Python tensorflow.AUTO_REUSE Examples

def __init__(self, epsilon=1e-4, shape=(), scope=''):
        sess = get_session()

        self._new_mean = tf.placeholder(shape=shape, dtype=tf.float64)
        self._new_var = tf.placeholder(shape=shape, dtype=tf.float64)
        self._new_count = tf.placeholder(shape=(), dtype=tf.float64)

        
        with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
            self._mean  = tf.get_variable('mean',  initializer=np.zeros(shape, 'float64'),      dtype=tf.float64)
            self._var   = tf.get_variable('std',   initializer=np.ones(shape, 'float64'),       dtype=tf.float64)    
            self._count = tf.get_variable('count', initializer=np.full((), epsilon, 'float64'), dtype=tf.float64)

        self.update_ops = tf.group([
            self._var.assign(self._new_var),
            self._mean.assign(self._new_mean),
            self._count.assign(self._new_count)
        ])

        sess.run(tf.variables_initializer([self._mean, self._var, self._count]))
        self.sess = sess
        self._set_mean_var_count() 

def conv2d(input_, output_dim, kernel_h=3, kernel_w=None, stride_h=1, stride_w=None, padding='SAME', reuse=False, initializer=None, use_bias = True, name="conv2d"):
  
  if kernel_w == None: kernel_w = kernel_h
  if stride_w == None: stride_w = stride_h
  if initializer == None: initializer = tf.contrib.layers.xavier_initializer()

  with tf.variable_scope(name, reuse = tf.AUTO_REUSE):
    if reuse==True: scope.reuse_variables()
    w = tf.get_variable('w', [kernel_h, kernel_w, input_.get_shape()[-1], output_dim],
      initializer=initializer)
    conv = tf.nn.conv2d(input_, w, strides=[1,stride_h, stride_w, 1], padding=padding)

    if use_bias:
      b = tf.get_variable('bias', [output_dim], initializer=tf.constant_initializer(0.0))
      conv = tf.nn.bias_add(conv, b)

    return conv 

def __init__(self, epsilon=1e-4, shape=(), scope=''):
        sess = get_session()

        self._new_mean = tf.placeholder(shape=shape, dtype=tf.float64)
        self._new_var = tf.placeholder(shape=shape, dtype=tf.float64)
        self._new_count = tf.placeholder(shape=(), dtype=tf.float64)


        with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
            self._mean  = tf.get_variable('mean',  initializer=np.zeros(shape, 'float64'),      dtype=tf.float64)
            self._var   = tf.get_variable('std',   initializer=np.ones(shape, 'float64'),       dtype=tf.float64)
            self._count = tf.get_variable('count', initializer=np.full((), epsilon, 'float64'), dtype=tf.float64)

        self.update_ops = tf.group([
            self._var.assign(self._new_var),
            self._mean.assign(self._new_mean),
            self._count.assign(self._new_count)
        ])

        sess.run(tf.variables_initializer([self._mean, self._var, self._count]))
        self.sess = sess
        self._set_mean_var_count() 

def res_block(input_, output_dim, name='res_block', is_dropout=False, drop_p=0.5):
  shortcut = input_
  num_input_c = shortcut.shape.as_list()[-1]

  with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
    conv = conv2d(input_, output_dim, name=name+'/conv1')
    conv = norm_layer(conv, 'instance')
    conv = tf.nn.relu(conv)

    if is_dropout:
      conv = tf.nn.dropout(conv, keep_prob = drop_p)

    #conv = conv2d(conv, output_dim, num_input_c * 2, name=name+'/conv2')
    conv = conv2d(conv, output_dim,  name=name+'/conv2')
    conv = norm_layer(conv, 'instance')

    conv = tf.identity(conv+shortcut, name='residual_block_output')

  return conv 

def fprop(self, x, **kwargs):
        del kwargs
        my_conv = functools.partial(tf.layers.conv2d,
                                    kernel_size=3,
                                    strides=2,
                                    padding='valid',
                                    activation=tf.nn.relu,
                                    kernel_initializer=HeReLuNormalInitializer)
        my_dense = functools.partial(
            tf.layers.dense, kernel_initializer=HeReLuNormalInitializer)

        with tf.variable_scope(self.scope, reuse=tf.AUTO_REUSE):
            for depth in [96, 256, 384, 384, 256]:
                x = my_conv(x, depth)
            y = tf.layers.flatten(x)
            y = my_dense(y, 4096, tf.nn.relu)
            y = fc7 = my_dense(y, 4096, tf.nn.relu)
            y = my_dense(y, 1000)
            return {'fc7': fc7,
                    self.O_LOGITS: y,
                    self.O_PROBS: tf.nn.softmax(logits=y)} 

def fc_layer(input_, output_dim, initializer = None, activation='linear', name=None):
  if initializer == None: initializer = tf.contrib.layers.xavier_initializer()
  shape = input_.get_shape().as_list()
  with tf.variable_scope(name or "Linear", reuse=tf.AUTO_REUSE) as scope:
    if len(shape) > 2 : input_ = tf.layers.flatten(input_)
    shape = input_.get_shape().as_list()

    w = tf.get_variable("fc_w", [shape[1], output_dim], dtype=tf.float32, initializer = initializer)
    b = tf.get_variable("fc_b", [output_dim], initializer = tf.constant_initializer(0.0))

    result = tf.matmul(input_, w) + b

    if activation == 'linear':
      return result
    elif activation == 'relu':
      return tf.nn.relu(result)
    elif activation == 'sigmoid':
      return tf.nn.sigmoid(result)
    elif activation == 'tanh':
      return tf.nn.tanh(result) 

def ln(inputs, epsilon = 1e-8, scope="ln"):
    '''Applies layer normalization. See https://arxiv.org/abs/1607.06450.
    inputs: A tensor with 2 or more dimensions, where the first dimension has `batch_size`.
    epsilon: A floating number. A very small number for preventing ZeroDivision Error.
    scope: Optional scope for `variable_scope`.
      
    Returns:
      A tensor with the same shape and data dtype as `inputs`.
    '''
    with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
        inputs_shape = inputs.get_shape()
        params_shape = inputs_shape[-1:]
    
        mean, variance = tf.nn.moments(inputs, [-1], keep_dims=True)
        beta= tf.get_variable("beta", params_shape, initializer=tf.zeros_initializer())
        gamma = tf.get_variable("gamma", params_shape, initializer=tf.ones_initializer())
        normalized = (inputs - mean) / ( (variance + epsilon) ** (.5) )
        outputs = gamma * normalized + beta
        
    return outputs 

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 __call__(self, obs, reuse=False):
        with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
            x = self.network_builder(obs)
            # x = tf.layers.dense(x, self.nb_actions, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            # x = tf.nn.tanh(x)

            x = tf.layers.dense(x, self.hidden_layer1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer2, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            # x = tf.layers.dense(x, self.hidden_layer3, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            # x = tf.nn.tanh(x)
            # x = tf.layers.dense(x, self.hidden_layer4, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            # x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.nb_actions, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            # x = self.step_activation(tf.nn.sigmoid(x))  # sigmoid ~ (0,1), tanh ~ (-1 , 1)
            '''1000 makes the binarization of output of sigmoid has smaller difference 
            that cannot be backpropagated with tensorflow'''
            x = tf.nn.sigmoid(1000*x)  # sigmoid ~ (0,1), tanh ~ (-1 , 1)

        return x 

def __call__(self, obs, action, reuse=False):
        with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
            x = tf.concat([obs, action], axis=-1) # this assumes observation and action can be concatenated
            x = self.network_builder(x)
            # x = tf.layers.dense(x, 1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))

            x = tf.layers.dense(x, self.hidden_layer1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer2, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer3, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer4, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x,  1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))

        return x 

def __call__(self, obs, reuse=False):
        with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
            x = self.network_builder(obs)
            # x = tf.layers.dense(x, self.nb_actions, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            # x = tf.nn.tanh(x)

            x = tf.layers.dense(x, self.hidden_layer1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer2, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer3, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer4, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.nb_actions, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            '''
            in order to discrete the action, apply the trick of sharp sigmoid.
            with the following line: x<0 will be near 0, x>0 will be near 1
            then the discrete step in action choice wont affect too much of accuracy (with only small change of action value!)
            the key point of this step is to keep most part of discrete operation in the tensor graph, so as to be back propagated
            '''

            x = tf.nn.sigmoid(1000*x)  # sigmoid ~ (0,1), tanh ~ (-1 , 1)

        return x 

def __call__(self, obs, reuse=False):
        with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
            x = self.network_builder(obs)
            # x = tf.layers.dense(x, self.nb_actions, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            # x = tf.nn.tanh(x)

            x = tf.layers.dense(x, self.hidden_layer1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer2, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer3, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer4, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.nb_actions, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            '''
            in order to discrete the action, apply the trick of sharp sigmoid.
            with the following line: x<0 will be near 0, x>0 will be near 1
            then the discrete step in action choice wont affect too much of accuracy (with only small change of action value!)
            the key point of this step is to keep most part of discrete operation in the tensor graph, so as to be back propagated
            '''

            x = tf.nn.sigmoid(1000*x)  # sigmoid ~ (0,1), tanh ~ (-1 , 1)

        return x 

def __init__(self, epsilon=1e-4, shape=(), scope=''):
        sess = get_session()

        self._new_mean = tf.placeholder(shape=shape, dtype=tf.float64)
        self._new_var = tf.placeholder(shape=shape, dtype=tf.float64)
        self._new_count = tf.placeholder(shape=(), dtype=tf.float64)


        with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
            self._mean  = tf.get_variable('mean',  initializer=np.zeros(shape, 'float64'),      dtype=tf.float64)
            self._var   = tf.get_variable('std',   initializer=np.ones(shape, 'float64'),       dtype=tf.float64)
            self._count = tf.get_variable('count', initializer=np.full((), epsilon, 'float64'), dtype=tf.float64)

        self.update_ops = tf.group([
            self._var.assign(self._new_var),
            self._mean.assign(self._new_mean),
            self._count.assign(self._new_count)
        ])

        sess.run(tf.variables_initializer([self._mean, self._var, self._count]))
        self.sess = sess
        self._set_mean_var_count() 

def extract_box_classifier_features(self, proposal_feature_maps, scope):
    """Extracts second stage box classifier features.

    Args:
      proposal_feature_maps: A 4-D float tensor with shape
        [batch_size * self.max_num_proposals, crop_height, crop_width, depth]
        representing the feature map cropped to each proposal.
      scope: A scope name.

    Returns:
      proposal_classifier_features: A 4-D float tensor with shape
        [batch_size * self.max_num_proposals, height, width, depth]
        representing box classifier features for each proposal.
    """
    with tf.variable_scope(
        scope, values=[proposal_feature_maps], reuse=tf.AUTO_REUSE):
      return self._extract_box_classifier_features(proposal_feature_maps, scope) 

def reward_prediction(
      self, input_image, input_reward, action, lstm_state, latent):
    """Builds a reward prediction network."""
    conv_size = self.tinyify([32, 32, 16, 4])
    lstm_size = self.tinyify([32, 64, 128, 64, 32])

    with tf.variable_scope("reward_pred", reuse=tf.AUTO_REUSE):
      hidden5, _ = self.bottom_part_tower(
          input_image, input_reward, action, latent,
          lstm_state, lstm_size, conv_size)

      x = hidden5
      x = slim.batch_norm(x, scope="reward_bn0")
      x = slim.conv2d(x, conv_size[1], [3, 3], scope="reward_conv1")
      x = slim.batch_norm(x, scope="reward_bn1")
      x = slim.conv2d(x, conv_size[2], [3, 3], scope="reward_conv2")
      x = slim.batch_norm(x, scope="reward_bn2")
      x = slim.conv2d(x, conv_size[3], [3, 3], scope="reward_conv3")

      pred_reward = self.decode_to_shape(
          x, input_reward.shape, "reward_dec")

      return pred_reward, lstm_state 

def _init_model(self, session):
        """Create Parser model and initialize with random or load parameters in session."""
        config_dict = load_config(self.model_config_path)
        config = get_config(config_dict, self.name)
        model_var_scope = get_model_var_scope(self.var_scope, self.name)
        print ("NOTICE: Initializing model var scope '%s'" % model_var_scope)
        # Check if self.model already exist
        if self.model is None:   # Create Graph Only once
            with tf.variable_scope(model_var_scope, reuse = tf.AUTO_REUSE):
                self.model = parse_model.NNParser(config=config)
        if len(glob.glob(self.ckpt_path + '.data*')) > 0: # file exist with pattern: 'parser.ckpt.data*'
            print("NOTICE: Loading model parameters from %s" % self.ckpt_path)
            all_vars = tf.global_variables()
            model_vars = [k for k in all_vars if model_var_scope in k.name.split("/")]   # Only Restore the Variable in Graph begin with parser/....
            tf.train.Saver(model_vars).restore(session, self.ckpt_path)
        else:
            print("NOTICE: Model not found, Try to run method: deepnlp.download('parse')")
            print("NOTICE: Created with fresh parameters.")
            session.run(tf.global_variables_initializer()) 

def _init_pos_model(self, session):
        """Create POS Tagger model and initialize with random or load parameters in session."""
        # initilize config
        config_dict = load_config(self.model_config_path)
        config = get_config(config_dict, self.name)
        config.batch_size = 1
        config.num_steps = 1 # iterator one token per time
        model_var_scope = get_model_var_scope(self.var_scope, self.name)
        print ("NOTICE: Input POS Model Var Scope Name '%s'" % model_var_scope)
        # Check if self.model already exist
        if self.model is None:
            with tf.variable_scope(model_var_scope, tf.AUTO_REUSE):
                self.model = pos_model.POSTagger(is_training=False, config=config) # save object after is_training
        # Load Specific .data* ckpt file
        if len(glob.glob(self.ckpt_path + '.data*')) > 0: # file exist with pattern: 'pos.ckpt.data*'
            print("NOTICE: Loading model parameters from %s" % self.ckpt_path)
            all_vars = tf.global_variables()
            model_vars = [k for k in all_vars if model_var_scope in k.name.split("/")]
            tf.train.Saver(model_vars).restore(session, self.ckpt_path)
        else:
            print("NOTICE: Model not found, Try to run method: deepnlp.download(module='pos', name='%s')" % self.name)
            print("NOTICE: Created with fresh parameters.")
            session.run(tf.global_variables_initializer()) 

def targets_bottom(self, x, summary_prefix="targets_bottom"):  # pylint: disable=arguments-differ
    inputs = x
    with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
      common_layers.summarize_video(inputs, summary_prefix)
      inputs_shape = common_layers.shape_list(inputs)
      # We embed each of 256=self.top_dimensionality possible pixel values.
      embedding_var = tf.get_variable(
          "pixel_embedding",
          [self.top_dimensionality, self.PIXEL_EMBEDDING_SIZE])
      hot_inputs = tf.one_hot(tf.to_int32(inputs), self.top_dimensionality)
      hot_inputs = tf.reshape(hot_inputs, [-1, self.top_dimensionality])
      embedded = tf.matmul(hot_inputs, embedding_var)
      # Let's now merge all channels that were embedded into a single vector.
      merged_size = self.PIXEL_EMBEDDING_SIZE * inputs_shape[4]
      embedded = tf.reshape(embedded, inputs_shape[:4] + [merged_size])
      transposed = common_layers.time_to_channels(embedded)
      return tf.layers.dense(
          transposed,
          self._body_input_depth,
          name="merge_pixel_embedded_frames") 

def get_vq_bottleneck(bottleneck_size, hidden_size):
  """Get lookup table for VQ bottleneck."""
  with tf.variable_scope("vq", reuse=tf.AUTO_REUSE):
    means = tf.get_variable(
        name="means",
        shape=[bottleneck_size, hidden_size],
        initializer=tf.uniform_unit_scaling_initializer())

    ema_count = tf.get_variable(
        name="ema_count",
        shape=[bottleneck_size],
        initializer=tf.constant_initializer(0),
        trainable=False)

    with tf.colocate_with(means):
      ema_means = tf.get_variable(
          name="ema_means",
          initializer=means.initialized_value(),
          trainable=False)

  return means, ema_means, ema_count 

def __call__(self, obs, action, reuse=False):
        with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
            x = tf.concat([obs, action], axis=-1) # this assumes observation and action can be concatenated
            x = self.network_builder(x)
            # x = tf.layers.dense(x, 1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))

            x = tf.layers.dense(x, self.hidden_layer1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer2, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer3, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x, self.hidden_layer4, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
            x = tf.nn.tanh(x)
            x = tf.layers.dense(x,  1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))

        return x 

def generator(self, inputs_condition):
        inputs = inputs_condition
        with tf.variable_scope("generator", reuse=tf.AUTO_REUSE):
            inputs1 = leaky_relu(conv2d("conv1", inputs, 64, 5, 2))#128x128x128
            inputs2 = leaky_relu(instanceNorm("in1", conv2d("conv2", inputs1, 128, 5, 2)))#64x64x256
            inputs3 = leaky_relu(instanceNorm("in2", conv2d("conv3", inputs2, 256, 5, 2)))#32x32x512
            inputs4 = leaky_relu(instanceNorm("in3", conv2d("conv4", inputs3, 512, 5, 2)))#16x16x512
            inputs5 = leaky_relu(instanceNorm("in4", conv2d("conv5", inputs4, 512, 5, 2)))#8x8x512
            inputs6 = leaky_relu(instanceNorm("in5", conv2d("conv6", inputs5, 512, 5, 2)))#4x4x512
            inputs7 = leaky_relu(instanceNorm("in6", conv2d("conv7", inputs6, 512, 5, 2)))#2x2x512
            inputs8 = leaky_relu(instanceNorm("in7", conv2d("conv8", inputs7, 512, 5, 2)))#1x1x512
            outputs1 = tf.nn.relu(tf.concat([tf.nn.dropout(instanceNorm("in9", deconv2d("dconv1", inputs8, 512, 5, 2)), 0.5), inputs7], axis=3))  # 2x2x512
            outputs2 = tf.nn.relu(tf.concat([tf.nn.dropout(instanceNorm("in10", deconv2d("dconv2", outputs1, 512, 5, 2)), 0.5), inputs6], axis=3))  # 4x4x512
            outputs3 = tf.nn.relu(tf.concat([tf.nn.dropout(instanceNorm("in11", deconv2d("dconv3", outputs2, 512, 5, 2)), 0.5), inputs5], axis=3))#8x8x512
            outputs4 = tf.nn.relu(tf.concat([instanceNorm("in12", deconv2d("dconv4", outputs3, 512, 5, 2)), inputs4], axis=3))#16x16x512
            outputs5 = tf.nn.relu(tf.concat([instanceNorm("in13", deconv2d("dconv5", outputs4, 256, 5, 2)), inputs3], axis=3))#32x32x256
            outputs6 = tf.nn.relu(tf.concat([instanceNorm("in14", deconv2d("dconv6", outputs5, 128, 5, 2)), inputs2], axis=3))#64x64x128
            outputs7 = tf.nn.relu(tf.concat([instanceNorm("in15", deconv2d("dconv7", outputs6, 64, 5, 2)), inputs1], axis=3))#128x128x64
            outputs8 = tf.nn.tanh((deconv2d("dconv8", outputs7, 3, 5, 2)))#256x256x3
            return outputs8 

def build_critic(x, h, output_size, scope, n_layers, size, gru_size, recurrent=True, activation=tf.tanh, output_activation=None, regularizer=None):
    """
    build recurrent critic

    arguments:
        regularizer: regularization for weights
        (see `build_policy()` for rest)

    n.b. the policy and critic should not share weights
    """
    with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
        if recurrent:
            x, h = build_rnn(x, h, gru_size, scope, n_layers, size, activation=activation, output_activation=output_activation, regularizer=regularizer)
        else:
            x = tf.reshape(x, (-1, x.get_shape()[1]*x.get_shape()[2]))
            x = build_mlp(x, gru_size, scope, n_layers + 1, size, activation=activation, output_activation=activation, regularizer=regularizer)
        x = tf.layers.dense(x, output_size, activation=output_activation, name='decoder', kernel_regularizer=regularizer, bias_regularizer=regularizer)
    return x 

def __init__(self,
                 U: List[Tensor],
                 priorU: List[Distribution],
                 likelihood: Likelihood,
                 dtype: tf.DType,
                 stopCriterion,
                 phase: Phase,
                 noiseUniformity: NoiseUniformity,
                 transform: bool = False) -> None:

        # setup the model
        self.dtype = dtype
        self.__transform = transform
        self.__noiseUniformity = noiseUniformity
        self.likelihood = likelihood
        self.stopCriterion = stopCriterion
        self.postU = []  # type: List[PostU]
        for f, priorUf in enumerate(priorU):
            postUf = PostU(likelihood, priorUf, f)
            self.postU.append(postUf)

        # create or reuse the variables for the filter banks
        for f, Uf in enumerate(copy(U)):
            if transform and (f == 0):
                paramName = "{}tr".format(f)
            else:
                paramName = "{}".format(f)
            with tf.variable_scope("U", reuse=tf.AUTO_REUSE):
                UfVar = tf.get_variable(paramName,
                                        dtype=dtype,
                                        initializer=Uf)
            U[f] = UfVar
        self.__U = tuple(U)
        if phase == Phase.EM or phase == Phase.INIT:
            self.__setEm()
        elif phase == Phase.BCD:
            self.__setBcd()
        else:
            raise ValueError 

def encode(self, xs, training=True):
        '''
        Returns
        memory: encoder outputs. (N, T1, d_model)
        '''
        with tf.variable_scope("encoder", reuse=tf.AUTO_REUSE):
            x, seqlens, sents1 = xs

            # src_masks
            src_masks = tf.math.equal(x, 0) # (N, T1)

            # embedding
            enc = tf.nn.embedding_lookup(self.embeddings, x) # (N, T1, d_model)
            enc *= self.hp.d_model**0.5 # scale

            enc += positional_encoding(enc, self.hp.maxlen1)
            enc = tf.layers.dropout(enc, self.hp.dropout_rate, training=training)

            ## Blocks
            for i in range(self.hp.num_blocks):
                with tf.variable_scope("num_blocks_{}".format(i), reuse=tf.AUTO_REUSE):
                    # self-attention
                    enc = multihead_attention(queries=enc,
                                              keys=enc,
                                              values=enc,
                                              key_masks=src_masks,
                                              num_heads=self.hp.num_heads,
                                              dropout_rate=self.hp.dropout_rate,
                                              training=training,
                                              causality=False)
                    # feed forward
                    enc = ff(enc, num_units=[self.hp.d_ff, self.hp.d_model])
        memory = enc
        return memory, sents1, src_masks 

def _init_ner_model(self, session):
        """Create ner Tagger model and initialize or load parameters in session."""
        # initilize config
        config_dict = load_config(self.model_config_path)
        config = get_config(config_dict, self.name)
        if config is None:
            print ("WARNING: Input model name %s has no configuration..." % self.name)
        config.batch_size = 1
        config.num_steps = 1 # iterator one token per time
        model_var_scope = get_model_var_scope(self.var_scope, self.name)
        print ("NOTICE: Input NER Model Var Scope Name '%s'" % model_var_scope)
        # Check if self.model already exist
        if self.model is None:
            with tf.variable_scope(model_var_scope, reuse = tf.AUTO_REUSE):
                self.model = ner_model.NERTagger(is_training=True, config=config) # save object after is_training
        #else:   # Model Graph Def already exist
        #    print ("DEBUG: Model Def already exists")
        # update model parameters
        if len(glob.glob(self.ckpt_path + '.data*')) > 0: # file exist with pattern: 'ner.ckpt.data*'
            print("NOTICE: Loading model parameters from %s" % self.ckpt_path)
            all_vars = tf.global_variables()
            model_vars = [k for k in all_vars if model_var_scope in k.name.split("/")]   # e.g. ner_var_scope_zh
            tf.train.Saver(model_vars).restore(session, self.ckpt_path)
        else:
            print("NOTICE: Model not found, Try to run method: deepnlp.download(module='ner', name='%s')" % self.name)
            print("NOTICE: Created with fresh parameters.")
            session.run(tf.global_variables_initializer()) 

def discriminator(self, inputs, inputs_condition):
        inputs = tf.concat([inputs, inputs_condition], axis=3)
        inputs = tf.random_crop(inputs, [1, 70, 70, 2])
        with tf.variable_scope("discriminator", reuse=tf.AUTO_REUSE):
            with tf.variable_scope("conv1"):
                inputs = leaky_relu(conv2d("conv1", inputs, 64, 5, 2))
            with tf.variable_scope("conv2"):
                inputs = leaky_relu(instanceNorm("in1", conv2d("conv2", inputs, 128, 5, 2)))
            with tf.variable_scope("conv3"):
                inputs = leaky_relu(instanceNorm("in2", conv2d("conv3", inputs, 256, 5, 2)))
            with tf.variable_scope("conv4"):
                inputs = leaky_relu(instanceNorm("in3", conv2d("conv4", inputs, 512, 5, 2)))
            with tf.variable_scope("outputs"):
                inputs = conv2d("conv5", inputs, 1, 5, 1)
            return inputs 

def discriminator(self, inputs, inputs_condition):
        inputs = tf.concat([inputs, inputs_condition], axis=3)
        inputs = tf.random_crop(inputs, [1, 70, 70, 2])
        with tf.variable_scope("discriminator", reuse=tf.AUTO_REUSE):
            with tf.variable_scope("conv1"):
                inputs = leaky_relu(conv2d("conv1", inputs, 64, 5, 2))
            with tf.variable_scope("conv2"):
                inputs = leaky_relu(instanceNorm("in1", conv2d("conv2", inputs, 128, 5, 2)))
            with tf.variable_scope("conv3"):
                inputs = leaky_relu(instanceNorm("in2", conv2d("conv3", inputs, 256, 5, 2)))
            with tf.variable_scope("conv4"):
                inputs = leaky_relu(instanceNorm("in3", conv2d("conv4", inputs, 512, 5, 2)))
            with tf.variable_scope("outputs"):
                inputs = conv2d("conv5", inputs, 1, 5, 1)
            return inputs 

def _build_net(self):
        with tf.variable_scope("placeholder"):
            self.input_x = tf.placeholder(tf.float32, shape=(None, self.vocab_size))
            self.input_y = tf.placeholder(tf.int32, shape=(None,))
        
        with tf.variable_scope("output", reuse=tf.AUTO_REUSE):
            W1 = tf.get_variable("W1", dtype=tf.float32,
                                 initializer=tf.truncated_normal((self.vocab_size, self.hidden_size)))
            b1 = tf.get_variable("b1", dtype=tf.float32,
                                initializer=tf.constant(0.1, shape=(self.hidden_size,)))
            W2 = tf.get_variable("W2", dtype=tf.float32,
                                initializer=tf.truncated_normal((self.hidden_size, self.n_class)))
            b2 = tf.get_variable("b2", dtype=tf.float32,
                                initializer=tf.constant(0.1, shape=(self.n_class,)))
            h = tf.nn.relu(tf.nn.xw_plus_b(self.input_x, W1, b1))
            logits = tf.nn.xw_plus_b(h, W2, b2)
            self.prob = tf.reduce_max(tf.nn.softmax(logits), axis=1)
            self.prediction = tf.cast(tf.argmax(logits, axis=1), tf.int32)
            
        with tf.variable_scope("loss"):
            self.loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=self.input_y))
        
        with tf.variable_scope("train", reuse=tf.AUTO_REUSE):
            optimizer = tf.train.AdamOptimizer(self.lr)
            self.train_op = optimizer.minimize(self.loss)

        with tf.variable_scope("accuracy"):
            correct = tf.equal(self.prediction, self.input_y)
            self.accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
        
        self.sess.run(tf.global_variables_initializer()) 

def positional_encoding(inputs,
                        maxlen,
                        masking=True,
                        scope="positional_encoding"):
    '''Sinusoidal Positional_Encoding. See 3.5
    inputs: 3d tensor. (N, T, E)
    maxlen: scalar. Must be >= T
    masking: Boolean. If True, padding positions are set to zeros.
    scope: Optional scope for `variable_scope`.

    returns
    3d tensor that has the same shape as inputs.
    '''

    E = inputs.get_shape().as_list()[-1] # static
    N, T = tf.shape(inputs)[0], tf.shape(inputs)[1] # dynamic
    with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
        # position indices
        position_ind = tf.tile(tf.expand_dims(tf.range(T), 0), [N, 1]) # (N, T)

        # First part of the PE function: sin and cos argument
        position_enc = np.array([
            [pos / np.power(10000, (i-i%2)/E) for i in range(E)]
            for pos in range(maxlen)])

        # Second part, apply the cosine to even columns and sin to odds.
        position_enc[:, 0::2] = np.sin(position_enc[:, 0::2])  # dim 2i
        position_enc[:, 1::2] = np.cos(position_enc[:, 1::2])  # dim 2i+1
        position_enc = tf.convert_to_tensor(position_enc, tf.float32) # (maxlen, E)

        # lookup
        outputs = tf.nn.embedding_lookup(position_enc, position_ind)

        # masks
        if masking:
            outputs = tf.where(tf.equal(inputs, 0), inputs, outputs)

        return tf.to_float(outputs) 

def __init__(self, name, include_top=False, weights='imagenet'):
        with tf.variable_scope(name, reuse=tf.AUTO_REUSE) as scope:
            if name.upper() == 'VGG19':
                self.vgg = tf.keras.applications.VGG19(include_top=include_top,
                                weights=weights)
            elif name.upper() == 'VGG16':
                self.vgg = tf.keras.applications.VGG16(include_top=include_top,
                                weights=weights)
            else:
                raise TypeError('Not supported model: VGG{}'.format(name))

            self.model = tf.keras.Model(inputs=self.vgg.input,
                                outputs = self.vgg.get_layer('block3_conv3').output)
            self.model.trainable=False
            print(" [*] ", name, " model was created")