python类squeeze()的实例源码

def _add_cross_entropy(labels, logits, pref):
    """Compute average cross entropy and add to loss collection.
    Args:
        labels: Single dimension labels from distorted_inputs() or inputs().
        logits: Output map from inference().
        pref: Either 'c' or 's', for contours or segments, respectively.
    """
    with tf.variable_scope('{}_cross_entropy'.format(pref)) as scope:
        class_prop = C_CLASS_PROP if pref == 'c' else S_CLASS_PROP
        weight_per_label = tf.scalar_mul(class_prop, tf.cast(tf.equal(labels, 0),
                                                             tf.float32)) + \
                           tf.scalar_mul(1.0 - class_prop, tf.cast(tf.equal(labels, 1),
                                                                   tf.float32))
        cross_entropy = tf.losses.sparse_softmax_cross_entropy(
            labels=tf.squeeze(labels, squeeze_dims=[3]), logits=logits)
        cross_entropy_weighted = tf.multiply(weight_per_label, cross_entropy)
        cross_entropy_mean = tf.reduce_mean(cross_entropy_weighted, name=scope.name)
        tf.add_to_collection('losses', cross_entropy_mean)

def _crop_pool_layer(self, bottom, rois, name):
    with tf.variable_scope(name) as scope:
      batch_ids = tf.squeeze(tf.slice(rois, [0, 0], [-1, 1], name="batch_id"), [1])
      # Get the normalized coordinates of bboxes
      bottom_shape = tf.shape(bottom)
      height = (tf.to_float(bottom_shape[1]) - 1.) * np.float32(self._feat_stride[0])
      width = (tf.to_float(bottom_shape[2]) - 1.) * np.float32(self._feat_stride[0])
      x1 = tf.slice(rois, [0, 1], [-1, 1], name="x1") / width
      y1 = tf.slice(rois, [0, 2], [-1, 1], name="y1") / height
      x2 = tf.slice(rois, [0, 3], [-1, 1], name="x2") / width
      y2 = tf.slice(rois, [0, 4], [-1, 1], name="y2") / height
      # Won't be back-propagated to rois anyway, but to save time
      bboxes = tf.stop_gradient(tf.concat([y1, x1, y2, x2], 1))
      if cfg.RESNET.MAX_POOL:
        pre_pool_size = cfg.POOLING_SIZE * 2
        crops = tf.image.crop_and_resize(bottom, bboxes, tf.to_int32(batch_ids), [pre_pool_size, pre_pool_size],
                                         name="crops")
        crops = slim.max_pool2d(crops, [2, 2], padding='SAME')
      else:
        crops = tf.image.crop_and_resize(bottom, bboxes, tf.to_int32(batch_ids), [cfg.POOLING_SIZE, cfg.POOLING_SIZE],
                                         name="crops")
    return crops

  # Do the first few layers manually, because 'SAME' padding can behave inconsistently
  # for images of different sizes: sometimes 0, sometimes 1

def _crop_pool_layer(self, bottom, rois, name):
    with tf.variable_scope(name) as scope:
      batch_ids = tf.squeeze(tf.slice(rois, [0, 0], [-1, 1], name="batch_id"), [1])
      # Get the normalized coordinates of bounding boxes
      bottom_shape = tf.shape(bottom)
      height = (tf.to_float(bottom_shape[1]) - 1.) * np.float32(self._feat_stride[0])
      width = (tf.to_float(bottom_shape[2]) - 1.) * np.float32(self._feat_stride[0])
      x1 = tf.slice(rois, [0, 1], [-1, 1], name="x1") / width
      y1 = tf.slice(rois, [0, 2], [-1, 1], name="y1") / height
      x2 = tf.slice(rois, [0, 3], [-1, 1], name="x2") / width
      y2 = tf.slice(rois, [0, 4], [-1, 1], name="y2") / height
      # Won't be back-propagated to rois anyway, but to save time
      bboxes = tf.stop_gradient(tf.concat([y1, x1, y2, x2], axis=1))
      pre_pool_size = cfg.POOLING_SIZE * 2
      crops = tf.image.crop_and_resize(bottom, bboxes, tf.to_int32(batch_ids), [pre_pool_size, pre_pool_size], name="crops")

    return slim.max_pool2d(crops, [2, 2], padding='SAME')

def _get_loss(self,labels):

        with tf.name_scope("Loss"):
            """
            with tf.name_scope("logloss"):
                logit = tf.squeeze(tf.nn.sigmoid(self.logit))
                self.loss = tf.reduce_mean(self._logloss(labels, logit))
            """
            with tf.name_scope("L2_loss"):
                if self.flags.lambdax:
                    lambdax = self.flags.lambdax
                else:
                    lambdax = 0
                self.l2loss = lambdax*tf.add_n(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))

            with tf.name_scope("dice_coef"):
                #yp_label = tf.cast(logit>self.flags.threshold, tf.float32)
                logit = tf.squeeze(self.logit)
                self.acc = tf.reduce_mean(self._dice_coef(labels,logit))
                self.metric = "dice_coef"
                self.loss = -self.acc

        with tf.name_scope("summary"):
            if self.flags.visualize:
                tf.summary.scalar(name='dice coef', tensor=self.acc, collections=[tf.GraphKeys.SCALARS])

def _build(self):
        V = self.V
        M = self.flags.embedding_size # 64
        H = self.flags.num_units 
        C = self.flags.classes

        netname = "CBOW"
        with tf.variable_scope(netname):
            self.inputs = tf.placeholder(dtype=tf.int32,shape=[None, None]) #[B,S]
            layer_name = "{}/embedding".format(netname)
            x = self._get_embedding(layer_name, self.inputs, V, M, reuse=False) # [B, S, M]

        netname = "RNN"
        cell_name = self.flags.cell
        with tf.variable_scope(netname):
            args = {"num_units":H,"num_proj":C}
            cell_f = self._get_rnn_cell(cell_name=cell_name, args=args)
            cell_b = self._get_rnn_cell(cell_name=cell_name, args=args)
            (out_f, out_b), _ = tf.nn.bidirectional_dynamic_rnn(cell_f,cell_b,x,dtype=tf.float32)
            #logit = (out_f[:,-1,:] + out_b[:,-1,:])*0.5  # [B,1,C]
            logit = tf.reduce_mean(out_f+out_b,axis=1)
            logit = tf.squeeze(logit) # [B,C]
            self.logit = logit

def rnn_story(self):
        """
        run rnn for story to get last hidden state
        input is:  story:                 [batch_size,story_length,embed_size]
        :return:   last hidden state.     [batch_size,embed_size]
        """
        # 1.split input to get lists.
        input_split=tf.split(self.story_embedding,self.story_length,axis=1) #a list.length is:story_length.each element is:[batch_size,1,embed_size]
        input_list=[tf.squeeze(x,axis=1) for x in input_split]           #a list.length is:story_length.each element is:[batch_size,embed_size]
        # 2.init keys(w_all) and values(h_all) of memory
        h_all=tf.get_variable("hidden_states",shape=[self.block_size,self.dimension],initializer=self.initializer)# [block_size,hidden_size]
        w_all=tf.get_variable("keys",          shape=[self.block_size,self.dimension],initializer=self.initializer)# [block_size,hidden_size]
        # 3.expand keys and values to prepare operation of rnn
        w_all_expand=tf.tile(tf.expand_dims(w_all,axis=0),[self.batch_size,1,1]) #[batch_size,block_size,hidden_size]
        h_all_expand=tf.tile(tf.expand_dims(h_all,axis=0),[self.batch_size,1,1]) #[batch_size,block_size,hidden_size]
        # 4. run rnn using input with cell.
        for i,input in enumerate(input_list):
            h_all_expand=self.cell(input,h_all_expand,w_all_expand,i) #w_all:[batch_size,block_size,hidden_size]; h_all:[batch_size,block_size,hidden_size]
        return h_all_expand #[batch_size,block_size,hidden_size]

def position_wise_feed_forward_fn(self):
        """
        x:       [batch,sequence_length,d_model]
        :return: [batch,sequence_length,d_model]
        """
        output=None
        #1.conv1
        input=tf.expand_dims(self.x,axis=3) #[batch,sequence_length,d_model,1]
        # conv2d.input:       [None,sentence_length,embed_size,1]. filter=[filter_size,self.embed_size,1,self.num_filters]
        # output with padding:[None,sentence_length,1,1]
        filter1 = tf.get_variable("filter1"+str(self.layer_index) , shape=[1, self.d_model, 1, 1],initializer=self.initializer)
        ouput_conv1=tf.nn.conv2d(input,filter1,strides=[1,1,1,1],padding="VALID",name="conv1") #[batch,sequence_length,1,1]
        print("output_conv1:",ouput_conv1)

        #2.conv2
        filter2 = tf.get_variable("filter2"+str(self.layer_index), [1, 1, 1, self.d_model], initializer=self.initializer)
        output_conv2=tf.nn.conv2d(ouput_conv1,filter2,strides=[1,1,1,1],padding="VALID",name="conv2") #[batch,sequence_length,1,d_model]
        output=tf.squeeze(output_conv2) #[batch,sequence_length,d_model]
        return output #[batch,sequence_length,d_model]

#test function of position_wise_feed_forward_fn
#time spent:OLD VERSION: length=8000,time spent:35.6s; NEW VERSION:0.03s

def gru_forward(self, embedded_words,gru_cell, reverse=False):
        """
        :param embedded_words:[None,sequence_length, self.embed_size]
        :return:forward hidden state: a list.length is sentence_length, each element is [batch_size,hidden_size]
        """
        # split embedded_words
        embedded_words_splitted = tf.split(embedded_words, self.sequence_length,axis=1)  # it is a list,length is sentence_length, each element is [batch_size,1,embed_size]
        embedded_words_squeeze = [tf.squeeze(x, axis=1) for x in embedded_words_splitted]  # it is a list,length is sentence_length, each element is [batch_size,embed_size]
        h_t = tf.ones((self.batch_size,self.hidden_size))
        h_t_list = []
        if reverse:
            embedded_words_squeeze.reverse()
        for time_step, Xt in enumerate(embedded_words_squeeze):  # Xt: [batch_size,embed_size]
            h_t = gru_cell(Xt,h_t) #h_t:[batch_size,embed_size]<------Xt:[batch_size,embed_size];h_t:[batch_size,embed_size]
            h_t_list.append(h_t)
        if reverse:
            h_t_list.reverse()
        return h_t_list  # a list,length is sentence_length, each element is [batch_size,hidden_size]

def gru_forward_word_level(self, embedded_words):
        """
        :param embedded_words:[batch_size*num_sentences,sentence_length,embed_size]
        :return:forward hidden state: a list.length is sentence_length, each element is [batch_size*num_sentences,hidden_size]
        """
        # split embedded_words
        embedded_words_splitted = tf.split(embedded_words, self.sequence_length,
                                           axis=1)  # it is a list,length is sentence_length, each element is [batch_size*num_sentences,1,embed_size]
        embedded_words_squeeze = [tf.squeeze(x, axis=1) for x in
                                  embedded_words_splitted]  # it is a list,length is sentence_length, each element is [batch_size*num_sentences,embed_size]
        # demension_1=embedded_words_squeeze[0].get_shape().dims[0]
        h_t = tf.ones((self.batch_size * self.num_sentences,
                       self.hidden_size))  #TODO self.hidden_size h_t =int(tf.get_shape(embedded_words_squeeze[0])[0]) # tf.ones([self.batch_size*self.num_sentences, self.hidden_size]) # [batch_size*num_sentences,embed_size]
        h_t_forward_list = []
        for time_step, Xt in enumerate(embedded_words_squeeze):  # Xt: [batch_size*num_sentences,embed_size]
            h_t = self.gru_single_step_word_level(Xt,h_t)  # [batch_size*num_sentences,embed_size]<------Xt:[batch_size*num_sentences,embed_size];h_t:[batch_size*num_sentences,embed_size]
            h_t_forward_list.append(h_t)
        return h_t_forward_list  # a list,length is sentence_length, each element is [batch_size*num_sentences,hidden_size]

    # backward gru for first level: word level

def gru_backward_word_level(self, embedded_words):
        """
        :param   embedded_words:[batch_size*num_sentences,sentence_length,embed_size]
        :return: backward hidden state:a list.length is sentence_length, each element is [batch_size*num_sentences,hidden_size]
        """
        # split embedded_words
        embedded_words_splitted = tf.split(embedded_words, self.sequence_length,
                                           axis=1)  # it is a list,length is sentence_length, each element is [batch_size*num_sentences,1,embed_size]
        embedded_words_squeeze = [tf.squeeze(x, axis=1) for x in
                                  embedded_words_splitted]  # it is a list,length is sentence_length, each element is [batch_size*num_sentences,embed_size]
        embedded_words_squeeze.reverse()  # it is a list,length is sentence_length, each element is [batch_size*num_sentences,embed_size]
        # demension_1=int(tf.get_shape(embedded_words_squeeze[0])[0]) #h_t = tf.ones([self.batch_size*self.num_sentences, self.hidden_size])
        h_t = tf.ones((self.batch_size * self.num_sentences, self.hidden_size))
        h_t_backward_list = []
        for time_step, Xt in enumerate(embedded_words_squeeze):
            h_t = self.gru_single_step_word_level(Xt, h_t)
            h_t_backward_list.append(h_t)
        h_t_backward_list.reverse() #ADD 2017.06.14
        return h_t_backward_list

    # forward gru for second level: sentence level

def gru_backward_sentence_level(self, sentence_representation):
        """
        :param sentence_representation: [batch_size,num_sentences,hidden_size*2]
        :return:forward hidden state: a list,length is num_sentences, each element is [batch_size,hidden_size]
        """
        # split embedded_words
        sentence_representation_splitted = tf.split(sentence_representation, self.num_sentences,
                                                    axis=1)  # it is a list.length is num_sentences,each element is [batch_size,1,hidden_size*2]
        sentence_representation_squeeze = [tf.squeeze(x, axis=1) for x in
                                           sentence_representation_splitted]  # it is a list.length is num_sentences,each element is [batch_size, hidden_size*2]
        sentence_representation_squeeze.reverse()
        # demension_1 = int(tf.get_shape(sentence_representation_squeeze[0])[0])  # scalar: batch_size
        h_t = tf.ones((self.batch_size, self.hidden_size * 2))
        h_t_forward_list = []
        for time_step, Xt in enumerate(sentence_representation_squeeze):  # Xt:[batch_size, hidden_size*2]
            h_t = self.gru_single_step_sentence_level(Xt,h_t)  # h_t:[batch_size,hidden_size]<---------Xt:[batch_size, hidden_size*2]; h_t:[batch_size, hidden_size*2]
            h_t_forward_list.append(h_t)
        h_t_forward_list.reverse() #ADD 2017.06.14
        return h_t_forward_list  # a list,length is num_sentences, each element is [batch_size,hidden_size]

def gru_forward_word_level(self, embedded_words):
        """
        :param embedded_words:[batch_size*num_sentences,sentence_length,embed_size]
        :return:forward hidden state: a list.length is sentence_length, each element is [batch_size*num_sentences,hidden_size]
        """
        # split embedded_words
        embedded_words_splitted = tf.split(embedded_words, self.sequence_length,
                                           axis=1)  # it is a list,length is sentence_length, each element is [batch_size*num_sentences,1,embed_size]
        embedded_words_squeeze = [tf.squeeze(x, axis=1) for x in
                                  embedded_words_splitted]  # it is a list,length is sentence_length, each element is [batch_size*num_sentences,embed_size]
        # demension_1=embedded_words_squeeze[0].get_shape().dims[0]
        h_t = tf.ones((self.batch_size * self.num_sentences,
                       self.hidden_size))  #TODO self.hidden_size h_t =int(tf.get_shape(embedded_words_squeeze[0])[0]) # tf.ones([self.batch_size*self.num_sentences, self.hidden_size]) # [batch_size*num_sentences,embed_size]
        h_t_forward_list = []
        for time_step, Xt in enumerate(embedded_words_squeeze):  # Xt: [batch_size*num_sentences,embed_size]
            h_t = self.gru_single_step_word_level(Xt,h_t)  # [batch_size*num_sentences,embed_size]<------Xt:[batch_size*num_sentences,embed_size];h_t:[batch_size*num_sentences,embed_size]
            h_t_forward_list.append(h_t)
        return h_t_forward_list  # a list,length is sentence_length, each element is [batch_size*num_sentences,hidden_size]

    # backward gru for first level: word level

def gru_backward_word_level(self, embedded_words):
        """
        :param   embedded_words:[batch_size*num_sentences,sentence_length,embed_size]
        :return: backward hidden state:a list.length is sentence_length, each element is [batch_size*num_sentences,hidden_size]
        """
        # split embedded_words
        embedded_words_splitted = tf.split(embedded_words, self.sequence_length,
                                           axis=1)  # it is a list,length is sentence_length, each element is [batch_size*num_sentences,1,embed_size]
        embedded_words_squeeze = [tf.squeeze(x, axis=1) for x in
                                  embedded_words_splitted]  # it is a list,length is sentence_length, each element is [batch_size*num_sentences,embed_size]
        embedded_words_squeeze.reverse()  # it is a list,length is sentence_length, each element is [batch_size*num_sentences,embed_size]
        # demension_1=int(tf.get_shape(embedded_words_squeeze[0])[0]) #h_t = tf.ones([self.batch_size*self.num_sentences, self.hidden_size])
        h_t = tf.ones((self.batch_size * self.num_sentences, self.hidden_size))
        h_t_backward_list = []
        for time_step, Xt in enumerate(embedded_words_squeeze):
            h_t = self.gru_single_step_word_level(Xt, h_t)
            h_t_backward_list.append(h_t)
        h_t_backward_list.reverse() #ADD 2017.06.14
        return h_t_backward_list

    # forward gru for second level: sentence level

def gru_forward_sentence_level(self, sentence_representation):
        """
        :param sentence_representation: [batch_size,num_sentences,hidden_size*2]
        :return:forward hidden state: a list,length is num_sentences, each element is [batch_size,hidden_size]
        """
        # split embedded_words
        sentence_representation_splitted = tf.split(sentence_representation, self.num_sentences,
                                                    axis=1)  # it is a list.length is num_sentences,each element is [batch_size,1,hidden_size*2]
        sentence_representation_squeeze = [tf.squeeze(x, axis=1) for x in
                                           sentence_representation_splitted]  # it is a list.length is num_sentences,each element is [batch_size, hidden_size*2]
        # demension_1 = int(tf.get_shape(sentence_representation_squeeze[0])[0]) #scalar: batch_size
        h_t = tf.ones((self.batch_size, self.hidden_size * 2))  # TODO
        h_t_forward_list = []
        for time_step, Xt in enumerate(sentence_representation_squeeze):  # Xt:[batch_size, hidden_size*2]
            h_t = self.gru_single_step_sentence_level(Xt,
                                                      h_t)  # h_t:[batch_size,hidden_size]<---------Xt:[batch_size, hidden_size*2]; h_t:[batch_size, hidden_size*2]
            h_t_forward_list.append(h_t)
        return h_t_forward_list  # a list,length is num_sentences, each element is [batch_size,hidden_size]

    # backward gru for second level: sentence level

def gru_backward_sentence_level(self, sentence_representation):
        """
        :param sentence_representation: [batch_size,num_sentences,hidden_size*2]
        :return:forward hidden state: a list,length is num_sentences, each element is [batch_size,hidden_size]
        """
        # split embedded_words
        sentence_representation_splitted = tf.split(sentence_representation, self.num_sentences,
                                                    axis=1)  # it is a list.length is num_sentences,each element is [batch_size,1,hidden_size*2]
        sentence_representation_squeeze = [tf.squeeze(x, axis=1) for x in
                                           sentence_representation_splitted]  # it is a list.length is num_sentences,each element is [batch_size, hidden_size*2]
        sentence_representation_squeeze.reverse()
        # demension_1 = int(tf.get_shape(sentence_representation_squeeze[0])[0])  # scalar: batch_size
        h_t = tf.ones((self.batch_size, self.hidden_size * 2))
        h_t_forward_list = []
        for time_step, Xt in enumerate(sentence_representation_squeeze):  # Xt:[batch_size, hidden_size*2]
            h_t = self.gru_single_step_sentence_level(Xt,h_t)  # h_t:[batch_size,hidden_size]<---------Xt:[batch_size, hidden_size*2]; h_t:[batch_size, hidden_size*2]
            h_t_forward_list.append(h_t)
        h_t_forward_list.reverse() #ADD 2017.06.14
        return h_t_forward_list  # a list,length is num_sentences, each element is [batch_size,hidden_size]

def SoftArgmin(outputLeft, outputRight, D=192):

    left_result_D = outputLeft
    right_result_D = outputRight
    left_result_D_squeeze = tf.squeeze(left_result_D, axis=[0, 4])
    right_result_D_squeeze = tf.squeeze(right_result_D, axis=[0, 4])  # 192 256 512
    left_result_softmax = tf.nn.softmax(left_result_D_squeeze, dim=0)
    right_result_softmax = tf.nn.softmax(right_result_D_squeeze, dim=0)  # 192 256 512

    d_grid = tf.cast(tf.range(D), tf.float32)
    d_grid = tf.reshape(d_grid, (-1, 1, 1))
    d_grid = tf.tile(d_grid, [1, 256, 512])

    left_softargmin = tf.reduce_sum(tf.multiply(left_result_softmax, d_grid), axis=0, keep_dims=True)
    right_softargmin = tf.reduce_sum(tf.multiply(right_result_softmax, d_grid), axis=0, keep_dims=True)

    return left_softargmin, right_softargmin

def prepare_label(self, input_batch, new_size):
        """Resize masks and perform one-hot encoding.

        Args:
          input_batch: input tensor of shape [batch_size H W 1].
          new_size: a tensor with new height and width.

        Returns:
          Outputs a tensor of shape [batch_size h w 21]
          with last dimension comprised of 0's and 1's only.
        """
        with tf.name_scope('label_encode'):
            input_batch = tf.image.resize_nearest_neighbor(input_batch, new_size) # As labels are integer numbers, need to use NN interp.
            input_batch = tf.squeeze(input_batch, axis=[3]) # Reducing the channel dimension.
            input_batch = tf.one_hot(input_batch, depth=21)
        return input_batch

def _build(self):
        assert(self.d is not None)
        assert(self.lr is not None)
        assert(self.l2_penalty is not None)
        assert(self.loss_function is not None)
        # Get input placeholders and sentence features
        self._create_placeholders()
        sentence_feats, save_kwargs = self._embed_sentences()
        # Define linear model
        s1, s2 = self.seed, (self.seed + 1 if self.seed is not None else None)
        w = tf.Variable(tf.random_normal((self.d, 1), stddev=SD, seed=s1))
        b = tf.Variable(tf.random_normal((1, 1), stddev=SD, seed=s2))
        h = tf.squeeze(tf.matmul(sentence_feats, w) + b)
        # Define training procedure
        self.loss       = self._get_loss(h, self.y)
        self.loss      += self.l2_penalty * tf.nn.l2_loss(w)
        self.prediction = tf.sigmoid(h)
        self.train_fn   = tf.train.AdamOptimizer(self.lr).minimize(self.loss)
        self.save_dict  = save_kwargs.update({'w': w, 'b': b})

def call(self,inputs):
        """
        inputs in as array which contains the support set the embeddings, the target embedding as the second last value in the array, and true class of target embedding as the last value in the array
        """ 
        similarities = []

        targetembedding = inputs[-2]
        numsupportset = len(inputs)-2
        for ii in range(numsupportset):
            supportembedding = inputs[ii]
            dd = tf.negative(tf.sqrt(tf.reduce_sum(tf.square(supportembedding-targetembedding),1,keep_dims=True)))

            similarities.append(dd)

        similarities = tf.concat(axis=1,values=similarities)
        softmax_similarities = tf.nn.softmax(similarities)
        preds = tf.squeeze(tf.matmul(tf.expand_dims(softmax_similarities,1),inputs[-1]))

        preds.set_shape((inputs[0].shape[0],self.nway))

        return preds

def _read_input(self, filename_queue):
        class DataRecord(object):
            pass

        reader = tf.WholeFileReader()
        key, value = reader.read(filename_queue)
        record = DataRecord()
        decoded_image = tf.image.decode_jpeg(value,
                                             channels=3)  # Assumption:Color images are read and are to be generated

        # decoded_image_4d = tf.expand_dims(decoded_image, 0)
        # resized_image = tf.image.resize_bilinear(decoded_image_4d, [self.target_image_size, self.target_image_size])
        # record.input_image = tf.squeeze(resized_image, squeeze_dims=[0])

        cropped_image = tf.cast(
            tf.image.crop_to_bounding_box(decoded_image, 55, 35, self.crop_image_size, self.crop_image_size),
            tf.float32)
        decoded_image_4d = tf.expand_dims(cropped_image, 0)
        resized_image = tf.image.resize_bilinear(decoded_image_4d, [self.resized_image_size, self.resized_image_size])
        record.input_image = tf.squeeze(resized_image, squeeze_dims=[0])
        return record

def run_bottleneck_on_image(sess, image_data, image_data_tensor,
                            bottleneck_tensor):
  """Runs inference on an image to extract the 'bottleneck' summary layer.

  Args:
    sess: Current active TensorFlow Session.
    image_data: String of raw JPEG data.
    image_data_tensor: Input data layer in the graph.
    bottleneck_tensor: Layer before the final softmax.

  Returns:
    Numpy array of bottleneck values.
  """
  bottleneck_values = sess.run(
      bottleneck_tensor,
      {image_data_tensor: image_data})
  bottleneck_values = np.squeeze(bottleneck_values)
  return bottleneck_values

def DCGRU(self, mem, kernel_width, prefix):
        """Convolutional diagonal GRU."""

        def conv_lin(input, suffix, bias_start):
            return self.conv_linear(input, kernel_width, self.num_units, self.num_units, bias_start,prefix + "/" + suffix)

        # perform shift
        mem_shifted = tf.squeeze(tf.nn.depthwise_conv2d(tf.expand_dims(mem,1), self.shift_filter,[1,1,1,1],'SAME'),[1])

        # calculate the new value
        reset = self.hard_sigmoid(conv_lin(mem, "r", 0.5))
        candidate = self.hard_tanh(conv_lin(reset * mem, "c", 0.0))
        gate = self.hard_sigmoid(conv_lin(mem, "g", 0.7))
        candidate =self.dropout(candidate)

        candidate =  gate*mem_shifted + (1 - gate)*candidate
        return candidate

def __call__(self, u_t, a, b, scope=None):
        """

        :param u_t: [N, M, d]
        :param a: [N, M. d]
        :param b: [N, M. d]
        :param mask:  [N, M]
        :return:
        """
        N, M, d = self.batch_size, self.mem_size, self.hidden_size
        L, sL = self.L, self.sL
        with tf.name_scope(scope or self.__class__.__name__):
            L = tf.tile(tf.expand_dims(tf.expand_dims(L, 0), 0), [N, d, 1, 1])
            sL = tf.tile(tf.expand_dims(tf.expand_dims(sL, 0), 0), [N, d, 1, 1])
            logb = tf.log(b + 1e-9)  # [N, M, d]
            logb = tf.concat(1, [tf.zeros([N, 1, d]), tf.slice(logb, [0, 1, 0], [-1, -1, -1])])  # [N, M, d]
            logb = tf.expand_dims(tf.transpose(logb, [0, 2, 1]), -1)  # [N, d, M, 1]
            left = L * tf.exp(tf.batch_matmul(L, logb * sL))  # [N, d, M, M]
            right = a * u_t  # [N, M, d]
            right = tf.expand_dims(tf.transpose(right, [0, 2, 1]), -1)  # [N, d, M, 1]
            u = tf.batch_matmul(left, right)  # [N, d, M, 1]
            u = tf.transpose(tf.squeeze(u, [3]), [0, 2, 1])  # [N, M, d]
        return u

def __call__(self, u_t, a, b, scope=None):
        """

        :param u_t: [N, M, d]
        :param a: [N, M. d]
        :param b: [N, M. d]
        :param mask:  [N, M]
        :return:
        """
        N, M, d = self.batch_size, self.mem_size, self.hidden_size
        L, sL = self.L, self.sL
        with tf.name_scope(scope or self.__class__.__name__):
            L = tf.tile(tf.expand_dims(tf.expand_dims(L, 0), 0), [N, d, 1, 1])
            sL = tf.tile(tf.expand_dims(tf.expand_dims(sL, 0), 0), [N, d, 1, 1])
            logb = tf.log(b + 1e-9)  # [N, M, d]
            logb = tf.concat(1, [tf.zeros([N, 1, d]), tf.slice(logb, [0, 1, 0], [-1, -1, -1])])  # [N, M, d]
            logb = tf.expand_dims(tf.transpose(logb, [0, 2, 1]), -1)  # [N, d, M, 1]
            left = L * tf.exp(tf.batch_matmul(L, logb * sL))  # [N, d, M, M]
            right = a * u_t  # [N, M, d]
            right = tf.expand_dims(tf.transpose(right, [0, 2, 1]), -1)  # [N, d, M, 1]
            u = tf.batch_matmul(left, right)  # [N, d, M, 1]
            u = tf.transpose(tf.squeeze(u, [3]), [0, 2, 1])  # [N, M, d]
        return u

def reinforce_baseline(decoder_states, reward):
    """
    Center the reward by computing a baseline reward over decoder states.

    :param decoder_states: internal states of the decoder, tensor of shape (batch_size, time_steps, state_size)
    :param reward: reward for each time step, tensor of shape (batch_size, time_steps)
    :return: reward - computed baseline, tensor of shape (batch_size, time_steps)
    """
    # batch_size = tf.shape(decoder_states)[0]
    # time_steps = tf.shape(decoder_states)[1]
    # state_size = decoder_states.get_shape()[2]
    # states = tf.reshape(decoder_states, shape=tf.stack([batch_size * time_steps, state_size]))

    baseline = dense(tf.stop_gradient(decoder_states), units=1, activation=None, name='reward_baseline',
                     kernel_initializer=tf.constant_initializer(0.01))
    baseline = tf.squeeze(baseline, axis=2)

    # baseline = tf.reshape(baseline, shape=tf.stack([batch_size, time_steps]))
    return reward - baseline

def data_augmentation(img, gt_bboxes, gt_cats, seg, config):
    params = config['train_augmentation']
    img = apply_with_random_selector(
        img,
        lambda x, ordering: photometric_distortions(x, ordering, params),
        num_cases=4)

    if seg is not None:
        img = tf.concat([img, tf.cast(seg, tf.float32)], axis=-1)

    img, gt_bboxes, gt_cats = scale_distortions(img, gt_bboxes, gt_cats,
                                                params)
    img, gt_bboxes = mirror_distortions(img, gt_bboxes, params)
    # XXX reference implementation also randomizes interpolation method
    img_size = config['image_size']
    img_out = tf.image.resize_images(img[..., :3], [img_size, img_size])
    gt_bboxes, gt_cats = filter_small_gt(gt_bboxes, gt_cats, 2/config['image_size'])

    if seg is not None:
        seg_shape = config['fm_sizes'][0]
        seg = tf.expand_dims(tf.expand_dims(img[..., 3], 0), -1)
        seg = tf.squeeze(tf.image.resize_nearest_neighbor(seg, [seg_shape, seg_shape]))
        seg = tf.cast(tf.round(seg), tf.int64)
    return img_out, gt_bboxes, gt_cats, seg

def save_np_image(image, output_file, save_format='jpeg'):
  """Saves an image to disk.

  Args:
    image: 3-D numpy array of shape [image_size, image_size, 3] and dtype
        float32, with values in [0, 1].
    output_file: str, output file.
    save_format: format for saving image (eg. jpeg).
  """
  image = np.uint8(image * 255.0)
  buf = io.BytesIO()
  scipy.misc.imsave(buf, np.squeeze(image, 0), format=save_format)
  buf.seek(0)
  f = tf.gfile.GFile(output_file, 'w')
  f.write(buf.getvalue())
  f.close()

def run_bottleneck_on_image(sess, image_data, image_data_tensor,
                            bottleneck_tensor):
  """Runs inference on an image to extract the 'bottleneck' summary layer.

  Args:
    sess: Current active TensorFlow Session.
    image_data: String of raw JPEG data.
    image_data_tensor: Input data layer in the graph.
    bottleneck_tensor: Layer before the final softmax.

  Returns:
    Numpy array of bottleneck values.
  """
  bottleneck_values = sess.run(
      bottleneck_tensor,
      {image_data_tensor: image_data})
  bottleneck_values = np.squeeze(bottleneck_values)
  return bottleneck_values

def _sample(self, n_samples):
        mean, cov_tril = self.mean, self.cov_tril
        if not self.is_reparameterized:
            mean = tf.stop_gradient(mean)
            cov_tril = tf.stop_gradient(cov_tril)

        def tile(t):
            new_shape = tf.concat([[n_samples], tf.ones_like(tf.shape(t))], 0)
            return tf.tile(tf.expand_dims(t, 0), new_shape)

        batch_mean = tile(mean)
        batch_cov = tile(cov_tril)
        # n_dim -> n_dim x 1 for matmul
        batch_mean = tf.expand_dims(batch_mean, -1)
        noise = tf.random_normal(tf.shape(batch_mean), dtype=self.dtype)
        samples = tf.matmul(batch_cov, noise) + batch_mean
        samples = tf.squeeze(samples, -1)
        # Update static shape
        static_n_samples = n_samples if isinstance(n_samples, int) else None
        samples.set_shape(tf.TensorShape([static_n_samples])
                          .concatenate(self.get_batch_shape())
                          .concatenate(self.get_value_shape()))
        return samples

def _log_prob(self, given):
        mean, cov_tril = (self.path_param(self.mean),
                          self.path_param(self.cov_tril))
        log_det = 2 * tf.reduce_sum(
            tf.log(tf.matrix_diag_part(cov_tril)), axis=-1)
        N = tf.cast(self._n_dim, self.dtype)
        logZ = - N / 2 * tf.log(2 * tf.constant(np.pi, dtype=self.dtype)) - \
            log_det / 2
        # logZ.shape == batch_shape
        if self._check_numerics:
            logZ = tf.check_numerics(logZ, "log[det(Cov)]")
        # (given-mean)' Sigma^{-1} (given-mean) =
        # (g-m)' L^{-T} L^{-1} (g-m) = |x|^2, where Lx = g-m =: y.
        y = tf.expand_dims(given - mean, -1)
        L, _ = maybe_explicit_broadcast(
            cov_tril, y, 'MultivariateNormalCholesky.cov_tril',
            'expand_dims(given, -1)')
        x = tf.matrix_triangular_solve(L, y, lower=True)
        x = tf.squeeze(x, -1)
        stoc_dist = -0.5 * tf.reduce_sum(tf.square(x), axis=-1)
        return logZ + stoc_dist