Python tensorflow.reduce_sum() Examples

def call(self, inputs, **kwargs):

        if K.ndim(inputs) != 3:
            raise ValueError(
                "Unexpected inputs dimensions %d, expect to be 3 dimensions"
                % (K.ndim(inputs)))

        concated_embeds_value = inputs

        square_of_sum = tf.square(tf.reduce_sum(
            concated_embeds_value, axis=1, keep_dims=True))
        sum_of_square = tf.reduce_sum(
            concated_embeds_value * concated_embeds_value, axis=1, keep_dims=True)
        cross_term = square_of_sum - sum_of_square
        cross_term = 0.5 * tf.reduce_sum(cross_term, axis=2, keep_dims=False)

        return cross_term 

def call(self, inputs, **kwargs):
        if K.ndim(inputs[0]) != 3:
            raise ValueError(
                "Unexpected inputs dimensions %d, expect to be 3 dimensions" % (K.ndim(inputs)))

        embed_list = inputs
        row = []
        col = []
        num_inputs = len(embed_list)

        for i in range(num_inputs - 1):
            for j in range(i + 1, num_inputs):
                row.append(i)
                col.append(j)
        p = tf.concat([embed_list[idx]
                       for idx in row], axis=1)  # batch num_pairs k
        q = tf.concat([embed_list[idx] for idx in col], axis=1)
        inner_product = p * q
        if self.reduce_sum:
            inner_product = tf.reduce_sum(
                inner_product, axis=2, keep_dims=True)
        return inner_product 

def _create_autosummary_var(name, value_expr):
    assert not _autosummary_finalized
    v = tf.cast(value_expr, tf.float32)
    if v.shape.ndims is 0:
        v = [v, np.float32(1.0)]
    elif v.shape.ndims is 1:
        v = [tf.reduce_sum(v), tf.cast(tf.shape(v)[0], tf.float32)]
    else:
        v = [tf.reduce_sum(v), tf.reduce_prod(tf.cast(tf.shape(v), tf.float32))]
    v = tf.cond(tf.is_finite(v[0]), lambda: tf.stack(v), lambda: tf.zeros(2))
    with tf.control_dependencies(None):
        var = tf.Variable(tf.zeros(2)) # [numerator, denominator]
    update_op = tf.cond(tf.is_variable_initialized(var), lambda: tf.assign_add(var, v), lambda: tf.assign(var, v))
    if name in _autosummary_vars:
        _autosummary_vars[name].append(var)
    else:
        _autosummary_vars[name] = [var]
    return update_op

#----------------------------------------------------------------------------
# Call filewriter.add_summary() with all summaries in the default graph,
# automatically finalizing and merging them on the first call. 

def set_input_shape(self, input_shape):
        batch_size, rows, cols, input_channels = input_shape
        kernel_shape = tuple(self.kernel_shape) + (input_channels,
                                                   self.output_channels)
        assert len(kernel_shape) == 4
        assert all(isinstance(e, int) for e in kernel_shape), kernel_shape
        init = tf.random_normal(kernel_shape, dtype=tf.float32)
        init = init / tf.sqrt(1e-7 + tf.reduce_sum(tf.square(init),
                                                   axis=(0, 1, 2)))
        self.kernels = tf.Variable(init)
        self.b = tf.Variable(
            np.zeros((self.output_channels,)).astype('float32'))
        input_shape = list(input_shape)
        input_shape[0] = 1
        dummy_batch = tf.zeros(input_shape)
        dummy_output = self.fprop(dummy_batch)
        output_shape = [int(e) for e in dummy_output.get_shape()]
        output_shape[0] = batch_size
        self.output_shape = tuple(output_shape) 

def set_input_shape(self, input_shape):
        batch_size, dim = input_shape
        self.input_shape = [batch_size, dim]
        self.output_shape = [batch_size, self.num_hid]
        if self.init_mode == "norm":
            init = tf.random_normal([dim, self.num_hid], dtype=tf.float32)
            init = init / tf.sqrt(1e-7 + tf.reduce_sum(tf.square(init), axis=0,
                                                       keep_dims=True))
            init = init * self.init_scale
        elif self.init_mode == "uniform_unit_scaling":
            scale = np.sqrt(3. / dim)
            init = tf.random_uniform([dim, self.num_hid], dtype=tf.float32,
                                     minval=-scale, maxval=scale)
        else:
            raise ValueError(self.init_mode)
        self.W = PV(init)
        if self.use_bias:
            self.b = PV((np.zeros((self.num_hid,))
                         + self.init_b).astype('float32')) 

def compute_column_softmax(self, column_controller_vector, time_step):
    #compute softmax over all the columns using column controller vector
    column_controller_vector = tf.tile(
        tf.expand_dims(column_controller_vector, 1),
        [1, self.num_cols + self.num_word_cols, 1])  #max_cols * bs * d
    column_controller_vector = nn_utils.apply_dropout(
        column_controller_vector, self.utility.FLAGS.dropout, self.mode)
    self.full_column_hidden_vectors = tf.concat(
        axis=1, values=[self.column_hidden_vectors, self.word_column_hidden_vectors])
    self.full_column_hidden_vectors += self.summary_text_entry_embeddings
    self.full_column_hidden_vectors = nn_utils.apply_dropout(
        self.full_column_hidden_vectors, self.utility.FLAGS.dropout, self.mode)
    column_logits = tf.reduce_sum(
        column_controller_vector * self.full_column_hidden_vectors, 2) + (
            self.params["word_match_feature_column_name"] *
            self.batch_column_exact_match) + self.full_column_mask
    column_softmax = tf.nn.softmax(column_logits)  #batch_size * max_cols
    return column_softmax 

def get_hash_slots(self, query):
    """Gets hashed-to buckets for batch of queries.

    Args:
      query: 2-d Tensor of query vectors.

    Returns:
      A list of hashed-to buckets for each hash function.
    """

    binary_hash = [
        tf.less(tf.matmul(query, self.hash_vecs[i], transpose_b=True), 0)
        for i in xrange(self.num_libraries)]
    hash_slot_idxs = [
        tf.reduce_sum(
            tf.to_int32(binary_hash[i]) *
            tf.constant([[2 ** i for i in xrange(self.num_hashes)]],
                        dtype=tf.int32), 1)
        for i in xrange(self.num_libraries)]
    return hash_slot_idxs 

def build_cross_entropy_loss(logits, gold):
  """Constructs a cross entropy from logits and one-hot encoded gold labels.

  Supports skipping rows where the gold label is the magic -1 value.

  Args:
    logits: float Tensor of scores.
    gold: int Tensor of one-hot labels.

  Returns:
    cost, correct, total: the total cost, the total number of correctly
        predicted labels, and the total number of valid labels.
  """
  valid = tf.reshape(tf.where(tf.greater(gold, -1)), [-1])
  gold = tf.gather(gold, valid)
  logits = tf.gather(logits, valid)
  correct = tf.reduce_sum(tf.to_int32(tf.nn.in_top_k(logits, gold, 1)))
  total = tf.size(gold)
  cost = tf.reduce_sum(
      tf.contrib.nn.deprecated_flipped_sparse_softmax_cross_entropy_with_logits(
          logits, tf.cast(gold, tf.int64))) / tf.cast(total, tf.float32)
  return cost, correct, total 

def l1_regularizer(weight=1.0, scope=None):
  """Define a L1 regularizer.

  Args:
    weight: scale the loss by this factor.
    scope: Optional scope for name_scope.

  Returns:
    a regularizer function.
  """
  def regularizer(tensor):
    with tf.name_scope(scope, 'L1Regularizer', [tensor]):
      l1_weight = tf.convert_to_tensor(weight,
                                       dtype=tensor.dtype.base_dtype,
                                       name='weight')
      return tf.multiply(l1_weight, tf.reduce_sum(tf.abs(tensor)), name='value')
  return regularizer 

def l1_l2_regularizer(weight_l1=1.0, weight_l2=1.0, scope=None):
  """Define a L1L2 regularizer.

  Args:
    weight_l1: scale the L1 loss by this factor.
    weight_l2: scale the L2 loss by this factor.
    scope: Optional scope for name_scope.

  Returns:
    a regularizer function.
  """
  def regularizer(tensor):
    with tf.name_scope(scope, 'L1L2Regularizer', [tensor]):
      weight_l1_t = tf.convert_to_tensor(weight_l1,
                                         dtype=tensor.dtype.base_dtype,
                                         name='weight_l1')
      weight_l2_t = tf.convert_to_tensor(weight_l2,
                                         dtype=tensor.dtype.base_dtype,
                                         name='weight_l2')
      reg_l1 = tf.multiply(weight_l1_t, tf.reduce_sum(tf.abs(tensor)),
                      name='value_l1')
      reg_l2 = tf.multiply(weight_l2_t, tf.nn.l2_loss(tensor),
                      name='value_l2')
      return tf.add(reg_l1, reg_l2, name='value')
  return regularizer 

def compute_mfcc(audio, **kwargs):
    """
    Compute the MFCC for a given audio waveform. This is
    identical to how DeepSpeech does it, but does it all in
    TensorFlow so that we can differentiate through it.
    """

    batch_size, size = audio.get_shape().as_list()
    audio = tf.cast(audio, tf.float32)

    # 1. Pre-emphasizer, a high-pass filter
    audio = tf.concat((audio[:, :1], audio[:, 1:] - 0.97*audio[:, :-1], np.zeros((batch_size,1000),dtype=np.float32)), 1)

    # 2. windowing into frames of 320 samples, overlapping
    windowed = tf.stack([audio[:, i:i+400] for i in range(0,size-320,160)],1)

    # 3. Take the FFT to convert to frequency space
    ffted = tf.spectral.rfft(windowed, [512])
    ffted = 1.0 / 512 * tf.square(tf.abs(ffted))

    # 4. Compute the Mel windowing of the FFT
    energy = tf.reduce_sum(ffted,axis=2)+1e-30
    filters = np.load("filterbanks.npy").T
    feat = tf.matmul(ffted, np.array([filters]*batch_size,dtype=np.float32))+1e-30

    # 5. Take the DCT again, because why not
    feat = tf.log(feat)
    feat = tf.spectral.dct(feat, type=2, norm='ortho')[:,:,:26]

    # 6. Amplify high frequencies for some reason
    _,nframes,ncoeff = feat.get_shape().as_list()
    n = np.arange(ncoeff)
    lift = 1 + (22/2.)*np.sin(np.pi*n/22)
    feat = lift*feat
    width = feat.get_shape().as_list()[1]

    # 7. And now stick the energy next to the features
    feat = tf.concat((tf.reshape(tf.log(energy),(-1,width,1)), feat[:, :, 1:]), axis=2)
    
    return feat 

def __init__(self, config):

		entity_total = config.entity
		relation_total = config.relation
		batch_size = config.batch_size
		size = config.hidden_size
		margin = config.margin

		self.pos_h = tf.placeholder(tf.int32, [None])
		self.pos_t = tf.placeholder(tf.int32, [None])
		self.pos_r = tf.placeholder(tf.int32, [None])

		self.neg_h = tf.placeholder(tf.int32, [None])
		self.neg_t = tf.placeholder(tf.int32, [None])
		self.neg_r = tf.placeholder(tf.int32, [None])

		with tf.name_scope("embedding"):
			self.ent_embeddings = tf.get_variable(name = "ent_embedding", shape = [entity_total, size], initializer = tf.contrib.layers.xavier_initializer(uniform = False))
			self.rel_embeddings = tf.get_variable(name = "rel_embedding", shape = [relation_total, size], initializer = tf.contrib.layers.xavier_initializer(uniform = False))
			pos_h_e = tf.nn.embedding_lookup(self.ent_embeddings, self.pos_h)
			pos_t_e = tf.nn.embedding_lookup(self.ent_embeddings, self.pos_t)
			pos_r_e = tf.nn.embedding_lookup(self.rel_embeddings, self.pos_r)
			neg_h_e = tf.nn.embedding_lookup(self.ent_embeddings, self.neg_h)
			neg_t_e = tf.nn.embedding_lookup(self.ent_embeddings, self.neg_t)
			neg_r_e = tf.nn.embedding_lookup(self.rel_embeddings, self.neg_r)

		if config.L1_flag:
			pos = tf.reduce_sum(abs(pos_h_e + pos_r_e - pos_t_e), 1, keep_dims = True)
			neg = tf.reduce_sum(abs(neg_h_e + neg_r_e - neg_t_e), 1, keep_dims = True)
			self.predict = pos
		else:
			pos = tf.reduce_sum((pos_h_e + pos_r_e - pos_t_e) ** 2, 1, keep_dims = True)
			neg = tf.reduce_sum((neg_h_e + neg_r_e - neg_t_e) ** 2, 1, keep_dims = True)
			self.predict = pos

		with tf.name_scope("output"):
			self.loss = tf.reduce_sum(tf.maximum(pos - neg + margin, 0)) 

def calc(self, e, t, r):
		return tf.nn.l2_normalize(e + tf.reduce_sum(e * t, 1, keep_dims = True) * r, 1) 

def calc(self, e, n):
		norm = tf.nn.l2_normalize(n, 1)
		return e - tf.reduce_sum(e * norm, 1, keep_dims = True) * norm 

def contrastive_loss(self, y,d,batch_size):
        tmp= y *tf.square(d)
        #tmp= tf.mul(y,tf.square(d))
        tmp2 = (1-y) *tf.square(tf.maximum((1 - d),0))
        return tf.reduce_sum(tmp +tmp2)/batch_size/2 

def __init__(
        self, sequence_length, vocab_size, embedding_size, hidden_units, l2_reg_lambda, batch_size, trainableEmbeddings):

        # Placeholders for input, output and dropout
        self.input_x1 = tf.placeholder(tf.int32, [None, sequence_length], name="input_x1")
        self.input_x2 = tf.placeholder(tf.int32, [None, sequence_length], name="input_x2")
        self.input_y = tf.placeholder(tf.float32, [None], name="input_y")
        self.dropout_keep_prob = tf.placeholder(tf.float32, name="dropout_keep_prob")

        # Keeping track of l2 regularization loss (optional)
        l2_loss = tf.constant(0.0, name="l2_loss")
          
        # Embedding layer
        with tf.name_scope("embedding"):
            self.W = tf.Variable(
                tf.constant(0.0, shape=[vocab_size, embedding_size]),
                trainable=trainableEmbeddings,name="W")
            self.embedded_words1 = tf.nn.embedding_lookup(self.W, self.input_x1)
            self.embedded_words2 = tf.nn.embedding_lookup(self.W, self.input_x2)
        print self.embedded_words1
        # Create a convolution + maxpool layer for each filter size
        with tf.name_scope("output"):
            self.out1=self.stackedRNN(self.embedded_words1, self.dropout_keep_prob, "side1", embedding_size, sequence_length, hidden_units)
            self.out2=self.stackedRNN(self.embedded_words2, self.dropout_keep_prob, "side2", embedding_size, sequence_length, hidden_units)
            self.distance = tf.sqrt(tf.reduce_sum(tf.square(tf.subtract(self.out1,self.out2)),1,keep_dims=True))
            self.distance = tf.div(self.distance, tf.add(tf.sqrt(tf.reduce_sum(tf.square(self.out1),1,keep_dims=True)),tf.sqrt(tf.reduce_sum(tf.square(self.out2),1,keep_dims=True))))
            self.distance = tf.reshape(self.distance, [-1], name="distance")
        with tf.name_scope("loss"):
            self.loss = self.contrastive_loss(self.input_y,self.distance, batch_size)
        #### Accuracy computation is outside of this class.
        with tf.name_scope("accuracy"):
            self.temp_sim = tf.subtract(tf.ones_like(self.distance),tf.rint(self.distance), name="temp_sim") #auto threshold 0.5
            correct_predictions = tf.equal(self.temp_sim, self.input_y)
            self.accuracy=tf.reduce_mean(tf.cast(correct_predictions, "float"), name="accuracy") 

def contrastive_loss(self, y,d,batch_size):
        tmp= y *tf.square(d)
        #tmp= tf.mul(y,tf.square(d))
        tmp2 = (1-y) *tf.square(tf.maximum((1 - d),0))
        return tf.reduce_sum(tmp +tmp2)/batch_size/2 

def __init__(
        self, sequence_length, vocab_size, embedding_size, hidden_units, l2_reg_lambda, batch_size):

        # Placeholders for input, output and dropout
        self.input_x1 = tf.placeholder(tf.int32, [None, sequence_length], name="input_x1")
        self.input_x2 = tf.placeholder(tf.int32, [None, sequence_length], name="input_x2")
        self.input_y = tf.placeholder(tf.float32, [None], name="input_y")
        self.dropout_keep_prob = tf.placeholder(tf.float32, name="dropout_keep_prob")

        # Keeping track of l2 regularization loss (optional)
        l2_loss = tf.constant(0.0, name="l2_loss")
          
        # Embedding layer
        with tf.name_scope("embedding"):
            self.W = tf.Variable(
                tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0),
                trainable=True,name="W")
            self.embedded_chars1 = tf.nn.embedding_lookup(self.W, self.input_x1)
            #self.embedded_chars_expanded1 = tf.expand_dims(self.embedded_chars1, -1)
            self.embedded_chars2 = tf.nn.embedding_lookup(self.W, self.input_x2)
            #self.embedded_chars_expanded2 = tf.expand_dims(self.embedded_chars2, -1)

        # Create a convolution + maxpool layer for each filter size
        with tf.name_scope("output"):
            self.out1=self.BiRNN(self.embedded_chars1, self.dropout_keep_prob, "side1", embedding_size, sequence_length, hidden_units)
            self.out2=self.BiRNN(self.embedded_chars2, self.dropout_keep_prob, "side2", embedding_size, sequence_length, hidden_units)
            self.distance = tf.sqrt(tf.reduce_sum(tf.square(tf.subtract(self.out1,self.out2)),1,keep_dims=True))
            self.distance = tf.div(self.distance, tf.add(tf.sqrt(tf.reduce_sum(tf.square(self.out1),1,keep_dims=True)),tf.sqrt(tf.reduce_sum(tf.square(self.out2),1,keep_dims=True))))
            self.distance = tf.reshape(self.distance, [-1], name="distance")
        with tf.name_scope("loss"):
            self.loss = self.contrastive_loss(self.input_y,self.distance, batch_size)
        #### Accuracy computation is outside of this class.
        with tf.name_scope("accuracy"):
            self.temp_sim = tf.subtract(tf.ones_like(self.distance),tf.rint(self.distance), name="temp_sim") #auto threshold 0.5
            correct_predictions = tf.equal(self.temp_sim, self.input_y)
            self.accuracy=tf.reduce_mean(tf.cast(correct_predictions, "float"), name="accuracy") 

def call(self, seq_value_len_list, mask=None, **kwargs):
        if self.supports_masking:
            if mask is None:
                raise ValueError(
                    "When supports_masking=True,input must support masking")
            uiseq_embed_list = seq_value_len_list
            mask = tf.to_float(mask)
            user_behavior_length = tf.reduce_sum(mask, axis=-1, keep_dims=True)
            mask = tf.expand_dims(mask, axis=2)
        else:
            uiseq_embed_list, user_behavior_length = seq_value_len_list

            mask = tf.sequence_mask(user_behavior_length,
                                    self.seq_len_max, dtype=tf.float32)
            mask = tf.transpose(mask, (0, 2, 1))

        embedding_size = uiseq_embed_list.shape[-1]

        mask = tf.tile(mask, [1, 1, embedding_size])

        uiseq_embed_list *= mask
        hist = uiseq_embed_list
        if self.mode == "max":
            return tf.reduce_max(hist, 1, keep_dims=True)

        hist = tf.reduce_sum(hist, 1, keep_dims=False)

        if self.mode == "mean":
            hist = tf.div(hist, user_behavior_length+self.eps)

        hist = tf.expand_dims(hist, axis=1)
        return hist 

def call(self, inputs, **kwargs):

        if K.ndim(inputs[0]) != 3:
            raise ValueError(
                "Unexpected inputs dimensions %d, expect to be 3 dimensions" % (K.ndim(inputs)))

        embeds_vec_list = inputs
        row = []
        col = []

        for r, c in itertools.combinations(embeds_vec_list, 2):
            row.append(r)
            col.append(c)

        p = tf.concat(row, axis=1)
        q = tf.concat(col, axis=1)
        inner_product = p * q

        bi_interaction = inner_product
        attention_temp = tf.nn.relu(tf.nn.bias_add(tf.tensordot(
            bi_interaction, self.attention_W, axes=(-1, 0)), self.attention_b))
        #  Dense(self.attention_factor,'relu',kernel_regularizer=l2(self.l2_reg_w))(bi_interaction)
        self.normalized_att_score = tf.nn.softmax(tf.tensordot(
            attention_temp, self.projection_h, axes=(-1, 0)), dim=1)
        attention_output = tf.reduce_sum(
            self.normalized_att_score*bi_interaction, axis=1)

        attention_output = tf.nn.dropout(
            attention_output, self.keep_prob, seed=1024)
        # Dropout(1-self.keep_prob)(attention_output)
        afm_out = tf.tensordot(
            attention_output, self.projection_p, axes=(-1, 0))

        return afm_out 

def call(self, inputs, **kwargs):

        if K.ndim(inputs) != 3:
            raise ValueError(
                "Unexpected inputs dimensions %d, expect to be 3 dimensions" % (K.ndim(inputs)))

        concated_embeds_value = inputs
        square_of_sum = tf.square(tf.reduce_sum(
            concated_embeds_value, axis=1, keep_dims=True))
        sum_of_square = tf.reduce_sum(
            concated_embeds_value * concated_embeds_value, axis=1, keep_dims=True)
        cross_term = 0.5*(square_of_sum - sum_of_square)

        return cross_term 

def compute_output_shape(self, input_shape):
        num_inputs = len(input_shape)
        num_pairs = int(num_inputs * (num_inputs - 1) / 2)
        input_shape = input_shape[0]
        embed_size = input_shape[-1]
        if self.reduce_sum:
            return (input_shape[0], num_pairs, 1)
        else:
            return (input_shape[0], num_pairs, embed_size) 

def get_config(self,):
        config = {'reduce_sum': self.reduce_sum, }
        base_config = super(InnerProductLayer, self).get_config()
        return dict(list(base_config.items()) + list(config.items())) 

def __setup_ops(self):
        cross_entropy = -tf.reduce_sum(self.actual_class * tf.log(self.output))
        self.summary = tf.scalar_summary(self.label, cross_entropy)
        self.train_op = tf.train.AdamOptimizer(0.0001).minimize(cross_entropy)
        self.merge_summaries = tf.merge_summary([self.summary])
        correct_prediction = tf.equal(tf.argmax(self.output,1), tf.argmax(self.actual_class,1))
        self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) 

def getCosineSimilarity(q, a):
        q1 = tf.sqrt(tf.reduce_sum(tf.multiply(q, q), 1))
        a1 = tf.sqrt(tf.reduce_sum(tf.multiply(a, a), 1))
        mul = tf.reduce_sum(tf.multiply(q, a), 1)
        cosSim = tf.div(mul, tf.multiply(q1, a1))
        return cosSim 

def getLoss(trueCosSim, falseCosSim, margin):
        zero = tf.fill(tf.shape(trueCosSim), 0.0)
        tfMargin = tf.fill(tf.shape(trueCosSim), margin)
        with tf.name_scope("loss"):
            losses = tf.maximum(zero, tf.subtract(tfMargin, tf.subtract(trueCosSim, falseCosSim)))
            loss = tf.reduce_sum(losses)
        return loss 

def lid_term(logits, batch_size=100):
    """Calculate LID loss term for a minibatch of logits

    :param logits: 
    :return: 
    """
    # y_pred = tf.nn.softmax(logits)
    y_pred = logits

    # calculate pairwise distance
    r = tf.reduce_sum(y_pred * y_pred, 1)
    # turn r into column vector
    r1 = tf.reshape(r, [-1, 1])
    D = r1 - 2 * tf.matmul(y_pred, tf.transpose(y_pred)) + tf.transpose(r1) + \
        tf.ones([batch_size, batch_size])

    # find the k nearest neighbor
    D1 = -tf.sqrt(D)
    D2, _ = tf.nn.top_k(D1, k=21, sorted=True)
    D3 = -D2[:, 1:]

    m = tf.transpose(tf.multiply(tf.transpose(D3), 1.0 / D3[:, -1]))
    v_log = tf.reduce_sum(tf.log(m + 1e-9), axis=1)  # to avoid nan
    lids = -20 / v_log

    ## batch normalize lids
    # lids = tf.nn.l2_normalize(lids, dim=0, epsilon=1e-12)

    return lids 

def lid_adv_term(clean_logits, adv_logits, batch_size=100):
    """Calculate LID loss term for a minibatch of advs logits

    :param logits: clean logits
    :param A_logits: adversarial logits
    :return: 
    """
    # y_pred = tf.nn.softmax(logits)
    c_pred = tf.reshape(clean_logits, (batch_size, -1))
    a_pred = tf.reshape(adv_logits, (batch_size, -1))

    # calculate pairwise distance
    r = tf.reduce_sum(c_pred * a_pred, 1)
    # turn r into column vector
    r1 = tf.reshape(r, [-1, 1])
    D = r1 - 2 * tf.matmul(c_pred, tf.transpose(a_pred)) + tf.transpose(r1) + \
        tf.ones([batch_size, batch_size])

    # find the k nearest neighbor
    D1 = -tf.sqrt(D)
    D2, _ = tf.nn.top_k(D1, k=21, sorted=True)
    D3 = -D2[:, 1:]

    m = tf.transpose(tf.multiply(tf.transpose(D3), 1.0 / D3[:, -1]))
    v_log = tf.reduce_sum(tf.log(m + 1e-9), axis=1)  # to avoid nan
    lids = -20 / v_log

    ## batch normalize lids
    lids = tf.nn.l2_normalize(lids, dim=0, epsilon=1e-12)

    return lids 

def set_input_shape(self, input_shape):
        batch_size, dim = input_shape
        self.input_shape = [batch_size, dim]
        self.output_shape = [batch_size, self.num_hid]
        init = tf.random_normal([dim, self.num_hid], dtype=tf.float32)
        init = init / tf.sqrt(1e-7 + tf.reduce_sum(tf.square(init), axis=0,
                                                   keep_dims=True))
        self.W = tf.Variable(init)
        self.b = tf.Variable(np.zeros((self.num_hid,)).astype('float32')) 

def set_input_shape(self, input_shape):
        batch_size, rows, cols, input_channels = input_shape
        assert len(self.kernel_shape) == 2
        kernel_shape = tuple(self.kernel_shape) + (input_channels,
                                                   self.output_channels)
        assert len(kernel_shape) == 4
        assert all(isinstance(e, int) for e in kernel_shape), kernel_shape
        if self.init_mode == "norm":
            init = tf.random_normal(kernel_shape, dtype=tf.float32)
            squared_norms = tf.reduce_sum(tf.square(init), axis=(0, 1, 2))
            denom = tf.sqrt(1e-7 + squared_norms)
            init = self.init_scale * init / denom
        elif self.init_mode == "inv_sqrt":
            fan_out = self.kernel_shape[0] * \
                self.kernel_shape[1] * self.output_channels
            init = tf.random_normal(kernel_shape, dtype=tf.float32,
                                    stddev=np.sqrt(2.0 / fan_out))
        else:
            raise ValueError(self.init_mode)
        self.kernels = PV(init, name=self.name + "_kernels")
        if self.use_bias:
            self.b = PV(np.zeros((self.output_channels,)).astype('float32'))
        input_shape = list(input_shape)
        orig_batch_size = input_shape[0]
        input_shape[0] = 1
        dummy_batch = tf.zeros(input_shape)
        dummy_output = self.fprop(dummy_batch)
        output_shape = [int(e) for e in dummy_output.get_shape()]
        output_shape[0] = orig_batch_size
        self.output_shape = tuple(output_shape)