _, __: True, one_step_attack,
                (input, tf.zeros_like(input)),
                back_prop=False,
                maximum_iterations=100,
                parallel_iterations=1)
        self.adv_image = adv
        return adv 

def _BuildLoss(self):
    # 1. reconstr_loss seems doesn't do better than l2 loss.
    # 2. Only works when using reduce_mean. reduce_sum doesn't work.
    # 3. It seems kl loss doesn't play an important role.
    self.loss = 0
    with tf.variable_scope('loss'):
      if self.params['l2_loss']:
        l2_loss = tf.reduce_mean(tf.square(self.diff_output - self.diffs[1]))
        tf.summary.scalar('l2_loss', l2_loss)
        self.loss += l2_loss
      if self.params['reconstr_loss']:
        reconstr_loss = (-tf.reduce_mean(
            self.diffs[1] * (1e-10 + self.diff_output) +
            (1-self.diffs[1]) * tf.log(1e-10 + 1 - self.diff_output)))
        reconstr_loss = tf.check_numerics(reconstr_loss, 'reconstr_loss')
        tf.summary.scalar('reconstr_loss', reconstr_loss)
        self.loss += reconstr_loss
      if self.params['kl_loss']:
        kl_loss = (0.5 * tf.reduce_mean(
            tf.square(self.z_mean) + tf.square(self.z_stddev) -
            2 * self.z_stddev_log - 1))
        tf.summary.scalar('kl_loss', kl_loss)
        self.loss += kl_loss

      tf.summary.scalar('loss', self.loss) 

def _define_experience(self, observ, action, reward):
    """Implement the branch of experience() entered during training."""
    update_filters = tf.summary.merge([
        self._observ_filter.update(observ),
        self._reward_filter.update(reward)])
    with tf.control_dependencies([update_filters]):
      if self._config.train_on_agent_action:
        # NOTE: Doesn't seem to change much.
        action = self._last_action
      batch = observ, action, self._last_mean, self._last_logstd, reward
      append = self._episodes.append(batch, tf.range(len(self._batch_env)))
    with tf.control_dependencies([append]):
      norm_observ = self._observ_filter.transform(observ)
      norm_reward = tf.reduce_mean(self._reward_filter.transform(reward))
      # pylint: disable=g-long-lambda
      summary = tf.cond(self._should_log, lambda: tf.summary.merge([
          update_filters,
          self._observ_filter.summary(),
          self._reward_filter.summary(),
          tf.summary.scalar('memory_size', self._memory_index),
          tf.summary.histogram('normalized_observ', norm_observ),
          tf.summary.histogram('action', self._last_action),
          tf.summary.scalar('normalized_reward', norm_reward)]), str)
      return summary 

def test_position_sensitive_with_single_bin(self):
    num_spatial_bins = [1, 1]
    image_shape = [2, 3, 3, 4]
    crop_size = [2, 2]

    image = tf.random_uniform(image_shape)
    boxes = tf.random_uniform((6, 4))
    box_ind = tf.constant([0, 0, 0, 1, 1, 1], dtype=tf.int32)

    # When a single bin is used, position-sensitive crop and pool should be
    # the same as non-position sensitive crop and pool.
    crop = tf.image.crop_and_resize(image, boxes, box_ind, crop_size)
    crop_and_pool = tf.reduce_mean(crop, [1, 2], keep_dims=True)

    ps_crop_and_pool = ops.position_sensitive_crop_regions(
        image, boxes, box_ind, crop_size, num_spatial_bins, global_pool=True)

    with self.test_session() as sess:
      expected_output, output = sess.run((crop_and_pool, ps_crop_and_pool))
      self.assertAllClose(output, expected_output) 

def fprop(self, x, y, **kwargs):
        kwargs.update(self.kwargs)
        if self.attack is not None:
            x = x, self.attack(x)
        else:
            x = x,

        # Catching RuntimeError: Variable -= value not supported by tf.eager.
        try:
            y -= self.smoothing * (y - 1. / tf.cast(y.shape[-1], y.dtype))
        except RuntimeError:
            y.assign_sub(self.smoothing * (y - 1. / tf.cast(y.shape[-1],
                                                            y.dtype)))

        logits = [self.model.get_logits(x, **kwargs) for x in x]
        loss = sum(
            tf.reduce_mean(softmax_cross_entropy_with_logits(labels=y,
                                                             logits=logit))
            for logit in logits)
        return loss 

def loss(logits, labels):
  """Add L2Loss to all the trainable variables.

  Add summary for "Loss" and "Loss/avg".
  Args:
    logits: Logits from inference().
    labels: Labels from distorted_inputs or inputs(). 1-D tensor
            of shape [batch_size]

  Returns:
    Loss tensor of type float.
  """
  # Calculate the average cross entropy loss across the batch.
  labels = tf.cast(labels, tf.int64)
  cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
      labels=labels, logits=logits, name='cross_entropy_per_example')
  cross_entropy_mean = tf.reduce_mean(cross_entropy, name='cross_entropy')
  tf.add_to_collection('losses', cross_entropy_mean)

  # The total loss is defined as the cross entropy loss plus all of the weight
  # decay terms (L2 loss).
  return tf.add_n(tf.get_collection('losses'), name='total_loss') 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 256, 256
    num_classes = 1000
    with self.test_session():
      train_inputs = tf.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = vgg.vgg_19(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      tf.get_variable_scope().reuse_variables()
      eval_inputs = tf.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = vgg.vgg_19(eval_inputs, is_training=False,
                             spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = tf.reduce_mean(logits, [1, 2])
      predictions = tf.argmax(logits, 1)
      self.assertEquals(predictions.get_shape().as_list(), [eval_batch_size]) 

def fprop(self, x, y, **kwargs):
        x_adv = self.attack(x)
        d1 = self.model.fprop(x, **kwargs)
        d2 = self.model.fprop(x_adv, **kwargs)
        pairing_loss = [tf.reduce_mean(tf.square(a - b))
                        for a, b in
                        zip(d1[Model.O_FEATURES], d2[Model.O_FEATURES])]
        pairing_loss = tf.reduce_mean(pairing_loss)
        loss = softmax_cross_entropy_with_logits(
            labels=y, logits=d1[Model.O_LOGITS])
        loss += softmax_cross_entropy_with_logits(
            labels=y, logits=d2[Model.O_LOGITS])
        warnings.warn("LossFeaturePairing is deprecated, switch to "
                      "FeaturePairing. LossFeaturePairing may be removed "
                      "on or after 2019-03-06.")
        return loss + self.weight * pairing_loss 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 256, 256
    num_classes = 1000
    with self.test_session():
      train_inputs = tf.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = vgg.vgg_16(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      tf.get_variable_scope().reuse_variables()
      eval_inputs = tf.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = vgg.vgg_16(eval_inputs, is_training=False,
                             spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = tf.reduce_mean(logits, [1, 2])
      predictions = tf.argmax(logits, 1)
      self.assertEquals(predictions.get_shape().as_list(), [eval_batch_size]) 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 256, 256
    num_classes = 1000
    with self.test_session():
      train_inputs = tf.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = vgg.vgg_a(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      tf.get_variable_scope().reuse_variables()
      eval_inputs = tf.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = vgg.vgg_a(eval_inputs, is_training=False,
                            spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = tf.reduce_mean(logits, [1, 2])
      predictions = tf.argmax(logits, 1)
      self.assertEquals(predictions.get_shape().as_list(), [eval_batch_size]) 

def _update_value(self, observ, reward, length):
    """Perform multiple update steps of the value baseline.

    We need to decide for the summary of one iteration, and thus choose the one
    after half of the iterations.

    Args:
      observ: Sequences of observations.
      reward: Sequences of reward.
      length: Batch of sequence lengths.

    Returns:
      Summary tensor.
    """
    with tf.name_scope('update_value'):
      loss, summary = tf.scan(
          lambda _1, _2: self._update_value_step(observ, reward, length),
          tf.range(self._config.update_epochs_value),
          [0., ''], parallel_iterations=1)
      print_loss = tf.Print(0, [tf.reduce_mean(loss)], 'value loss: ')
      with tf.control_dependencies([loss, print_loss]):
        return summary[self._config.update_epochs_value // 2] 

def createLinearModel(dimension):
    np.random.seed(1024)
    # 定义 x 和 y
    x = tf.placeholder(tf.float64, shape=[None, dimension], name='x')
    # 写成矩阵形式会大大加快运算速度
    y = tf.placeholder(tf.float64, shape=[None, 1], name='y')
    # 定义参数估计值和预测值
    betaPred = tf.Variable(np.random.random([dimension, 1]))
    yPred = tf.matmul(x, betaPred, name='y_pred')
    # 定义损失函数
    loss = tf.reduce_mean(tf.square(yPred - y))
    model = {
        'loss_function': loss,
        'independent_variable': x,
        'dependent_variable': y,
        'prediction': yPred,
        'model_params': betaPred
    }
    return model 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 231, 231
    eval_height, eval_width = 281, 281
    num_classes = 1000
    with self.test_session():
      train_inputs = tf.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = overfeat.overfeat(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      tf.get_variable_scope().reuse_variables()
      eval_inputs = tf.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = overfeat.overfeat(eval_inputs, is_training=False,
                                    spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = tf.reduce_mean(logits, [1, 2])
      predictions = tf.argmax(logits, 1)
      self.assertEquals(predictions.get_shape().as_list(), [eval_batch_size]) 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 300, 400
    num_classes = 1000
    with self.test_session():
      train_inputs = tf.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = alexnet.alexnet_v2(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      tf.get_variable_scope().reuse_variables()
      eval_inputs = tf.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = alexnet.alexnet_v2(eval_inputs, is_training=False,
                                     spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 4, 7, num_classes])
      logits = tf.reduce_mean(logits, [1, 2])
      predictions = tf.argmax(logits, 1)
      self.assertEquals(predictions.get_shape().as_list(), [eval_batch_size]) 

def test_position_sensitive_with_equal_channels(self):
    num_spatial_bins = [2, 2]
    image_shape = [1, 3, 3, 4]
    crop_size = [2, 2]

    image = tf.constant(range(1, 3 * 3 + 1), dtype=tf.float32,
                        shape=[1, 3, 3, 1])
    tiled_image = tf.tile(image, [1, 1, 1, image_shape[3]])
    boxes = tf.random_uniform((3, 4))
    box_ind = tf.constant([0, 0, 0], dtype=tf.int32)

    # All channels are equal so position-sensitive crop and resize should
    # work as the usual crop and resize for just one channel.
    crop = tf.image.crop_and_resize(image, boxes, box_ind, crop_size)
    crop_and_pool = tf.reduce_mean(crop, [1, 2], keep_dims=True)

    ps_crop_and_pool = ops.position_sensitive_crop_regions(
        tiled_image,
        boxes,
        box_ind,
        crop_size,
        num_spatial_bins,
        global_pool=True)

    with self.test_session() as sess:
      expected_output, output = sess.run((crop_and_pool, ps_crop_and_pool))
      self.assertAllClose(output, expected_output) 

def __init__(self, n_input, n_hidden, optimizer = tf.train.AdamOptimizer()):
        self.n_input = n_input
        self.n_hidden = n_hidden

        network_weights = self._initialize_weights()
        self.weights = network_weights

        # model
        self.x = tf.placeholder(tf.float32, [None, self.n_input])
        self.z_mean = tf.add(tf.matmul(self.x, self.weights['w1']), self.weights['b1'])
        self.z_log_sigma_sq = tf.add(tf.matmul(self.x, self.weights['log_sigma_w1']), self.weights['log_sigma_b1'])

        # sample from gaussian distribution
        eps = tf.random_normal(tf.stack([tf.shape(self.x)[0], self.n_hidden]), 0, 1, dtype = tf.float32)
        self.z = tf.add(self.z_mean, tf.multiply(tf.sqrt(tf.exp(self.z_log_sigma_sq)), eps))

        self.reconstruction = tf.add(tf.matmul(self.z, self.weights['w2']), self.weights['b2'])

        # cost
        reconstr_loss = 0.5 * tf.reduce_sum(tf.pow(tf.subtract(self.reconstruction, self.x), 2.0))
        latent_loss = -0.5 * tf.reduce_sum(1 + self.z_log_sigma_sq
                                           - tf.square(self.z_mean)
                                           - tf.exp(self.z_log_sigma_sq), 1)
        self.cost = tf.reduce_mean(reconstr_loss + latent_loss)
        self.optimizer = optimizer.minimize(self.cost)

        init = tf.global_variables_initializer()
        self.sess = tf.Session()
        self.sess.run(init) 

def _tower_fn(is_training, weight_decay, feature, label, tower_losses,
              tower_gradvars, tower_preds, is_cpu):
  """Build computation tower for each device (CPU or GPU).

  Args:
    is_training: true if is for training graph.
    weight_decay: weight regularization strength, a float.
    feature: a Tensor.
    label: a Tensor.
    tower_losses: a list to be appended with current tower's loss.
    tower_gradvars: a list to be appended with current tower's gradients.
    tower_preds: a list to be appended with current tower's predictions.
    is_cpu: true if build tower on CPU.
  """
  data_format = 'channels_last' if is_cpu else 'channels_first'
  model = cifar10_model.ResNetCifar10(
      FLAGS.num_layers, is_training=is_training, data_format=data_format)
  logits = model.forward_pass(feature, input_data_format='channels_last')
  tower_pred = {
      'classes': tf.argmax(input=logits, axis=1),
      'probabilities': tf.nn.softmax(logits)
  }
  tower_preds.append(tower_pred)

  tower_loss = tf.losses.sparse_softmax_cross_entropy(
      logits=logits, labels=label)
  tower_loss = tf.reduce_mean(tower_loss)
  tower_losses.append(tower_loss)

  model_params = tf.trainable_variables()
  tower_loss += weight_decay * tf.add_n(
      [tf.nn.l2_loss(v) for v in model_params])
  tower_losses.append(tower_loss)

  tower_grad = tf.gradients(tower_loss, model_params)
  tower_gradvars.append(zip(tower_grad, model_params)) 

def _summarize_vars_and_grads(grads_and_vars):
  tf.logging.info('Trainable variables:')
  tf.logging.info('-' * 60)
  for grad, var in grads_and_vars:
    tf.logging.info(var)

    def tag(name, v=var):
      return v.op.name + '_' + name

    # Variable summary
    mean = tf.reduce_mean(var)
    tf.summary.scalar(tag('mean'), mean)
    with tf.name_scope(tag('stddev')):
      stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean)))
    tf.summary.scalar(tag('stddev'), stddev)
    tf.summary.scalar(tag('max'), tf.reduce_max(var))
    tf.summary.scalar(tag('min'), tf.reduce_min(var))
    tf.summary.histogram(tag('histogram'), var)

    # Gradient summary
    if grad is not None:
      if isinstance(grad, tf.IndexedSlices):
        grad_values = grad.values
      else:
        grad_values = grad

      tf.summary.histogram(tag('gradient'), grad_values)
      tf.summary.scalar(tag('gradient_norm'), tf.global_norm([grad_values]))
    else:
      tf.logging.info('Var %s has no gradient', var.op.name) 

def difference_loss(private_samples, shared_samples, weight=1.0, name=''):
  """Adds the difference loss between the private and shared representations.

  Args:
    private_samples: a tensor of shape [num_samples, num_features].
    shared_samples: a tensor of shape [num_samples, num_features].
    weight: the weight of the incoherence loss.
    name: the name of the tf summary.
  """
  private_samples -= tf.reduce_mean(private_samples, 0)
  shared_samples -= tf.reduce_mean(shared_samples, 0)

  private_samples = tf.nn.l2_normalize(private_samples, 1)
  shared_samples = tf.nn.l2_normalize(shared_samples, 1)

  correlation_matrix = tf.matmul(
      private_samples, shared_samples, transpose_a=True)

  cost = tf.reduce_mean(tf.square(correlation_matrix)) * weight
  cost = tf.where(cost > 0, cost, 0, name='value')

  tf.summary.scalar('losses/Difference Loss {}'.format(name),
                                       cost)
  assert_op = tf.Assert(tf.is_finite(cost), [cost])
  with tf.control_dependencies([assert_op]):
    tf.losses.add_loss(cost)


################################################################################
# TASK LOSS
################################################################################ 

def accuracy(predictions, labels):
  """Calculates the classificaton accuracy.

  Args:
    predictions: the predicted values, a tensor whose size matches 'labels'.
    labels: the ground truth values, a tensor of any size.

  Returns:
    a tensor whose value on evaluation returns the total accuracy.
  """
  return tf.reduce_mean(tf.cast(tf.equal(predictions, labels), tf.float32)) 

def variable_summaries(name,var, with_max_min=False):
  with tf.name_scope(name):
    mean = tf.reduce_mean(var)
    tf.summary.scalar('mean', mean)
    with tf.name_scope('stddev'):
      stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean)))
    tf.summary.scalar('stddev', stddev)
    if with_max_min == True:
        tf.summary.scalar('max', tf.reduce_max(var))
        tf.summary.scalar('min', tf.reduce_min(var)) 

def predict_rewards(self,input_):

        with tf.variable_scope("encoder"):

            Encoder = Attentive_encoder(self.config)
            encoder_output = Encoder.encode(input_)
            frame = tf.reduce_mean(encoder_output, 1) # [Batch size, Sequence Length, Num_neurons] to [Batch size, Num_neurons]

        with tf.variable_scope("ffn"):
            # ffn 1
            h0 = tf.layers.dense(frame, self.num_neurons, activation=tf.nn.relu, kernel_initializer=self.initializer)
            # ffn 2
            w1 =tf.get_variable("w1", [self.num_neurons, 1], initializer=self.initializer)
            b1 = tf.Variable(self.init_baseline, name="b1")
            self.predictions = tf.squeeze(tf.matmul(h0, w1)+b1) 

def build_optim(self):
        # Update moving_mean and moving_variance for batch normalization layers
        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
        with tf.control_dependencies(update_ops):

            with tf.name_scope('reinforce'):
                # Actor learning rate
                self.lr1 = tf.train.exponential_decay(self.lr1_start, self.global_step, self.lr1_decay_step,self.lr1_decay_rate, staircase=False, name="learning_rate1")
                # Optimizer
                self.opt1 = tf.train.AdamOptimizer(learning_rate=self.lr1,beta1=0.9,beta2=0.99, epsilon=0.0000001)
                # Discounted reward
                self.reward_baseline = tf.stop_gradient(self.reward - self.critic.predictions) # [Batch size, 1]
                variable_summaries('reward_baseline',self.reward_baseline, with_max_min = True)
                # Loss
                self.loss1 = tf.reduce_mean(self.reward_baseline*self.log_softmax,0)
                tf.summary.scalar('loss1', self.loss1)
                # Minimize step
                gvs = self.opt1.compute_gradients(self.loss1)
                capped_gvs = [(tf.clip_by_norm(grad, 1.), var) for grad, var in gvs if grad is not None] # L2 clip
                self.train_step1 = self.opt1.apply_gradients(capped_gvs, global_step=self.global_step)

            with tf.name_scope('state_value'):
                # Critic learning rate
                self.lr2 = tf.train.exponential_decay(self.lr2_start, self.global_step2, self.lr2_decay_step,self.lr2_decay_rate, staircase=False, name="learning_rate1")
                # Optimizer
                self.opt2 = tf.train.AdamOptimizer(learning_rate=self.lr2,beta1=0.9,beta2=0.99, epsilon=0.0000001)
                # Loss
                self.loss2 = tf.losses.mean_squared_error(self.reward, self.critic.predictions, weights = 1.0)
                tf.summary.scalar('loss2', self.loss1)
                # Minimize step
                gvs2 = self.opt2.compute_gradients(self.loss2)
                capped_gvs2 = [(tf.clip_by_norm(grad, 1.), var) for grad, var in gvs2 if grad is not None] # L2 clip
                self.train_step2 = self.opt1.apply_gradients(capped_gvs2, global_step=self.global_step2) 

def variable_summaries(name,var, with_max_min=False):
  with tf.name_scope(name):
    mean = tf.reduce_mean(var)
    tf.summary.scalar('mean', mean)
    with tf.name_scope('stddev'):
      stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean)))
    tf.summary.scalar('stddev', stddev)
    if with_max_min == True:
        tf.summary.scalar('max', tf.reduce_max(var))
        tf.summary.scalar('min', tf.reduce_min(var)) 

def gradientDescent(X, Y, model, learningRate=0.01, maxIter=10000, tol=1.e-6):
    # 确定最优算法
    method = tf.train.GradientDescentOptimizer(learning_rate=learningRate)
    optimizer = method.minimize(model['loss_function'])
    # 增加日志
    tf.summary.scalar('loss_function', model['loss_function'])
    tf.summary.histogram('params', model['model_params'])
    tf.summary.scalar('first_param', tf.reduce_mean(model['model_params'][0]))
    tf.summary.scalar('last_param', tf.reduce_mean(model['model_params'][-1]))
    summary = tf.summary.merge_all()
    # 程序运行结束后执行 tensorboard --logdir logs/
    summaryWriter = createSummaryWriter('logs/gradient_descent')

    # TF 开始运行
    sess = tf.Session()
    # 产生初始参数
    init = tf.global_variables_initializer()
    sess.run(init)

    # 迭代梯度下降
    step = 0
    prevLoss = np.inf
    diff = np.inf
    # 当损失函数的变动小于阈值或达到最大循环次数，则停止迭代
    while (step < maxIter) & (diff > tol):
        _, summaryStr, loss = sess.run(
            [optimizer, summary, model['loss_function']],
            feed_dict={
                model['independent_variable']: X,
                model['dependent_variable']: Y
            }
        )
        summaryWriter.add_summary(summaryStr, step)
        # 计算损失函数的变动
        diff = abs(prevLoss - loss)
        prevLoss = loss
        step += 1
    summaryWriter.close() 

def get(self, rewards, pads, values, final_values,
          log_probs, prev_log_probs, target_log_probs,
          entropies, logits):
    seq_length = tf.shape(rewards)[0]

    not_pad = tf.reshape(1 - pads, [seq_length, -1, self.num_samples])
    rewards = not_pad * tf.reshape(rewards, [seq_length, -1, self.num_samples])
    log_probs = not_pad * tf.reshape(sum(log_probs), [seq_length, -1, self.num_samples])

    total_rewards = tf.reduce_sum(rewards, 0)
    total_log_probs = tf.reduce_sum(log_probs, 0)

    rewards_and_bonus = (total_rewards +
                         self.bonus_weight *
                         self.get_bonus(total_rewards, total_log_probs))

    baseline = tf.reduce_mean(rewards_and_bonus, 1, keep_dims=True)

    loss = -tf.stop_gradient(rewards_and_bonus - baseline) * total_log_probs
    loss = tf.reduce_mean(loss)
    raw_loss = loss  # TODO

    gradient_ops = self.training_ops(
        loss, learning_rate=self.learning_rate)

    tf.summary.histogram('log_probs', total_log_probs)
    tf.summary.histogram('rewards', total_rewards)
    tf.summary.scalar('avg_rewards',
                      tf.reduce_mean(total_rewards))
    tf.summary.scalar('loss', loss)

    return loss, raw_loss, baseline, gradient_ops, tf.summary.merge_all()