Python tensorflow.stack() Examples

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 softmax_unet(input_tensor, instruments, params={}):
    """ Apply softmax to multitrack unet in order to have mask suming to one.

    :param input_tensor: Tensor to apply blstm to.
    :param instruments: Iterable that provides a collection of instruments.
    :param params: (Optional) dict of BLSTM parameters.
    :returns: Created output tensor dict.
    """
    logit_mask_list = []
    for instrument in instruments:
        out_name = f'{instrument}_spectrogram'
        logit_mask_list.append(
            apply_unet(
                input_tensor,
                output_name=out_name,
                params=params,
                output_mask_logit=True))
    masks = Softmax(axis=4)(tf.stack(logit_mask_list, axis=4))
    output_dict = {}
    for i, instrument in enumerate(instruments):
        out_name = f'{instrument}_spectrogram'
        output_dict[out_name] = Multiply(name=out_name)([
            masks[..., i],
            input_tensor])
    return output_dict 

def loop_decode(self):
        # decoder_initial_state: Tuple Tensor (c,h) of size [batch_size x cell.state_size]
        # decoder_first_input: Tensor [batch_size x cell.state_size]

        # Loop the decoding process and collect results
        s,i = self.decoder_initial_state,  tf.cast(self.decoder_first_input,tf.float32)
        for step in range(self.seq_length):
            s, i = self.decode(s,i,step)

        # Return to start
        self.positions.append(self.first_city)

        # Stack visited indices
        self.positions=tf.stack(self.positions,axis=1)  # [Batch,seq_length+1]

        # Sum log_softmax over output steps
        self.log_softmax=tf.add_n(self.log_softmax)  # [Batch,seq_length]

        # Stack attending & pointing distribution
        self.attending=tf.stack(self.attending,axis=1) # [Batch,seq_length,seq_length]
        self.pointing=tf.stack(self.pointing,axis=1) # [Batch,seq_length,seq_length]
        
        # Return stacked lists of visited_indices and log_softmax for backprop
        return self.positions,self.log_softmax 

def images_to_sequence(tensor):
  """Convert a batch of images into a batch of sequences.

  Args:
    tensor: a (num_images, height, width, depth) tensor

  Returns:
    (width, num_images*height, depth) sequence tensor
  """
  transposed = tf.transpose(tensor, [2, 0, 1, 3])

  shapeT = tf.shape(transposed)
  shapeL = transposed.get_shape().as_list()
  # Calculate the ouput size of the upsampled tensor
  n_shape = tf.stack([
      shapeT[0],
      shapeT[1]*shapeT[2],
      shapeL[3]
  ])
  reshaped = tf.reshape(transposed, n_shape)
  return reshaped 

def sequence_to_images(tensor, num_batches):
  """Convert a batch of sequences into a batch of images.

  Args:
    tensor: (num_steps, num_batchesRNN, depth) sequence tensor
    num_batches: the number of image batches

  Returns:
    (num_batches, height, width, depth) tensor
  """

  shapeT = tf.shape(tensor)
  shapeL = tensor.get_shape().as_list()
  # Calculate the ouput size of the upsampled tensor
  height = tf.to_int32(shapeT[1] / num_batches)
  n_shape = tf.stack([
      shapeT[0],
      num_batches,
      height,
      shapeL[2]
  ])

  reshaped = tf.reshape(tensor, n_shape)
  return tf.transpose(reshaped, [1, 2, 0, 3]) 

def lstm_online(cell_fn, num_steps, inputs, state, varscope):
  # inputs is B x num_steps x C, C channels.
  # state is 2 tuple with B x 1 x C1, B x 1 x C2 
  # Output state is always B x 1 x C
  inputs = tf.unstack(inputs, axis=1, num=num_steps)
  state = tf.unstack(state, axis=1, num=1)[0]
  outputs = [] 
  
  if num_steps > 1: 
    varscope.reuse_variables()
  
  for s in range(num_steps):
    output, state = cell_fn(inputs[s], state)
    outputs.append(output)
  outputs = tf.stack(outputs, axis=1)
  state = tf.stack([state], axis=1)
  return outputs, state 

def convert_network_state_tensorarray(tensorarray):
  """Converts a source TensorArray to a source Tensor.

  Performs a permutation between the steps * [stride, D] shape of a
  source TensorArray and the (flattened) [stride * steps, D] shape of
  a source Tensor.

  The TensorArrays used during recurrence have an additional zeroth step that
  needs to be removed.

  Args:
    tensorarray: TensorArray object to be converted.

  Returns:
    Tensor object after conversion.
  """
  tensor = tensorarray.stack()  # Results in a [steps, stride, D] tensor.
  tensor = tf.slice(tensor, [1, 0, 0], [-1, -1, -1])  # Lop off the 0th step.
  tensor = tf.transpose(tensor, [1, 0, 2])  # Switch steps and stride.
  return tf.reshape(tensor, [-1, tf.shape(tensor)[2]]) 

def one_hot_encoding(labels, num_classes, scope=None):
  """Transform numeric labels into onehot_labels.

  Args:
    labels: [batch_size] target labels.
    num_classes: total number of classes.
    scope: Optional scope for name_scope.
  Returns:
    one hot encoding of the labels.
  """
  with tf.name_scope(scope, 'OneHotEncoding', [labels]):
    batch_size = labels.get_shape()[0]
    indices = tf.expand_dims(tf.range(0, batch_size), 1)
    labels = tf.cast(tf.expand_dims(labels, 1), indices.dtype)
    concated = tf.concat(axis=1, values=[indices, labels])
    onehot_labels = tf.sparse_to_dense(
        concated, tf.stack([batch_size, num_classes]), 1.0, 0.0)
    onehot_labels.set_shape([batch_size, num_classes])
    return onehot_labels 

def test_batch_decode(self):
    mock_anchor_corners = tf.constant(
        [[0, 0.1, 0.2, 0.3], [0.2, 0.4, 0.4, 0.6]], tf.float32)
    mock_anchors = box_list.BoxList(mock_anchor_corners)
    mock_box_coder = MockBoxCoder()

    expected_boxes = [[[0.0, 0.1, 0.5, 0.6], [0.5, 0.6, 0.7, 0.8]],
                      [[0.1, 0.2, 0.3, 0.4], [0.7, 0.8, 0.9, 1.0]]]

    encoded_boxes_list = [mock_box_coder.encode(
        box_list.BoxList(tf.constant(boxes)), mock_anchors)
                          for boxes in expected_boxes]
    encoded_boxes = tf.stack(encoded_boxes_list)
    decoded_boxes = box_coder.batch_decode(
        encoded_boxes, mock_box_coder, mock_anchors)

    with self.test_session() as sess:
      decoded_boxes_result = sess.run(decoded_boxes)
      self.assertAllClose(expected_boxes, decoded_boxes_result) 

def __init__(self, x_bxu, z_size, name, var_min=0.0):
    """Create an input dependent diagonal Gaussian distribution.

    Args:
      x: The input tensor from which the mean and variance are computed,
        via a linear transformation of x.  I.e.
          mu = Wx + b, log(var) = Mx + c
      z_size: The size of the distribution.
      name:  The name to prefix to learned variables.
      var_min (optional): Minimal variance allowed.  This is an additional
        way to control the amount of information getting through the stochastic
        layer.
    """
    size_bxn = tf.stack([tf.shape(x_bxu)[0], z_size])
    self.mean_bxn = mean_bxn = linear(x_bxu, z_size, name=(name+"/mean"))
    logvar_bxn = linear(x_bxu, z_size, name=(name+"/logvar"))
    if var_min > 0.0:
      logvar_bxn = tf.log(tf.exp(logvar_bxn) + var_min)
    self.logvar_bxn = logvar_bxn

    self.noise_bxn = noise_bxn = tf.random_normal(size_bxn)
    self.noise_bxn.set_shape([None, z_size])
    self.sample_bxn = mean_bxn + tf.exp(0.5 * logvar_bxn) * noise_bxn 

def combine(self, expert_out, multiply_by_gates=True):
    """Sum together the expert output, multiplied by the corresponding gates.

    Args:
      expert_out: a list of `num_experts` `Tensor`s, each with shape
        `[expert_batch_size_i, <extra_output_dims>]`.
      multiply_by_gates: a boolean.

    Returns:
      a list of num_datashards `Tensor`s with shapes
        `[batch_size[d], <extra_output_dims>]`.
    """
    expert_part_sizes = tf.unstack(
        tf.stack([d.part_sizes for d in self._dispatchers]),
        num=self._ep.n,
        axis=1)
    # list of lists of shape [num_experts][num_datashards]
    expert_output_parts = self._ep(tf.split, expert_out, expert_part_sizes)
    expert_output_parts_t = transpose_list_of_lists(expert_output_parts)
    def my_combine(dispatcher, parts):
      return dispatcher.combine(
          common_layers.convert_gradient_to_tensor(tf.concat(parts, 0)),
          multiply_by_gates=multiply_by_gates)
    return self._dp(my_combine, self._dispatchers, expert_output_parts_t) 

def select_dim_value(x, indices, name=None):
    with tf.name_scope(name, "select-dim-value", values=[x, indices]):
        # x.shape = (rest..., dims)
        rest = tf.shape(x)[:-1]
        dims = tf.shape(x)[-1]
        size = tf.size(indices, out_type=indices.dtype)

        # reshape to (size, dims)
        t = tf.reshape(x, shape=[-1, dims])
        # then index as ([1,2,3,...,size], indices.ravel())
        nd_indices = tf.stack([
            tf.range(0, size, dtype=indices.dtype),
            tf.reshape(indices, shape=[-1])
        ], axis=1)
        t = tf.gather_nd(t, indices=nd_indices)

        # reshape back to (rest...)
        t = tf.reshape(t, rest)
        t.set_shape(x.get_shape()[:-1])
        return t 

def test_batch_decode(self):
    mock_anchor_corners = tf.constant(
        [[0, 0.1, 0.2, 0.3], [0.2, 0.4, 0.4, 0.6]], tf.float32)
    mock_anchors = box_list.BoxList(mock_anchor_corners)
    mock_box_coder = MockBoxCoder()

    expected_boxes = [[[0.0, 0.1, 0.5, 0.6], [0.5, 0.6, 0.7, 0.8]],
                      [[0.1, 0.2, 0.3, 0.4], [0.7, 0.8, 0.9, 1.0]]]

    encoded_boxes_list = [mock_box_coder.encode(
        box_list.BoxList(tf.constant(boxes)), mock_anchors)
                          for boxes in expected_boxes]
    encoded_boxes = tf.stack(encoded_boxes_list)
    decoded_boxes = box_coder.batch_decode(
        encoded_boxes, mock_box_coder, mock_anchors)

    with self.test_session() as sess:
      decoded_boxes_result = sess.run(decoded_boxes)
      self.assertAllClose(expected_boxes, decoded_boxes_result) 

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 get_logits(new_input, length, first=[]):
    """
    Compute the logits for a given waveform.

    First, preprocess with the TF version of MFC above,
    and then call DeepSpeech on the features.
    """
    # new_input = tf.Print(new_input, [tf.shape(new_input)])

    # We need to init DeepSpeech the first time we're called
    if first == []:
        first.append(False)
        # Okay, so this is ugly again.
        # We just want it to not crash.
        tf.app.flags.FLAGS.alphabet_config_path = "DeepSpeech/data/alphabet.txt"
        DeepSpeech.initialize_globals()
        print('initialized deepspeech globals')

    batch_size = new_input.get_shape()[0]

    # 1. Compute the MFCCs for the input audio
    # (this is differentable with our implementation above)
    empty_context = np.zeros((batch_size, 9, 26), dtype=np.float32)
    new_input_to_mfcc = compute_mfcc(new_input)[:, ::2]
    features = tf.concat((empty_context, new_input_to_mfcc, empty_context), 1)

    # 2. We get to see 9 frames at a time to make our decision,
    # so concatenate them together.
    features = tf.reshape(features, [new_input.get_shape()[0], -1])
    features = tf.stack([features[:, i:i+19*26] for i in range(0,features.shape[1]-19*26+1,26)],1)
    features = tf.reshape(features, [batch_size, -1, 19*26])

    # 3. Whiten the data
    mean, var = tf.nn.moments(features, axes=[0,1,2])
    features = (features-mean)/(var**.5)

    # 4. Finally we process it with DeepSpeech
    logits = DeepSpeech.BiRNN(features, length, [0]*10)

    return logits 

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

        querys = tf.tensordot(inputs, self.W_Query,
                              axes=(-1, 0))  # None F D*head_num
        keys = tf.tensordot(inputs, self.W_key, axes=(-1, 0))
        values = tf.tensordot(inputs, self.W_Value, axes=(-1, 0))

        # head_num None F D
        querys = tf.stack(tf.split(querys, self.head_num, axis=2))
        keys = tf.stack(tf.split(keys, self.head_num, axis=2))
        values = tf.stack(tf.split(values, self.head_num, axis=2))

        inner_product = tf.matmul(
            querys, keys, transpose_b=True)  # head_num None F F
        self.normalized_att_scores = tf.nn.softmax(inner_product)

        result = tf.matmul(self.normalized_att_scores,
                           values)  # head_num None F D
        result = tf.concat(tf.split(result, self.head_num, ), axis=-1)
        result = tf.squeeze(result, axis=0)  # None F D*head_num

        if self.use_res:
            result += tf.tensordot(inputs, self.W_Res, axes=(-1, 0))
        result = tf.nn.relu(result)

        return result 

def extract_patches(matrix, ws, h, w):

    val = []
    for yo in range(ws):
        for xo in range(ws):
            MN = matrix[:, yo:yo+h-ws+1, xo:xo+w-ws+1, :]
            val.append(MN)

    win_ids = tf.stack(val, 3)
    
    return win_ids

#----------------------------------------------------------------------------
# Compute structural loss 

def biLSTMCell(x, hiddenSize):
        input_x = tf.transpose(x, [1, 0, 2])
        input_x = tf.unstack(input_x)
        lstm_fw_cell = tf.contrib.rnn.BasicLSTMCell(hiddenSize, forget_bias=1.0, state_is_tuple=True)
        lstm_bw_cell = tf.contrib.rnn.BasicLSTMCell(hiddenSize, forget_bias=1.0, state_is_tuple=True)
        output, _, _ = tf.contrib.rnn.static_bidirectional_rnn(lstm_fw_cell, lstm_bw_cell, input_x, dtype=tf.float32)
        output = tf.stack(output)
        output = tf.transpose(output, [1, 0, 2])
        return output 

def _compute_gradients(self, loss_fn, x, unused_optim_state):
        """Compute gradient estimates using SPSA."""
        # Assumes `x` is a list, containing a [1, H, W, C] image
        assert len(x) == 1 and x[0].get_shape().as_list()[0] == 1
        x = x[0]
        x_shape = x.get_shape().as_list()

        def body(i, grad_array):
            delta = self._delta
            delta_x = self._get_delta(x, delta)
            delta_x = tf.concat([delta_x, -delta_x], axis=0)
            loss_vals = tf.reshape(
                loss_fn(x + delta_x),
                [2 * self._num_samples] + [1] * (len(x_shape) - 1))
            avg_grad = reduce_mean(loss_vals * delta_x, axis=0) / delta
            avg_grad = tf.expand_dims(avg_grad, axis=0)
            new_grad_array = grad_array.write(i, avg_grad)
            return i + 1, new_grad_array

        def cond(i, _):
            return i < self._num_iters

        _, all_grads = tf.while_loop(
            cond,
            body,
            loop_vars=[
                0, tf.TensorArray(size=self._num_iters, dtype=tf_dtype)
            ],
            back_prop=False,
            parallel_iterations=1)
        avg_grad = reduce_sum(all_grads.stack(), axis=0)
        return [avg_grad] 

def _build_mwf_output_waveform(self):
        """ Perform separation with multichannel Wiener Filtering using Norbert.
        Note: multichannel Wiener Filtering is not coded in Tensorflow and thus
        may be quite slow.

        :returns: dictionary of separated waveforms (key: instrument name,
            value: estimated waveform of the instrument)
        """
        import norbert  # pylint: disable=import-error
        output_dict = self.model_outputs
        x = self.stft_feature
        v = tf.stack(
            [
                pad_and_reshape(
                    output_dict[f'{instrument}_spectrogram'],
                    self._frame_length,
                    self._F)[:tf.shape(x)[0], ...]
                for instrument in self._instruments
            ],
            axis=3)
        input_args = [v, x]
        stft_function = tf.py_function(
            lambda v, x: norbert.wiener(v.numpy(), x.numpy()),
            input_args,
            tf.complex64),
        return {
            instrument: self._inverse_stft(stft_function[0][:, :, :, k])
            for k, instrument in enumerate(self._instruments)
        } 

def loop_decode(self, decoder_initial_state):
        # decoder_initial_state: Tuple Tensor (c,h) of size [batch_size x cell.state_size]
        # decoder_first_input: Tensor [batch_size x cell.state_size]

        # Decoder initial input is depot (start)
        decoder_first_input = tf.gather(self.h,self.depot_position)[0]

        # Loop the decoding process and collect results
        s,i = decoder_initial_state, decoder_first_input
        for step in range(self.seq_length-1):
            s, i = self.decode(s,i,step)

        # Return to depot
        self.positions.append(self.depot_position)

        # Stack visited indices
        self.positions=tf.stack(self.positions,axis=1)  # [Batch,seq_length+1]

        # Sum log_softmax over output steps
        self.log_softmax=tf.add_n(self.log_softmax)  # [Batch,seq_length-1]

        # Stack attending & pointing distribution
        self.attending=tf.stack(self.attending,axis=1) # [Batch,seq_length-1,seq_length]
        self.pointing=tf.stack(self.pointing,axis=1) # [Batch,seq_length-1,seq_length]
        
        # Return stacked lists of visited_indices and log_softmax for backprop
        return self.positions,self.log_softmax, self.attending, self.pointing 

def deconv2d_bn_lrn_drop(scope_or_name, inputs, kernel_shape, out_shape, subS=2, activation=tf.nn.relu,
                       use_bn=False,
                       use_mvn=False,
                       is_training=True,
                       use_lrn=False,
                       keep_prob=1.0,
                       dropout_maps=False,
                       initOpt=0):
    with tf.variable_scope(scope_or_name):
        if initOpt == 0:
            stddev = np.sqrt(2.0 / (kernel_shape[0] * kernel_shape[1] * kernel_shape[2] + kernel_shape[3]))
        if initOpt == 1:
            stddev = 5e-2
        if initOpt == 2:
            stddev = min(np.sqrt(2.0 / (kernel_shape[0] * kernel_shape[1] * kernel_shape[2])),5e-2)
        kernel = tf.get_variable("weights", kernel_shape,
                                 initializer=tf.random_normal_initializer(stddev=stddev))
        bias = tf.get_variable("bias", kernel_shape[2],
                               initializer=tf.constant_initializer(value=0.1))
        conv=tf.nn.conv2d_transpose(inputs, kernel, out_shape, strides=[1, subS, subS, 1], padding='SAME', name='conv')
        outputs = tf.nn.bias_add(conv, bias, name='preActivation')
        if use_bn:
            # outputs = tf.layers.batch_normalization(outputs, axis=3, training=is_training, name="batchNorm")
            outputs = batch_norm(outputs, is_training=is_training, scale=True, fused=True, scope="batchNorm")
        if use_mvn:
            outputs = feat_norm(outputs, kernel_shape[3])
        if activation:
            outputs = activation(outputs, name='activation')
        if use_lrn:
            outputs = tf.nn.local_response_normalization(outputs, name='localResponseNorm')
        if dropout_maps:
            conv_shape = tf.shape(outputs)
            n_shape = tf.stack([conv_shape[0], 1, 1, conv_shape[3]])
            outputs = tf.nn.dropout(outputs, keep_prob, noise_shape=n_shape)
        else:
            outputs = tf.nn.dropout(outputs, keep_prob)
        return outputs 

def _crop(image, offset_height, offset_width, crop_height, crop_width):
  """Crops the given image using the provided offsets and sizes.

  Note that the method doesn't assume we know the input image size but it does
  assume we know the input image rank.

  Args:
    image: an image of shape [height, width, channels].
    offset_height: a scalar tensor indicating the height offset.
    offset_width: a scalar tensor indicating the width offset.
    crop_height: the height of the cropped image.
    crop_width: the width of the cropped image.

  Returns:
    the cropped (and resized) image.

  Raises:
    InvalidArgumentError: if the rank is not 3 or if the image dimensions are
      less than the crop size.
  """
  original_shape = tf.shape(image)

  rank_assertion = tf.Assert(
      tf.equal(tf.rank(image), 3),
      ['Rank of image must be equal to 3.'])
  with tf.control_dependencies([rank_assertion]):
    cropped_shape = tf.stack([crop_height, crop_width, original_shape[2]])

  size_assertion = tf.Assert(
      tf.logical_and(
          tf.greater_equal(original_shape[0], crop_height),
          tf.greater_equal(original_shape[1], crop_width)),
      ['Crop size greater than the image size.'])

  offsets = tf.to_int32(tf.stack([offset_height, offset_width, 0]))

  # Use tf.slice instead of crop_to_bounding box as it accepts tensors to
  # define the crop size.
  with tf.control_dependencies([size_assertion]):
    image = tf.slice(image, offsets, cropped_shape)
  return tf.reshape(image, cropped_shape) 

def running_combine(fss_logits, confs_probs, incremental_locs,
                    incremental_thetas, previous_sum_num, previous_sum_denom,
                    previous_max_denom, map_size, num_steps):
  # fss_logits is B x N x H x W x C
  # confs_logits is B x N x H x W x C
  # incremental_locs is B x N x 2
  # incremental_thetas is B x N x 1
  # previous_sum_num etc is B x 1 x H x W x C

  with tf.name_scope('combine_{:d}'.format(num_steps)):
    running_sum_nums_ = []; running_sum_denoms_ = [];
    running_max_denoms_ = [];

    fss_logits_ = tf.unstack(fss_logits, axis=1, num=num_steps)
    confs_probs_ = tf.unstack(confs_probs, axis=1, num=num_steps)
    incremental_locs_ = tf.unstack(incremental_locs, axis=1, num=num_steps)
    incremental_thetas_ = tf.unstack(incremental_thetas, axis=1, num=num_steps)
    running_sum_num = tf.unstack(previous_sum_num, axis=1, num=1)[0]
    running_sum_denom = tf.unstack(previous_sum_denom, axis=1, num=1)[0]
    running_max_denom = tf.unstack(previous_max_denom, axis=1, num=1)[0]

    for i in range(num_steps):
      # Rotate the previous running_num and running_denom
      running_sum_num, running_sum_denom, running_max_denom = rotate_preds(
          incremental_locs_[i], incremental_thetas_[i], map_size,
          [running_sum_num, running_sum_denom, running_max_denom],
          output_valid_mask=False)[0]
      # print i, num_steps, running_sum_num.get_shape().as_list()
      running_sum_num = running_sum_num + fss_logits_[i] * confs_probs_[i]
      running_sum_denom = running_sum_denom + confs_probs_[i]
      running_max_denom = tf.maximum(running_max_denom, confs_probs_[i])
      running_sum_nums_.append(running_sum_num)
      running_sum_denoms_.append(running_sum_denom)
      running_max_denoms_.append(running_max_denom)

    running_sum_nums = tf.stack(running_sum_nums_, axis=1)
    running_sum_denoms = tf.stack(running_sum_denoms_, axis=1)
    running_max_denoms = tf.stack(running_max_denoms_, axis=1)
    return running_sum_nums, running_sum_denoms, running_max_denoms 

def create(self,
             fixed_embeddings,
             linked_embeddings,
             context_tensor_arrays,
             attention_tensor,
             during_training,
             stride=None):
    """Requires |stride|; otherwise see base class."""
    # TODO(googleuser): Normalize the arguments to create(). 'stride'
    # is unused by the recurrent network units, while 'context_tensor_arrays'
    # and 'attenion_tensor_array' is unused by bulk network units. b/33587044
    if stride is None:
      raise ValueError("PairwiseConvNetwork needs 'stride'")

    input_tensor = get_input_tensor_with_stride(fixed_embeddings,
                                                linked_embeddings, stride)

    # TODO(googleuser): Add dropout.
    del context_tensor_arrays, attention_tensor, during_training  # Unused.

    num_steps = tf.shape(input_tensor)[1]
    arg1 = tf.expand_dims(input_tensor, 1)
    arg1 = tf.tile(arg1, tf.stack([1, num_steps, 1, 1]))
    arg2 = tf.expand_dims(input_tensor, 2)
    arg2 = tf.tile(arg2, tf.stack([1, 1, num_steps, 1]))
    conv = tf.concat([arg1, arg2], 3)
    for i in xrange(self._num_layers):
      with tf.variable_scope('conv%d' % i, reuse=True) as scope:
        conv = tf.nn.conv2d(
            conv,
            self._component.get_variable('weights'), [1, 1, 1, 1],
            padding='SAME')
        conv = tf.nn.bias_add(conv, self._component.get_variable('biases'))
        if i in self._relu_layers:
          conv = tf.nn.relu(conv, name=scope.name)
    return [tf.reshape(conv, [-1, num_steps], name='reshape_activations')] 

def __call__(self, w, z_grads):
    idx = list(self.op.inputs).index(w)
    # Make sure that `op` was actually applied to `w`
    assert idx != -1
    assert len(z_grads) == len(self.op.outputs)
    # The following assert may be removed when we are ready to use this
    # for general purpose code.
    # This assert is only expected to hold in the contex of our preliminary
    # MNIST experiments.
    assert idx == 1  # We expect convolution weights to be arg 1

    images, filters = self.op.inputs
    strides = self.op.get_attr("strides")
    padding = self.op.get_attr("padding")
    # Currently assuming that one specifies at most these four arguments and
    # that all other arguments to conv2d are set to default.

    conv, w_px = self._PxConv2DBuilder(images, filters, strides, padding)
    z_grads, = z_grads

    gradients_list = tf.gradients(conv, w_px, z_grads,
                                  colocate_gradients_with_ops=
                                  self.colocate_gradients_with_ops,
                                  gate_gradients=self.gate_gradients)

    return tf.stack(gradients_list) 

def dense_to_sparse_boxes(dense_locations, dense_num_boxes, num_classes):
  """Converts bounding boxes from dense to sparse form.

  Args:
    dense_locations:  a [max_num_boxes, 4] tensor in which only the first k rows
      are valid bounding box location coordinates, where k is the sum of
      elements in dense_num_boxes.
    dense_num_boxes: a [max_num_classes] tensor indicating the counts of
       various bounding box classes e.g. [1, 0, 0, 2] means that the first
       bounding box is of class 0 and the second and third bounding boxes are
       of class 3. The sum of elements in this tensor is the number of valid
       bounding boxes.
    num_classes: number of classes

  Returns:
    box_locations: a [num_boxes, 4] tensor containing only valid bounding
       boxes (i.e. the first num_boxes rows of dense_locations)
    box_classes: a [num_boxes] tensor containing the classes of each bounding
       box (e.g. dense_num_boxes = [1, 0, 0, 2] => box_classes = [0, 3, 3]
  """

  num_valid_boxes = tf.reduce_sum(dense_num_boxes)
  box_locations = tf.slice(dense_locations,
                           tf.constant([0, 0]), tf.stack([num_valid_boxes, 4]))
  tiled_classes = [tf.tile([i], tf.expand_dims(dense_num_boxes[i], 0))
                   for i in range(num_classes)]
  box_classes = tf.concat(tiled_classes, 0)
  box_locations.set_shape([None, 4])
  return box_locations, box_classes 

def loss(self, prediction_dict):
    """Compute scalar loss tensors with respect to provided groundtruth.

    Calling this function requires that groundtruth tensors have been
    provided via the provide_groundtruth function.

    Args:
      prediction_dict: a dictionary holding predicted tensors

    Returns:
      a dictionary mapping strings (loss names) to scalar tensors representing
        loss values.
    """
    batch_reg_targets = tf.stack(
        self.groundtruth_lists(fields.BoxListFields.boxes))
    batch_cls_targets = tf.stack(
        self.groundtruth_lists(fields.BoxListFields.classes))
    weights = tf.constant(
        1.0, dtype=tf.float32,
        shape=[len(self.groundtruth_lists(fields.BoxListFields.boxes)), 1])

    location_losses = self._localization_loss(
        prediction_dict['box_encodings'], batch_reg_targets,
        weights=weights)
    cls_losses = self._classification_loss(
        prediction_dict['class_predictions_with_background'], batch_cls_targets,
        weights=weights)

    loss_dict = {
        'localization_loss': tf.reduce_sum(location_losses),
        'classification_loss': tf.reduce_sum(cls_losses),
    }
    return loss_dict 

def _summarize_input(self, groundtruth_boxes_list, match_list):
    """Creates tensorflow summaries for the input boxes and anchors.

    This function creates four summaries corresponding to the average
    number (over images in a batch) of (1) groundtruth boxes, (2) anchors
    marked as positive, (3) anchors marked as negative, and (4) anchors marked
    as ignored.

    Args:
      groundtruth_boxes_list: a list of 2-D tensors of shape [num_boxes, 4]
        containing corners of the groundtruth boxes.
      match_list: a list of matcher.Match objects encoding the match between
        anchors and groundtruth boxes for each image of the batch,
        with rows of the Match objects corresponding to groundtruth boxes
        and columns corresponding to anchors.
    """
    num_boxes_per_image = tf.stack(
        [tf.shape(x)[0] for x in groundtruth_boxes_list])
    pos_anchors_per_image = tf.stack(
        [match.num_matched_columns() for match in match_list])
    neg_anchors_per_image = tf.stack(
        [match.num_unmatched_columns() for match in match_list])
    ignored_anchors_per_image = tf.stack(
        [match.num_ignored_columns() for match in match_list])
    tf.summary.scalar('Input/AvgNumGroundtruthBoxesPerImage',
                      tf.reduce_mean(tf.to_float(num_boxes_per_image)))
    tf.summary.scalar('Input/AvgNumPositiveAnchorsPerImage',
                      tf.reduce_mean(tf.to_float(pos_anchors_per_image)))
    tf.summary.scalar('Input/AvgNumNegativeAnchorsPerImage',
                      tf.reduce_mean(tf.to_float(neg_anchors_per_image)))
    tf.summary.scalar('Input/AvgNumIgnoredAnchorsPerImage',
                      tf.reduce_mean(tf.to_float(ignored_anchors_per_image))) 

def _create_regression_targets(self, anchors, groundtruth_boxes, match):
    """Returns a regression target for each anchor.

    Args:
      anchors: a BoxList representing N anchors
      groundtruth_boxes: a BoxList representing M groundtruth_boxes
      match: a matcher.Match object

    Returns:
      reg_targets: a float32 tensor with shape [N, box_code_dimension]
    """
    matched_anchor_indices = match.matched_column_indices()
    unmatched_ignored_anchor_indices = (match.
                                        unmatched_or_ignored_column_indices())
    matched_gt_indices = match.matched_row_indices()
    matched_anchors = box_list_ops.gather(anchors,
                                          matched_anchor_indices)
    matched_gt_boxes = box_list_ops.gather(groundtruth_boxes,
                                           matched_gt_indices)
    matched_reg_targets = self._box_coder.encode(matched_gt_boxes,
                                                 matched_anchors)
    unmatched_ignored_reg_targets = tf.tile(
        self._default_regression_target(),
        tf.stack([tf.size(unmatched_ignored_anchor_indices), 1]))
    reg_targets = tf.dynamic_stitch(
        [matched_anchor_indices, unmatched_ignored_anchor_indices],
        [matched_reg_targets, unmatched_ignored_reg_targets])
    # TODO: summarize the number of matches on average.
    return reg_targets