Ejemplo n.º 1
0
def read_cnn(filename_queue):
  
  class CNNRecord(object):
    pass
  result = CNNRecord()

  # each example in the file is ('megabatch') is label + (300,42) data in float32 (4 bytes per value). (1+12600)*4 = 50404 bytes per example
  bytes_per_value = 4 # float32
  label_bytes = 1 * bytes_per_value
  result.height = EXAMPLE_HEIGHT
  result.width = EXAMPLE_WIDTH
  result.depth = 1
  example_bytes = result.height * result.width * result.depth * bytes_per_value
  # Every record consists of a label followed by the image, with a
  # fixed number of bytes for each.
  record_bytes = label_bytes + example_bytes
  # print 'record_bytes number is ', record_bytes # should be 4 + 300*42*4 = 50404
  reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
  result.key, value_bytes = reader.read(filename_queue)
  # Convert from a string to a vector of float32 that is record_bytes long.
  value = tf.decode_raw(value_bytes, tf.float32)
  # The first bytes represent the label, which we convert from uint8->int32.
  result.label = tf.slice(value, [0], [label_bytes/bytes_per_value])
  
  # The remaining bytes after the label represent the image, which we reshape
  # from [depth * height * width] to [depth, height, width].
  result.example = tf.slice(value, [label_bytes/bytes_per_value], [example_bytes/bytes_per_value])

  return result
Ejemplo n.º 2
0
def cell_locate(size, bbox, S):

    """ 
    locate the center of ground truth in which grid cell

    """
    x = tf.cast(tf.slice(bbox, [0,0], [-1,1]), tf.float32)
    y = tf.cast(tf.slice(bbox, [0,1], [-1,1]), tf.float32)
    w = tf.cast(tf.slice(bbox, [0,2], [-1,1]), tf.float32)
    h = tf.cast(tf.slice(bbox, [0,3], [-1,1]), tf.float32)


    height, width = size

    cell_w = width / S
    cell_h = height / S

    center_y = tf.add(y, tf.mul(h, 0.5))
    center_x = tf.add(x, tf.mul(w, 0.5))

    cell_coord_x = tf.cast(tf.div(center_x, cell_w), tf.int32)
    cell_coord_y = tf.cast(tf.div(center_y, cell_h), tf.int32)

    cell_num = tf.add(tf.mul(cell_coord_y, S), cell_coord_x)

    return cell_num
Ejemplo n.º 3
0
def get_idx_map(shape):
    """Get index map for a image.
    Args:
        shape: [B, T, H, W] or [B, H, W]
    Returns:
        idx: [B, T, H, W, 2], or [B, H, W, 2]
    """
    s = shape
    ndims = tf.shape(s)
    wdim = ndims - 1
    hdim = ndims - 2
    idx_shape = tf.concat(0, [s, tf.constant([1])])
    ones_h = tf.ones(hdim - 1, dtype='int32')
    ones_w = tf.ones(wdim - 1, dtype='int32')
    h_shape = tf.concat(0, [ones_h, tf.constant([-1]), tf.constant([1, 1])])
    w_shape = tf.concat(0, [ones_w, tf.constant([-1]), tf.constant([1])])

    idx_y = tf.zeros(idx_shape, dtype='float')
    idx_x = tf.zeros(idx_shape, dtype='float')

    h = tf.slice(s, ndims - 2, [1])
    w = tf.slice(s, ndims - 1, [1])
    idx_y += tf.reshape(tf.to_float(tf.range(h[0])), h_shape)
    idx_x += tf.reshape(tf.to_float(tf.range(w[0])), w_shape)
    idx = tf.concat(ndims[0], [idx_y, idx_x])

    return idx
Ejemplo n.º 4
0
def read_record(filename_queue):
    class FCNRecord(object):
        pass
    result = FCNRecord()
    result.mask_height = int(420/DOWNSAMPLE_FACTOR)
    result.mask_width = int(580/DOWNSAMPLE_FACTOR)
    result.mask_depth = 1
    result.img_depth = 1
    img_len = result.mask_height*result.mask_width*result.img_depth
    mask_len = result.mask_height*result.mask_width*result.mask_depth
    record_len = img_len + mask_len

    reader = tf.FixedLengthRecordReader(record_bytes=record_len)
    result.key, value = reader.read(filename_queue)
    record_bytes = tf.decode_raw(value, tf.uint8)
    #print(record_bytes.get_shape())
    int_image = tf.reshape(tf.slice(record_bytes, [0], [img_len]),[result.mask_height, result.mask_width])
    rgb_image = tf.pack([int_image,int_image,int_image])
    rgb_img = tf.transpose(rgb_image,(1,2,0))
    result.image = tf.cast(rgb_img,tf.float32)
    bool_mask = tf.cast( tf.reshape(tf.slice(record_bytes, [img_len], [mask_len]),[result.mask_height, result.mask_width]), tf.bool)
    hot_mask= tf.pack( [bool_mask, tf.logical_not(bool_mask)])
    h_mask = tf.transpose(hot_mask,(1,2,0))
    result.mask = tf.cast(h_mask, tf.float32)
    return result
Ejemplo n.º 5
0
def BatchClipByL2norm(t, upper_bound, name=None):
  """Clip an array of tensors by L2 norm.

  Shrink each dimension-0 slice of tensor (for matrix it is each row) such
  that the l2 norm is at most upper_bound. Here we clip each row as it
  corresponds to each example in the batch.

  Args:
    t: the input tensor.
    upper_bound: the upperbound of the L2 norm.
    name: optional name.
  Returns:
    the clipped tensor.
  """

  assert upper_bound > 0
  with tf.op_scope([t, upper_bound], name, "batch_clip_by_l2norm") as name:
    saved_shape = tf.shape(t)
    batch_size = tf.slice(saved_shape, [0], [1])
    t2 = tf.reshape(t, tf.concat(0, [batch_size, [-1]]))
    upper_bound_inv = tf.fill(tf.slice(saved_shape, [0], [1]),
                              tf.constant(1.0/upper_bound))
    # Add a small number to avoid divide by 0
    l2norm_inv = tf.rsqrt(tf.reduce_sum(t2 * t2, [1]) + 0.000001)
    scale = tf.minimum(l2norm_inv, upper_bound_inv) * upper_bound
    clipped_t = tf.matmul(tf.diag(scale), t2)
    clipped_t = tf.reshape(clipped_t, saved_shape, name=name)
  return clipped_t
Ejemplo n.º 6
0
 def embed_sequences(self, embed_sequence_batch):
     """Return sentence embeddings as a tensor with with shape
     [batch_size, hidden_size * 2]
     """
     forward_values = embed_sequence_batch.values
     forward_mask = embed_sequence_batch.mask
     backward_values = tf.reverse(forward_values, [False, True, False])
     backward_mask = tf.reverse(forward_mask, [False, True])
     # Initialize LSTMs
     self._forward_lstm = LSTM(self.hidden_size, return_sequences=True)
     self._backward_lstm = LSTM(self.hidden_size, return_sequences=True)
     # Pass input through the LSTMs
     # Shape: (batch_size, seq_length, hidden_size)
     forward_seq = self._forward_lstm(forward_values, forward_mask)
     forward_seq.set_shape((None, self.seq_length, self.hidden_size))
     backward_seq = self._backward_lstm(backward_values, backward_mask)
     backward_seq.set_shape((None, self.seq_length, self.hidden_size))
     # Stitch the outputs together --> hidden states (for computing attention)
     # Final dimension: (batch_size, seq_length, hidden_size * 2)
     lstm_states = tf.concat(2, [forward_seq, tf.reverse(backward_seq, [False, True, False])])
     self._hidden_states = SequenceBatch(lstm_states, forward_mask)
     # Stitch the final outputs together --> sequence embedding
     # Final dimension: (batch_size, hidden_size * 2)
     seq_length = tf.shape(forward_values)[1]
     forward_final = tf.slice(forward_seq, [0, seq_length - 1, 0], [-1, 1, self.hidden_size])
     backward_final = tf.slice(backward_seq, [0, seq_length - 1, 0], [-1, 1, self.hidden_size])
     return tf.squeeze(tf.concat(2, [forward_final, backward_final]), [1])
Ejemplo n.º 7
0
def knn_point(k, xyz1, xyz2):
    '''
    Input:
        k: int32, number of k in k-nn search
        xyz1: (batch_size, ndataset, c) float32 array, input points
        xyz2: (batch_size, npoint, c) float32 array, query points
    Output:
        val: (batch_size, npoint, k) float32 array, L2 distances
        idx: (batch_size, npoint, k) int32 array, indices to input points
    '''
    b = xyz1.get_shape()[0].value
    n = xyz1.get_shape()[1].value
    c = xyz1.get_shape()[2].value
    m = xyz2.get_shape()[1].value
    print b, n, c, m
    print xyz1, (b,1,n,c)
    xyz1 = tf.tile(tf.reshape(xyz1, (b,1,n,c)), [1,m,1,1])
    xyz2 = tf.tile(tf.reshape(xyz2, (b,m,1,c)), [1,1,n,1])
    dist = tf.reduce_sum((xyz1-xyz2)**2, -1)
    print dist, k
    outi, out = select_top_k(k, dist)
    idx = tf.slice(outi, [0,0,0], [-1,-1,k])
    val = tf.slice(out, [0,0,0], [-1,-1,k])
    print idx, val
    #val, idx = tf.nn.top_k(-dist, k=k) # ONLY SUPPORT CPU
    return val, idx
 def input_fn():
   starts = tf.random_uniform([batch_size], maxval=(2 * np.pi), seed=seed)
   sin_curves = tf.map_fn(_sin_fn, (starts,), dtype=tf.float32)
   inputs = tf.expand_dims(
       tf.slice(sin_curves, [0, 0], [batch_size, sequence_length]), 2)
   labels = tf.slice(sin_curves, [0, 1], [batch_size, sequence_length])
   return {'inputs': inputs}, labels
Ejemplo n.º 9
0
def diff(x, axis=-1):
  """Take the finite difference of a tensor along an axis.

  Args:
    x: Input tensor of any dimension.
    axis: Axis on which to take the finite difference.

  Returns:
    d: Tensor with size less than x by 1 along the difference dimension.

  Raises:
    ValueError: Axis out of range for tensor.
  """
  shape = x.get_shape()
  if axis >= len(shape):
    raise ValueError('Invalid axis index: %d for tensor with only %d axes.' %
                     (axis, len(shape)))

  begin_back = [0 for unused_s in range(len(shape))]
  begin_front = [0 for unused_s in range(len(shape))]
  begin_front[axis] = 1

  size = shape.as_list()
  size[axis] -= 1
  slice_front = tf.slice(x, begin_front, size)
  slice_back = tf.slice(x, begin_back, size)
  d = slice_front - slice_back
  return d
Ejemplo n.º 10
0
    def fast_rcnn_minibatch(self, reference_boxes):
        with tf.variable_scope('fast_rcnn_minibatch'):

            reference_boxes_mattached_gtboxes, object_mask, label = \
                self.fast_rcnn_find_positive_negative_samples(reference_boxes)

            positive_indices = tf.reshape(tf.where(tf.not_equal(object_mask, 0.)), [-1])

            num_of_positives = tf.minimum(tf.shape(positive_indices)[0],
                                          tf.cast(self.fast_rcnn_minibatch_size*self.fast_rcnn_positives_ratio, tf.int32))

            positive_indices = tf.random_shuffle(positive_indices)
            positive_indices = tf.slice(positive_indices, begin=[0], size=[num_of_positives])

            negative_indices = tf.reshape(tf.where(tf.equal(object_mask, 0.)), [-1])
            num_of_negatives = tf.minimum(tf.shape(negative_indices)[0],
                                          self.fast_rcnn_minibatch_size - num_of_positives)

            negative_indices = tf.random_shuffle(negative_indices)
            negative_indices = tf.slice(negative_indices, begin=[0], size=[num_of_negatives])

            minibatch_indices = tf.concat([positive_indices, negative_indices], axis=0)
            minibatch_indices = tf.random_shuffle(minibatch_indices)

            minibatch_reference_boxes_mattached_gtboxes = tf.gather(reference_boxes_mattached_gtboxes,
                                                                    minibatch_indices)
            object_mask = tf.gather(object_mask, minibatch_indices)
            label = tf.gather(label, minibatch_indices)
            label_one_hot = tf.one_hot(label, self.num_classes + 1)

            return minibatch_indices, minibatch_reference_boxes_mattached_gtboxes, object_mask, label_one_hot
Ejemplo n.º 11
0
    def forward(self, state, autoencoder):
        '''
        state: vector
        '''

        if autoencoder is None:
            _input = state
        else:
            _input, _ = autoencoder.forward(state)

        state_ = _input
        # clip observation variables from full state
        x_H_ = tf.slice(state_, [0, 0], [-1, 6])
        v_ct = tf.slice(state_, [0, 1], [-1, 1]) - tf.slice(state_, [0, 3], [-1, 1])
        # x_H_ = tf.concat(concat_dim=1, values=[x_H_, v_ct])

        h0 = tf.nn.xw_plus_b(x_H_, self.weights['0'], self.biases['0'], name='h0')
        relu0 = tf.nn.relu(h0)

        h1 = tf.nn.xw_plus_b(relu0, self.weights['1'], self.biases['1'], name='h1')
        relu1 = tf.nn.relu(h1)

        relu1_do = tf.nn.dropout(relu1, self.arch_params['do_keep_prob'])

        a = tf.nn.xw_plus_b(relu1_do, self.weights['c'], self.biases['c'], name='a')

        return a
def read_cifar_files(filename_queue, distort_images = True):
    reader = tf.FixedLengthRecordReader(record_bytes=record_length)
    key, record_string = reader.read(filename_queue)
    record_bytes = tf.decode_raw(record_string, tf.uint8)
    image_label = tf.cast(tf.slice(record_bytes, [0], [1]), tf.int32)
  
    # Extract image
    image_extracted = tf.reshape(tf.slice(record_bytes, [1], [image_vec_length]),
                                 [num_channels, image_height, image_width])
    
    # Reshape image
    image_uint8image = tf.transpose(image_extracted, [1, 2, 0])
    reshaped_image = tf.cast(image_uint8image, tf.float32)
    # Randomly Crop image
    final_image = tf.image.resize_image_with_crop_or_pad(reshaped_image, crop_width, crop_height)
    
    if distort_images:
        # Randomly flip the image horizontally, change the brightness and contrast
        final_image = tf.image.random_flip_left_right(final_image)
        final_image = tf.image.random_brightness(final_image,max_delta=63)
        final_image = tf.image.random_contrast(final_image,lower=0.2, upper=1.8)

    # Normalize whitening
    final_image = tf.image.per_image_standardization(final_image)
    return(final_image, image_label)
def get_image(filename_queue):
  #CIFAR10Record is a 'C struct' bundling tensorflow input data
  class CIFAR10Record(object):
    pass
  #
  result = CIFAR10Record()
  label_bytes = 1  # 2 for CIFAR-100
  result.height = 32
  result.width = 32
  result.depth = 3
  image_bytes = result.height * result.width * result.depth
  # Every record consists of a label followed by the image, with a
  # fixed number of bytes for each.
  record_bytes = label_bytes + image_bytes

  # Read a record, getting filenames from the filename_queue.  No
  # header or footer in the CIFAR-10 format, so we leave header_bytes
  # and footer_bytes at their default of 0.
  reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
  result.key, value = reader.read(filename_queue)

  # Convert from a string to a vector of uint8 that is record_bytes long.
  record_bytes = tf.decode_raw(value, tf.uint8)

  # The first bytes represent the label, which we convert from uint8->int32.
  result.label = tf.cast(
      tf.slice(record_bytes, [0], [label_bytes]), tf.int32)

  # The remaining bytes after the label represent the image, which we reshape
  # from [depth * height * width] to [depth, height, width].
  depth_major = tf.reshape(tf.slice(record_bytes, [label_bytes], [image_bytes]),
                           [result.depth, result.height, result.width])
  # Convert from [depth, height, width] to [height, width, depth].
  result.uint8image = tf.transpose(depth_major, [1, 2, 0])
  return result
Ejemplo n.º 14
0
def read_cifar10(filename_queue):
    class CIFAR10Record(object):
        pass

    result = CIFAR10Record()

    label_bytes = 1
    result.height = 32
    result.width = 32
    result.depth = 3
    images_bytes = result.height * result.width * result.depth
    record_bytes = label_bytes + images_bytes
    print(record_bytes)

    reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
    result.key, value = reader.read(filename_queue)

    record = tf.decode_raw(value, tf.uint8)

    result.label = tf.cast(tf.slice(record, [0], [label_bytes]), tf.int32)

    depth_major = tf.reshape(
        tf.slice(record, [label_bytes], [images_bytes]), [result.depth, result.height, result.width]
    )
    result.uint8image = tf.transpose(depth_major, [1, 2, 0])

    return result
 def input_fn():
     random_sequence = tf.random_uniform([batch_size, sequence_length + 1], 0, 2, dtype=tf.int32, seed=seed)
     labels = tf.slice(random_sequence, [0, 0], [batch_size, sequence_length])
     inputs = tf.expand_dims(
         tf.to_float(tf.slice(random_sequence, [0, 1], [batch_size, sequence_length])), 2
     )
     return {"inputs": inputs}, labels
Ejemplo n.º 16
0
def read_cifar10(filename_queue):
    """
    CIFAR10のデータを読み込む

    Args:
        filename_queue: A queue of strings with the filenames to read from.
    """

    class CIFAR10Record(object):
        pass
    result = CIFAR10Record()

    # CIFAR-10データのサイズ
    label_bytes = 1
    result.height = 32
    result.width = 32
    result.depth = 3
    image_bytes = result.height + result.width * result.depth

    record_bytes = label_bytes + image_bytes

    reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
    result.key, value = reader.read(filename_queue)

    record_bytes = tf.decode_raw(value, tf.uint8)

    result.label = tf.cast(
        tf.slice(record_bytes, [0], [label_bytes]), tf.ini32)

    depth_major = tf.reshape(tf.slice(record_bytes, [label_bytes], [image_bytes]),
                             [result.depth, result.height, result.width])

    result.unit8image = tf.transpose(depth_major, [1, 2, 0])

    return result
Ejemplo n.º 17
0
def read_cifar10(filename_queue):
    """Reads and parses examples from CIFAR10 data files.

    Recommendation: if you want N-way read parallelism, call this function
    N times.  This will give you N independent Readers reading different
    files & positions within those files, which will give better mixing of
    examples.

    Args:
      filename_queue: A queue of strings with the filenames to read from.

    Returns:
      An object representing a single example, with the following fields:
        height: number of rows in the result (32)
        width: number of columns in the result (32)
        depth: number of color channels in the result (3)
        key: a scalar string Tensor describing the filename & record number
          for this example.
        label: an int32 Tensor with the label in the range 0..9.
        uint8image: a [height, width, depth] uint8 Tensor with the image data
    """

    class CIFAR10Record(object):
        pass

    result = CIFAR10Record()

    # Dimensions of the images in the CIFAR-10 dataset.
    # See http://www.cs.toronto.edu/~kriz/cifar.html for a description of the
    # input format.
    label_bytes = 1  # 2 for CIFAR-100
    result.height = 32
    result.width = 32
    result.depth = 3
    image_bytes = result.height * result.width * result.depth
    # Every record consists of a label followed by the image, with a
    # fixed number of bytes for each.
    record_bytes = label_bytes + image_bytes

    # Read a record, getting filenames from the filename_queue.  No
    # header or footer in the CIFAR-10 format, so we leave header_bytes
    # and footer_bytes at their default of 0.
    reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
    result.key, value = reader.read(filename_queue)

    # Convert from a string to a vector of uint8 that is record_bytes long.
    record_bytes = tf.decode_raw(value, tf.uint8)

    # The first bytes represent the label, which we convert from uint8->int32.
    result.label = tf.cast(
        tf.slice(record_bytes, [0], [label_bytes]), tf.int32)

    # The remaining bytes after the label represent the image, which we reshape
    # from [depth * height * width] to [depth, height, width].
    depth_major = tf.reshape(tf.slice(record_bytes, [label_bytes], [image_bytes]),
                             [result.depth, result.height, result.width])
    # Convert from [depth, height, width] to [height, width, depth].
    result.uint8image = tf.transpose(depth_major, [1, 2, 0])

    return result
Ejemplo n.º 18
0
    def _concat_top_nodes(self, old_beam_path, cur_top_nodes):

        cur_path_id = tf.reshape(tf.slice(cur_top_nodes, [0, 0], [-1, 1]), [-1])
        cur_node = tf.reshape(tf.slice(cur_top_nodes, [0, 1], [-1, 1]), [-1,1])
        old_path = tf.gather(old_beam_path, cur_path_id)
        new_beam_path = tf.concat(1, [old_path, cur_node])
        return new_beam_path
Ejemplo n.º 19
0
def slice_constant(data, batch_size=32, name='constant_data', global_step=None):
  """Provide a slice based on the global_step.

  This is useful when the entire data array can be stored in memory because it
  allows you to feed the data very efficiently.

  Args:
    data: A numpy array or tensor.
    batch_size: The batch size for the produced data.
    name: An optional name for this data.
    global_step: A global step variable that is used to read the data. If None
      then the default prettytensor global_step is used.
  Returns:
    A tensor that produces the given data.
  """
  with tf.name_scope(name):
    all_data = tf.convert_to_tensor(data)
    global_step = global_step or bookkeeper.global_step()

    count = len(data) / batch_size
    extra = len(data) - count * batch_size

    if extra:
      offset = tf.mod(global_step, count)
      return tf.slice(all_data, offset * batch_size, batch_size)
    else:
      offset = tf.mod(global_step, count + 1)
      return tf.slice(all_data, offset * batch_size,
                      tf.where(tf.equal(offset, count), extra, batch_size))
Ejemplo n.º 20
0
    def make_minibatch(self, valid_anchors):
        with tf.variable_scope('rpn_minibatch'):

            # in labels(shape is [N, ]): 1 is positive, 0 is negative, -1 is ignored
            labels, anchor_matched_gtboxes, object_mask = \
                self.rpn_find_positive_negative_samples(valid_anchors)  # [num_of_valid_anchors, ]

            positive_indices = tf.reshape(tf.where(tf.equal(labels, 1.0)), [-1])  # use labels is same as object_mask

            num_of_positives = tf.minimum(tf.shape(positive_indices)[0],
                                          tf.cast(self.rpn_mini_batch_size * self.rpn_positives_ratio, tf.int32))

            # num of positives <= minibatch_size * 0.5
            positive_indices = tf.random_shuffle(positive_indices)
            positive_indices = tf.slice(positive_indices, begin=[0], size=[num_of_positives])
            # positive_anchors = tf.gather(self.anchors, positive_indices)

            negative_indices = tf.reshape(tf.where(tf.equal(labels, 0.0)), [-1])
            num_of_negatives = tf.minimum(self.rpn_mini_batch_size - num_of_positives,
                                          tf.shape(negative_indices)[0])

            negative_indices = tf.random_shuffle(negative_indices)
            negative_indices = tf.slice(negative_indices, begin=[0], size=[num_of_negatives])
            # negative_anchors = tf.gather(self.anchors, negative_indices)

            minibatch_indices = tf.concat([positive_indices, negative_indices], axis=0)
            minibatch_indices = tf.random_shuffle(minibatch_indices)

            minibatch_anchor_matched_gtboxes = tf.gather(anchor_matched_gtboxes, minibatch_indices)
            object_mask = tf.gather(object_mask, minibatch_indices)
            labels = tf.cast(tf.gather(labels, minibatch_indices), tf.int32)
            labels_one_hot = tf.one_hot(labels, depth=2)
            return minibatch_indices, minibatch_anchor_matched_gtboxes, object_mask, labels_one_hot
Ejemplo n.º 21
0
def diagonal_bilstm(inputs, conf, scope='diagonal_bilstm'):
  with tf.variable_scope(scope):
    def reverse(inputs):
      return tf.reverse(inputs, [False, False, True, False])

    output_state_fw = diagonal_lstm(inputs, conf, scope='output_state_fw')
    output_state_bw = reverse(diagonal_lstm(reverse(inputs), conf, scope='output_state_bw'))

    tf.add_to_collection('output_state_fw', output_state_fw)
    tf.add_to_collection('output_state_bw', output_state_bw)

    if conf.use_residual:
      residual_state_fw = conv2d(output_state_fw, conf.hidden_dims * 2, [1, 1], "B", scope="residual_fw")
      output_state_fw = residual_state_fw + inputs

      residual_state_bw = conv2d(output_state_bw, conf.hidden_dims * 2, [1, 1], "B", scope="residual_bw")
      output_state_bw = residual_state_bw + inputs

      tf.add_to_collection('residual_state_fw', residual_state_fw)
      tf.add_to_collection('residual_state_bw', residual_state_bw)
      tf.add_to_collection('residual_output_state_fw', output_state_fw)
      tf.add_to_collection('residual_output_state_bw', output_state_bw)

    batch, height, width, channel = get_shape(output_state_bw)

    output_state_bw_except_last = tf.slice(output_state_bw, [0, 0, 0, 0], [-1, height-1, -1, -1])
    output_state_bw_only_last = tf.slice(output_state_bw, [0, height-1, 0, 0], [-1, 1, -1, -1])
    dummy_zeros = tf.zeros_like(output_state_bw_only_last)

    output_state_bw_with_last_zeros = tf.concat(1, [output_state_bw_except_last, dummy_zeros])

    tf.add_to_collection('output_state_bw_with_last_zeros', output_state_bw_with_last_zeros)

    return output_state_fw + output_state_bw_with_last_zeros
Ejemplo n.º 22
0
def process(_, current):
    count = tf.cast(current[0], tf.int32)
    current = tf.slice(current, [1], [-1])
    max = tf.shape(current)[0]
    sm = tf.expand_dims(tf.slice(current, [max - count], [-1]), 0)
    sm = tf.nn.softmax(sm)
    return tf.concat(0, [tf.zeros([max-count]), tf.squeeze(sm, [0])])
Ejemplo n.º 23
0
def get_model(point_cloud, is_training, bn_decay=None):
    """ Part segmentation PointNet, input is BxNx6 (XYZ NormalX NormalY NormalZ), output Bx50 """
    batch_size = point_cloud.get_shape()[0].value
    num_point = point_cloud.get_shape()[1].value
    end_points = {}
    l0_xyz = tf.slice(point_cloud, [0,0,0], [-1,-1,3])
    l0_points = tf.slice(point_cloud, [0,0,3], [-1,-1,3])

    # Set Abstraction layers
    l1_xyz, l1_points, l1_indices = pointnet_sa_module(l0_xyz, l0_points, npoint=512, radius=0.2, nsample=64, mlp=[64,64,128], mlp2=None, group_all=False, is_training=is_training, bn_decay=bn_decay, scope='layer1')
    l2_xyz, l2_points, l2_indices = pointnet_sa_module(l1_xyz, l1_points, npoint=128, radius=0.4, nsample=64, mlp=[128,128,256], mlp2=None, group_all=False, is_training=is_training, bn_decay=bn_decay, scope='layer2')
    l3_xyz, l3_points, l3_indices = pointnet_sa_module(l2_xyz, l2_points, npoint=None, radius=None, nsample=None, mlp=[256,512,1024], mlp2=None, group_all=True, is_training=is_training, bn_decay=bn_decay, scope='layer3')

    # Feature Propagation layers
    l2_points = pointnet_fp_module(l2_xyz, l3_xyz, l2_points, l3_points, [256,256], is_training, bn_decay, scope='fa_layer1')
    l1_points = pointnet_fp_module(l1_xyz, l2_xyz, l1_points, l2_points, [256,128], is_training, bn_decay, scope='fa_layer2')
    l0_points = pointnet_fp_module(l0_xyz, l1_xyz, tf.concat([l0_xyz,l0_points],axis=-1), l1_points, [128,128,128], is_training, bn_decay, scope='fa_layer3')

    # FC layers
    net = tf_util.conv1d(l0_points, 128, 1, padding='VALID', bn=True, is_training=is_training, scope='fc1', bn_decay=bn_decay)
    end_points['feats'] = net 
    net = tf_util.dropout(net, keep_prob=0.5, is_training=is_training, scope='dp1')
    net = tf_util.conv1d(net, 50, 1, padding='VALID', activation_fn=None, scope='fc2')

    return net, end_points
def read_cifar10(filename_queue):
    result = CIFAR10Record()
    label_bytes = 1 # 2 for CIFAR-100
    result.height = 32
    result.width = 32
    result.depth = 3
    image_bytes = result.height * result.width * result.depth
    record_bytes = label_bytes + image_bytes
    
    # 固定長データのReader
    reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
    result.key, value = reader.read(filename_queue)

    # valueをdecode_rowでデコード
    record_bytes = tf.decode_raw(value, tf.uint8)

    # labelデータ(最初の1byteをスライス)
    result.label = tf.cast(tf.slice(record_bytes, [0], [label_bytes]), tf.int32)

    # 画像データ
    depth_major = tf.reshape(tf.slice(record_bytes, [label_bytes], [image_bytes]), [result.depth, result.height, result.width])

    # transposeで[1,2,0]の順に並べる
    result.unit8image = tf.transpose(depth_major, [1,2,0])

    return result
Ejemplo n.º 25
0
Archivo: nn.py Proyecto: febert/DeepRL
        def make_input_output_histogramm(inp, labels):
            x_slice = tf.slice(input_= inp, begin= [0,0], size= [-1,1])
            v_slice = tf.slice(input_= inp, begin= [0,1], size= [-1,1])

            tf.histogram_summary('x_input', x_slice)
            tf.histogram_summary('v_input', v_slice)
            tf.histogram_summary('labels hist', labels)
Ejemplo n.º 26
0
	def __call__(self, inputs, state, scope=None):
		'''Runs vanilla LSTM Cell and applies zoneout.
		'''
		#Apply vanilla LSTM
		output, new_state = self._cell(inputs, state, scope)

		if self.state_is_tuple:
			(prev_c, prev_h) = state
			(new_c, new_h) = new_state
		else:
			num_proj = self._cell._num_units if self._cell._num_proj is None else self._cell._num_proj
			prev_c = tf.slice(state, [0, 0], [-1, self._cell._num_units])
			prev_h = tf.slice(state, [0, self._cell._num_units], [-1, num_proj])
			new_c = tf.slice(new_state, [0, 0], [-1, self._cell._num_units])
			new_h = tf.slice(new_state, [0, self._cell._num_units], [-1, num_proj])

		#Apply zoneout
		if self.is_training:
			#nn.dropout takes keep_prob (probability to keep activations) not drop_prob (probability to mask activations)!
			c = (1 - self._zoneout_cell) * tf.nn.dropout(new_c - prev_c, (1 - self._zoneout_cell)) + prev_c
			h = (1 - self._zoneout_outputs) * tf.nn.dropout(new_h - prev_h, (1 - self._zoneout_outputs)) + prev_h

		else:
			c = (1 - self._zoneout_cell) * new_c + self._zoneout_cell * prev_c
			h = (1 - self._zoneout_outputs) * new_h + self._zoneout_outputs * prev_h

		new_state = tf.nn.rnn_cell.LSTMStateTuple(c, h) if self.state_is_tuple else tf.concat(1, [c, h])

		return output, new_state
Ejemplo n.º 27
0
 def compute_first_or_last(self, select, first=True):
   #perform first ot last operation on row select with probabilistic row selection
   answer = tf.zeros_like(select)
   running_sum = tf.zeros([self.batch_size, 1], self.data_type)
   for i in range(self.max_elements):
     if (first):
       current = tf.slice(select, [0, i], [self.batch_size, 1])
     else:
       current = tf.slice(select, [0, self.max_elements - 1 - i],
                          [self.batch_size, 1])
     curr_prob = current * (1 - running_sum)
     curr_prob = curr_prob * tf.cast(curr_prob >= 0.0, self.data_type)
     running_sum += curr_prob
     temp_ans = []
     curr_prob = tf.expand_dims(tf.reshape(curr_prob, [self.batch_size]), 0)
     for i_ans in range(self.max_elements):
       if (not (first) and i_ans == self.max_elements - 1 - i):
         temp_ans.append(curr_prob)
       elif (first and i_ans == i):
         temp_ans.append(curr_prob)
       else:
         temp_ans.append(tf.zeros_like(curr_prob))
     temp_ans = tf.transpose(tf.concat(axis=0, values=temp_ans))
     answer += temp_ans
   return answer
def tf_format_mnist_images(X, Y, Y_, n=100, lines=10):
    correct_prediction = tf.equal(tf.argmax(Y,1), tf.argmax(Y_,1))
    correctly_recognised_indices = tf.squeeze(tf.where(correct_prediction), [1])  # indices of correctly recognised images
    incorrectly_recognised_indices = tf.squeeze(tf.where(tf.logical_not(correct_prediction)), [1]) # indices of incorrectly recognised images
    everything_incorrect_first = tf.concat([incorrectly_recognised_indices, correctly_recognised_indices], 0) # images reordered with indeces of unrecognised images first
    everything_incorrect_first = tf.slice(everything_incorrect_first, [0], [n]) # compute first 100 only - no space to display more anyway
    # compute n=100 digits to display only
    Xs = tf.gather(X, everything_incorrect_first)
    Ys = tf.gather(Y, everything_incorrect_first)
    Ys_ = tf.gather(Y_, everything_incorrect_first)
    correct_prediction_s = tf.gather(correct_prediction, everything_incorrect_first)

    digits_left = tf.image.grayscale_to_rgb(tensorflowvisu_digits.digits_left())
    correct_tags = tf.gather(digits_left, tf.argmax(Ys_, 1)) # correct digits to be printed on the images
    digits_right = tf.image.grayscale_to_rgb(tensorflowvisu_digits.digits_right())
    computed_tags = tf.gather(digits_right, tf.argmax(Ys, 1)) # computed digits to be printed on the images
    #superimposed_digits = correct_tags+computed_tags
    superimposed_digits = tf.where(correct_prediction_s, tf.zeros_like(correct_tags),correct_tags+computed_tags) # only pring the correct and computed digits on unrecognised images
    correct_bkg   = tf.reshape(tf.tile([1.3,1.3,1.3], [28*28]), [1, 28,28,3]) # white background
    incorrect_bkg = tf.reshape(tf.tile([1.3,1.0,1.0], [28*28]), [1, 28,28,3]) # red background
    recognised_bkg = tf.gather(tf.concat([incorrect_bkg, correct_bkg], 0), tf.cast(correct_prediction_s, tf.int32)) # pick either the red or the white background depending on recognised status

    I = tf.image.grayscale_to_rgb(Xs)
    I = ((1-(I+superimposed_digits))*recognised_bkg)/1.3 # stencil extra data on top of images and reorder them unrecognised first
    I = tf.image.convert_image_dtype(I, tf.uint8, saturate=True)
    Islices = [] # 100 images => 10x10 image block
    for imslice in range(lines):
        Islices.append(tf.concat(tf.unstack(tf.slice(I, [imslice*n//lines,0,0,0], [n//lines,28,28,3])), 1))
    I = tf.concat(Islices, 0)
    return I
Ejemplo n.º 29
0
    def _transform(theta, input_dim, out_size):
        num_batch = tf.shape(input=input_dim)[0]
        num_channels = tf.shape(input=input_dim)[3]
        theta = tf.reshape(theta, (-1, 2, 3))
        theta = tf.cast(theta, 'float32')

        # grid of (x_t, y_t, 1), eq (1) in ref [1]
        out_height = out_size[0]
        out_width = out_size[1]
        grid = _meshgrid(out_height, out_width)
        grid = tf.expand_dims(grid, 0)
        grid = tf.reshape(grid, [-1])
        grid = tf.tile(grid, tf.stack([num_batch]))
        grid = tf.reshape(grid, tf.stack([num_batch, 3, -1]))

        # Transform A x (x_t, y_t, 1)^T -> (x_s, y_s)
        T_g = tf.matmul(theta, grid)
        x_s = tf.slice(T_g, [0, 0, 0], [-1, 1, -1])
        y_s = tf.slice(T_g, [0, 1, 0], [-1, 1, -1])
        x_s_flat = tf.reshape(x_s, [-1])
        y_s_flat = tf.reshape(y_s, [-1])

        input_transformed = _interpolate(input_dim, x_s_flat, y_s_flat, out_size)

        output = tf.reshape(input_transformed, tf.stack([num_batch, out_height, out_width, num_channels]))
        return output
Ejemplo n.º 30
0
def ApplyPcaAndWhitening(data,
                         pca_matrix,
                         pca_mean,
                         output_dim,
                         use_whitening=False,
                         pca_variances=None):
  """Applies PCA/whitening to data.

  Args:
    data: [N, dim] float tensor containing data which undergoes PCA/whitening.
    pca_matrix: [dim, dim] float tensor PCA matrix, row-major.
    pca_mean: [dim] float tensor, mean to subtract before projection.
    output_dim: Number of dimensions to use in output data, of type int.
    use_whitening: Whether whitening is to be used.
    pca_variances: [dim] float tensor containing PCA variances. Only used if
      use_whitening is True.

  Returns:
    output: [N, output_dim] float tensor with output of PCA/whitening operation.
  """
  output = tf.matmul(
      tf.subtract(data, pca_mean),
      tf.slice(pca_matrix, [0, 0], [output_dim, -1]),
      transpose_b=True,
      name='pca_matmul')

  # Apply whitening if desired.
  if use_whitening:
    output = tf.divide(
        output,
        tf.sqrt(tf.slice(pca_variances, [0], [output_dim])),
        name='whitening')

  return output
Ejemplo n.º 31
0
    def _compute_loss(self, input_dict):
        """
    Computes cross entropy based sequence-to-sequence loss
    :param input_dict: inputs to compute loss
    {
          "logits": logits tensor of shape [batch_size, T, dim]
          "target_sequence": tensor of shape [batch_size, T]
          "tgt_lengths": tensor of shape [batch_size] or None
    }
    :return: Singleton loss tensor
    """
        logits = input_dict["decoder_output"]["logits"]
        #target_sequence = input_dict["tgt_sequence"]
        #tgt_lengths = input_dict["tgt_length"]
        target_sequence = input_dict['target_tensors'][0]
        tgt_lengths = input_dict['target_tensors'][1]

        if self._offset_target_by_one:
            # this is necessary for auto-regressive models
            current_ts = tf.to_int32(
                tf.minimum(
                    tf.shape(target_sequence)[1],
                    tf.shape(logits)[1],
                )) - 1

            logits = tf.slice(
                logits,
                begin=[0, 0, 0],
                size=[-1, current_ts, -1],
            )
            target_sequence = tf.slice(target_sequence,
                                       begin=[0, 1],
                                       size=[-1, current_ts])
        else:
            current_ts = tf.to_int32(
                tf.minimum(
                    tf.shape(target_sequence)[1],
                    tf.shape(logits)[1],
                ))

        # Cast logits after potential slice
        if logits.dtype.base_dtype != tf.float32:
            logits = tf.cast(logits, tf.float32)

        if self._do_mask:
            if tgt_lengths is None:
                raise ValueError(
                    "If you are masking loss, tgt_lengths can't be None")
            mask = tf.sequence_mask(
                lengths=tgt_lengths - 1, maxlen=current_ts,
                dtype=logits.dtype)  # TODO: why store in float?
        else:
            mask = tf.cast(tf.ones_like(target_sequence), logits.dtype)
        """
    if self._average_across_timestep:
      loss = tf.contrib.seq2seq.sequence_loss(
        logits=logits,
        targets=target_sequence,
        weights=mask,
        average_across_timesteps=self._average_across_timestep,
        average_across_batch=True,
        softmax_loss_function=tf.nn.sparse_softmax_cross_entropy_with_logits,
      )
    else:
      crossent = tf.nn.sparse_softmax_cross_entropy_with_logits(
        labels=tf.reshape(target_sequence, shape=[-1]),
        logits=tf.reshape(logits, shape=[-1, self._tgt_vocab_size]),
      )
      loss = tf.reduce_sum(crossent * tf.reshape(mask, shape=[-1]))      
      loss /= self._batch_size_per_gpu
    """
        crossent = tf.nn.sparse_softmax_cross_entropy_with_logits(
            labels=tf.reshape(target_sequence, shape=[-1]),
            logits=tf.reshape(logits, shape=[-1, self._tgt_vocab_size]),
        )
        if self._average_across_timestep:
            loss = tf.reduce_mean(crossent * tf.reshape(mask, shape=[-1]))
        else:
            loss = tf.reduce_sum(crossent * tf.reshape(mask, shape=[-1]))
            loss /= self._batch_size
        return loss
Ejemplo n.º 32
0
def embedding_postprocessor(input_tensor,
                            use_token_type=False,
                            token_type_ids=None,
                            token_type_vocab_size=16,
                            token_type_embedding_name="token_type_embeddings",
                            use_position_embeddings=True,
                            position_embedding_name="position_embeddings",
                            initializer_range=0.02,
                            max_position_embeddings=512,
                            dropout_prob=0.1):
    """Performs various post-processing on a word embedding tensor.

  Args:
    input_tensor: float Tensor of shape [batch_size, seq_length,
      embedding_size].
    use_token_type: bool. Whether to add embeddings for `token_type_ids`.
    token_type_ids: (optional) int32 Tensor of shape [batch_size, seq_length].
      Must be specified if `use_token_type` is True.
    token_type_vocab_size: int. The vocabulary size of `token_type_ids`.
    token_type_embedding_name: string. The name of the embedding table variable
      for token type ids.
    use_position_embeddings: bool. Whether to add position embeddings for the
      position of each token in the sequence.
    position_embedding_name: string. The name of the embedding table variable
      for positional embeddings.
    initializer_range: float. Range of the weight initialization.
    max_position_embeddings: int. Maximum sequence length that might ever be
      used with this model. This can be longer than the sequence length of
      input_tensor, but cannot be shorter.
    dropout_prob: float. Dropout probability applied to the final output tensor.

  Returns:
    float tensor with same shape as `input_tensor`.

  Raises:
    ValueError: One of the tensor shapes or input values is invalid.
  """
    input_shape = get_shape_list(input_tensor, expected_rank=3)
    batch_size = input_shape[0]
    seq_length = input_shape[1]
    width = input_shape[2]

    output = input_tensor

    if use_token_type:
        if token_type_ids is None:
            raise ValueError("`token_type_ids` must be specified if"
                             "`use_token_type` is True.")
        token_type_table = tf.get_variable(
            name=token_type_embedding_name,
            shape=[token_type_vocab_size, width],
            initializer=create_initializer(initializer_range))
        # This vocab will be small so we always do one-hot here, since it is always
        # faster for a small vocabulary.
        flat_token_type_ids = tf.reshape(token_type_ids, [-1])
        one_hot_ids = tf.one_hot(flat_token_type_ids,
                                 depth=token_type_vocab_size)
        token_type_embeddings = tf.matmul(one_hot_ids, token_type_table)
        token_type_embeddings = tf.reshape(token_type_embeddings,
                                           [batch_size, seq_length, width])
        output += token_type_embeddings

    if use_position_embeddings:
        assert_op = tf.assert_less_equal(seq_length, max_position_embeddings)
        with tf.control_dependencies([assert_op]):
            full_position_embeddings = tf.get_variable(
                name=position_embedding_name,
                shape=[max_position_embeddings, width],
                initializer=create_initializer(initializer_range))
            # Since the position embedding table is a learned variable, we create it
            # using a (long) sequence length `max_position_embeddings`. The actual
            # sequence length might be shorter than this, for faster training of
            # tasks that do not have long sequences.
            #
            # So `full_position_embeddings` is effectively an embedding table
            # for position [0, 1, 2, ..., max_position_embeddings-1], and the current
            # sequence has positions [0, 1, 2, ... seq_length-1], so we can just
            # perform a slice.
            position_embeddings = tf.slice(full_position_embeddings, [0, 0],
                                           [seq_length, -1])
            num_dims = len(output.shape.as_list())

            # Only the last two dimensions are relevant (`seq_length` and `width`), so
            # we broadcast among the first dimensions, which is typically just
            # the batch size.
            position_broadcast_shape = []
            for _ in range(num_dims - 2):
                position_broadcast_shape.append(1)
            position_broadcast_shape.extend([seq_length, width])
            position_embeddings = tf.reshape(position_embeddings,
                                             position_broadcast_shape)
            output += position_embeddings

    output = layer_norm_and_dropout(output, dropout_prob)
    return output
def winograd_conv(image, kernel, transformed_image, output_channel, mask=None):
        batch_size_d, image_width_d, image_height_d, input_channel_d = image.get_shape()
        batch_size = batch_size_d.value
        image_width = image_width_d.value
        image_height = image_height_d.value
        input_channel = input_channel_d.value
        n_patch_width = (image_width + 1 + 2 - 4) / 2 + 1
        n_patch_height = (image_height + 1 + 2 - 4) / 2 + 1
        if batch_size is None:
                batch_size = -1

        CTdC_0 = tf.reshape(tf.strided_slice(transformed_image, [0, 0, 0], [1, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_1 = tf.reshape(tf.strided_slice(transformed_image, [1, 0, 0], [2, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_2 = tf.reshape(tf.strided_slice(transformed_image, [2, 0, 0], [3, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_3 = tf.reshape(tf.strided_slice(transformed_image, [3, 0, 0], [4, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_4 = tf.reshape(tf.strided_slice(transformed_image, [4, 0, 0], [5, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_5 = tf.reshape(tf.strided_slice(transformed_image, [5, 0, 0], [6, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_6 = tf.reshape(tf.strided_slice(transformed_image, [6, 0, 0], [7, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_7 = tf.reshape(tf.strided_slice(transformed_image, [7, 0, 0], [8, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_8 = tf.reshape(tf.strided_slice(transformed_image, [8, 0, 0], [9, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_9 = tf.reshape(tf.strided_slice(transformed_image, [9, 0, 0], [10, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_10 = tf.reshape(tf.strided_slice(transformed_image, [10, 0, 0], [11, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_11 = tf.reshape(tf.strided_slice(transformed_image, [11, 0, 0], [12, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_12 = tf.reshape(tf.strided_slice(transformed_image, [12, 0, 0], [13, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_13 = tf.reshape(tf.strided_slice(transformed_image, [13, 0, 0], [14, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_14 = tf.reshape(tf.strided_slice(transformed_image, [14, 0, 0], [15, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])
        CTdC_15 = tf.reshape(tf.strided_slice(transformed_image, [15, 0, 0], [16, 0, 0], [1, 1, 1], end_mask=6), [-1, input_channel])

        GgGT = kernel # (16, input_channel, output_channel)

        GgGT_0 = tf.reshape(tf.slice(GgGT, [0, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_1 = tf.reshape(tf.slice(GgGT, [1, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_2 = tf.reshape(tf.slice(GgGT, [2, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_3 = tf.reshape(tf.slice(GgGT, [3, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_4 = tf.reshape(tf.slice(GgGT, [4, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_5 = tf.reshape(tf.slice(GgGT, [5, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_6 = tf.reshape(tf.slice(GgGT, [6, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_7 = tf.reshape(tf.slice(GgGT, [7, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_8 = tf.reshape(tf.slice(GgGT, [8, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_9 = tf.reshape(tf.slice(GgGT, [9, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_10 = tf.reshape(tf.slice(GgGT, [10, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_11 = tf.reshape(tf.slice(GgGT, [11, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_12 = tf.reshape(tf.slice(GgGT, [12, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_13 = tf.reshape(tf.slice(GgGT, [13, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_14 = tf.reshape(tf.slice(GgGT, [14, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])
        GgGT_15 = tf.reshape(tf.slice(GgGT, [15, 0, 0], [1, input_channel, output_channel]), [input_channel, output_channel])

        prod_0 = tf.matmul(CTdC_0, GgGT_0)
        prod_1 = tf.matmul(CTdC_1, GgGT_1)
        prod_2 = tf.matmul(CTdC_2, GgGT_2)
        prod_3 = tf.matmul(CTdC_3, GgGT_3)
        prod_4 = tf.matmul(CTdC_4, GgGT_4)
        prod_5 = tf.matmul(CTdC_5, GgGT_5)
        prod_6 = tf.matmul(CTdC_6, GgGT_6)
        prod_7 = tf.matmul(CTdC_7, GgGT_7)
        prod_8 = tf.matmul(CTdC_8, GgGT_8)
        prod_9 = tf.matmul(CTdC_9, GgGT_9)
        prod_10 = tf.matmul(CTdC_10, GgGT_10)
        prod_11 = tf.matmul(CTdC_11, GgGT_11)
        prod_12 = tf.matmul(CTdC_12, GgGT_12)
        prod_13 = tf.matmul(CTdC_13, GgGT_13)
        prod_14 = tf.matmul(CTdC_14, GgGT_14)
        prod_15 = tf.matmul(CTdC_15, GgGT_15)

        ATprodA_0 = prod_0 + prod_1 + prod_2 + prod_4 + prod_5 + prod_6 + prod_8 + prod_9 + prod_10
        ATprodA_1 = prod_1 - prod_2 - prod_3 + prod_5 - prod_6 - prod_7 + prod_9 - prod_10 - prod_11
        ATprodA_2 = prod_4 + prod_5 + prod_6 - prod_8 - prod_9 - prod_10 - prod_12 - prod_13 - prod_14
        ATprodA_3 = prod_5 - prod_6 - prod_7 - prod_9 + prod_10 + prod_11 - prod_13 + prod_14 + prod_15
        ATprodA = tf.concat([ ATprodA_0, ATprodA_1, ATprodA_2, ATprodA_3 ], 0)
        ATprodA = tf.reshape(ATprodA, [2, 2, batch_size, n_patch_width, n_patch_height, output_channel])
        ATprodA = tf.transpose(ATprodA, perm = [3, 0, 4, 1, 5, 2])
        Image = tf.reshape(ATprodA, [2 * n_patch_width, 2 * n_patch_height, -1])
        if image_width % 2 is not 0 or image_height % 2 is not 0:
                Image = tf.image.crop_to_bounding_box(Image, 0, 0, image_width, image_height)
        Image = tf.reshape(Image, [image_width, image_height, output_channel, batch_size])
        Image = tf.transpose(Image, perm=[3, 0, 1, 2])
        return Image
Ejemplo n.º 34
0
def model_fn(features, labels, mode, params):
    """model_fn constructs the ML model used to predict handwritten digits."""

    del params

    targets = labels
    targets = ((1.0 - 0.1) * targets) + (1.0 / targets.shape[1].value
                                         )  # label smoothing

    print('inputs: {}'.format(features))

    print('features shape: {}'.format(features.shape))
    e1_idx = tf.slice(features, [0, 0], [features.shape[0].value, 1])
    print('e1_idx: {}'.format(e1_idx))
    r_idx = tf.slice(features, [0, 1], [features.shape[0].value, 1])
    labels = tf.slice(features, [0, 2], [features.shape[0].value, 1])

    data = Data(dataset='WN18', reverse=False)
    model = HyperER(len(data.entities), len(data.relations))

    if mode == tf.estimator.ModeKeys.PREDICT:

        logits = model(e1_idx, r_idx, training=False)
        predictions = {
            'class_ids': tf.argmax(logits, axis=1),
            'probabilities': tf.sigmoid(logits),
        }

        return tf.contrib.tpu.TPUEstimatorSpec(mode, predictions=predictions)

    logits = model(e1_idx,
                   r_idx,
                   training=(mode == tf.estimator.ModeKeys.TRAIN))
    predictions = tf.sigmoid(logits)

    print('number of inputs: {}'.format(features.shape[0].value))

    loss = tf.keras.losses.binary_crossentropy(
        tf.cast(targets, tf.float32), tf.cast(predictions, tf.float32))

    loss = tf.reduce_mean(loss)
    global_step = tf.train.get_or_create_global_step()

    if mode == tf.estimator.ModeKeys.TRAIN:

        learning_rate = tf.train.exponential_decay(FLAGS.learning_rate,
                                                   global_step,
                                                   decay_steps=1337,
                                                   decay_rate=0.99,
                                                   staircase=True)

        optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)

        if FLAGS.use_tpu:
            optimizer = tf.contrib.tpu.CrossShardOptimizer(optimizer)

        # Batch normalization requires UPDATE_OPS to be added as a dependency to
        # the train operation.
        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
        with tf.control_dependencies(update_ops):
            train_op = optimizer.minimize(loss, global_step=global_step)

        return tf.contrib.tpu.TPUEstimatorSpec(mode=mode,
                                               loss=loss,
                                               train_op=train_op)

    if mode == tf.estimator.ModeKeys.EVAL:
        return tf.contrib.tpu.TPUEstimatorSpec(
            mode=mode,
            loss=loss,
            eval_metrics=(metric_fn, [features, labels, logits]))
Ejemplo n.º 35
0
    def __init__(self, buckets, isTraining, max_gradient_norm, batch_size,
                 learning_rate, learning_rate_decay_factor, encoder_attribs,
                 decoder_attribs):
        """Initializer of class that defines the computational graph.

        Args:
            buckets: List of input-output sizes that limit the amount of
                sequence padding (http://goo.gl/d8ybpl).
            isTraining: boolean that denotes training v/s evaluation.
            max_gradient_norm: Maximum value of gradient norm.
            batch_size: Minibatch size used for doing SGD.
            learning_rate: Initial learning rate of optimizer
            learning_rate_decay_factor: Multiplicative learning rate decay
                factor
            {encoder, decoder}_attribs: Dictionary containing attributes for
                {encoder, decoder} RNN.
        """
        self.buckets = buckets
        self.isTraining = isTraining
        self.batch_size = batch_size

        self.learning_rate = tf.Variable(float(learning_rate), trainable=False)
        self.learning_rate_decay_op = self.learning_rate.assign(
            self.learning_rate * learning_rate_decay_factor)

        # Number of gradient updates performed
        self.global_step = tf.Variable(0, trainable=False)
        # Number of epochs done
        self.epoch = tf.Variable(0, trainable=False)
        self.epoch_incr = self.epoch.assign(self.epoch + 1)

        self.encoder_attribs = encoder_attribs
        self.decoder_attribs = decoder_attribs
        # Placeholder for encoder input IDs
        self.encoder_inputs = {}
        for feat_type in self.encoder_attribs.feat_types:
            if feat_type == "speech_frames":
                self.encoder_inputs[feat_type] = tf.placeholder(
                    tf.float32,
                    # T * B * num_frame_per_word * frame_dimension
                    shape=[
                        None, None, self.encoder_attribs.fixed_word_length,
                        self.encoder_attribs.feat_dim
                    ],
                    name=feat_type + '_encoder')
            elif feat_type == "word_dur":
                self.encoder_inputs[feat_type] = tf.placeholder(
                    tf.float32,
                    shape=[None, None],
                    name=feat_type + '_encoder')
            else:
                self.encoder_inputs[feat_type] = tf.placeholder(
                    tf.int32, shape=[None, None], name=feat_type + '_encoder')

        _batch_size = self.encoder_inputs["word"].get_shape()[1].value
        # Input sequence length placeholder
        self.seq_len = tf.placeholder(tf.int32,
                                      shape=[_batch_size],
                                      name="seq_len")
        # Output sequence length placeholder
        self.seq_len_target = tf.placeholder(tf.int32,
                                             shape=[_batch_size],
                                             name="seq_len_target")

        # Input to decoder RNN. This input has an initial extra symbol - GO -
        # that initiates the decoding process.
        self.decoder_inputs = tf.placeholder(tf.int32,
                                             shape=[None, None],
                                             name="decoder")
        # Targets are decoder inputs shifted by one thus, ignoring GO symbol
        self.targets = tf.slice(self.decoder_inputs, [1, 0], [-1, -1])

        # Initialize the encoder and decoder RNNs
        self.encoder = Encoder(isTraining, encoder_attribs)
        self.decoder = Decoder(isTraining, decoder_attribs)

        # First encode input
        self.encoder_hidden_states, self.final_state = \
            self.encoder.encode_input(self.encoder_inputs, self.seq_len)
        # Then decode
        self.outputs = \
            self.decoder.decode(self.decoder_inputs, self.seq_len_target,
                                self.encoder_hidden_states, self.final_state,
                                self.seq_len)
        if isTraining:
            # Training outputs and losses.
            self.losses = self.seq2seq_loss(self.outputs, self.targets,
                                            self.seq_len_target)
            # Gradients and parameter updation for training the model.
            params = tf.trainable_variables()
            print("\nModel parameters:\n")
            for var in params:
                print(("{0}: {1}").format(var.name, var.get_shape()))
            print
            # Initialize optimizer
            opt = tf.train.AdamOptimizer(self.learning_rate)
            # Get gradients from loss
            gradients = tf.gradients(self.losses, params)
            # Clip the gradients to avoid the problem of gradient explosion
            # possible early in training
            clipped_gradients, norm = tf.clip_by_global_norm(
                gradients, max_gradient_norm)
            self.gradient_norms = norm
            # Apply gradients
            self.updates = opt.apply_gradients(zip(clipped_gradients, params),
                                               global_step=self.global_step)

        # Model saver function
        self.saver = tf.train.Saver(tf.global_variables(), max_to_keep=2)
        self.best_saver = tf.train.Saver(tf.global_variables(), max_to_keep=2)
# rows
x = tf.placeholder(dtype='float', shape=[None, 3])
y = x * 2

with tf.Session() as session:
    x_data = [[1, 2, 3], [4, 5, 6]]
    result = session.run(y, feed_dict={x: x_data})
    print(result)

import matplotlib.image as mpimg
import matplotlib.pyplot as plt

raw_image_data = mpimg.imread('MarshOrchid.jpg')

image = tf.placeholder(dtype='uint8', shape=[None, None, 3])
sliced = tf.slice(image, begin=[1000, 0, 0], size=[3000, -1, -1])

# 注意这里没有global_variables_initializer(),因为没有Variable
with tf.Session() as session:
    result = session.run(sliced, feed_dict={image: raw_image_data})

plt.imshow(result)

# =============================================================================
# Exercise
# =============================================================================

# 1) Take a look at the other functions for arrays in TensorFlow at the official
#    documentation.

# 2) Break the image apart into four “corners”, then stitch it back together
Ejemplo n.º 37
0
def create_model (inputs, backbone_fn):
    #box_ft, mask_ft, gt_masks, gt_anchors, gt_anchors_weight, gt_params, gt_params_weight, gt_boxes, config):
    # ft:           B * H' * W' * 3     input feature, H' W' is feature map size
    # gt_counts:    B                   number of boxes in each sample of the batch
    # gt_boxes:     ? * 4               boxes
    bb, _ = backbone_fn(inputs.X-PIXEL_MEANS, global_pool=False, output_stride=FLAGS.backbone_stride)
    #bb2, _ = backbone_fn(inputs.X-PIXEL_MEANS, global_pool=False, output_stride=FLAGS.backbone_stride, scope='bb2')

    gt_matcher = cpp.GTMatcher(FLAGS.match_th, FLAGS.max_masks)
    mask_extractor = cpp.MaskExtractor(FLAGS.mask_size, FLAGS.mask_size)

    end_points = {}

    with tf.variable_scope('boxnet'):

        assert FLAGS.backbone_stride % FLAGS.anchor_stride == 0
        ss = FLAGS.backbone_stride // FLAGS.anchor_stride
        # generate anchor feature
        anchor_logits_ft = slim.conv2d_transpose(bb, FLAGS.anchor_logit_filters, ss*2, ss)
        anchor_params_ft = slim.conv2d_transpose(bb, FLAGS.anchor_params_filters, ss*2, ss)

        assert FLAGS.backbone_stride % FLAGS.mask_stride == 0
        ss = FLAGS.backbone_stride // FLAGS.mask_stride

        mask_ft = slim.conv2d_transpose(bb, FLAGS.mask_filters, ss*2, ss)

        anchor_logits = slim.conv2d(anchor_logits_ft, 2 * len(PRIORS), 3, 1, activation_fn=None) 
        anchor_logits2 = tf.reshape(anchor_logits, (-1, 2))   # ? * 2
        # anchor probabilities
        anchor_prob = tf.squeeze(tf.slice(tf.nn.softmax(anchor_logits2), [0, 1], [-1, 1]), 1)

        gt_anchors = tf.reshape(inputs.gt_anchors, (-1, ))
        gt_anchors_weight = tf.reshape(inputs.gt_anchors_weight, (-1,))

        # anchor cross-entropy
        axe = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=anchor_logits2, labels=gt_anchors)
        axe = axe * gt_anchors_weight
        axe = tf.reduce_sum(axe) / (tf.reduce_sum(gt_anchors_weight) + 1)

        params = slim.conv2d(anchor_params_ft, 4 * len(PRIORS), 3, 1, activation_fn=None)
        params = tf.reshape(params, (-1, 4))     # ? * 4
        gt_params = tf.reshape(inputs.gt_params, (-1, 4))
        gt_params_weight = tf.reshape(inputs.gt_params_weight, (-1,))
        # params loss
        if True:
            dxy, wh = tf.split(params, [2,2], 1)
            dxy_gt, wh_gt = tf.split(gt_params, [2,2], 1)

            #wh = tf.log(tf.nn.relu(wh) + 1)
            wh_gt = tf.log(wh_gt + 1)

            pl = tf.losses.huber_loss(dxy, dxy_gt, reduction=tf.losses.Reduction.NONE) + \
                 tf.losses.huber_loss(wh, wh_gt, reduction=tf.losses.Reduction.NONE)
            pl = tf.reduce_sum(pl, axis=1)

        pl = tf.reduce_sum(pl * gt_params_weight) / (tf.reduce_sum(gt_params_weight) + 1)

        # generate boxes from anchor params
        boxes, box_ind = anchors2boxes(tf.shape(anchor_logits_ft), params)
        boxes_pre = boxes

        sel = tf.greater_equal(anchor_prob, inputs.anchor_th)
        # sel is a boolean mask

        # select only boxes with prob > th for nms
        anchor_prob = tf.boolean_mask(anchor_prob, sel)
        boxes = tf.boolean_mask(boxes, sel)
        box_ind = tf.boolean_mask(box_ind, sel)

        sel = tf.image.non_max_suppression(shift_boxes(boxes, box_ind), anchor_prob, 100000, iou_threshold=inputs.nms_th)
        # sel is a list of indices

        if True:    # prediction head, not used in training
            psel = tf.slice(sel, [0], [tf.minimum(inputs.nms_max, tf.shape(sel)[0])])
            boxes_predicted = tf.gather(boxes, psel)
            box_ind_predicted = tf.gather(box_ind, psel)

            mlogits = mask_net(inputs.X, mask_ft, boxes_predicted, box_ind_predicted)
            masks_predicted = tf.squeeze(tf.slice(tf.nn.softmax(mlogits), [0, 0, 0, 1], [-1, -1, -1, 1]), 3)
            pass

        anchor_prob = None  # discard
        boxes = tf.gather(boxes, sel)
        box_ind = tf.gather(box_ind, sel)

        hit, index, gt_index = tf.py_func(gt_matcher.apply, [boxes, box_ind, inputs.gt_boxes], [tf.float32, tf.int32, tf.int32])


        # % boxes found
        precision = hit / tf.cast(tf.shape(boxes)[0] + 1, tf.float32);
        recall = hit / tf.cast(tf.shape(inputs.gt_boxes)[0] + 1, tf.float32);

        boxes = tf.gather(boxes, index)
        box_ind = tf.gather(box_ind, index)
        gt_boxes = tf.gather(inputs.gt_boxes, gt_index)

        # normalize boxes to [0-1]
        nboxes = normalize_boxes(tf.shape(inputs.X), boxes)
        mlogits = mask_net(inputs.X, mask_ft, boxes, box_ind)

        gt_masks, = tf.py_func(mask_extractor.apply, [inputs.gt_masks, gt_boxes, boxes], [tf.float32])
        #gt_masks, = tf.py_func(mask_extractor.apply, [inputs.gt_masks, gt_boxes, tf.slice(gt_boxes, [0, 3], [-1, 4])], [tf.float32])
        end_points['gt_boxes'] = gt_boxes
        end_points['boxes'] = boxes
        gt_masks = tf.cast(tf.round(gt_masks), tf.int32)
        end_points['gt_masks'] = gt_masks
        # mask cross entropy
        mxe = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=mlogits, labels=gt_masks)
        mxe = tf.reshape(mxe, (-1, ))
        mxe = tf.reduce_sum(mxe) / tf.cast(tf.shape(mxe)[0] + 1, tf.float32)

    #tf.identity(logits, name='logits')
    #tf.identity(params, name='params')
    #tf.identity(boxes_pre, name='boxes_pre')
    tf.identity(boxes_predicted, name='boxes')
    tf.identity(masks_predicted, name='masks')
    #tf.identity(mlogits, name='mlogits')
    axe = tf.identity(axe, name='ax') # cross-entropy
    mxe = tf.identity(mxe, name='mx') # cross-entropy
    pl = tf.identity(pl * FLAGS.pl_weight, name='pl') # params-loss
    reg = tf.identity(tf.reduce_sum(tf.losses.get_regularization_losses()) * FLAGS.re_weight, name='re')
    precision = tf.identity(precision, name='p')
    recall = tf.identity(recall, name='r')

    loss = tf.identity(axe + mxe + pl + reg, name='lo')

    return loss, [axe, mxe, pl, reg, precision, recall], end_points
Ejemplo n.º 38
0
    def build(self, for_deploy, variants=""):
        conf = self.conf
        name = self.name
        job_type = self.job_type
        dtype = self.dtype
        self.beam_size = 1 if (not for_deploy or variants == "score") else sum(
            self.conf.beam_splits)

        graphlg.info("Creating placeholders...")
        self.enc_str_inps = tf.placeholder(tf.string,
                                           shape=(None, conf.input_max_len),
                                           name="enc_inps")
        self.enc_lens = tf.placeholder(tf.int32, shape=[None], name="enc_lens")
        self.dec_str_inps = tf.placeholder(
            tf.string, shape=[None, conf.output_max_len + 2], name="dec_inps")
        self.dec_lens = tf.placeholder(tf.int32, shape=[None], name="dec_lens")
        self.down_wgts = tf.placeholder(tf.float32,
                                        shape=[None],
                                        name="down_wgts")

        with tf.name_scope("TableLookup"):
            # lookup tables
            self.in_table = lookup.MutableHashTable(key_dtype=tf.string,
                                                    value_dtype=tf.int64,
                                                    default_value=UNK_ID,
                                                    shared_name="in_table",
                                                    name="in_table",
                                                    checkpoint=True)

            self.out_table = lookup.MutableHashTable(key_dtype=tf.int64,
                                                     value_dtype=tf.string,
                                                     default_value="_UNK",
                                                     shared_name="out_table",
                                                     name="out_table",
                                                     checkpoint=True)
            self.enc_inps = self.in_table.lookup(self.enc_str_inps)
            self.dec_inps = self.in_table.lookup(self.dec_str_inps)

        # Create encode graph and get attn states
        graphlg.info("Creating embeddings and embedding enc_inps.")
        with ops.device("/cpu:0"):
            self.embedding = variable_scope.get_variable(
                "embedding", [conf.output_vocab_size, conf.embedding_size])

        with tf.name_scope("Embed") as scope:
            dec_inps = tf.slice(self.dec_inps, [0, 0],
                                [-1, conf.output_max_len + 1])
            with ops.device("/cpu:0"):
                self.emb_inps = embedding_lookup_unique(
                    self.embedding, self.enc_inps)
                emb_dec_inps = embedding_lookup_unique(self.embedding,
                                                       dec_inps)
        # output projector (w, b)
        with tf.variable_scope("OutProj"):
            if conf.out_layer_size:
                w = tf.get_variable(
                    "proj_w", [conf.out_layer_size, conf.output_vocab_size],
                    dtype=dtype)
            elif conf.bidirectional:
                w = tf.get_variable(
                    "proj_w", [conf.num_units * 2, conf.output_vocab_size],
                    dtype=dtype)
            else:
                w = tf.get_variable("proj_w",
                                    [conf.num_units, conf.output_vocab_size],
                                    dtype=dtype)
            b = tf.get_variable("proj_b", [conf.output_vocab_size],
                                dtype=dtype)

        graphlg.info("Creating dynamic rnn...")
        self.enc_outs, self.enc_states, mem_size, enc_state_size = DynRNN(
            conf.cell_model,
            conf.num_units,
            conf.num_layers,
            self.emb_inps,
            self.enc_lens,
            keep_prob=1.0,
            bidi=conf.bidirectional,
            name_scope="DynRNNEncoder")
        batch_size = tf.shape(self.enc_outs)[0]
        # Do vae on the state of the last layer of the encoder
        final_enc_states = []
        KLDs = 0.0
        for each in self.enc_states:
            z, KLD, l2 = CreateVAE([each],
                                   self.conf.enc_latent_dim,
                                   name_scope="VAE")
            if isinstance(each, LSTMStateTuple):
                final_enc_states.append(
                    LSTMStateTuple(each.c, tf.concat([each.h, z], 1)))
            else:
                final_enc_state.append(tf.concat([z, each], 1))
            KLDs += KLD / self.conf.num_layers

        with tf.name_scope("DynRNNDecode") as scope:
            with tf.name_scope("ShapeToBeam") as scope:
                beam_memory = tf.reshape(
                    tf.tile(self.enc_outs, [1, 1, self.beam_size]),
                    [-1, conf.input_max_len, mem_size])
                beam_memory_lens = tf.squeeze(
                    tf.reshape(
                        tf.tile(tf.expand_dims(self.enc_lens, 1),
                                [1, self.beam_size]), [-1, 1]), 1)

                def _to_beam(t):
                    return tf.reshape(tf.tile(t, [1, self.beam_size]),
                                      [-1, int(t.get_shape()[1])])

                beam_init_states = tf.contrib.framework.nest.map_structure(
                    _to_beam, final_enc_states)
            max_mem_size = self.conf.input_max_len + self.conf.output_max_len + 2
            cell = AttnCell(cell_model=conf.cell_model,
                            num_units=mem_size,
                            num_layers=conf.num_layers,
                            attn_type=self.conf.attention,
                            memory=beam_memory,
                            mem_lens=beam_memory_lens,
                            max_mem_size=max_mem_size,
                            addmem=self.conf.addmem,
                            keep_prob=conf.keep_prob,
                            dtype=tf.float32,
                            name_scope="AttnCell")

            dec_init_state = DecStateInit(all_enc_states=beam_init_states,
                                          decoder_cell=cell,
                                          batch_size=batch_size *
                                          self.beam_size,
                                          init_type="each2each")

            if not for_deploy:
                hp_train = helper.ScheduledEmbeddingTrainingHelper(
                    inputs=emb_dec_inps,
                    sequence_length=self.dec_lens,
                    embedding=self.embedding,
                    sampling_probability=self.conf.sample_prob,
                    out_proj=(w, b))
                output_layer = layers_core.Dense(
                    self.conf.out_layer_size,
                    use_bias=True) if self.conf.out_layer_size else None
                my_decoder = basic_decoder.BasicDecoder(
                    cell=cell,
                    helper=hp_train,
                    initial_state=dec_init_state,
                    output_layer=output_layer)
                cell_outs, final_state = decoder.dynamic_decode(
                    decoder=my_decoder,
                    impute_finished=False,
                    maximum_iterations=conf.output_max_len + 1,
                    scope=scope)
            elif variants == "score":
                dec_init_state = zero_attn_states
                hp_train = helper.ScheduledEmbeddingTrainingHelper(
                    inputs=emb_dec_inps,
                    sequence_length=self.dec_lens,
                    embedding=self.embedding,
                    sampling_probability=0.0,
                    out_proj=(w, b))
                output_layer = layers_core.Dense(
                    self.conf.out_layer_size,
                    use_bias=True) if self.conf.out_layer_size else None
                my_decoder = score_decoder.ScoreDecoder(
                    cell=cell,
                    helper=hp_train,
                    out_proj=(w, b),
                    initial_state=dec_init_state,
                    output_layer=output_layer)
                cell_outs, final_state = decoder.dynamic_decode(
                    decoder=my_decoder,
                    scope=scope,
                    maximum_iterations=self.conf.output_max_len,
                    impute_finished=False)
            else:
                hp_infer = helper.GreedyEmbeddingHelper(
                    embedding=self.embedding,
                    start_tokens=tf.ones(shape=[batch_size * self.beam_size],
                                         dtype=tf.int32),
                    end_token=EOS_ID,
                    out_proj=(w, b))

                output_layer = layers_core.Dense(
                    self.conf.out_layer_size,
                    use_bias=True) if self.conf.out_layer_size else None
                my_decoder = beam_decoder.BeamDecoder(
                    cell=cell,
                    helper=hp_infer,
                    out_proj=(w, b),
                    initial_state=dec_init_state,
                    beam_splits=self.conf.beam_splits,
                    max_res_num=self.conf.max_res_num,
                    output_layer=output_layer)
                cell_outs, final_state = decoder.dynamic_decode(
                    decoder=my_decoder,
                    scope=scope,
                    maximum_iterations=self.conf.output_max_len,
                    impute_finished=True)

        if not for_deploy:
            outputs = cell_outs.rnn_output
            # Output ouputprojected to logits
            L = tf.shape(outputs)[1]
            outputs = tf.reshape(outputs, [-1, int(w.shape[0])])
            outputs = tf.matmul(outputs, w) + b
            logits = tf.reshape(outputs, [-1, L, int(w.shape[1])])

            # branch 1 for debugging, doesn't have to be called
            with tf.name_scope("DebugOutputs") as scope:
                self.outputs = tf.argmax(logits, axis=2)
                self.outputs = tf.reshape(self.outputs, [-1, L])
                self.outputs = self.out_table.lookup(
                    tf.cast(self.outputs, tf.int64))

            with tf.name_scope("Loss") as scope:
                tars = tf.slice(self.dec_inps, [0, 1], [-1, L])
                wgts = tf.cumsum(tf.one_hot(self.dec_lens, L),
                                 axis=1,
                                 reverse=True)
                #wgts = wgts * tf.expand_dims(self.down_wgts, 1)
                self.loss = loss.sequence_loss(logits=logits,
                                               targets=tars,
                                               weights=wgts,
                                               average_across_timesteps=False,
                                               average_across_batch=False)
                example_losses = tf.reduce_sum(self.loss, 1)

                batch_wgt = tf.reduce_sum(self.down_wgts)
                see_KLD = tf.reduce_sum(KLDs * self.down_wgts) / batch_wgt
                see_loss = tf.reduce_sum(example_losses / tf.cast(
                    self.dec_lens, tf.float32) * self.down_wgts) / batch_wgt

                # not average over length
                self.loss = tf.reduce_sum(
                    (example_losses + self.conf.kld_ratio * KLDs) *
                    self.down_wgts) / batch_wgt

            with tf.name_scope(self.model_kind):
                tf.summary.scalar("loss", see_loss)
                tf.summary.scalar("kld", see_KLD)

            graph_nodes = {
                "loss": self.loss,
                "inputs": {},
                "outputs": {},
                "debug_outputs": self.outputs
            }

        elif variants == "score":
            L = tf.shape(cell_outs.logprobs)[1]
            one_hot = tf.one_hot(tf.slice(self.dec_inps, [0, 1], [-1, L]),
                                 depth=self.conf.output_vocab_size,
                                 axis=-1,
                                 on_value=1.0,
                                 off_value=0.0)
            outputs = tf.reduce_sum(cell_outs.logprobs * one_hot, 2)
            outputs = tf.reduce_sum(outputs, axis=1)
            inputs = {
                "enc_inps:0": self.enc_str_inps,
                "enc_lens:0": self.enc_lens,
                "dec_inps:0": self.dec_str_inps,
                "dec_lens:0": self.dec_lens
            }
            graph_nodes = {
                "loss": None,
                "inputs": inputs,
                "outputs": {
                    "logprobs": outputs
                },
                "visualize": None
            }
        else:
            L = tf.shape(cell_outs.beam_ends)[1]
            beam_symbols = cell_outs.beam_symbols
            beam_parents = cell_outs.beam_parents

            beam_ends = cell_outs.beam_ends
            beam_end_parents = cell_outs.beam_end_parents
            beam_end_probs = cell_outs.beam_end_probs
            alignments = cell_outs.alignments

            beam_ends = tf.reshape(tf.transpose(beam_ends, [0, 2, 1]), [-1, L])
            beam_end_parents = tf.reshape(
                tf.transpose(beam_end_parents, [0, 2, 1]), [-1, L])
            beam_end_probs = tf.reshape(
                tf.transpose(beam_end_probs, [0, 2, 1]), [-1, L])

            ## Creating tail_ids
            batch_size = tf.Print(batch_size, [batch_size],
                                  message="VAERNN2 batch")
            batch_offset = tf.expand_dims(
                tf.cumsum(
                    tf.ones([batch_size, self.beam_size], dtype=tf.int32) *
                    self.beam_size,
                    axis=0,
                    exclusive=True), 2)
            offset2 = tf.expand_dims(
                tf.cumsum(
                    tf.ones([batch_size, self.beam_size * 2], dtype=tf.int32) *
                    self.beam_size,
                    axis=0,
                    exclusive=True), 2)

            out_len = tf.shape(beam_symbols)[1]
            self.beam_symbol_strs = tf.reshape(
                self.out_table.lookup(tf.cast(beam_symbols, tf.int64)),
                [batch_size, self.beam_size, -1])
            self.beam_parents = tf.reshape(
                beam_parents, [batch_size, self.beam_size, -1]) - batch_offset

            self.beam_ends = tf.reshape(beam_ends,
                                        [batch_size, self.beam_size * 2, -1])
            self.beam_end_parents = tf.reshape(
                beam_end_parents,
                [batch_size, self.beam_size * 2, -1]) - offset2
            self.beam_end_probs = tf.reshape(
                beam_end_probs, [batch_size, self.beam_size * 2, -1])
            self.beam_attns = tf.reshape(
                alignments, [batch_size, self.beam_size, out_len, -1])

            inputs = {
                "enc_inps:0": self.enc_str_inps,
                "enc_lens:0": self.enc_lens
            }
            outputs = {
                "beam_symbols": self.beam_symbol_strs,
                "beam_parents": self.beam_parents,
                "beam_ends": self.beam_ends,
                "beam_end_parents": self.beam_end_parents,
                "beam_end_probs": self.beam_end_probs,
                "beam_attns": self.beam_attns
            }
            graph_nodes = {
                "loss": None,
                "inputs": inputs,
                "outputs": outputs,
                "visualize": {
                    "z": z
                }
            }
        return graph_nodes
Ejemplo n.º 39
0
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)

batch_size = 512
learning_rate = 1e-3
epoch = 10000
discrete_latent_size = 10
contin_latent_size = 2

x = tf.placeholder(tf.float32, shape=[None, 784])
x_image = tf.reshape(x, [-1, 28, 28, 1])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
z_in = tf.placeholder(tf.float32, shape=[batch_size, 112])
z_label_check = tf.slice(z_in, begin=[0, 100], size=[batch_size, 10])
z_contin_check = tf.slice(z_in, begin=[0, 110], size=[batch_size, 2])

initializer = tf.truncated_normal_initializer(stddev=0.02)


def int_to_onehot(z_label):

    one_hot_array = np.zeros(shape=[len(z_label), discrete_latent_size])
    one_hot_array[np.arange(len(z_label)), z_label] = 1

    return one_hot_array


def lrelu(x, leak=0.2, name="lrelu"):
Ejemplo n.º 40
0
    def _compute_loss(self, input_dict):
        """
    Computes cross entropy based sequence-to-sequence loss with label smoothing
    :param input_dict: inputs to compute loss
    {
          "logits": logits tensor of shape [batch_size, T, dim]
          "target_sequence": tensor of shape [batch_size, T]
          "tgt_lengths": tensor of shape [batch_size] or None
    }
    :return: Singleton loss tensor
    """
        logits = input_dict["decoder_output"]["logits"]
        target_sequence = input_dict["target_tensors"][0]
        tgt_lengths = input_dict["target_tensors"][1]
        #target_sequence = input_dict["tgt_sequence"]
        #tgt_lengths = input_dict["tgt_length"]

        if self._offset_target_by_one:
            # this is necessary for auto-regressive models
            current_ts = tf.to_int32(
                tf.minimum(
                    tf.shape(target_sequence)[1],
                    tf.shape(logits)[1],
                )) - 1

            logits = tf.slice(
                logits,
                begin=[0, 0, 0],
                size=[-1, current_ts, -1],
            )
            target_sequence = tf.slice(target_sequence,
                                       begin=[0, 1],
                                       size=[-1, current_ts])
        else:
            current_ts = tf.to_int32(
                tf.minimum(
                    tf.shape(target_sequence)[1],
                    tf.shape(logits)[1],
                ))

        # Cast logits after potential slice
        if logits.dtype.base_dtype != tf.float32:
            logits = tf.cast(logits, tf.float32)

        if self._do_mask:
            if tgt_lengths is None:
                raise ValueError(
                    "If you are masking loss, tgt_lengths can't be None")
            mask = tf.sequence_mask(lengths=tgt_lengths - 1,
                                    maxlen=current_ts,
                                    dtype=tf.float32)
        else:
            mask = tf.cast(tf.ones_like(target_sequence), logits.dtype)

        labels = tf.one_hot(indices=tf.reshape(target_sequence, shape=[-1]),
                            depth=self._tgt_vocab_size)
        logits = tf.reshape(logits, shape=[-1, self._tgt_vocab_size])
        mask = tf.reshape(mask, shape=[-1])

        loss = tf.losses.softmax_cross_entropy(
            onehot_labels=labels,
            logits=logits,
            weights=mask,
            label_smoothing=self._label_smoothing,
            reduction=tf.losses.Reduction.NONE)

        loss = tf.reduce_sum(loss * tf.reshape(mask, shape=[-1]))
        if self._average_across_timestep:
            loss /= tf.reduce_sum(mask)
        else:
            loss /= self._batch_size
        return loss
Ejemplo n.º 41
0
def inference_mem(images,
                  cams,
                  depth_num,
                  depth_start,
                  depth_interval,
                  is_master_gpu=True):
    """ infer depth image from multi-view images and cameras """

    # dynamic gpu params
    depth_end = depth_start + (tf.cast(depth_num, tf.float32) -
                               1) * depth_interval
    feature_c = 32
    feature_h = FLAGS.max_h / 4
    feature_w = FLAGS.max_w / 4

    # reference image
    ref_image = tf.squeeze(tf.slice(images, [0, 0, 0, 0, 0],
                                    [-1, 1, -1, -1, 3]),
                           axis=1)
    ref_cam = tf.squeeze(tf.slice(cams, [0, 0, 0, 0, 0], [-1, 1, 2, 4, 4]),
                         axis=1)

    # image feature extraction
    if is_master_gpu:
        ref_tower = UniNetDS2({'data': ref_image},
                              is_training=True,
                              reuse=False)
    else:
        ref_tower = UniNetDS2({'data': ref_image},
                              is_training=True,
                              reuse=True)
    ref_feature = ref_tower.get_output()
    ref_feature2 = tf.square(ref_feature)

    view_features = []
    for view in range(1, FLAGS.view_num):
        view_image = tf.squeeze(tf.slice(images, [0, view, 0, 0, 0],
                                         [-1, 1, -1, -1, -1]),
                                axis=1)
        view_tower = UniNetDS2({'data': view_image},
                               is_training=True,
                               reuse=True)
        view_features.append(view_tower.get_output())
    view_features = tf.stack(view_features, axis=0)

    # get all homographies
    view_homographies = []
    for view in range(1, FLAGS.view_num):
        view_cam = tf.squeeze(tf.slice(cams, [0, view, 0, 0, 0],
                                       [-1, 1, 2, 4, 4]),
                              axis=1)
        homographies = get_homographies(ref_cam,
                                        view_cam,
                                        depth_num=depth_num,
                                        depth_start=depth_start,
                                        depth_interval=depth_interval)
        view_homographies.append(homographies)
    view_homographies = tf.stack(view_homographies, axis=0)

    # build cost volume by differentialble homography
    with tf.name_scope('cost_volume_homography'):
        depth_costs = []

        for d in range(depth_num):
            # compute cost (standard deviation feature)
            ave_feature = tf.Variable(
                tf.zeros([FLAGS.batch_size, feature_h, feature_w, feature_c]),
                name='ave',
                trainable=False,
                collections=[tf.GraphKeys.LOCAL_VARIABLES])
            ave_feature2 = tf.Variable(
                tf.zeros([FLAGS.batch_size, feature_h, feature_w, feature_c]),
                name='ave2',
                trainable=False,
                collections=[tf.GraphKeys.LOCAL_VARIABLES])
            ave_feature = tf.assign(ave_feature, ref_feature)
            ave_feature2 = tf.assign(ave_feature2, ref_feature2)

            def body(view, ave_feature, ave_feature2):
                """Loop body."""
                homography = tf.slice(view_homographies[view],
                                      begin=[0, d, 0, 0],
                                      size=[-1, 1, 3, 3])
                homography = tf.squeeze(homography, axis=1)
                warped_view_feature = homography_warping(
                    view_features[view], homography)
                ave_feature = tf.assign_add(ave_feature, warped_view_feature)
                ave_feature2 = tf.assign_add(ave_feature2,
                                             tf.square(warped_view_feature))
                view = tf.add(view, 1)
                return view, ave_feature, ave_feature2

            view = tf.constant(0)
            cond = lambda view, *_: tf.less(view, FLAGS.view_num - 1)
            _, ave_feature, ave_feature2 = tf.while_loop(
                cond,
                body, [view, ave_feature, ave_feature2],
                back_prop=False,
                parallel_iterations=1)

            ave_feature = tf.assign(
                ave_feature,
                tf.square(ave_feature) / (FLAGS.view_num * FLAGS.view_num))
            ave_feature2 = tf.assign(
                ave_feature2, ave_feature2 / FLAGS.view_num - ave_feature)
            depth_costs.append(ave_feature2)
        cost_volume = tf.stack(depth_costs, axis=1)

    # filtered cost volume, size of (B, D, H, W, 1)
    if is_master_gpu:
        filtered_cost_volume_tower = RegNetUS0({'data': cost_volume},
                                               is_training=True,
                                               reuse=False)
    else:
        filtered_cost_volume_tower = RegNetUS0({'data': cost_volume},
                                               is_training=True,
                                               reuse=True)
    filtered_cost_volume = tf.squeeze(filtered_cost_volume_tower.get_output(),
                                      axis=-1)

    # depth map by softArgmin
    with tf.name_scope('soft_arg_min'):
        # probability volume by soft max
        probability_volume = tf.nn.softmax(tf.scalar_mul(
            -1, filtered_cost_volume),
                                           axis=1,
                                           name='prob_volume')

        # depth image by soft argmin
        volume_shape = tf.shape(probability_volume)
        soft_2d = []
        for i in range(FLAGS.batch_size):
            soft_1d = tf.linspace(depth_start[i], depth_end[i],
                                  tf.cast(depth_num, tf.int32))
            soft_2d.append(soft_1d)
        soft_2d = tf.reshape(tf.stack(soft_2d, axis=0),
                             [volume_shape[0], volume_shape[1], 1, 1])
        soft_4d = tf.tile(soft_2d, [1, 1, volume_shape[2], volume_shape[3]])
        estimated_depth_map = tf.reduce_sum(soft_4d * probability_volume,
                                            axis=1)
        estimated_depth_map = tf.expand_dims(estimated_depth_map, axis=3)

    # probability map
    prob_map = get_propability_map(probability_volume, estimated_depth_map,
                                   depth_start, depth_interval)

    filtered_depth_map = tf.cast(tf.greater_equal(prob_map, 0.8),
                                 dtype='float32') * estimated_depth_map

    return filtered_depth_map, prob_map
Ejemplo n.º 42
0
    def _init_graph(self):
        self._init_placeholders()

        self.word_emb = tf.get_variable("word_emb",
                                        initializer=tf.constant(
                                            self.init_word_emb,
                                            dtype=tf.float32),
                                        trainable=False)
        self.char_emb = tf.get_variable("char_emb",
                                        initializer=tf.constant(
                                            self.init_char_emb,
                                            dtype=tf.float32),
                                        trainable=self.opt['train_char_emb'])

        self.c_mask = tf.cast(self.c_ph, tf.bool)
        self.q_mask = tf.cast(self.q_ph, tf.bool)
        self.c_len = tf.reduce_sum(tf.cast(self.c_mask, tf.int32), axis=1)
        self.q_len = tf.reduce_sum(tf.cast(self.q_mask, tf.int32), axis=1)

        bs = tf.shape(self.c_ph)[0]
        self.c_maxlen = tf.reduce_max(self.c_len)
        self.q_maxlen = tf.reduce_max(self.q_len)
        self.c = tf.slice(self.c_ph, [0, 0], [bs, self.c_maxlen])
        self.q = tf.slice(self.q_ph, [0, 0], [bs, self.q_maxlen])
        self.c_mask = tf.slice(self.c_mask, [0, 0], [bs, self.c_maxlen])
        self.q_mask = tf.slice(self.q_mask, [0, 0], [bs, self.q_maxlen])
        self.cc = tf.slice(self.cc_ph, [0, 0, 0],
                           [bs, self.c_maxlen, self.char_limit])
        self.qc = tf.slice(self.qc_ph, [0, 0, 0],
                           [bs, self.q_maxlen, self.char_limit])
        self.cc_len = tf.reshape(
            tf.reduce_sum(tf.cast(tf.cast(self.cc, tf.bool), tf.int32),
                          axis=2), [-1])
        self.qc_len = tf.reshape(
            tf.reduce_sum(tf.cast(tf.cast(self.qc, tf.bool), tf.int32),
                          axis=2), [-1])
        self.y1 = tf.one_hot(self.y1_ph, depth=self.context_limit)
        self.y2 = tf.one_hot(self.y2_ph, depth=self.context_limit)
        self.y1 = tf.slice(self.y1, [0, 0], [bs, self.c_maxlen])
        self.y2 = tf.slice(self.y2, [0, 0], [bs, self.c_maxlen])

        with tf.variable_scope("emb"):
            with tf.variable_scope("char"):
                cc_emb = tf.reshape(
                    tf.nn.embedding_lookup(self.char_emb, self.cc),
                    [bs * self.c_maxlen, self.char_limit, self.char_emb_dim])
                qc_emb = tf.reshape(
                    tf.nn.embedding_lookup(self.char_emb, self.qc),
                    [bs * self.q_maxlen, self.char_limit, self.char_emb_dim])

                cc_emb = variational_dropout(cc_emb,
                                             keep_prob=self.keep_prob_ph)
                qc_emb = variational_dropout(qc_emb,
                                             keep_prob=self.keep_prob_ph)

                _, (state_fw, state_bw) = cudnn_bi_gru(cc_emb,
                                                       self.char_hidden_size,
                                                       seq_lengths=self.cc_len)
                cc_emb = tf.concat([state_fw, state_bw], axis=1)

                _, (state_fw, state_bw) = cudnn_bi_gru(qc_emb,
                                                       self.char_hidden_size,
                                                       seq_lengths=self.qc_len,
                                                       reuse=True)
                qc_emb = tf.concat([state_fw, state_bw], axis=1)

                cc_emb = tf.reshape(
                    cc_emb, [bs, self.c_maxlen, 2 * self.char_hidden_size])
                qc_emb = tf.reshape(
                    qc_emb, [bs, self.q_maxlen, 2 * self.char_hidden_size])

            with tf.name_scope("word"):
                c_emb = tf.nn.embedding_lookup(self.word_emb, self.c)
                q_emb = tf.nn.embedding_lookup(self.word_emb, self.q)

            c_emb = tf.concat([c_emb, cc_emb], axis=2)
            q_emb = tf.concat([q_emb, qc_emb], axis=2)

        with tf.variable_scope("encoding"):
            rnn = CudnnGRU(num_layers=3,
                           num_units=self.hidden_size,
                           batch_size=bs,
                           input_size=c_emb.get_shape().as_list()[-1],
                           keep_prob=self.keep_prob_ph)
            c = rnn(c_emb, seq_len=self.c_len)
            q = rnn(q_emb, seq_len=self.q_len)

        with tf.variable_scope("attention"):
            qc_att = dot_attention(c,
                                   q,
                                   mask=self.q_mask,
                                   att_size=self.attention_hidden_size,
                                   keep_prob=self.keep_prob_ph)
            rnn = CudnnGRU(num_layers=1,
                           num_units=self.hidden_size,
                           batch_size=bs,
                           input_size=qc_att.get_shape().as_list()[-1],
                           keep_prob=self.keep_prob_ph)
            att = rnn(qc_att, seq_len=self.c_len)

        with tf.variable_scope("match"):
            self_att = dot_attention(att,
                                     att,
                                     mask=self.c_mask,
                                     att_size=self.attention_hidden_size,
                                     keep_prob=self.keep_prob_ph)
            rnn = CudnnGRU(num_layers=1,
                           num_units=self.hidden_size,
                           batch_size=bs,
                           input_size=self_att.get_shape().as_list()[-1],
                           keep_prob=self.keep_prob_ph)
            match = rnn(self_att, seq_len=self.c_len)

        with tf.variable_scope("pointer"):
            init = simple_attention(q,
                                    self.hidden_size,
                                    mask=self.q_mask,
                                    keep_prob=self.keep_prob_ph)
            pointer = PtrNet(cell_size=init.get_shape().as_list()[-1],
                             keep_prob=self.keep_prob_ph)
            logits1, logits2 = pointer(init, match, self.hidden_size,
                                       self.c_mask)

        with tf.variable_scope("predict"):
            outer = tf.matmul(tf.expand_dims(tf.nn.softmax(logits1), axis=2),
                              tf.expand_dims(tf.nn.softmax(logits2), axis=1))
            outer = tf.matrix_band_part(
                outer, 0, tf.cast(tf.minimum(15, self.c_maxlen), tf.int64))
            self.yp1 = tf.argmax(tf.reduce_max(outer, axis=2), axis=1)
            self.yp2 = tf.argmax(tf.reduce_max(outer, axis=1), axis=1)
            loss_1 = tf.nn.softmax_cross_entropy_with_logits(logits=logits1,
                                                             labels=self.y1)
            loss_2 = tf.nn.softmax_cross_entropy_with_logits(logits=logits2,
                                                             labels=self.y2)
            self.loss = tf.reduce_mean(loss_1 + loss_2)

        if self.weight_decay < 1.0:
            self.var_ema = tf.train.ExponentialMovingAverage(self.weight_decay)
            ema_op = self.var_ema.apply(tf.trainable_variables())
            with tf.control_dependencies([ema_op]):
                self.loss = tf.identity(self.loss)

                self.shadow_vars = []
                self.global_vars = []
                for var in tf.global_variables():
                    v = self.var_ema.average(var)
                    if v:
                        self.shadow_vars.append(v)
                        self.global_vars.append(var)
                self.assign_vars = []
                for g, v in zip(self.global_vars, self.shadow_vars):
                    self.assign_vars.append(tf.assign(g, v))
Ejemplo n.º 43
0
 def _call(self, inputs):
     ids, num_samples = inputs
     adj_lists = tf.nn.embedding_lookup(self.adj_info, ids)
     adj_lists = tf.transpose(tf.random_shuffle(tf.transpose(adj_lists)))
     adj_lists = tf.slice(adj_lists, [0,0], [-1, num_samples])
     return adj_lists
Ejemplo n.º 44
0
def inference(images,
              cams,
              depth_num,
              depth_start,
              depth_interval,
              is_master_gpu=True):
    """ infer depth image from multi-view images and cameras """

    # dynamic gpu params
    depth_end = depth_start + (tf.cast(depth_num, tf.float32) -
                               1) * depth_interval

    # reference image
    ref_image = tf.squeeze(tf.slice(images, [0, 0, 0, 0, 0],
                                    [-1, 1, -1, -1, 3]),
                           axis=1)
    ref_cam = tf.squeeze(tf.slice(cams, [0, 0, 0, 0, 0], [-1, 1, 2, 4, 4]),
                         axis=1)

    # image feature extraction
    if is_master_gpu:
        ref_tower = UniNetDS2({'data': ref_image},
                              is_training=True,
                              reuse=False)
    else:
        ref_tower = UniNetDS2({'data': ref_image},
                              is_training=True,
                              reuse=True)
    view_towers = []
    for view in range(1, FLAGS.view_num):
        view_image = tf.squeeze(tf.slice(images, [0, view, 0, 0, 0],
                                         [-1, 1, -1, -1, -1]),
                                axis=1)
        view_tower = UniNetDS2({'data': view_image},
                               is_training=True,
                               reuse=True)
        view_towers.append(view_tower)

    # get all homographies
    view_homographies = []
    for view in range(1, FLAGS.view_num):
        view_cam = tf.squeeze(tf.slice(cams, [0, view, 0, 0, 0],
                                       [-1, 1, 2, 4, 4]),
                              axis=1)
        homographies = get_homographies(ref_cam,
                                        view_cam,
                                        depth_num=depth_num,
                                        depth_start=depth_start,
                                        depth_interval=depth_interval)
        view_homographies.append(homographies)

    # build cost volume by differentialble homography
    with tf.name_scope('cost_volume_homography'):
        depth_costs = []
        for d in range(depth_num):
            # compute cost (variation metric)
            ave_feature = ref_tower.get_output()
            ave_feature2 = tf.square(ref_tower.get_output())
            for view in range(0, FLAGS.view_num - 1):
                homography = tf.slice(view_homographies[view],
                                      begin=[0, d, 0, 0],
                                      size=[-1, 1, 3, 3])
                homography = tf.squeeze(homography, axis=1)
                warped_view_feature = homography_warping(
                    view_towers[view].get_output(), homography)
                ave_feature = ave_feature + warped_view_feature
                ave_feature2 = ave_feature2 + tf.square(warped_view_feature)
            ave_feature = ave_feature / FLAGS.view_num
            ave_feature2 = ave_feature2 / FLAGS.view_num
            cost = ave_feature2 - tf.square(ave_feature)
            depth_costs.append(cost)
        cost_volume = tf.stack(depth_costs, axis=1)

    # filtered cost volume, size of (B, D, H, W, 1)
    if is_master_gpu:
        filtered_cost_volume_tower = RegNetUS0({'data': cost_volume},
                                               is_training=True,
                                               reuse=False)
    else:
        filtered_cost_volume_tower = RegNetUS0({'data': cost_volume},
                                               is_training=True,
                                               reuse=True)
    filtered_cost_volume = tf.squeeze(filtered_cost_volume_tower.get_output(),
                                      axis=-1)

    # depth map by softArgmin
    with tf.name_scope('soft_arg_min'):
        # probability volume by soft max
        probability_volume = tf.nn.softmax(tf.scalar_mul(
            -1, filtered_cost_volume),
                                           axis=1,
                                           name='prob_volume')
        # depth image by soft argmin
        volume_shape = tf.shape(probability_volume)
        soft_2d = []
        for i in range(FLAGS.batch_size):
            soft_1d = tf.linspace(depth_start[i], depth_end[i],
                                  tf.cast(depth_num, tf.int32))
            soft_2d.append(soft_1d)
        soft_2d = tf.reshape(tf.stack(soft_2d, axis=0),
                             [volume_shape[0], volume_shape[1], 1, 1])
        soft_4d = tf.tile(soft_2d, [1, 1, volume_shape[2], volume_shape[3]])
        estimated_depth_map = tf.reduce_sum(soft_4d * probability_volume,
                                            axis=1)
        estimated_depth_map = tf.expand_dims(estimated_depth_map, axis=3)

    # probability map
    prob_map = get_propability_map(probability_volume, estimated_depth_map,
                                   depth_start, depth_interval)

    return estimated_depth_map, prob_map  #, filtered_depth_map, probability_volume
Ejemplo n.º 45
0
def read_cifar10(filename_queue):
    """Reads and parses examples from CIFAR10 data files.

  Recommendation: if you want N-way read parallelism, call this function
  N times.  This will give you N independent Readers reading different
  files & positions within those files, which will give better mixing of
  examples.

  Args:
    filename_queue: A queue of strings with the filenames to read from.

  Returns:
    An object representing a single example, with the following fields:
      height: number of rows in the result (32)
      width: number of columns in the result (32)
      depth: number of color channels in the result (3)
      key: a scalar string Tensor describing the filename & record number
        for this example.
      label: an int32 Tensor with the label in the range 0..9.
      uint8image: a [height, width, depth] uint8 Tensor with the image data
  """

    class CIFAR10Record(object):
        pass

    result = CIFAR10Record()

    # Dimensions of the images in the CIFAR-10 dataset.
    # See http://www.cs.toronto.edu/~kriz/cifar.html for a description of the
    # input format.
    label_bytes = 1  # 2 for CIFAR-100
    result.height = 32
    result.width = 32
    result.depth = 3
    image_bytes = result.height * result.width * result.depth
    # Every record consists of a label followed by the image, with a
    # fixed number of bytes for each.
    record_bytes = label_bytes + image_bytes

    # Read a record, getting filenames from the filename_queue.  No
    # header or footer in the CIFAR-10 format, so we leave header_bytes
    # and footer_bytes at their default of 0.
    '''读取一条记录麽???,然后指针移动到下一个待读取的位置'''
    reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
    result.key, value = reader.read(filename_queue)

    # Convert from a string to a vector of uint8 that is record_bytes long.
    '''原始数据为二进制,所以需要一个转换,成八位无符号整形'''
    record_bytes = tf.decode_raw(value, tf.uint8)

    # The first bytes represent the label, which we convert from uint8->int32.
    result.label = tf.cast(
        tf.slice(record_bytes, [0], [label_bytes]), tf.int32)

    # The remaining bytes after the label represent the image, which we reshape
    # from [depth * height * width] to [depth, height, width].
    depth_major = tf.reshape(tf.slice(record_bytes, [label_bytes], [image_bytes]),
                             [result.depth, result.height, result.width])
    # Convert from [depth, height, width] to [height, width, depth].
    '''原来是[0,1,2],现在是[1,2,0]'''
    result.uint8image = tf.transpose(depth_major, [1, 2, 0])

    return result
Ejemplo n.º 46
0
def eval_loop(preprocess_fn,
              network_factory,
              data_x,
              data_y,
              camera_indices,
              log_dir,
              eval_log_dir,
              image_shape=None,
              run_id=None,
              loss_mode="cosine-softmax",
              num_galleries=10,
              random_seed=4321):
    """Evaluate a running training session using CMC metric averaged over
    `num_galleries` galleries where each gallery contains for every identity a
    randomly selected image-pair.

    A call to this function will block indefinitely, monitoring the
    `log_dir/run_id` for saved checkpoints. Then, creates summaries in
    `eval_log_dir/run_id` that can be monitored with tensorboard.

    Parameters
    ----------
    preprocess_fn : Callable[tf.Tensor] -> tf.Tensor
        A callable that applies preprocessing to a given input image tensor of
        dtype tf.uint8 and returns a floating point representation (tf.float32).
    network_factory : Callable[tf.Tensor] -> (tf.Tensor, tf.Tensor)
        A callable that takes as argument a preprocessed input image of dtype
        tf.float32 and returns the feature representation as well as a logits
        tensors. The logits may be set to None if not required by the loss.
    data_x : List[str] | np.ndarray
        A list of image filenames or a tensor of images.
    data_y : List[int] | np.ndarray
        A list or one-dimensional array of labels for the images in `data_x`.
    camera_indices: Optional[List[int] | np.ndarray]
        A list or one-dimensional array of camera indices for the images in
        `data_x`. If not None, CMC galleries are created such that image pairs
        are collected from different cameras.
    log_dir: str
        Should be equivalent to the `log_dir` passed into `train_loop` of the
        training run to monitor.
    eval_log_dir:
        Used to construct the tensorboard log directory where metrics are
        summarized.
    image_shape : Tuple[int, int, int] | NoneType
        Image shape (height, width, channels) or None. If None, `train_x` must
        be an array of images such that the shape can be queries from this
        variable.
    run_id : str
        A string that identifies the training run; must be set to the same
        `run_id` passed into `train_loop`.
    loss_mode : Optional[str]
        A string that identifies the loss function used for training; must be
        one of 'cosine-softmax', 'magnet', 'triplet'. This value defaults to
        'cosine-softmax'.
    num_galleries: int
        The number of galleries to be constructed for evaluation of CMC
        metrics.
    random_seed: Optional[int]
        If not None, the NumPy random seed is fixed to this number; can be used
        to produce the same galleries over multiple runs.

    """
    if image_shape is None:
        # If image_shape is not set, train_x must be an image array. Here we
        # query the image shape from the array of images.
        assert type(data_x) == np.ndarray
        image_shape = data_x.shape[1:]
    elif type(data_x) == np.ndarray:
        assert data_x.shape[1:] == image_shape
    read_from_file = type(data_x) != np.ndarray

    # Create num_galleries random CMC galleries to average CMC top-k over.
    probes, galleries = [], []
    for i in range(num_galleries):
        probe_indices, gallery_indices = util.create_cmc_probe_and_gallery(
            data_y, camera_indices, seed=random_seed + i)
        probes.append(probe_indices)
        galleries.append(gallery_indices)
    probes, galleries = np.asarray(probes), np.asarray(galleries)

    # Set up the data feed.
    with tf.device("/cpu:0"):
        # Feed probe and gallery indices to the trainer.
        num_probes, num_gallery_images = probes.shape[1], galleries.shape[1]
        probe_idx_var = tf.placeholder(tf.int64, (None, num_probes))
        gallery_idx_var = tf.placeholder(tf.int64, (None, num_gallery_images))
        trainer = queued_trainer.QueuedTrainer(
            [probe_idx_var, gallery_idx_var])

        # Retrieve indices from trainer and gather data from constant memory.
        data_x_var = tf.constant(data_x)
        data_y_var = tf.constant(data_y)

        probe_idx_var, gallery_idx_var = trainer.get_input_vars(batch_size=1)
        probe_idx_var = tf.squeeze(probe_idx_var)
        gallery_idx_var = tf.squeeze(gallery_idx_var)

        # Apply preprocessing.
        probe_x_var = tf.gather(data_x_var, probe_idx_var)
        if read_from_file:
            # NOTE(nwojke): tf.image.decode_jpg handles various image types.
            num_channels = image_shape[-1] if len(image_shape) == 3 else 1
            probe_x_var = tf.map_fn(lambda x: tf.image.decode_jpeg(
                tf.read_file(x), channels=num_channels),
                                    probe_x_var,
                                    dtype=tf.uint8)
            probe_x_var = tf.image.resize_images(probe_x_var, image_shape[:2])
        probe_x_var = tf.map_fn(lambda x: preprocess_fn(x, is_training=False),
                                probe_x_var,
                                back_prop=False,
                                dtype=tf.float32)
        probe_y_var = tf.gather(data_y_var, probe_idx_var)

        gallery_x_var = tf.gather(data_x_var, gallery_idx_var)
        if read_from_file:
            # NOTE(nwojke): tf.image.decode_jpg handles various image types.
            num_channels = image_shape[-1] if len(image_shape) == 3 else 1
            gallery_x_var = tf.map_fn(lambda x: tf.image.decode_jpeg(
                tf.read_file(x), channels=num_channels),
                                      gallery_x_var,
                                      dtype=tf.uint8)
            gallery_x_var = tf.image.resize_images(gallery_x_var,
                                                   image_shape[:2])
        gallery_x_var = tf.map_fn(
            lambda x: preprocess_fn(x, is_training=False),
            gallery_x_var,
            back_prop=False,
            dtype=tf.float32)
        gallery_y_var = tf.gather(data_y_var, gallery_idx_var)

    # Construct the network and compute features.
    probe_and_gallery_x_var = tf.concat(axis=0,
                                        values=[probe_x_var, gallery_x_var])
    probe_and_gallery_x_var, _ = network_factory(probe_and_gallery_x_var)

    num_probe = tf.shape(probe_x_var)[0]
    probe_x_var = tf.slice(probe_and_gallery_x_var, [0, 0], [num_probe, -1])
    gallery_x_var = tf.slice(probe_and_gallery_x_var, [num_probe, 0], [-1, -1])

    # Set up the metrics.
    distance_measure = (metrics.cosine_distance
                        if loss_mode == "cosine-softmax" else metrics.pdist)

    def cmc_metric_at_k(k):
        return metrics.streaming_mean_cmc_at_k(probe_x_var,
                                               probe_y_var,
                                               gallery_x_var,
                                               gallery_y_var,
                                               k=k,
                                               measure=distance_measure)

    names_to_values, names_to_updates = slim.metrics.aggregate_metric_map(
        {"Precision@%d" % k: cmc_metric_at_k(k)
         for k in [1, 5, 10, 20]})
    for metric_name, metric_value in names_to_values.items():
        tf.summary.scalar(metric_name, metric_value)

    # Start evaluation loop.
    trainer.evaluate((probes, galleries),
                     log_dir,
                     eval_log_dir,
                     run_id=run_id,
                     eval_op=list(names_to_updates.values()),
                     eval_interval_secs=60)
Ejemplo n.º 47
0
nce_weights = tf.Variable(tf.truncated_normal([vocabulary_size, concatenated_size],
                                               stddev=1.0 / np.sqrt(concatenated_size)))
nce_biases = tf.Variable(tf.zeros([vocabulary_size]))

# Create data/target placeholders
x_inputs = tf.placeholder(tf.int32, shape=[None, window_size + 1]) # plus 1 for doc index
y_target = tf.placeholder(tf.int32, shape=[None, 1])
valid_dataset = tf.constant(valid_examples, dtype=tf.int32)

# Lookup the word embedding
# Add together element embeddings in window:
embed = tf.zeros([batch_size, embedding_size])
for element in range(window_size):
    embed += tf.nn.embedding_lookup(embeddings, x_inputs[:, element])

doc_indices = tf.slice(x_inputs, [0,window_size],[batch_size,1])
doc_embed = tf.nn.embedding_lookup(doc_embeddings,doc_indices)

#累加词嵌入向量后将文档嵌入向量连接到后面
# concatenate embeddings
final_embed = tf.concat(1, [embed, tf.squeeze(doc_embed)])

# Get loss from prediction
loss = tf.reduce_mean(tf.nn.nce_loss(nce_weights, nce_biases, y_target,final_embed,
                                     num_sampled, vocabulary_size))
                                     
# Create optimizer
optimizer = tf.train.GradientDescentOptimizer(learning_rate=model_learning_rate)
train_step = optimizer.minimize(loss)

# Cosine similarity between words
Ejemplo n.º 48
0
def pnn_rec(features, labels, mode, params):
    ##################################################
    #weight variables for pnn modle
    fm_b = tf.get_variable(name='fm_b',
                           shape=[1],
                           initializer=tf.glorot_normal_initializer())
    fm_w = tf.get_variable(name='fm_w',
                           shape=[params['feature_size']],
                           initializer=tf.glorot_normal_initializer())
    fm_v = tf.get_variable(
        name='fm_v',
        shape=[params['feature_size'], params['embedding_size']],
        initializer=tf.glorot_normal_initializer())
    ##################################################
    #feature index and vals
    feature_index = tf.reshape(features['feature_index'],
                               shape=[-1, params['field_size']])
    feature_value = tf.reshape(features['feature_value'],
                               shape=[-1, params['field_size']])

    #linear part
    embeddings_w = tf.nn.embedding_lookup(fm_w, feature_index)  #N,F
    embeddings_w = tf.multiply(embeddings_w, feature_value)
    embeddings_w = tf.reshape(embeddings_w,
                              shape=[-1, params['field_size'], 1])  #N,F,1

    #cross part
    embeddings_v = tf.nn.embedding_lookup(fm_v, feature_index)  #N, F, k
    feature_value = tf.reshape(feature_value,
                               shape=[-1, params['field_size'], 1])
    embeddings_v = tf.multiply(embeddings_v, feature_value)  #N, F, k

    #inner product pnn
    if params['product_type'] == 'in_product':
        embeddings_p = tf.matmul(embeddings_v, embeddings_v, transpose_b=True)
        feat_arr = []
        for i in range(params['field_size'] - 1):
            row_len = params['field_size'] - 1 - i
            row = tf.slice(embeddings_p, [0, i, i + 1], [-1, 1, row_len])
            feat = tf.reshape(row, [-1, row_len])
            feat_arr.append(feat)
        product = tf.concat(feat_arr, 1)  # N, F*(F-1)
    #outer product pnn
    else:
        assert params['product_type'] == 'out_product'
        embeddings_p = tf.reshape(
            embeddings_v,
            [-1, params['field_size'] * params['embedding_size']])
        embeddings_r = tf.reshape(
            tf.einsum('ni,nj->nij', embeddings_p, embeddings_p), [
                -1, params['field_size'], params['embedding_size'],
                params['field_size'], params['embedding_size']
            ])
        embeddings_r = tf.transpose(embeddings_r, perm=[0, 1, 3, 2, 4])
        embeddings_r = tf.reshape(embeddings_r, [
            -1, params['field_size'], params['field_size'],
            params['embedding_size'] * params['embedding_size']
        ])

        feat_arr = []
        for i in range(params['field_size'] - 1):
            row_len = params['field_size'] - 1 - i
            row = tf.slice(embeddings_r, [0, i, i + 1, 0],
                           [-1, 1, row_len, -1])
            feat = tf.reshape(row, [
                -1,
                row_len * params['embedding_size'] * params['embedding_size']
            ])
            feat_arr.append(feat)
        product = tf.concat(feat_arr, 1)  #N,F*(F-1)*K*K/2

    #concat linear part and cross part
    embeddings = tf.concat([embeddings_w, embeddings_v], 2)

    #reshape concat result
    net = tf.reshape(
        embeddings,
        shape=[-1, params['field_size'] * (params['embedding_size'] + 1)])
    b = fm_b * tf.ones_like(tf.reshape(tf.slice(net, [0, 0], [-1, 1]),
                                       [-1, 1]))
    net = tf.concat([net, b, product], 1)

    ##################################################
    #deep part of pnn model, perform weight decay, batch normlization, drop_out with each hidden layer
    hidden_index = 0
    for units in params['hidden_units']:
        net = tf.layers.dense(
            net,
            units=int(units),
            activation=tf.nn.relu,
            kernel_regularizer=tf.contrib.layers.l2_regularizer(
                params['l2_reg']))
        if params['batch_norm']:
            net = tf.layers.batch_normalization(
                net, training=(mode == tf.estimator.ModeKeys.TRAIN))
        if params['drop_out']:
            net = tf.layers.dropout(
                net,
                rate=float(params['drop_rates'][hidden_index]),
                training=(mode == tf.estimator.ModeKeys.TRAIN))

        hidden_index = hidden_index + 1
    logits = tf.layers.dense(net, 1, activation=tf.nn.relu)

    #predict and export
    if mode == tf.estimator.ModeKeys.PREDICT:
        predictions = {'prob': tf.nn.sigmoid(logits), 'logits': logits}
        export_outputs = {
            tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
            tf.estimator.export.PredictOutput(predictions)
        }
        return tf.estimator.EstimatorSpec(mode,
                                          predictions=predictions,
                                          export_outputs=export_outputs)
    #evaluate and train
    labels = tf.reshape(labels, (-1, 1))
    update_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
    with tf.control_dependencies(update_op):
        loss = tf.reduce_mean(
            tf.nn.sigmoid_cross_entropy_with_logits(labels=labels,
                                                    logits=logits))
        auc = tf.metrics.auc(labels=labels,
                             predictions=tf.nn.sigmoid(logits),
                             name='auc_op')
        metrics = {'auc': auc}
        tf.summary.scalar('auc', auc[1])
        if mode == tf.estimator.ModeKeys.EVAL:
            return tf.estimator.EstimatorSpec(mode,
                                              loss=loss,
                                              eval_metric_ops=metrics)

        # train.
        assert mode == tf.estimator.ModeKeys.TRAIN
        if 'Adam' == params['optimizer']:
            optimizer = tf.train.AdamOptimizer(
                learning_rate=params['learning_rate'],
                beta1=0.9,
                beta2=0.999,
                epsilon=1e-08,
                use_locking=False,
                name='Adam')
        elif 'RMSProp' == params['optimizer']:
            optimizer = tf.train.RMSPropOptimizer(
                learning_rate=params['learning_rate'],
                decay=0.9,
                momentum=0.0,
                epsilon=1e-10)
        elif 'Momentum' == params['optimizer']:
            optimizer = tf.train.MomentumOptimizer(
                learning_rate=params['learning_rate'], momentum=0.9)
        elif 'Adagrad' == params['optimizer']:
            optimizer = tf.train.AdagradOptimizer(
                learning_rate=params['learning_rate'],
                initial_accumulator_value=0.1)
        else:
            assert 'GD' == params['optimizer']
            optimizer = tf.train.GradientDescentOptimizer(
                learning_rate=params['learning_rate'])
        train_op = optimizer.minimize(loss,
                                      global_step=tf.train.get_global_step())
        return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op)
Ejemplo n.º 49
0
def run_inference_for_image_list(ordered_test_set, graph):
    output_dict_list = []
    with graph.as_default():
        with tf.Session() as sess:

            for i in range(len(ordered_test_set)):

                im_current_path = os.path.join(PATH_TO_LABEL_IMAGES_DIR,
                                               ordered_test_set[i])
                im_prev_path = im_current_path if i == 0 else os.path.join(
                    PATH_TO_LABEL_IMAGES_DIR, ordered_test_set[i - 1])

                current_frame = skimage.io.imread(im_current_path)
                prev_frame = skimage.io.imread(im_prev_path)
                image_s = cv2.subtract(current_frame, prev_frame)
                six_channels_im = np.concatenate((current_frame, image_s),
                                                 axis=2)

                # Get handles to input and output tensors
                ops = tf.get_default_graph().get_operations()
                all_tensor_names = {
                    output.name
                    for op in ops for output in op.outputs
                }
                tensor_dict = {}
                for key in [
                        'num_detections', 'detection_boxes',
                        'detection_scores', 'detection_classes',
                        'detection_masks'
                ]:
                    tensor_name = key + ':0'
                    if tensor_name in all_tensor_names:
                        tensor_dict[key] = tf.get_default_graph(
                        ).get_tensor_by_name(tensor_name)
                if 'detection_masks' in tensor_dict:
                    print("we have masks! :)")
                    # The following processing is only for single image
                    detection_boxes = tf.squeeze(
                        tensor_dict['detection_boxes'], [0])
                    detection_masks = tf.squeeze(
                        tensor_dict['detection_masks'], [0])
                    # Reframe is required to translate mask from box coordinates to image coordinates and fit the image size.
                    real_num_detection = tf.cast(
                        tensor_dict['num_detections'][0], tf.int32)
                    detection_boxes = tf.slice(detection_boxes, [0, 0],
                                               [real_num_detection, -1])
                    detection_masks = tf.slice(detection_masks, [0, 0, 0],
                                               [real_num_detection, -1, -1])
                    detection_masks_reframed = utils_ops.reframe_box_masks_to_image_masks(
                        detection_masks, detection_boxes, image_s.shape[0],
                        image_s.shape[1])
                    detection_masks_reframed = tf.cast(
                        tf.greater(detection_masks_reframed, 0.5), tf.uint8)
                    # Follow the convention by adding back the batch dimension
                    tensor_dict['detection_masks'] = tf.expand_dims(
                        detection_masks_reframed, 0)
                image_tensor = tf.get_default_graph().get_tensor_by_name(
                    'image_tensor:0')

                # Run inference
                print("running {}/{}".format(i + 1, len(ordered_test_set)))
                output_dict = sess.run(tensor_dict,
                                       feed_dict={
                                           image_tensor:
                                           np.expand_dims(six_channels_im, 0)
                                       })

                # all outputs are float32 numpy arrays, so convert types as appropriate
                output_dict['num_detections'] = int(
                    output_dict['num_detections'][0])
                output_dict['detection_classes'] = output_dict[
                    'detection_classes'][0].astype(np.uint8)
                output_dict['detection_boxes'] = output_dict[
                    'detection_boxes'][0]
                output_dict['detection_scores'] = output_dict[
                    'detection_scores'][0]
                if 'detection_masks' in output_dict:
                    output_dict['detection_masks'] = output_dict[
                        'detection_masks'][0]

                output_dict_list.append(output_dict)
    return output_dict_list
Ejemplo n.º 50
0
    def createComputationalGraph(self):
        """
            Create the computational graph in tensorflow
        """

        ######################################
        # Variables that are needed
        self.u = tf.Variable(np.zeros(self.N), dtype=self.dtype)
        self.rLowPass = tf.Variable(np.zeros(self.N), dtype=self.dtype)
        uNoise = tf.Variable(np.zeros(self.N), dtype=self.dtype)
        self.uNoiseLowPass = tf.Variable(np.zeros(self.N), dtype=self.dtype)
        self.uDotOld = tf.Variable(np.zeros(self.N), dtype=self.dtype)
        self.eligNow = tf.Variable(
            np.zeros((self.N, self.N)), dtype=self.dtype)
        self.eligibility = tf.Variable(
            np.zeros((self.N, self.N)), dtype=self.dtype)
        self.regEligibility = tf.Variable(
            np.zeros((self.N, self.N)), dtype=self.dtype)
        self.regEligibilityBorder = tf.Variable(
            np.zeros((self.N, self.N)), dtype=self.dtype)
        self.wTfNoWta = tf.Variable(self.WnoWta, dtype=self.dtype)
        self.wTfOnlyWta = tf.Variable(self.onlyWta, dtype=self.dtype)
        inputMask = self.input.getInput(0.)[2]
        self.inputMaskTf = tf.Variable(inputMask,
                                       dtype=self.dtype)
        outputMask = self.target.getTarget(self.T)[2]
        self.outputMaskTf = tf.Variable(outputMask,
                                        dtype=self.dtype)
        self.biasTf = tf.Variable(self.biasVector, dtype=self.dtype)
        # set up a mask for the learned weights in self.wTfNoWta
        # note that W.mask must omit the WTA network
        self.wNoWtaMask = tf.Variable(self.Wplastic.astype(float), dtype=self.dtype)

        #####################################
        # Variables for debugging
        self.error = tf.Variable(np.zeros(self.N), dtype=self.dtype)


        #####################################
        # Placeholders
        # The only datatransfer between the GPU and the CPU should be the
        # input to the input layer and the modulatory signal
        self.inputTf = tf.placeholder(dtype=self.dtype,
                                      shape=(self.N))
        self.inputPrimeTf = tf.placeholder(dtype=self.dtype,
                                           shape=(self.N))
        self.modulator = tf.placeholder(dtype=self.dtype,
                                        shape=())

        ####################################
        # Aux variables for the calculations
        nInput = len(np.where(inputMask == 1)[0])
        nOutput = len(np.where(outputMask == 1)[0])
        nFull = len(inputMask)

        #####################################################
        # Start the actual calculations for the comp graph  #
        #####################################################

        ####################################
        # Calculate the activations functions using the updated values
        self.rho = self.actFunc(self.u)
        rhoPrime = self.actFuncPrime(self.u)
        rhoPrimePrime = self.actFuncPrimePrime(self.u)
        self.rhoOutput = self.actFunc(self.u)

        ###################################
        # Update the exploration noise on the output neurons
        uNoiseOut = tf.slice(uNoise, [nFull - nOutput], [-1])
        duOutNoise = self.noiseTheta * (self.noiseMean - uNoiseOut) * self.timeStep + self.noiseSigma * \
            np.sqrt(self.timeStep) * \
            tf.random_normal([nOutput], mean=0., stddev=1.0, dtype=self.dtype)
        updateNoise = tf.scatter_update(uNoise,
                                        np.arange(nFull - nOutput, nFull),
                                        uNoiseOut + duOutNoise)
        # Update the low-pass noise
        with tf.control_dependencies([updateNoise]):
            self.uNoiseLowPass = self.uNoiseLowPass + (self.timeStep/self.tau) * (uNoise - self.uNoiseLowPass)

        ####################################
        # Calculate the updates for the membrane potential and for the
        # eligibility trace

        with tf.control_dependencies([self.uNoiseLowPass, updateNoise, rhoPrime, rhoPrimePrime]):

            # frequently used tensors are claculated early on
            wNoWtaT = tf.transpose(self.wTfNoWta)
            wNoWtaRho = tfTools.tf_mat_vec_dot(self.wTfNoWta, self.rho)
            c = tfTools.tf_mat_vec_dot(wNoWtaT, self.u - wNoWtaRho - self.biasTf - self.inputTf - self.uNoiseLowPass)

            # get the matrix side of the equation
            A1 = tf.matmul(self.wTfNoWta, tf.diag(rhoPrime))
            A2 = tf.matmul(tf.diag(c), tf.diag(rhoPrimePrime))
            A3 = tf.matmul(tf.diag(rhoPrime), wNoWtaT)
            A4 = tf.matmul(tf.matmul(tf.diag(rhoPrime),wNoWtaT),
                           tf.matmul(self.wTfNoWta, tf.diag(rhoPrime)))
            A5 = self.beta * self.alphaWna * tf.matmul(self.wTfOnlyWta, tf.diag(rhoPrime))
            A = self.tau*(tf.eye(self.N) - A1 - A2 - A3 + A4 - A5)

            # get the vector side of the equation
            y1 = wNoWtaRho + self.biasTf + self.inputTf + self.alphaNoise * self.beta * uNoise - self.u
            y2 = self.tau * self.inputPrimeTf
            y3 = rhoPrime * c
            y4 = self.tau * rhoPrime * tfTools.tf_mat_vec_dot(
                                                        wNoWtaT, 
                                                        self.inputPrimeTf)
            y5 = self.beta * self.alphaWna * tfTools.tf_mat_vec_dot(
                                        self.wTfOnlyWta, self.rho)
            y6 = rhoPrime * tfTools.tf_mat_vec_dot(wNoWtaT, uNoise - self.uNoiseLowPass)
            y = y1 + y2 + y3 - y4 + y5 - y6

            # Solve the equation for uDot
            #self.uDiff = (1. / self.tau) * tf.linalg.solve(A, y)
            uDiff = tf.linalg.solve(A, tf.expand_dims(y, 1))[:, 0]
            #chol = tf.cholesky(A)
            #uDiff = tf.cholesky_solve(tf.cholesky(A), tf.expand_dims(y, 1))[:, 0]

            """
            # The regular component with lookahead
            reg = tfTools.tf_mat_vec_dot(
                self.wTfNoWta, self.rho + rhoPrime * self.uDotOld * self.tau) - self.u + self.biasTf

            # Error term from the vanilla lagrange
            eVfirst = rhoPrime * c
            eVsecond = (rhoPrimePrime * self.uDotOld) * c
            eVthird = rhoPrime * \
                tfTools.tf_mat_vec_dot(
                    wNoWtaT,
                    self.uDotOld - tfTools.tf_mat_vec_dot(
                        self.wTfNoWta,
                        rhoPrime * self.uDotOld)
                )
            eV = eVfirst + self.tau * (eVsecond + eVthird)

            # terms from the winner nudges all circuit
            regWna = self.beta * self.alphaWna * tfTools.tf_mat_vec_dot(
                self.wTfOnlyWta, self.rho + self.tau * rhoPrime * self.uDotOld)

            # Terms from the exploration noise term
            #eNoise = self.alphaNoise * self.beta * \
            #    ((uNoise) - (uOut + self.tau * uDotOut))
            eNoise = self.alphaNoise * self.beta * uNoise
            """

        #uDiff = (1. / self.tau) * (reg + eV + regWna + eNoise)
        saveOldUDot = self.uDotOld.assign(uDiff)
        updateLowPassActivity = self.rLowPass.assign((self.rLowPass + self.timeStep / self.tauEligibility * self.rho) * tf.exp(-1. * self.timeStep / self.tauEligibility))

        self.eligNowUpdate = self.eligNow.assign(tfTools.tf_outer_product(self.u - tfTools.tf_mat_vec_dot(self.wTfNoWta, self.rho) - self.biasTf, self.rho))
        errorUpdate = self.error.assign(self.u - tfTools.tf_mat_vec_dot(self.wTfNoWta, self.rho) - self.biasTf - self.inputTf)

        with tf.control_dependencies([saveOldUDot, updateLowPassActivity, self.eligNowUpdate, errorUpdate]):

            self.updateEligiblity = self.eligibility.assign(
                (self.eligibility + self.timeStep * tfTools.tf_outer_product(
                    self.u - tfTools.tf_mat_vec_dot(self.wTfNoWta, self.rho) - self.biasTf - self.inputTf - self.uNoiseLowPass, self.rho)) * tf.exp(-1. * self.timeStep / self.tauEligibility)
            )
            
            self.updateRegEligibility = self.regEligibility.assign(
                (self.regEligibility + self.timeStep * tfTools.tf_outer_product(
                    tf.nn.relu(self.uTarget - self.u),
                    self.rho)) * tf.exp(-1. * self.timeStep / self.tauEligibility)
            )

            self.updateRegEligibilityBorder = self.regEligibilityBorder.assign(
                (self.regEligibilityBorder + self.timeStep * tfTools.tf_outer_product(
                    tf.nn.relu(self.uLow - self.u) -
                    tf.nn.relu(self.u - self.uHigh),
                    self.rho)) * tf.exp(-1. * self.timeStep / self.tauEligibility)
            )

            #self.applyMembranePot = tf.scatter_update(self.u, np.arange(
            #    nInput, nFull), tf.slice(self.u, [nInput], [-1]) + self.timeStep * tf.slice(uDiff, [nInput], [-1]))

        with tf.control_dependencies([saveOldUDot, updateLowPassActivity, self.eligNowUpdate, errorUpdate, self.updateEligiblity, self.updateRegEligibility, self.updateRegEligibilityBorder]):
            self.applyMembranePot = self.u.assign(self.u + self.timeStep * uDiff)

        ###############################################
        ## Node to update the weights of the network ##
        ###############################################

        self.updateW = self.wTfNoWta.assign(self.wTfNoWta + ( 1. / self.tauEligibility) * (
            self.modulator * self.learningRate * self.eligibility * self.Wplastic + tf.math.abs(self.modulator) * self.learningRateH * self.regEligibility * self.noWnaMask + self.learningRateB * self.regEligibilityBorder * self.wNoWtaMask))
    tf.reset_default_graph()


#These two lines established the feed-forward part of the network.  This does
#the actual choosing.
weights = tf.Variable(tf.ones([num_bandits]))
chosen_action = tf.argmax(weights, 0)

#The next six lines establish the training proceedure.  We feed the reward and
#chosen action into the network
#to compute the loss, and use it to update the network.
reward_holder = tf.placeholder(shape=[1], dtype=tf.float32)
action_holder = tf.placeholder(
    shape=[1], dtype=tf.int32)  #index of slot machine to choose
responsible_weight = tf.slice(weights, action_holder,
                              [1])  #scalar weight responsible for reward
loss = -(tf.log(responsible_weight) * reward_holder)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
update = optimizer.minimize(loss)

total_episodes = 2000  #Set total number of episodes to train agent on.
total_reward = np.zeros(num_bandits)  #Set scoreboard for bandits to 0.
choices = np.zeros(num_bandits)  #How many times each bandit was chosen
e = 0.1  #Set the chance of taking a random action (exploration rate)

init = tf.global_variables_initializer()

# Launch the tensorflow graph
with tf.Session() as sess:
    sess.run(init)
    for i in range(total_episodes):
Ejemplo n.º 52
0
    def convLSTM_anim(self, graph, log_count):
        print("Python version    :", sys.version)
        print("TensorFlow version: ", tf.VERSION)
        print("Current directory : ", os.getcwd())
        LOGDIR = "/tmp/convLSTM/"

        NUM_UNROLLINGS = 3
        NUM_TRAINING_STEPS = 801
        counter = 4

        anim = pickle.load(open(self.file, 'rb'))
        anim = tf.expand_dims(anim, axis=-1)
        image = tf.slice(anim, [counter, 0, 0, 0],
                         [1, self.IM_SZ_LEN, self.IM_SZ_WID, 1])

        def newEmpty4Dtensor_1channel():
            emptyTensor = tf.zeros([self.IM_SZ_LEN, self.IM_SZ_WID])
            emptyTensor = tf.expand_dims(emptyTensor, 0)
            emptyTensor = tf.expand_dims(emptyTensor, -1)
            return emptyTensor

        def newEmpty4Dtensor_2channels():
            emptyTensor = tf.zeros([self.IM_SZ_LEN, self.IM_SZ_WID, 2])
            emptyTensor = tf.expand_dims(emptyTensor, 0)
            return emptyTensor

    # create some initializations

        initial_lstm_state = newEmpty4Dtensor_1channel()
        initial_lstm_output = newEmpty4Dtensor_1channel()
        initial_err_input = newEmpty4Dtensor_2channels()

        # The above weights are global to this definition.
        def convLstmLayer(err_inp, prev_s, prev_h):
            with tf.name_scope("LSTM"):
                inp = tf.nn.tanh(
                    tf.nn.conv2d(err_inp, self.U, [1, 1, 1, 1], padding='SAME')
                    + tf.nn.conv2d(
                        prev_h, self.W, [1, 1, 1, 1], padding='SAME') + self.B,
                    name="inp")
                g_gate = tf.nn.tanh(
                    tf.nn.conv2d(
                        err_inp, self.Ug, [1, 1, 1, 1], padding='SAME') +
                    tf.nn.conv2d(prev_h, self.Wg, [1, 1, 1, 1],
                                 padding='SAME') + self.Bg,
                    name="g_gate")  # i_gate is more common name
                f_gate = tf.nn.tanh(
                    tf.nn.conv2d(
                        err_inp, self.Uf, [1, 1, 1, 1], padding='SAME') +
                    tf.nn.conv2d(prev_h, self.Wf, [1, 1, 1, 1],
                                 padding='SAME') + self.Bf,
                    name="f_gate")
                q_gate = tf.nn.tanh(
                    tf.nn.conv2d(
                        err_inp, self.Uo, [1, 1, 1, 1], padding='SAME') +
                    tf.nn.conv2d(prev_h, self.Wo, [1, 1, 1, 1],
                                 padding='SAME') + self.Bo,
                    name="q_gate")  # o_gate is more common name
                s = tf.add(tf.multiply(f_gate, prev_s),
                           tf.multiply(g_gate, inp),
                           name="state")
                h = tf.multiply(q_gate, tf.nn.relu6(s),
                                name="output")  # Also try relu
                return s, h  # normally above is tanh

    # errorModule doesn't use variables, so doesn't undergo training

        def errorModule(image, predict):
            with tf.name_scope("ErrMod"):
                err1 = tf.nn.relu(image - predict, name="E1")
                err2 = tf.nn.relu(predict - image, name="E2")
                tensor5D = tf.stack([err1, err2], axis=3)
                tensor4D = tf.reshape(tensor5D,
                                      [1, self.IM_SZ_LEN, self.IM_SZ_WID, 2],
                                      name="PrdErr")
                return tensor4D

    # Build LSTM

        lstm_state = initial_lstm_state
        lstm_output = initial_lstm_output
        err_input = initial_err_input
        with tf.name_scope("full_model"):
            for _ in range(NUM_UNROLLINGS):  # three unrollings
                lstm_state, lstm_output = convLstmLayer(
                    err_input, lstm_state, lstm_output)
                err_input = errorModule(image, lstm_output)
                counter = counter + 1
                image = tf.slice(anim, [counter, 0, 0, 0],
                                 [1, self.IM_SZ_LEN, self.IM_SZ_WID, 1])

    #New optimizer block, uses exp decay on learning rate, added clip_by_global_norm
        loss = tf.reduce_sum(
            err_input)  # sums the values across each component of the tensor
        global_step = tf.Variable(0)

        #learning rate starts at 10, decreases by 90% every 300 steps
        learning_rate = tf.train.exponential_decay(10.0,
                                                   global_step,
                                                   300,
                                                   0.1,
                                                   staircase=True,
                                                   name='LearningRate')
        optimizer = tf.train.GradientDescentOptimizer(learning_rate)
        gradients, v = zip(*optimizer.compute_gradients(loss))
        gradients, _ = tf.clip_by_global_norm(gradients, 1.25)
        optimizer = optimizer.apply_gradients(zip(gradients, v),
                                              global_step=global_step)

        with tf.name_scope("initializations"):
            tf.summary.image("initial_lstm_state", initial_lstm_state, 3)
            tf.summary.image("initial_lstm_output", initial_lstm_output, 3)
            tf.summary.image(
                "initial_error1",
                tf.slice(initial_err_input, [0, 0, 0, 0],
                         [1, self.IM_SZ_LEN, self.IM_SZ_WID, 1]), 3)
            tf.summary.image(
                "initial_error2",
                tf.slice(initial_err_input, [0, 0, 0, 1],
                         [1, self.IM_SZ_LEN, self.IM_SZ_WID, 1]), 3)

        with tf.name_scope("input"):
            tf.summary.image("image", image, 3)

        with tf.name_scope("lstm"):
            tf.summary.image("lstm_out", lstm_output, 3)
            tf.summary.image("lstm_state", lstm_state, 3)

        with tf.name_scope("error"):
            tf.summary.image(
                "perror_1",
                tf.slice(err_input, [0, 0, 0, 0],
                         [1, self.IM_SZ_LEN, self.IM_SZ_WID, 1]), 3)
            tf.summary.image(
                "perror_2",
                tf.slice(err_input, [0, 0, 0, 1],
                         [1, self.IM_SZ_LEN, self.IM_SZ_WID, 1]), 3)

        with tf.name_scope('optimizer'):
            tf.summary.scalar('loss', loss)
            tf.summary.scalar('learning_rate', learning_rate)

        with tf.name_scope('weights'):
            with tf.name_scope('input_update'):
                newU1 = tf.slice(self.U, [0, 0, 0, 0], [5, 5, 1, 1])
                newU2 = tf.slice(self.U, [0, 0, 1, 0], [5, 5, 1, 1])
                newW = tf.slice(self.W, [0, 0, 0, 0], [5, 5, 1, 1])
                newU1 = tf.squeeze(newU1)  #now a viewable [5x5] matrix
                newU2 = tf.squeeze(newU2)
                newW = tf.squeeze(newW)
                newU1 = tf.reshape(newU1, [1, 5, 5, 1])
                newU2 = tf.reshape(newU2, [1, 5, 5, 1])
                newW = tf.reshape(newW, [1, 5, 5, 1])
                tf.summary.image('U1', newU1)
                tf.summary.image('U2', newU2)
                tf.summary.image('W', newW)
                tf.summary.image('B', self.B)

            with tf.name_scope('input_gate'):
                newUg1 = tf.slice(self.Ug, [0, 0, 0, 0], [5, 5, 1, 1])
                newUg2 = tf.slice(self.Ug, [0, 0, 1, 0], [5, 5, 1, 1])
                newWg = tf.slice(self.Wg, [0, 0, 0, 0], [5, 5, 1, 1])
                newUg1 = tf.squeeze(newUg1)  #now a viewable [5x5] matrix
                newUg2 = tf.squeeze(newUg2)
                newWg = tf.squeeze(newWg)
                newUg1 = tf.reshape(newUg1, [1, 5, 5, 1])
                newUg2 = tf.reshape(newUg2, [1, 5, 5, 1])
                newWg = tf.reshape(newWg, [1, 5, 5, 1])
                tf.summary.image('Ug1', newUg1)
                tf.summary.image('Ug2', newUg2)
                tf.summary.image('Wg', newWg)
                tf.summary.image('Bg', self.Bg)

            with tf.name_scope('forget_gate'):
                newUf1 = tf.slice(self.Uf, [0, 0, 0, 0], [5, 5, 1, 1])
                newUf2 = tf.slice(self.Uf, [0, 0, 1, 0], [5, 5, 1, 1])
                newWf = tf.slice(self.Wf, [0, 0, 0, 0], [5, 5, 1, 1])
                newUf1 = tf.squeeze(newUf1)  #now a viewable [5x5] matrix
                newUf2 = tf.squeeze(newUf2)
                newWf = tf.squeeze(newWf)
                newUf1 = tf.reshape(newUf1, [1, 5, 5, 1])
                newUf2 = tf.reshape(newUf2, [1, 5, 5, 1])
                newWf = tf.reshape(newWf, [1, 5, 5, 1])
                tf.summary.image('Uf1', newUf1)
                tf.summary.image('Uf2', newUf2)
                tf.summary.image('Wf', newWf)
                tf.summary.image('Bf', self.Bf)

            with tf.name_scope('output_gate'):
                newUo1 = tf.slice(self.Uo, [0, 0, 0, 0], [5, 5, 1, 1])
                newUo2 = tf.slice(self.Uo, [0, 0, 1, 0], [5, 5, 1, 1])
                newWo = tf.slice(self.Wo, [0, 0, 0, 0], [5, 5, 1, 1])
                newUo1 = tf.squeeze(newUo1)  #now a viewable [5x5] matrix
                newUo2 = tf.squeeze(newUo2)
                newWo = tf.squeeze(newWo)
                newUo1 = tf.reshape(newUo1, [1, 5, 5, 1])
                newUo2 = tf.reshape(newUo2, [1, 5, 5, 1])
                newWo = tf.reshape(newWo, [1, 5, 5, 1])
                tf.summary.image('Uo1', newUo1)
                tf.summary.image('Uo2', newUo2)
                tf.summary.image('Wo', newWo)
                tf.summary.image('Bo', self.Bo)


# Start training
        with tf.Session(graph=graph) as sess:
            tf.global_variables_initializer().run()

            # Create graph summary
            # Use a different log file each time you run the program.
            msumm = tf.summary.merge_all()
            writer = tf.summary.FileWriter(
                LOGDIR +
                str(log_count))  # += 1 for each run till /tmp is cleard
            log_count += 1
            writer.add_graph(sess.graph)

            for step in range(NUM_TRAINING_STEPS):  # 0 to 100
                if step % 1 == 0:
                    ms = sess.run(msumm)
                    writer.add_summary(ms, step)
                _, l, predictions = sess.run([optimizer, loss, lstm_output])

                print("Step: ", step)
                print("Loss: ", l)
Ejemplo n.º 53
0
def main(unused_argv):
    tf.logging.set_verbosity(tf.logging.INFO)

    # Get dataset-dependent information.
    dataset = data_generator.Dataset(
        dataset_name=FLAGS.dataset,
        split_name=FLAGS.vis_split,
        dataset_dir=FLAGS.dataset_dir,
        batch_size=FLAGS.vis_batch_size,
        crop_size=[int(sz) for sz in FLAGS.vis_crop_size],
        min_resize_value=FLAGS.min_resize_value,
        max_resize_value=FLAGS.max_resize_value,
        resize_factor=FLAGS.resize_factor,
        model_variant=FLAGS.model_variant,
        is_training=False,
        should_shuffle=False,
        should_repeat=False)

    train_id_to_eval_id = None
    if dataset.dataset_name == data_generator.get_cityscapes_dataset_name():
        tf.logging.info('Cityscapes requires converting train_id to eval_id.')
        train_id_to_eval_id = _CITYSCAPES_TRAIN_ID_TO_EVAL_ID

    # Prepare for visualization.
    tf.gfile.MakeDirs(FLAGS.vis_logdir)
    save_dir = os.path.join(FLAGS.vis_logdir, _SEMANTIC_PREDICTION_SAVE_FOLDER)
    tf.gfile.MakeDirs(save_dir)
    raw_save_dir = os.path.join(FLAGS.vis_logdir,
                                _RAW_SEMANTIC_PREDICTION_SAVE_FOLDER)
    tf.gfile.MakeDirs(raw_save_dir)

    tf.logging.info('Visualizing on %s set', FLAGS.vis_split)

    with tf.Graph().as_default():
        samples = dataset.get_one_shot_iterator().get_next()

        model_options = common.ModelOptions(
            outputs_to_num_classes={
                common.OUTPUT_TYPE: dataset.num_of_classes
            },
            crop_size=[int(sz) for sz in FLAGS.vis_crop_size],
            atrous_rates=FLAGS.atrous_rates,
            output_stride=FLAGS.output_stride)

        if tuple(FLAGS.eval_scales) == (1.0, ):
            tf.logging.info('Performing single-scale test.')
            predictions = model.predict_labels(
                samples[common.IMAGE],
                model_options=model_options,
                image_pyramid=FLAGS.image_pyramid)
        else:
            tf.logging.info('Performing multi-scale test.')
            if FLAGS.quantize_delay_step >= 0:
                raise ValueError(
                    'Quantize mode is not supported with multi-scale test.')
            predictions = model.predict_labels_multi_scale(
                samples[common.IMAGE],
                model_options=model_options,
                eval_scales=FLAGS.eval_scales,
                add_flipped_images=FLAGS.add_flipped_images)
        predictions = predictions[common.OUTPUT_TYPE]

        if FLAGS.min_resize_value and FLAGS.max_resize_value:
            # Only support batch_size = 1, since we assume the dimensions of original
            # image after tf.squeeze is [height, width, 3].
            assert FLAGS.vis_batch_size == 1

            # Reverse the resizing and padding operations performed in preprocessing.
            # First, we slice the valid regions (i.e., remove padded region) and then
            # we resize the predictions back.
            original_image = tf.squeeze(samples[common.ORIGINAL_IMAGE])
            original_image_shape = tf.shape(original_image)
            predictions = tf.slice(
                predictions, [0, 0, 0],
                [1, original_image_shape[0], original_image_shape[1]])
            resized_shape = tf.to_int32([
                tf.squeeze(samples[common.HEIGHT]),
                tf.squeeze(samples[common.WIDTH])
            ])
            predictions = tf.squeeze(
                tf.image.resize_images(
                    tf.expand_dims(predictions, 3),
                    resized_shape,
                    method=tf.image.ResizeMethod.NEAREST_NEIGHBOR,
                    align_corners=True), 3)

        tf.train.get_or_create_global_step()
        if FLAGS.quantize_delay_step >= 0:
            contrib_quantize.create_eval_graph()

        num_iteration = 0
        max_num_iteration = FLAGS.max_number_of_iterations

        checkpoints_iterator = contrib_training.checkpoints_iterator(
            FLAGS.checkpoint_dir, min_interval_secs=FLAGS.eval_interval_secs)
        for checkpoint_path in checkpoints_iterator:
            num_iteration += 1
            tf.logging.info('Starting visualization at ' +
                            time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime()))
            tf.logging.info('Visualizing with model %s', checkpoint_path)

            scaffold = tf.train.Scaffold(
                init_op=tf.global_variables_initializer())
            session_creator = tf.train.ChiefSessionCreator(
                scaffold=scaffold,
                master=FLAGS.master,
                checkpoint_filename_with_path=checkpoint_path)
            with tf.train.MonitoredSession(session_creator=session_creator,
                                           hooks=None) as sess:
                batch = 0
                image_id_offset = 0

                while not sess.should_stop():
                    tf.logging.info('Visualizing batch %d', batch + 1)
                    _process_batch(
                        sess=sess,
                        original_images=samples[common.ORIGINAL_IMAGE],
                        semantic_predictions=predictions,
                        image_names=samples[common.IMAGE_NAME],
                        image_heights=samples[common.HEIGHT],
                        image_widths=samples[common.WIDTH],
                        image_id_offset=image_id_offset,
                        save_dir=save_dir,
                        raw_save_dir=raw_save_dir,
                        train_id_to_eval_id=train_id_to_eval_id)
                    image_id_offset += FLAGS.vis_batch_size
                    batch += 1

            tf.logging.info('Finished visualization at ' +
                            time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime()))
            if max_num_iteration > 0 and num_iteration >= max_num_iteration:
                break
Ejemplo n.º 54
0
    def _build_model(self):
        # filename_queue = tf.train.string_input_producer(['american_train.tfrecords', 'british_train.tfrecords'], num_epochs=200)
        # reader = tf.TFRecordReader()
        count = tf.Variable(0, name='counter', dtype=tf.float32)
        inc = tf.assign_add(count, 1, name='increment')
        # self.token = tf.Variable(0, name='token', dtype=tf.float32)
        # self.X_source = train_source_x
        # tf.Print(self.X_source, [self.X_source])
        # self.y_source = train_source_y
        # self.X_target = train_target_x
        # self.y_target = train_target_y
        # self.X_source = source_batch_x
        # self.y_source = source_batch_y
        # self.X_target = target_batch_x
        # self.y_target = target_batch_y

        # source_x, source_y = custom_runner.queue.dequeue_many(128)
        # target_x, target_y = custom_runner.queue2.dequeue_many(128)
        # # self.X_source = source_x
        # self.token2 = source_x + 1
        # self.y_source = source_batch_y
        # self.X_target = target_batch_x
        # self.y_target = target_batch_y
        # source_x, source_y, target_x, target_y = input_pipeline()

        self.X = tf.concat([source_x, target_x], 0)
        self.y = tf.concat([source_y, target_y], 0)
        self.y = tf.cast(self.y, dtype=tf.int32)
        # self.y = tf.cast(self.y, dtype=tf.int32)
        self.y = tf.one_hot(self.y, depth=1677)
        # self.X = tf.reshape(self.X, [None, 31*160, 1, 1])
        # self.domain = tf.concat([self.domain_0, self.domain_1], 0)
        self.train = tf.constant(True, dtype=tf.bool)
        self.l = 2 / (1 + tf.exp(-10 * (inc / num_steps))) - 1
        self.lr = 0.005 / (tf.pow(1 + 10 * (inc / num_steps), 0.75))
        # self.l = tf.constant(0.4, dtype=tf.float32)
        # self.lr = tf.constant(0.01, dtype=tf.float32)
        # self.domain_0 = tf.zeros([128], dtype=tf.int32)
        # self.domain_1 = tf.ones([128], dtype=tf.int32)
        self.domain_0 = tf.zeros([128], dtype=tf.int32)
        self.domain_1 = tf.ones([128], dtype=tf.int32)
        rand = tf.random_uniform(shape=[], maxval=1)
        self.domain1 = lambda: tf.concat([self.domain_0, self.domain_1], 0)
        self.domain2 = lambda: tf.concat([self.domain_1, self.domain_0], 0)

        self.domain = tf.cond(tf.greater(rand, 0.9), self.domain2,
                              self.domain1)
        # self.domain = tf.concat([self.domain_0, self.domain_1], 0)
        self.domain = tf.one_hot(self.domain, depth=2)
        self.keep_prob = tf.constant(0.5, dtype=tf.float32)
        # X_input = (tf.cast(self.X, tf.float32) - pixel_mean) / 255.
        # CNN model for feature extractio
        with tf.name_scope('feature_extractor'):
            self.W_conv0 = weight_variable([256, 1, 1, 64], 'W_Conv0')
            self.b_conv0 = bias_variable([64], 'b_Conv0')
            # h_conv0 = conv2d(self.X, W_conv0, [1, 31,1,1])+ b_conv0
            # h_conv0 = tf.contrib.layers.batch_norm(h_conv0, center=True, scale=True, is_training=self.train)
            # h_conv0 = tf.nn.relu(h_conv0)
            # print(self.X.get_shape().as_list())
            h_conv0 = tf.nn.relu(
                conv2d(self.X, self.W_conv0, [1, 31, 1, 1]) + self.b_conv0)
            tf.Print(h_conv0, [h_conv0])
            h_pool0 = max_pool_2x2(h_conv0, 2, 2)
            tf.summary.histogram('Conv 0 weight', self.W_conv0)
            tf.summary.histogram('Conv 0 bias', self.b_conv0)
            tf.summary.histogram('Conv 0 activation', h_conv0)

            self.W_conv1 = weight_variable([15, 1, 64, 128], 'W_conv1')
            self.b_conv1 = bias_variable([128], 'b_conv1')
            # h_conv1 = conv2d(h_pool0, W_conv1, [1,1,1,1])+b_conv1
            # h_conv1 = tf.contrib.layers.batch_norm(h_conv1, center=True, scale=True, is_training=self.train)
            # h_conv1 = tf.nn.relu(h_conv1)
            h_conv1 = tf.nn.tanh(
                conv2d(h_pool0, self.W_conv1, [1, 1, 1, 1]) + self.b_conv1)
            h_pool1 = max_pool_2x2(h_conv1, 2, 2)
            tf.summary.histogram('Conv 1 weight', self.W_conv1)
            tf.summary.histogram('Conv 1 bias', self.b_conv1)
            tf.summary.histogram('Conv 1 activation', h_conv1)
            a, b, c, d = h_pool1.get_shape().as_list()

            # The domain-invariant feature
            self.feature = tf.reshape(h_pool1, [-1, b * c * d])

        # MLP for class prediction
        with tf.name_scope('label_predictor'):
            # Switches to route target examples (second half of batch) differently
            # depending on train or test mode.
            all_features = lambda: self.feature
            source_features = lambda: tf.slice(self.feature, [0, 0],
                                               [batch_size, -1])
            #            classify_feats = tf.cond(self.train, source_features, all_features)
            classify_feats = self.feature
            all_labels = lambda: self.y
            source_labels = lambda: tf.slice(self.y, [0, 0], [batch_size, -1])
            #            self.classify_labels = tf.cond(self.train, source_labels, all_labels)
            self.classify_labels = self.y
            self.W_fc0 = weight_variable([b * c * d, 1024 * 2], 'W_fc0')
            self.b_fc0 = bias_variable([1024 * 2], 'b_fc0')
            h_fc0 = tf.nn.relu(
                tf.matmul(classify_feats, self.W_fc0) + self.b_fc0)
            tf.summary.histogram('fc0 weight', self.W_fc0)
            tf.summary.histogram('fc0 bias', self.b_fc0)
            tf.summary.histogram('fc0 activation', h_fc0)
            h_fc0 = tf.nn.dropout(h_fc0, self.keep_prob)

            self.W_fc1 = weight_variable([1024 * 2, 1024 * 2], 'W_fc1')
            self.b_fc1 = bias_variable([1024 * 2], 'b_fc1')
            h_fc1 = tf.nn.relu(tf.matmul(h_fc0, self.W_fc1) + self.b_fc1)
            tf.summary.histogram('fc1 weight', self.W_fc1)
            tf.summary.histogram('fc1 bias', self.b_fc1)
            tf.summary.histogram('fc1 activation', h_fc1)
            h_fc1 = tf.nn.dropout(h_fc1, self.keep_prob)

            self.W_fc2 = weight_variable([1024 * 2, 1024 * 2], 'W_fc2')
            self.b_fc2 = bias_variable([1024 * 2], 'b_fc2')
            h_fc2 = tf.nn.relu(tf.matmul(h_fc1, self.W_fc2) + self.b_fc2)
            tf.summary.histogram('fc2 weight', self.W_fc2)
            tf.summary.histogram('fc2 bias', self.b_fc2)
            tf.summary.histogram('fc2 activation', h_fc2)
            h_fc2 = tf.nn.dropout(h_fc2, self.keep_prob)

            self.W_fc3 = weight_variable([1024 * 2, 1024 * 2], 'W_fc3')
            self.b_fc3 = bias_variable([1024 * 2], 'b_fc3')
            h_fc3 = tf.nn.relu(tf.matmul(h_fc2, self.W_fc3) + self.b_fc3)
            tf.summary.histogram('fc3 weight', self.W_fc3)
            tf.summary.histogram('fc3 bias', self.b_fc3)
            tf.summary.histogram('fc3 activation', h_fc3)
            h_fc3 = tf.nn.dropout(h_fc3, self.keep_prob)

            self.W_fc4 = weight_variable([1024 * 2, 1024 * 2], 'W_fc4')
            self.b_fc4 = bias_variable([1024 * 2], 'b_fc4')
            h_fc4 = tf.nn.relu(tf.matmul(h_fc3, self.W_fc4) + self.b_fc4)
            tf.summary.histogram('fc4 weight', self.W_fc4)
            tf.summary.histogram('fc4 bias', self.b_fc4)
            tf.summary.histogram('fc4 activation', h_fc4)
            h_fc4 = tf.nn.dropout(h_fc4, self.keep_prob)

            self.W_fc5 = weight_variable([1024 * 2, 1024 * 2], 'W_fc5')
            self.b_fc5 = bias_variable([1024 * 2], 'b_fc5')
            h_fc5 = tf.nn.relu(tf.matmul(h_fc4, self.W_fc5) + self.b_fc5)
            tf.summary.histogram('fc5 weight', self.W_fc5)
            tf.summary.histogram('fc5 bias', self.b_fc5)
            tf.summary.histogram('fc5 activation', h_fc5)
            h_fc5 = tf.nn.dropout(h_fc5, self.keep_prob)

            self.W_fc6 = weight_variable([1024 * 2, 1024 * 2], 'W_fc6')
            self.b_fc6 = bias_variable([1024 * 2], 'b_fc6')
            h_fc6 = tf.nn.relu(tf.matmul(h_fc5, self.W_fc6) + self.b_fc6)
            tf.summary.histogram('fc6 weight', self.W_fc6)
            tf.summary.histogram('fc6 bias', self.b_fc6)
            tf.summary.histogram('fc6 activation', h_fc6)
            h_fc6 = tf.nn.dropout(h_fc6, self.keep_prob)

            self.W_fc7 = weight_variable([1024 * 2, 1677], 'W_fc7')
            self.b_fc7 = bias_variable([1677], 'b_fc7')

            logits = tf.matmul(h_fc6, self.W_fc7) + self.b_fc7
            tf.summary.histogram('fc7 weight', self.W_fc7)
            tf.summary.histogram('fc7 bias', self.b_fc7)
            tf.summary.histogram('fc7 activation', logits)
        with tf.name_scope('predicted_loss'):

            self.pred = tf.nn.softmax(logits)
            self.pred_loss = tf.nn.softmax_cross_entropy_with_logits(
                logits=logits, labels=self.classify_labels)

        # Small MLP for domain prediction with adversarial loss
        with tf.name_scope('domain_predictor'):
            # Flip the gradient when backpropagating through this operation
            feat = flip_gradient(self.feature, self.l)

            d_W_fc0 = weight_variable([b * c * d, 1024 * 2], 'd_W_fc0')
            d_b_fc0 = bias_variable([1024 * 2], 'd_b_fc0')
            d_h_fc0 = tf.nn.relu(tf.matmul(feat, d_W_fc0) + d_b_fc0)
            tf.summary.histogram('domain fc0 weights', d_W_fc0)
            tf.summary.histogram('domain fc0 bias', d_b_fc0)
            tf.summary.histogram('domain fc0 activation', d_h_fc0)
            d_h_fc0 = tf.nn.dropout(d_h_fc0, keep_prob=self.keep_prob)

            d_W_fc1 = weight_variable([1024 * 2, 1024 * 2], 'd_W_fc1')
            d_b_fc1 = bias_variable([1024 * 2], 'd_b_fc1')
            d_h_fc1 = tf.nn.relu(tf.matmul(d_h_fc0, d_W_fc1) + d_b_fc1)
            tf.summary.histogram('domain fc1 weights', d_W_fc1)
            tf.summary.histogram('domain fc1 bias', d_b_fc1)
            tf.summary.histogram('domain fc1 activation', d_h_fc1)
            d_h_fc1 = tf.nn.dropout(d_h_fc1, keep_prob=self.keep_prob)

            d_W_fc2 = weight_variable([1024 * 2, 1024 * 2], 'd_W_fc2')
            d_b_fc2 = bias_variable([1024 * 2], 'd_b_fc2')
            d_h_fc2 = tf.nn.relu(tf.matmul(d_h_fc1, d_W_fc2) + d_b_fc2)
            tf.summary.histogram('domain fc2 weights', d_W_fc2)
            tf.summary.histogram('domain fc2 bias', d_b_fc2)
            tf.summary.histogram('domain fc2 activation', d_h_fc2)
            d_h_fc2 = tf.nn.dropout(d_h_fc2, keep_prob=self.keep_prob)

            d_W_fc3 = weight_variable([1024 * 2, 1024 * 2], 'd_W_fc3')
            d_b_fc3 = bias_variable([1024 * 2], 'd_b_fc3')
            d_h_fc3 = tf.nn.relu(tf.matmul(d_h_fc2, d_W_fc3) + d_b_fc3)
            tf.summary.histogram('domain fc3 weights', d_W_fc3)
            tf.summary.histogram('domain fc3 bias', d_b_fc3)
            tf.summary.histogram('domain fc3 activation', d_h_fc3)
            d_h_fc3 = tf.nn.dropout(d_h_fc3, keep_prob=self.keep_prob)

            d_W_fc4 = weight_variable([1024 * 2, 1024 * 2], 'd_W_fc4')
            d_b_fc4 = bias_variable([1024 * 2], 'd_b_fc4')
            d_h_fc4 = tf.nn.relu(tf.matmul(d_h_fc3, d_W_fc4) + d_b_fc4)
            tf.summary.histogram('domain fc4 weights', d_W_fc4)
            tf.summary.histogram('domain fc4 bias', d_b_fc4)
            tf.summary.histogram('domain fc4 activation', d_h_fc4)
            d_h_fc4 = tf.nn.dropout(d_h_fc4, keep_prob=self.keep_prob)

            d_W_fc5 = weight_variable([1024 * 2, 2], 'd_W_fc5')
            d_b_fc5 = bias_variable([2], 'd_b_fc5')
            d_logits = tf.matmul(d_h_fc4, d_W_fc5) + d_b_fc5
            tf.summary.histogram('domain fc5 weigth', d_W_fc5)
            tf.summary.histogram('domain fc5 bias', d_b_fc5)
            tf.summary.histogram('domain fc5 activation', d_logits)
        with tf.name_scope('domain_loss'):

            self.domain_pred = tf.nn.softmax(d_logits)
            self.domain_loss = tf.nn.softmax_cross_entropy_with_logits(
                logits=d_logits, labels=self.domain)
Ejemplo n.º 55
0
def _dynamic_slice(operand, *start_indices, slice_sizes=None):
  return tf.slice(operand, tf.stack(start_indices), slice_sizes)
Ejemplo n.º 56
0
def decode_record_seg(record, level_num, is_point_feature):
    mesh_name = record['mesh_name']
    label = tf.decode_raw(record['label'], tf.int32)
    label = tf.cast(label, tf.int32)
    shape_list = tf.decode_raw(record['shape'], tf.int32)
    shape_list = tf.reshape(shape_list, (-1, 4))

    if (is_point_feature):
        input_feature = tf.decode_raw(record['input_feature'], tf.float32)
        feature_channel = record['feature_channel']
        feature_channel = tf.cast(feature_channel, tf.int32)
        input_feature = tf.reshape(input_feature,
                                   (shape_list[0][0], feature_channel))
    else:
        feature_channel = record['feature_channel']
        feature_channel = tf.cast(feature_channel, tf.int32)
        input_feature = tf.decode_raw(record['input_feature'], tf.float32)
        input_feature = tf.reshape(
            input_feature,
            (shape_list[0][0] * shape_list[0][1] * shape_list[0][2] *
             shape_list[0][3], feature_channel))

    conv_offset = tf.decode_raw(record['conv/offset'], tf.int32)
    conv_indices_list = tf.decode_raw(record['conv/indices'], tf.int32)
    conv_weights_list = tf.decode_raw(record['conv/weights'], tf.float32)

    conv_begin = 0
    conv_matrix_list = []
    for level_i in range(level_num):
        conv_indices = tf.slice(conv_indices_list, [2 * conv_begin],
                                [2 * conv_offset[level_i]])
        conv_indices = tf.cast(
            tf.reshape(conv_indices, (conv_offset[level_i], 2)), tf.int64)
        conv_weights = tf.slice(conv_weights_list, [conv_begin],
                                [conv_offset[level_i]])
        conv_matrix_list.append(
            Conv_Matrix(shape_list[level_i][0], shape_list[level_i][1],
                        shape_list[level_i][2], shape_list[level_i][3],
                        conv_indices, conv_weights))
        conv_begin = conv_begin + conv_offset[level_i]

    maxpooling_offset = tf.decode_raw(record['maxpool/offset'], tf.int32)
    maxpooling_arg_list = tf.decode_raw(record['maxpool/arg'], tf.int32)
    maxpooling_indices_list = tf.decode_raw(record['maxpool/indices'],
                                            tf.int32)
    unpooling_indices_list = tf.decode_raw(record['unpooling/indices'],
                                           tf.int32)

    maxpooling_matrix_list = []
    unpooling_list = []

    pool_begin = 0
    for level_i in range(level_num - 1):
        maxpooling_indices = tf.slice(maxpooling_indices_list,
                                      [2 * pool_begin],
                                      [2 * maxpooling_offset[level_i]])
        maxpooling_indices = tf.cast(
            tf.reshape(maxpooling_indices, (maxpooling_offset[level_i], 2)),
            tf.int64)
        maxpooling_values = tf.ones(shape=[maxpooling_offset[level_i]],
                                    dtype=tf.float32)

        maxpooling_matrix_list.append(
            MaxPooling_Matrix(shape_list[level_i][0],
                              shape_list[level_i + 1][0],
                              shape_list[level_i][1], maxpooling_indices,
                              maxpooling_values, maxpooling_arg_list[level_i]))

        unpooling_indices = tf.slice(unpooling_indices_list, [2 * pool_begin],
                                     [2 * maxpooling_offset[level_i]])
        unpooling_values = tf.ones(shape=[maxpooling_offset[level_i]],
                                   dtype=tf.float32)
        unpooling_indices = tf.cast(
            tf.reshape(unpooling_indices, (maxpooling_offset[level_i], 2)),
            tf.int64)
        unpooling_list.append(
            UnPooling_Matrix(shape_list[level_i][0],
                             shape_list[level_i + 1][0],
                             shape_list[level_i][1], unpooling_indices,
                             unpooling_values))

        pool_begin = pool_begin + maxpooling_offset[level_i]
    return mesh_name, label, shape_list, input_feature, maxpooling_matrix_list, maxpooling_arg_list, conv_matrix_list, unpooling_list
Ejemplo n.º 57
0
def train_image(image_buffer,
                height,
                width,
                bbox,
                batch_position,
                resize_method,
                distortions,
                scope=None,
                summary_verbosity=0,
                distort_color_in_yiq=False,
                fuse_decode_and_crop=False):
    """Distort one image for training a network.

  Distorting images provides a useful technique for augmenting the data
  set during training in order to make the network invariant to aspects
  of the image that do not effect the label.

  Args:
    image_buffer: scalar string Tensor representing the raw JPEG image buffer.
    height: integer
    width: integer
    bbox: 3-D float Tensor of bounding boxes arranged [1, num_boxes, coords]
      where each coordinate is [0, 1) and the coordinates are arranged
      as [ymin, xmin, ymax, xmax].
    batch_position: position of the image in a batch, which affects how images
      are distorted and resized. NOTE: this argument can be an integer or a
      tensor
    resize_method: round_robin, nearest, bilinear, bicubic, or area.
    distortions: If true, apply full distortions for image colors.
    scope: Optional scope for op_scope.
    summary_verbosity: Verbosity level for summary ops. Pass 0 to disable both
      summaries and checkpoints.
    distort_color_in_yiq: distort color of input images in YIQ space.
    fuse_decode_and_crop: fuse the decode/crop operation.
  Returns:
    3-D float Tensor of distorted image used for training.
  """
    # with tf.op_scope([image, height, width, bbox], scope, 'distort_image'):
    # with tf.name_scope(scope, 'distort_image', [image, height, width, bbox]):
    with tf.name_scope(scope or 'distort_image'):
        # A large fraction of image datasets contain a human-annotated bounding box
        # delineating the region of the image containing the object of interest.  We
        # choose to create a new bounding box for the object which is a randomly
        # distorted version of the human-annotated bounding box that obeys an
        # allowed range of aspect ratios, sizes and overlap with the human-annotated
        # bounding box. If no box is supplied, then we assume the bounding box is
        # the entire image.
        sample_distorted_bounding_box = tf.image.sample_distorted_bounding_box(
            tf.image.extract_jpeg_shape(image_buffer),
            bounding_boxes=bbox,
            min_object_covered=0.1,
            aspect_ratio_range=[0.75, 1.33],
            area_range=[0.05, 1.0],
            max_attempts=100,
            use_image_if_no_bounding_boxes=True)
        bbox_begin, bbox_size, distort_bbox = sample_distorted_bounding_box
        if summary_verbosity >= 3:
            image = tf.image.decode_jpeg(image_buffer,
                                         channels=3,
                                         dct_method='INTEGER_FAST')
            image = tf.image.convert_image_dtype(image, dtype=tf.float32)
            image_with_distorted_box = tf.image.draw_bounding_boxes(
                tf.expand_dims(image, 0), distort_bbox)
            tf.summary.image('images_with_distorted_bounding_box',
                             image_with_distorted_box)

        # Crop the image to the specified bounding box.
        if fuse_decode_and_crop:
            offset_y, offset_x, _ = tf.unstack(bbox_begin)
            target_height, target_width, _ = tf.unstack(bbox_size)
            crop_window = tf.stack(
                [offset_y, offset_x, target_height, target_width])
            image = tf.image.decode_and_crop_jpeg(image_buffer,
                                                  crop_window,
                                                  channels=3)
        else:
            image = tf.image.decode_jpeg(image_buffer,
                                         channels=3,
                                         dct_method='INTEGER_FAST')
            image = tf.slice(image, bbox_begin, bbox_size)

        distorted_image = tf.image.random_flip_left_right(image)

        # This resizing operation may distort the images because the aspect
        # ratio is not respected.
        image_resize_method = get_image_resize_method(resize_method,
                                                      batch_position)
        if cnn_util.tensorflow_version() >= 11:
            distorted_image = tf.image.resize_images(distorted_image,
                                                     [height, width],
                                                     image_resize_method,
                                                     align_corners=False)
        else:
            distorted_image = tf.image.resize_images(distorted_image,
                                                     height,
                                                     width,
                                                     image_resize_method,
                                                     align_corners=False)
        # Restore the shape since the dynamic slice based upon the bbox_size loses
        # the third dimension.
        distorted_image.set_shape([height, width, 3])
        if summary_verbosity >= 3:
            tf.summary.image('cropped_resized_maybe_flipped_image',
                             tf.expand_dims(distorted_image, 0))

        if distortions:
            distorted_image = tf.cast(distorted_image, dtype=tf.float32)
            # Images values are expected to be in [0,1] for color distortion.
            distorted_image /= 255.
            # Randomly distort the colors.
            distorted_image = distort_color(
                distorted_image,
                batch_position,
                distort_color_in_yiq=distort_color_in_yiq)

            # Note: This ensures the scaling matches the output of eval_image
            distorted_image *= 255

        if summary_verbosity >= 3:
            tf.summary.image('final_distorted_image',
                             tf.expand_dims(distorted_image, 0))
        return distorted_image
Ejemplo n.º 58
0
                           axis=1)
gen_loss = tf.reduce_mean(gen_loss0)  #average over N
gen_opt = tf.train.AdamOptimizer(lr)
gen_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='decoder')
gen_gradvars = gen_opt.compute_gradients(gen_loss, var_list=gen_vars)
gen_train_op = gen_opt.apply_gradients(gen_gradvars)

#provide encoder q(b|x) gradient by data augmentation
Dir = Dirichlet([1.0] * n_class)
pai = Dir.sample(sample_shape=[N, K_u, n_cv])  #[N,K_u,n_cv,n_class]

x_star_u = tf.tile(tf.expand_dims(x_binary, axis=1), [1, K_u, 1])  #N*K_u*d_x

jj = 0

pai_slice_j = tf.slice(pai, begin=[0, 0, 0, jj], size=[-1, -1, -1,
                                                       1])  #N,K_u,n_cv,1
pai_j = swap(pai, jj, 0)
F_j = Fn(pai_j, prior_logit0, z_concate, x_star_u)  #N*K_u
for mm in range(1, n_class):
    pai_m = swap(pai, jj, mm)
    F_m = Fn(pai_m, prior_logit0, z_concate, x_star_u)
    grad_m = tf.expand_dims(tf.expand_dims(F_m - F_j, 2),
                            3) * (1 - n_class * pai_slice_j)
    if mm == 1:
        alpha_grads = grad_m
    else:
        alpha_grads = tf.concat([alpha_grads, grad_m], 3)

alpha_grads = tf.reduce_mean(alpha_grads,
                             axis=1)  #N*n_cv*d_b, expectation over pai
alpha_grads = tf.reshape(alpha_grads, [-1, z_dim])
Ejemplo n.º 59
0
def eval_image(image,
               height,
               width,
               batch_position,
               resize_method,
               summary_verbosity=0):
    """Get the image for model evaluation.

  We preprocess the image simiarly to Slim, see
  https://github.com/tensorflow/models/blob/master/slim/preprocessing/vgg_preprocessing.py
  Validation images do not have bounding boxes, so to crop the image, we first
  resize the image such that the aspect ratio is maintained and the resized
  height and width are both at least 1.15 times `height` and `width`
  respectively. Then, we do a central crop to size (`height`, `width`).

  Args:
    image: 3-D float Tensor representing the image.
    height: The height of the image that will be returned.
    width: The width of the image that will be returned.
    batch_position: position of the image in a batch, which affects how images
      are distorted and resized. NOTE: this argument can be an integer or a
      tensor
    resize_method: one of the strings 'round_robin', 'nearest', 'bilinear',
      'bicubic', or 'area'.
    summary_verbosity: Verbosity level for summary ops. Pass 0 to disable both
      summaries and checkpoints.
  Returns:
    An image of size (output_height, output_width, 3) that is resized and
    cropped as described above.
  """
    # TODO(reedwm): Currently we resize then crop. Investigate if it's faster to
    # crop then resize.
    with tf.name_scope('eval_image'):
        if summary_verbosity >= 3:
            tf.summary.image('original_image', tf.expand_dims(image, 0))

        shape = tf.shape(image)
        image_height = shape[0]
        image_width = shape[1]
        image_height_float = tf.cast(image_height, tf.float32)
        image_width_float = tf.cast(image_width, tf.float32)

        scale_factor = 1.15

        # Compute resize_height and resize_width to be the minimum values such that
        #   1. The aspect ratio is maintained (i.e. resize_height / resize_width is
        #      image_height / image_width), and
        #   2. resize_height >= height * `scale_factor`, and
        #   3. resize_width >= width * `scale_factor`
        max_ratio = tf.maximum(height / image_height_float,
                               width / image_width_float)
        resize_height = tf.cast(image_height_float * max_ratio * scale_factor,
                                tf.int32)
        resize_width = tf.cast(image_width_float * max_ratio * scale_factor,
                               tf.int32)

        # Resize the image to shape (`resize_height`, `resize_width`)
        image_resize_method = get_image_resize_method(resize_method,
                                                      batch_position)
        distorted_image = tf.image.resize_images(image,
                                                 [resize_height, resize_width],
                                                 image_resize_method,
                                                 align_corners=False)

        # Do a central crop of the image to size (height, width).
        total_crop_height = (resize_height - height)
        crop_top = total_crop_height // 2
        total_crop_width = (resize_width - width)
        crop_left = total_crop_width // 2
        distorted_image = tf.slice(distorted_image, [crop_top, crop_left, 0],
                                   [height, width, 3])

        distorted_image.set_shape([height, width, 3])
        if summary_verbosity >= 3:
            tf.summary.image('cropped_resized_image',
                             tf.expand_dims(distorted_image, 0))
        image = distorted_image
    return image
Ejemplo n.º 60
0
def main(argv=None):
    keep_probability = tf.placeholder(tf.float32, name="keep_probabilty")
    image = tf.placeholder(tf.float32, shape=[None, IMAGE_SIZE, IMAGE_SIZE, 1], name="input_image")
    annotation = tf.placeholder(tf.float32, shape=[None, IMAGE_SIZE, IMAGE_SIZE, 2], name="annotation")
    z = tf.placeholder(tf.float32, shape=[None, 1, 1, 10], name="z")
  #  mask = tf.placeholder(tf.float32, shape=[None, 64, 64, 1], name="mask")
  #  mask2 = tf.placeholder(tf.float32, shape=[None, 64, 64, 1], name="mask2")
    z_new = tf.placeholder(tf.float32, shape=[None, 1, 1, 10], name="z_new")
    istrain = tf.placeholder(tf.bool)
   #z_lip = tf.placeholder(tf.float32, shape=[None, 1, 1, 10], name="z_lip")
   #z_lip_inv = tf.placeholder(tf.float32, shape=[None, 1, 1, 10], name="z_lip_inv")
    e = tf.placeholder(tf.float32, shape=[None, 4, 4, 522], name="e")
    e_p = tf.placeholder(tf.float32, shape=[None, 1, 1, 8202], name="e_p")
    
    save_itr = 0
    # pred_annotation, logits = inference(image, keep_probability,z)
 #   tf.summary.image("input_image", image, max_outputs=2)
 #   tf.summary.image("ground_truth", tf.cast(annotation, tf.uint8), max_outputs=2)
 #   tf.summary.image("pred_annotation", tf.cast(pred_annotation, tf.uint8), max_outputs=2)
#    loss = tf.reduce_mean((tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,
 #                                                                         labels=tf.squeeze(annotation, squeeze_dims=[3]),
  #                                                                    name="entropy")))
    
    
   # mask_ = tf.ones([FLAGS.batch_size,32,64,3])
   # mask = tf.pad(mask_, [[0,0],[0,32],[0,0],[0,0]])

 #   mask2__ = tf.ones([FLAGS.batch_size,78,78,3])
  #  mask2_ = tf.pad(mask2__, [[0,0],[25,25],[25,25],[0,0]])
   # mask2 = mask2_ - mask
    zero = tf.zeros([FLAGS.batch_size,1,1,8202])   
    logits, h  = inference(image, keep_probability,z,0.0,istrain)
    logits_e, h_e = inference(image, keep_probability,z,e,istrain)
    #logits_lip,_  = inference((1-mask)*image + mask*0.0, keep_probability,z_lip,istrain   ) 
    #logits_lip_inv,_  = inference((1-mask)*image + mask*0.0, keep_probability,z_lip_inv,istrain   )

    z_pred = predictor(h,z,zero,istrain)
    z_pred_e = predictor(h,z,e_p,istrain)
  #  z_pred_lip =  predictor(h,z_lip,istrain)
  #  z_pred_lip_inv =  predictor(h,z_lip_inv,istrain)
   # logits = inference(image, keep_probability,z,istrain)
    tf.summary.image("input_image", image, max_outputs=2)
    tf.summary.image("ground_truth", tf.cast(annotation, tf.uint8), max_outputs=2)
   # tf.summary.image("pred_annotation", tf.cast(pred_annotation, tf.uint8), max_outputs=2)

    
   # lossz = 0.1 * tf.reduce_mean(tf.reduce_sum(tf.abs(z),[1,2,3]))
  #  lossz = 0.1 * tf.reduce_mean(tf.abs(z))
   # loss_all = tf.reduce_mean(tf.sqrt(tf.reduce_sum(tf.square((image - logits)),[1,2,3])))
   # loss_all = tf.reduce_mean(tf.reduce_sum(tf.contrib.layers.flatten(tf.abs(image - logits)),1))
    
  #  loss_mask = 0.8*tf.reduce_mean(tf.sqrt(tf.reduce_sum(tf.square((image - logits)*mask),[1,2,3])))
    g_k = gaussian_kernel(2,0.0,1.0)
    
    gauss_kernel = tf.tile(g_k[ :, :, tf.newaxis, tf.newaxis],[1,1,1,1])
    
# Convolve.
    logits_a = tf.slice(logits, [0,0,0,0],[-1,-1,-1,1])
    logits_b = tf.slice(logits, [0,0,0,1],[-1,-1,-1,1])


    logits_smooth_a_ = tf.nn.conv2d(tf.abs(logits_a), gauss_kernel, strides=[1, 1, 1, 1], padding="SAME")
    logits_smooth_b_ = tf.nn.conv2d(tf.abs(logits_b), gauss_kernel, strides=[1, 1, 1, 1], padding="SAME")

    #logits_smooth_ = tf.concat([logits_smooth_a,logits_smooth_b], axis = 3)
   
    logits_smooth_a = tf.maximum(logits_smooth_a_, 0.0001)
    logits_smooth_b = tf.maximum(logits_smooth_b_, 0.0001)
  #  logits_smooth_norm = tf.contrib.layers.flatten(logits_smooth)/tf.reduce_sum(logits_smooth,axis=[1,2,3], keep_dims = True)
    logits_smooth_norm_a = tf.nn.softmax(tf.contrib.layers.flatten(logits_smooth_a), axis = 1)
    logits_smooth_norm_b = tf.nn.softmax(tf.contrib.layers.flatten(logits_smooth_b), axis = 1)


    ones = tf.ones([FLAGS.batch_size,IMAGE_SIZE,IMAGE_SIZE,1])
    zeros = tf.zeros([FLAGS.batch_size,IMAGE_SIZE,IMAGE_SIZE,1])
   
    normal_dist_d = tf.distributions.Uniform(low = zeros, high = ones)
    normal_dist_a_ = normal_dist_d.sample()
    normal_dist_a = tf.maximum(normal_dist_a_, 0.0001)
   #normal_dist_norm = tf.contrib.layers.flatten(normal_dist)/tf.reduce_sum(normal_dist,axis=[1,2,3], keep_dims = True)
    normal_dist_norm_a =tf.nn.softmax(tf.contrib.layers.flatten(normal_dist_a), axis = 1)
    
    normal_dist_b_ = normal_dist_d.sample()
    normal_dist_b = tf.maximum(normal_dist_b_, 0.0001)
    normal_dist_norm_b =tf.nn.softmax(tf.contrib.layers.flatten(normal_dist_b), axis = 1)

    X_a = tf.distributions.Categorical(probs=logits_smooth_norm_a)
    X_b = tf.distributions.Categorical(probs=logits_smooth_norm_b)
    Y_a = tf.distributions.Categorical(probs=normal_dist_norm_a)
    Y_b = tf.distributions.Categorical(probs=normal_dist_norm_b)    

    kl_dist_a = tf.reduce_mean(tf.distributions.kl_divergence(X_a, Y_a))
    kl_dist_b = tf.reduce_mean(tf.distributions.kl_divergence(X_b, Y_b))
#    kl_dist = tf.reduce_sum(logits_smooth_norm * tf.log(logits_smooth_norm/normal_dist_norm))
   # kl_dist = tf.contrib.distributions.kl_divergence(logits_smooth_norm,normal_dist_norm)
   # logits_std = tf.reduce_std(logits_smooth, axis =[1,2],keep_dims=True )
   # logits_mean = tf.reduce_mean(logits_smooth, axis =[1,2], keep_dims=True)
   # logits_normalized = (logits_smooth - logits_mean)/logits_std       

    
 
 #   annotation_weights_norm =  tf.reduce_sum(tf.exp(tf.abs(annotation))/tf.exp(1.0))
#    annotation_weights = (tf.exp(tf.abs(annotation))/tf.exp(1.0))
    loss_ = tf.reduce_mean(tf.reduce_sum(tf.contrib.layers.flatten(tf.abs((annotation - logits))    ),1)) +0.5* kl_dist_a + 0.5*kl_dist_b
    
   # loss_ =  0.4*loss_mask + loss_mask2
  #  loss = tf.reduce_mean(tf.squared_difference(logits ,annotation ))
    loss_summary = tf.summary.scalar("entropy", loss_)
   # zloss = tf.reduce_mean(tf.losses.cosine_distance(tf.contrib.layers.flatten(z_new) ,tf.contrib.layers.flatten(z_pred),axis =1)) 
    zloss_ = tf.reduce_mean(tf.reduce_sum(tf.contrib.layers.flatten(tf.abs((z_pred - z_new))),1))
 #   zloss_lip = tf.reduce_mean(tf.reduce_sum(tf.contrib.layers.flatten(tf.abs((z_pred - z_pred_lip))),1))
#    zloss_lip_inv = -tf.reduce_mean(tf.reduce_sum(tf.contrib.layers.flatten(tf.abs((z_pred - z_pred_lip_inv))),1))
 
#    z_loss = zloss_ + 0.1* zloss_lip# + zloss_lip_inv
        
    
    lip_loss_dec = tf.reduce_mean(tf.reduce_sum(tf.contrib.layers.flatten(tf.abs((logits - logits_e))),1))
    loss = loss_ + 0.1*lip_loss_dec 
   
    lip_loss_pred = tf.reduce_mean(tf.reduce_sum(tf.contrib.layers.flatten(tf.abs((z_pred - z_pred_e))),1))     
    zloss = zloss_ + 0.1*lip_loss_pred
   
    grads = train_z(loss_,z)    

    trainable_var = tf.trainable_variables()
    trainable_z_pred_var = tf.trainable_variables(scope="predictor")
    trainable_d_pred_var = tf.trainable_variables(scope="decoder")


    print(trainable_z_pred_var)
    if FLAGS.debug:
        for var in trainable_var:
            utils.add_to_regularization_and_summary(var)
    train_op = train(loss, trainable_var)
    train_pred = train_predictor(zloss,trainable_z_pred_var)
    

    print("Setting up summary op...")
    summary_op = tf.summary.merge_all()

    print("Setting up image reader...")
    train_records, valid_records = scene_parsing.read_dataset(FLAGS.data_dir)
    print(len(train_records))
    print(len(valid_records))

    print("Setting up dataset reader")
    image_options = {'resize': True, 'resize_size': IMAGE_SIZE}
    if FLAGS.mode == 'train':
        train_dataset_reader = dataset.BatchDatset(train_records, image_options)
    validation_dataset_reader = dataset.BatchDatset(valid_records, image_options)

    sess = tf.Session()

    print("Setting up Saver...")
    saver = tf.train.Saver()

    # create two summary writers to show training loss and validation loss in the same graph
    # need to create two folders 'train' and 'validation' inside FLAGS.logs_dir
    train_writer = tf.summary.FileWriter(FLAGS.logs_dir + '/train', sess.graph)
    validation_writer = tf.summary.FileWriter(FLAGS.logs_dir + '/validation')

    sess.run(tf.global_variables_initializer())
    ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
    if ckpt and ckpt.model_checkpoint_path:
        saver.restore(sess, ckpt.model_checkpoint_path)
        print("Model restored...")
    
    saved =True
    if FLAGS.mode == "train":
        for itr in xrange(MAX_ITERATION):
            
            train_images, train_annotations = train_dataset_reader.next_batch(FLAGS.batch_size)
            print("$$$$$$$$$$$$$$$$$$$$")
            print(train_images.shape)
          #  z_ = np.reshape(signal.gaussian(200, std=1),(FLAGS.batch_size,1,1,10))-0.5
            z_ = np.random.uniform(low=-1.0, high=1.0, size=(FLAGS.batch_size,1,1,10))
 #           train_images[train_images < 0.] = -1.
  #          train_annotations[train_annotations < 0.] = -1.
   #         train_images[train_images >= 0.] = 1.0
    #        train_annotations[train_annotations >= 0.] = 1.0
                    
            x1 = random.randint(0, 10) 
            w1 = random.randint(30, 54)
            y1 = random.randint(0, 10)
            h1 = random.randint(30, 54)

            cond = random.randint(0, 10)
           # saved = True   
            if False:
                saved = False
                train_images_m, train_annotations_m = train_dataset_reader.get_random_batch(FLAGS.batch_size)
                train_images_m[train_images_m < 0.] = -1.
                train_annotations_m[train_annotations_m < 0.] = -1.
                train_images_m[train_images_m >= 0.] = 1.0
                train_annotations_m[train_annotations_m >= 0.] = 1.0

                train_images = (train_images + 1.)/2.0*255.0
                train_annotations = (train_annotations + 1.)/2.0*255.0
                train_images_m = (train_images_m + 1.)/2.0*255.0
                train_annotations_m = (train_annotations_m +  1.)/2.0*255.0

                train_images_m[:,32:,:,:] = 0
                train_annotations_m[:,32:,:,:] = 0
                train_images = np.clip((train_images + train_images_m),0.0,255.0)
                train_annotations =  np.clip((train_annotations + train_annotations_m),0.0,255.0)
                '''
                train_images[train_images < 0.] = -1.
                train_annotations[train_annotations < 0.] = -1.
                train_images[train_images >= 0.] = 1.0
                train_annotations[train_annotations >= 0.] = 1.0
                '''

                train_annotations_ = np.squeeze(train_annotations,axis = 3)
                train_images_ = train_images

                train_images = train_images/127.5 - 1.0
                train_annotations = train_annotations/127.5 - 1.0                     

               # for itr_ in range(FLAGS.batch_size):
                #    utils.save_image(train_images_[itr_].astype(np.uint8), FLAGS.logs_dir, name="inp_" + str(5+itr_) )
                 #   utils.save_image(train_annotations_[itr_].astype(np.uint8), FLAGS.logs_dir, name="gt_" + str(5+itr_) )
    #        train_images[:,x1:w1,y1:h1,:] = 0
            
          #  print(train_images)
            r_m, r_m2 = random_mask(64)
           #feed_dict = {image: train_images, annotation: train_annotations, keep_probability: 0.85, z: z_,mask:r_m, istrain:True }
           #train_images[:,50:100,50:100,:] =0
            v = 0
           # print(train_images)    
            error_dec =  np.random.normal(0.0,0.001,(FLAGS.batch_size,4,4,522))
            error_dec_ =  np.random.normal(0.0,0.001,(FLAGS.batch_size,1,1,8202))
           # z_l_inv = z_ + np.random.normal(0.0,0.1)
          #  feed_dict = {image: train_images, annotation: train_annotations, keep_probability: 0.85, z: z_, e:error_dec, mask:r_m, istrain:True }
          #  z_l = z_ + np.random.normal(0.0,0.001)
      #     lloss,_ = sess.run([lip_loss, train_lip ], feed_dict=feed_dict)
           # z_l = z_ + np.random.normal(0.0,0.001)
           # print("Step: %d, lip_loss:%g" % (itr,lloss))

            for p in range(20):
                z_ol = np.copy(z_)
            #    z_l = z_ol + np.random.normal(0.0,0.001)
               # print("666666666666666666666666666666666666666")
                feed_dict = {image: train_images, annotation: train_annotations, keep_probability: 0.85, z: z_,e:error_dec, istrain:True }         
               # lloss,_ = sess.run([lip_loss, train_lip ], feed_dict=feed_dict)
               # print("Step: %d, z_step: %d, lip_loss:%g" % (itr,p,lloss))     
                z_loss, summ = sess.run([loss,loss_summary], feed_dict=feed_dict)
                print("Step: %d, z_step: %d, Train_loss:%g" % (itr,p,z_loss))
#                print(z_) 
                g = sess.run([grads],feed_dict=feed_dict)
                v_prev = np.copy(v)
               # print(g[0][0].shape)
                v = 0.001*v - 0.1*g[0][0]
                z_ += 0.001 * v_prev + (1+0.001)*v
               #z_ = np.clip(z_, -10.0, 10.0)
                
                '''
                m = interp1d([-10.0,10.0],[-1.0,1.0])
                print(np.max(z_))
                print(np.min(z_))
                z_ol_interp = m(z_ol)
                z_interp = m(z_)
                _,z_pred_loss =sess.run([train_pred,zloss],feed_dict={image: train_images,mask:r_m,z:z_ol_interp,z_new:z_interp,e_p:error_dec_,istrain:True,keep_probability: 0.85})
                print("Step: %d, z_step: %d, z_pred_loss:%g" % (itr,p,z_pred_loss))
                '''

               # _,z_pred_loss =sess.run([train_pred,zloss],feed_dict={image: train_images,mask:r_m,z:z_ol,z_new:z_,istrain:True,keep_probability: 0.85})
               # print("Step: %d, z_step: %d, z_pred_loss:%g" % (itr,p,z_pred_loss))
               # z_ = np.clip(z_, -1.0, 1.0)
               # print(v.shape)
               # print(z_.shape)
            feed_dict = {image: train_images, annotation: train_annotations, keep_probability:0.85,e:error_dec, z: z_, istrain:True }
            sess.run(train_op, feed_dict=feed_dict)

            if itr % 10 == 0:
                train_loss, summary_str = sess.run([loss, loss_summary], feed_dict=feed_dict)
                print("Step: %d, Train_loss:%g" % (itr, train_loss))
                
                train_writer.add_summary(summary_str, itr)
              

            if itr % 500 == 0:
                valid_images, valid_annotations = validation_dataset_reader.next_batch(FLAGS.batch_size)
     #           valid_annotations[valid_annotations < 0.] = -1.
      #          valid_images[valid_images < 0.] = -1.
       #         valid_annotations[valid_annotations >= 0.] = 1.0
        #        valid_images[valid_images >= 0.] = 1.0
                
                x1 = random.randint(0, 10)
                w1 = random.randint(30, 54)
                y1 = random.randint(0, 10)
                h1 = random.randint(30, 54)

     #           valid_images[:,x1:w1,y1:h1,:] = 0
                 
                valid_loss, summary_sva = sess.run([loss, loss_summary], feed_dict={image: valid_images, annotation: valid_annotations,
                                                       keep_probability: 1.0, z: z_,e:error_dec, istrain:False })
                print("%s ---> Validation_loss: %g" % (datetime.datetime.now(), valid_loss))

                # add validation loss to TensorBoard
                validation_writer.add_summary(summary_sva, itr)
                if itr % 3000 == 0:
                    save_itr = save_itr + 3000
               
                saver.save(sess, FLAGS.logs_dir + "model_fuse.ckpt", save_itr)

                

    elif FLAGS.mode == "visualize":
        valid_images, valid_annotations = validation_dataset_reader.get_random_batch(FLAGS.batch_size)
      #  valid_annotations[valid_annotations < 0.] = -1.0
      #  valid_images[valid_images < 0.] = -1.0
      #  valid_annotations[valid_annotations >= 0.] = 1.0
      #  valid_images[valid_images >= 0.] = 1.0
        
        x1 = random.randint(0, 10)
        w1 = random.randint(30, 54)
        y1 = random.randint(0, 10)
        h1 = random.randint(30, 54)

      #  valid_images[:,x1:w1,y1:h1,:] = 0
        r_m, r_m2 = random_mask(64)      
       # z_ = np.zeros(low=-1.0, high=1.0, size=(FLAGS.batch_size,1,1,10))
       # z_ = np.reshape(signal.gaussian(200, std=1),(FLAGS.batch_size,1,1,10))-0.5
        z_ = np.random.uniform(low=-1.0, high=1.0, size=(FLAGS.batch_size,1,1,10))
        feed_dict = {image: valid_images, annotation: valid_annotations, keep_probability: 0.85, z: z_, istrain:False }
        v= 0
        m__ = interp1d([-10.0,10.0],[-1.0,1.0])
        z_ = m__(z_)
  #      feed_dict = {image: valid_images, annotation: valid_annotations, keep_probability: 0.85, z: z_, istrain:False,mask:r_m }
        for p in range(20):
                z_ol = np.copy(z_)
               # print("666666666666666666666666666666666666666")
#                print(z_)
               # feed_dict = {image: valid_images, annotation: valid_annotations, keep_probability: 0.85, z: z_, istrain:False,mask:r_m }
              #  z_loss, summ = sess.run([loss,loss_summary], feed_dict=feed_dict)
              #  print("z_step: %d, Train_loss:%g" % (p,z_loss))
               # z_, z_pred_loss = sess.run(z_pred,zlossfeed_dict = {image: valid_images, annotation: valid_annotations, keep_probability: 1.0, z:z_ol, istrain:False,mask:r_m})
                
#                print(z_)
                g = sess.run([grads],feed_dict=feed_dict)
                v_prev = np.copy(v)
               # print(g[0][0].shape)
                v = 0.001*v - 0.1*g[0][0]
                z_ = z_ol +  0.001 * v_prev + (1+0.001)*v
              #  z_ = z_ol + 0.001 * v_prev + (1+0.001)*v
               # print("z_____________")
               # print(z__)
               # print("z_")
               # print(z_)
            #    m__ = interp1d([-10.0,10.0],[-1.0,1.0])
           #     z_ol = m__(z_ol)
         #       z_ = sess.run(z_pred,feed_dict = {image: valid_images, annotation: valid_annotations, keep_probability: 0.85, z:z_ol, istrain:False,mask:r_m})
             #   m_ = interp1d([-1.0,1.0],[-10.0,10.0])
              #  z_ = m_(z_)             
               # z_ = np.clip(z_, -1.0, 1.0)
               # print(z_pred_loss)
       # m_ = interp1d([-1.0,1.0],[-10.0,10.0])
       # z_ = m_(z_)
       
        pred = sess.run(logits, feed_dict={image: valid_images, annotation: valid_annotations,z:z_, istrain:False,
                                                    keep_probability: 0.85})
        
      #  print(sess.run(logits_smooth_norm, feed_dict={image: valid_images, annotation: valid_annotations,z:z_, istrain:False,
        #                                            keep_probability: 0.85}    ) )
        print("#######################")
       # print(sess.run(normal_dist_norm))
                
        valid_images = (valid_images +1.)/2.0*100.0
       # predicted_patch = sess.run(mask) * pred
       # pred = valid_images_masked + predicted_patch
        pred_ = pred * 128.0
#        pred = pred + 1./2.0*255
        print(np.max(pred_))
        print(np.min(pred_))
        pred = np.reshape(np.concatenate((valid_images,pred_), axis=3),(-1,64,64,3))

        valid_annotations_ = valid_annotations * 128.0
        valid_annotations = np.reshape(np.concatenate((valid_images,  valid_annotations_), axis=3),(-1,64,64,3))
        valid_images_gray = np.squeeze(valid_images)  
       # for itr in range(FLAGS.batch_size):
           # utils.save_image(valid_images_masked[itr].astype(np.uint8), FLAGS.logs_dir, name="inp_" + str(5+itr))
           # utils.save_image(valid_annotations[itr].astype(np.uint8), FLAGS.logs_dir, name="gt_" + str(5+itr))
           # utils.save_image(pred[itr].astype(np.uint8), FLAGS.logs_dir, name="predz_" + str(5+itr))
            #        utils.save_image(valid_images_masked[itr].astype(np.uint8), FLAGS.logs_dir, name="inp_" + str(5+itr)+'_' + str(p) )
             #       utils.save_image(valid_annotations_[itr].astype(np.uint8), FLAGS.logs_dir, name="gt_" + str(5+itr)+'_' + str(p)  )
             #  utils.save_image(pred[itr].astype(np.uint8), FLAGS.logs_dir, name="predz_" + str(5+itr)+'_' + str(p)  )
            #   print("Saved image: %d" % itr)
        
        for itr in range(FLAGS.batch_size):
            utils.save_image(valid_images_gray[itr], FLAGS.logs_dir, name="inp_" + str(5+itr) )
            utils.save_image(color.lab2rgb(pred[itr]), FLAGS.logs_dir, name="predz_" + str(5+itr) )
            utils.save_image(color.lab2rgb(valid_annotations[itr]), FLAGS.logs_dir, name="gt_" + str(5+itr) )