Python tensorflow.contrib.slim.get_variables_to_restore() Examples

def main():
    args = parse_args()

    with tf.Session(graph=tf.Graph()) as session:
        input_var = tf.placeholder(
            tf.uint8, (None, 128, 64, 3), name="images")
        image_var = tf.map_fn(
            lambda x: _preprocess(x), tf.cast(input_var, tf.float32),
            back_prop=False)

        factory_fn = _network_factory()
        features, _ = factory_fn(image_var, reuse=None)
        features = tf.identity(features, name="features")

        saver = tf.train.Saver(slim.get_variables_to_restore())
        saver.restore(session, args.checkpoint_in)

        output_graph_def = tf.graph_util.convert_variables_to_constants(
            session, tf.get_default_graph().as_graph_def(),
            [features.name.split(":")[0]])
        with tf.gfile.GFile(args.graphdef_out, "wb") as file_handle:
            file_handle.write(output_graph_def.SerializeToString()) 

def build_pretrained_graph(
      self, images, resnet_layer, checkpoint, is_training, reuse=False):
    """See baseclass."""
    with slim.arg_scope(resnet_v2.resnet_arg_scope()):
      _, endpoints = resnet_v2.resnet_v2_50(
          images, is_training=is_training, reuse=reuse)
      resnet_layer = 'resnet_v2_50/block%d' % resnet_layer
      resnet_output = endpoints[resnet_layer]
      resnet_variables = slim.get_variables_to_restore()
      resnet_variables = [
          i for i in resnet_variables if 'global_step' not in i.name]
      if is_training and not reuse:
        init_saver = tf.train.Saver(resnet_variables)
        def init_fn(scaffold, sess):
          del scaffold
          init_saver.restore(sess, checkpoint)
      else:
        init_fn = None

      return resnet_output, resnet_variables, init_fn 

def build_pretrained_graph(
      self, images, resnet_layer, checkpoint, is_training, reuse=False):
    """See baseclass."""
    with slim.arg_scope(resnet_v2.resnet_arg_scope()):
      _, endpoints = resnet_v2.resnet_v2_50(
          images, is_training=is_training, reuse=reuse)
      resnet_layer = 'resnet_v2_50/block%d' % resnet_layer
      resnet_output = endpoints[resnet_layer]
      resnet_variables = slim.get_variables_to_restore()
      resnet_variables = [
          i for i in resnet_variables if 'global_step' not in i.name]
      if is_training and not reuse:
        init_saver = tf.train.Saver(resnet_variables)
        def init_fn(scaffold, sess):
          del scaffold
          init_saver.restore(sess, checkpoint)
      else:
        init_fn = None

      return resnet_output, resnet_variables, init_fn 

def build_pretrained_graph(
      self, images, resnet_layer, checkpoint, is_training, reuse=False):
    """See baseclass."""
    with slim.arg_scope(resnet_v2.resnet_arg_scope()):
      _, endpoints = resnet_v2.resnet_v2_50(
          images, is_training=is_training, reuse=reuse)
      resnet_layer = 'resnet_v2_50/block%d' % resnet_layer
      resnet_output = endpoints[resnet_layer]
      resnet_variables = slim.get_variables_to_restore()
      resnet_variables = [
          i for i in resnet_variables if 'global_step' not in i.name]
      if is_training and not reuse:
        init_saver = tf.train.Saver(resnet_variables)
        def init_fn(scaffold, sess):
          del scaffold
          init_saver.restore(sess, checkpoint)
      else:
        init_fn = None

      return resnet_output, resnet_variables, init_fn 

def build_pretrained_graph(
      self, images, resnet_layer, checkpoint, is_training, reuse=False):
    """See baseclass."""
    with slim.arg_scope(resnet_v2.resnet_arg_scope()):
      _, endpoints = resnet_v2.resnet_v2_50(
          images, is_training=is_training, reuse=reuse)
      resnet_layer = 'resnet_v2_50/block%d' % resnet_layer
      resnet_output = endpoints[resnet_layer]
      resnet_variables = slim.get_variables_to_restore()
      resnet_variables = [
          i for i in resnet_variables if 'global_step' not in i.name]
      if is_training and not reuse:
        init_saver = tf.train.Saver(resnet_variables)
        def init_fn(scaffold, sess):
          del scaffold
          init_saver.restore(sess, checkpoint)
      else:
        init_fn = None

      return resnet_output, resnet_variables, init_fn 

def main():
    args = parse_args()

    with tf.Session(graph=tf.Graph()) as session:
        input_var = tf.placeholder(
            tf.uint8, (None, 128, 64, 3), name="images")
        image_var = tf.map_fn(
            lambda x: _preprocess(x), tf.cast(input_var, tf.float32),
            back_prop=False)

        factory_fn = _network_factory()
        features, _ = factory_fn(image_var, reuse=None)
        features = tf.identity(features, name="features")

        saver = tf.train.Saver(slim.get_variables_to_restore())
        saver.restore(session, args.checkpoint_in)

        output_graph_def = tf.graph_util.convert_variables_to_constants(
            session, tf.get_default_graph().as_graph_def(),
            [features.name.split(":")[0]])
        with tf.gfile.GFile(args.graphdef_out, "wb") as file_handle:
            file_handle.write(output_graph_def.SerializeToString()) 

def main():
    args = parse_args()

    with tf.Session(graph=tf.Graph()) as session:
        input_var = tf.placeholder(
            tf.uint8, (None, 128, 64, 3), name="images")
        image_var = tf.map_fn(
            lambda x: _preprocess(x), tf.cast(input_var, tf.float32),
            back_prop=False)

        factory_fn = _network_factory()
        features, _ = factory_fn(image_var, reuse=None)
        features = tf.identity(features, name="features")

        saver = tf.train.Saver(slim.get_variables_to_restore())
        saver.restore(session, args.checkpoint_in)

        output_graph_def = tf.graph_util.convert_variables_to_constants(
            session, tf.get_default_graph().as_graph_def(),
            [features.name.split(":")[0]])
        with tf.gfile.GFile(args.graphdef_out, "wb") as file_handle:
            file_handle.write(output_graph_def.SerializeToString()) 

def build_pretrained_graph(
      self, images, resnet_layer, checkpoint, is_training, reuse=False):
    """See baseclass."""
    with slim.arg_scope(resnet_v2.resnet_arg_scope()):
      _, endpoints = resnet_v2.resnet_v2_50(
          images, is_training=is_training, reuse=reuse)
      resnet_layer = 'resnet_v2_50/block%d' % resnet_layer
      resnet_output = endpoints[resnet_layer]
      resnet_variables = slim.get_variables_to_restore()
      resnet_variables = [
          i for i in resnet_variables if 'global_step' not in i.name]
      if is_training and not reuse:
        init_saver = tf.train.Saver(resnet_variables)
        def init_fn(scaffold, sess):
          del scaffold
          init_saver.restore(sess, checkpoint)
      else:
        init_fn = None

      return resnet_output, resnet_variables, init_fn 

def main():
    args = parse_args()

    with tf.Session(graph=tf.Graph()) as session:
        input_var = tf.placeholder(
            tf.uint8, (None, 128, 64, 3), name="images")
        image_var = tf.map_fn(
            lambda x: _preprocess(x), tf.cast(input_var, tf.float32),
            back_prop=False)

        factory_fn = _network_factory()
        features, _ = factory_fn(image_var, reuse=None)
        features = tf.identity(features, name="features")

        saver = tf.train.Saver(slim.get_variables_to_restore())
        saver.restore(session, args.checkpoint_in)

        output_graph_def = tf.graph_util.convert_variables_to_constants(
            session, tf.get_default_graph().as_graph_def(),
            [features.name.split(":")[0]])
        with tf.gfile.GFile(args.graphdef_out, "wb") as file_handle:
            file_handle.write(output_graph_def.SerializeToString()) 

def restore(self, path = None, restore_opt = True, restore_resnet18_blocks = True, restore_dilation_blocks = True):
    if path is None:
      path = tf.train.latest_checkpoint(self.pr.train_dir)      
    print 'Restoring from:', path
    var_list = slim.get_variables_to_restore()
    opt_names = ['Adam', 'beta1_power', 'beta2_power', 'Momentum', 'cache']
    if not restore_resnet18_blocks:
      opt_names += ['conv2_2_', 'conv3_2_', 'conv4_2_', 'conv5_2_']

    if not restore_opt:
      var_list = [x for x in var_list if not any(name in x.name for name in opt_names)]

    print 'Restoring:'
    for x in var_list:
      print x.name
    print
    tf.train.Saver(var_list).restore(self.sess, path) 

def main():
    args = parse_args()

    with tf.Session(graph=tf.Graph()) as session:
        input_var = tf.placeholder(
            tf.uint8, (None, 128, 64, 3), name="images")
        image_var = tf.map_fn(
            lambda x: _preprocess(x), tf.cast(input_var, tf.float32),
            back_prop=False)

        factory_fn = _network_factory()
        features, _ = factory_fn(image_var, reuse=None)
        features = tf.identity(features, name="features")

        saver = tf.train.Saver(slim.get_variables_to_restore())
        saver.restore(session, args.checkpoint_in)

        output_graph_def = tf.graph_util.convert_variables_to_constants(
            session, tf.get_default_graph().as_graph_def(),
            [features.name.split(":")[0]])
        with tf.gfile.GFile(args.graphdef_out, "wb") as file_handle:
            file_handle.write(output_graph_def.SerializeToString()) 

def restore(self, path = None, restore_opt = True, ul_only = False):
    if path is None:
      path = tf.train.latest_checkpoint(self.pr.train_dir)      
    print 'Restoring:', path
    var_list = slim.get_variables_to_restore()
    for x in var_list:
      print x.name
    print
    var_list = slim.get_variables_to_restore()
    if not restore_opt:
      opt_names = ['Adam', 'beta1_power', 'beta2_power', 'Momentum'] + ['cls']# + ['renorm_mean_weight', 'renorm_stddev_weight', 'moving_mean', 'renorm']
      print 'removing bn gamma'
      opt_names += ['gamma']
      var_list = [x for x in var_list if not any(name in x.name for name in opt_names)]
    if ul_only:
      var_list = [x for x in var_list if not x.name.startswith('lb/') and ('global_step' not in x.name)]
    #var_list = [x for x in var_list if ('global_step' not in x.name)]
    print 'Restoring variables:'
    for x in var_list:
      print x.name
    tf.train.Saver(var_list).restore(self.sess, path)
    # print 'TEST: restoring all'
    # tf.train.Saver().restore(self.sess, path) 

def build_pretrained_graph(
      self, images, resnet_layer, checkpoint, is_training, reuse=False):
    """See baseclass."""
    with slim.arg_scope(resnet_v2.resnet_arg_scope()):
      _, endpoints = resnet_v2.resnet_v2_50(
          images, is_training=is_training, reuse=reuse)
      resnet_layer = 'resnet_v2_50/block%d' % resnet_layer
      resnet_output = endpoints[resnet_layer]
      resnet_variables = slim.get_variables_to_restore()
      resnet_variables = [
          i for i in resnet_variables if 'global_step' not in i.name]
      if is_training and not reuse:
        init_saver = tf.train.Saver(resnet_variables)
        def init_fn(scaffold, sess):
          del scaffold
          init_saver.restore(sess, checkpoint)
      else:
        init_fn = None

      return resnet_output, resnet_variables, init_fn 

def main():
    args = parse_args()

    with tf.Session(graph=tf.Graph()) as session:
        input_var = tf.placeholder(
            tf.uint8, (None, 128, 64, 3), name="images")
        image_var = tf.map_fn(
            lambda x: _preprocess(x), tf.cast(input_var, tf.float32),
            back_prop=False)

        factory_fn = _network_factory()
        features, _ = factory_fn(image_var, reuse=None)
        features = tf.identity(features, name="features")

        saver = tf.train.Saver(slim.get_variables_to_restore())
        saver.restore(session, args.checkpoint_in)

        output_graph_def = tf.graph_util.convert_variables_to_constants(
            session, tf.get_default_graph().as_graph_def(),
            [features.name.split(":")[0]])
        with tf.gfile.GFile(args.graphdef_out, "wb") as file_handle:
            file_handle.write(output_graph_def.SerializeToString()) 

def main():
    args = parse_args()

    with tf.Session(graph=tf.Graph()) as session:
        input_var = tf.placeholder(
            tf.uint8, (None, 128, 64, 3), name="images")
        image_var = tf.map_fn(
            lambda x: _preprocess(x), tf.cast(input_var, tf.float32),
            back_prop=False)

        factory_fn = _network_factory()
        features, _ = factory_fn(image_var, reuse=None)
        features = tf.identity(features, name="features")

        saver = tf.train.Saver(slim.get_variables_to_restore())
        saver.restore(session, args.checkpoint_in)

        output_graph_def = tf.graph_util.convert_variables_to_constants(
            session, tf.get_default_graph().as_graph_def(),
            [features.name.split(":")[0]])
        with tf.gfile.GFile(args.graphdef_out, "wb") as file_handle:
            file_handle.write(output_graph_def.SerializeToString()) 

def get_repr_from_image(images_reshaped, modalities, data_augment, encoder,
                        freeze_conv, wt_decay, is_training):
  # Pass image through lots of convolutional layers, to obtain pool5
  if modalities == ['rgb']:
    with tf.name_scope('pre_rgb'):
      x = (images_reshaped + 128.) / 255. # Convert to brightness between 0 and 1.
      if data_augment.relight and is_training:
        x = tf_utils.distort_image(x, fast_mode=data_augment.relight_fast)
      x = (x-0.5)*2.0
    scope_name = encoder
  elif modalities == ['depth']:
    with tf.name_scope('pre_d'):
      d_image = images_reshaped
      x = 2*(d_image[...,0] - 80.0)/100.0
      y = d_image[...,1]
      d_image = tf.concat([tf.expand_dims(x, -1), tf.expand_dims(y, -1)], 3)
      x = d_image
    scope_name = 'd_'+encoder

  resnet_is_training = is_training and (not freeze_conv)
  with slim.arg_scope(resnet_v2.resnet_utils.resnet_arg_scope(resnet_is_training)):
    fn = getattr(tf_utils, encoder)
    x, end_points = fn(x, num_classes=None, global_pool=False,
                       output_stride=None, reuse=None,
                       scope=scope_name)
  vars_ = slim.get_variables_to_restore()

  conv_feat = x
  return conv_feat, vars_ 

def build_inceptionv3_graph(images, endpoint, is_training, checkpoint,
                            reuse=False):
  """Builds an InceptionV3 model graph.

  Args:
    images: A 4-D float32 `Tensor` of batch images.
    endpoint: String, name of the InceptionV3 endpoint.
    is_training: Boolean, whether or not to build a training or inference graph.
    checkpoint: String, path to the pretrained model checkpoint.
    reuse: Boolean, whether or not we are reusing the embedder.
  Returns:
    inception_output: `Tensor` holding the InceptionV3 output.
    inception_variables: List of inception variables.
    init_fn: Function to initialize the weights (if not reusing, then None).
  """
  with slim.arg_scope(inception.inception_v3_arg_scope()):
    _, endpoints = inception.inception_v3(
        images, num_classes=1001, is_training=is_training)
    inception_output = endpoints[endpoint]
    inception_variables = slim.get_variables_to_restore()
    inception_variables = [
        i for i in inception_variables if 'global_step' not in i.name]
    if is_training and not reuse:
      init_saver = tf.train.Saver(inception_variables)
      def init_fn(scaffold, sess):
        del scaffold
        init_saver.restore(sess, checkpoint)
    else:
      init_fn = None
    return inception_output, inception_variables, init_fn 

def restore(self, path = None, restore_opt = True, init_type = None):
    if path is None:
      path = tf.train.latest_checkpoint(self.pr.train_dir)      
    print 'Restoring from:', path
    var_list = slim.get_variables_to_restore()
    opt_names = ['Adam', 'beta1_power', 'beta2_power', 'Momentum', 'cache']

    if init_type == 'shift':
      # gamma is reinitialized
      opt_names += ['gen/', 'discrim/', 'global_step', 'gamma']
    elif init_type == 'sep':
      #opt_names += ['global_step']
      opt_names += ['global_step', 'discrim']
    elif init_type is None:
      pass
    else:
      raise RuntimeError()

    if not restore_opt or init_type is not None:
      var_list = [x for x in var_list if not any(name in x.name for name in opt_names)]

    print 'Restoring:'
    for x in var_list:
      print x.name
    print
    
    tf.train.Saver(var_list).restore(self.sess, path) 

def variables_to_restore(scope=None, strip_scope=False):
  """Returns a list of variables to restore for the specified list of methods.

  It is supposed that variable name starts with the method's scope (a prefix
  returned by _method_scope function).

  Args:
    methods_names: a list of names of configurable methods.
    strip_scope: if True will return variable names without method's scope.
      If methods_names is None will return names unchanged.
    model_scope: a scope for a whole model.

  Returns:
    a dictionary mapping variable names to variables for restore.
  """
  if scope:
    variable_map = {}
    method_variables = slim.get_variables_to_restore(include=[scope])
    for var in method_variables:
      if strip_scope:
        var_name = var.op.name[len(scope) + 1:]
      else:
        var_name = var.op.name
      variable_map[var_name] = var

    return variable_map
  else:
    return {v.op.name: v for v in slim.get_variables_to_restore()} 

def get_repr_from_image(images_reshaped, modalities, data_augment, encoder,
                        freeze_conv, wt_decay, is_training):
  # Pass image through lots of convolutional layers, to obtain pool5
  if modalities == ['rgb']:
    with tf.name_scope('pre_rgb'):
      x = (images_reshaped + 128.) / 255. # Convert to brightness between 0 and 1.
      if data_augment.relight and is_training:
        x = tf_utils.distort_image(x, fast_mode=data_augment.relight_fast)
      x = (x-0.5)*2.0
    scope_name = encoder
  elif modalities == ['depth']:
    with tf.name_scope('pre_d'):
      d_image = images_reshaped
      x = 2*(d_image[...,0] - 80.0)/100.0
      y = d_image[...,1]
      d_image = tf.concat([tf.expand_dims(x, -1), tf.expand_dims(y, -1)], 3)
      x = d_image
    scope_name = 'd_'+encoder

  resnet_is_training = is_training and (not freeze_conv)
  with slim.arg_scope(resnet_v2.resnet_utils.resnet_arg_scope(resnet_is_training)):
    fn = getattr(tf_utils, encoder)
    x, end_points = fn(x, num_classes=None, global_pool=False,
                       output_stride=None, reuse=None,
                       scope=scope_name)
  vars_ = slim.get_variables_to_restore()

  conv_feat = x
  return conv_feat, vars_ 

def get_repr_from_image(images_reshaped, modalities, data_augment, encoder,
                        freeze_conv, wt_decay, is_training):
  # Pass image through lots of convolutional layers, to obtain pool5
  if modalities == ['rgb']:
    with tf.name_scope('pre_rgb'):
      x = (images_reshaped + 128.) / 255. # Convert to brightness between 0 and 1.
      if data_augment.relight and is_training:
        x = tf_utils.distort_image(x, fast_mode=data_augment.relight_fast)
      x = (x-0.5)*2.0
    scope_name = encoder
  elif modalities == ['depth']:
    with tf.name_scope('pre_d'):
      d_image = images_reshaped
      x = 2*(d_image[...,0] - 80.0)/100.0
      y = d_image[...,1]
      d_image = tf.concat([tf.expand_dims(x, -1), tf.expand_dims(y, -1)], 3)
      x = d_image
    scope_name = 'd_'+encoder

  resnet_is_training = is_training and (not freeze_conv)
  with slim.arg_scope(resnet_v2.resnet_utils.resnet_arg_scope(resnet_is_training)):
    fn = getattr(tf_utils, encoder)
    x, end_points = fn(x, num_classes=None, global_pool=False,
                       output_stride=None, reuse=None,
                       scope=scope_name)
  vars_ = slim.get_variables_to_restore()

  conv_feat = x
  return conv_feat, vars_ 

def build_inceptionv3_graph(images, endpoint, is_training, checkpoint,
                            reuse=False):
  """Builds an InceptionV3 model graph.

  Args:
    images: A 4-D float32 `Tensor` of batch images.
    endpoint: String, name of the InceptionV3 endpoint.
    is_training: Boolean, whether or not to build a training or inference graph.
    checkpoint: String, path to the pretrained model checkpoint.
    reuse: Boolean, whether or not we are reusing the embedder.
  Returns:
    inception_output: `Tensor` holding the InceptionV3 output.
    inception_variables: List of inception variables.
    init_fn: Function to initialize the weights (if not reusing, then None).
  """
  with slim.arg_scope(inception.inception_v3_arg_scope()):
    _, endpoints = inception.inception_v3(
        images, num_classes=1001, is_training=is_training)
    inception_output = endpoints[endpoint]
    inception_variables = slim.get_variables_to_restore()
    inception_variables = [
        i for i in inception_variables if 'global_step' not in i.name]
    if is_training and not reuse:
      init_saver = tf.train.Saver(inception_variables)
      def init_fn(scaffold, sess):
        del scaffold
        init_saver.restore(sess, checkpoint)
    else:
      init_fn = None
    return inception_output, inception_variables, init_fn 

def variables_to_restore(scope=None, strip_scope=False):
  """Returns a list of variables to restore for the specified list of methods.

  It is supposed that variable name starts with the method's scope (a prefix
  returned by _method_scope function).

  Args:
    methods_names: a list of names of configurable methods.
    strip_scope: if True will return variable names without method's scope.
      If methods_names is None will return names unchanged.
    model_scope: a scope for a whole model.

  Returns:
    a dictionary mapping variable names to variables for restore.
  """
  if scope:
    variable_map = {}
    method_variables = slim.get_variables_to_restore(include=[scope])
    for var in method_variables:
      if strip_scope:
        var_name = var.op.name[len(scope) + 1:]
      else:
        var_name = var.op.name
      variable_map[var_name] = var

    return variable_map
  else:
    return {v.op.name: v for v in slim.get_variables_to_restore()} 

def get_repr_from_image(images_reshaped, modalities, data_augment, encoder,
                        freeze_conv, wt_decay, is_training):
  # Pass image through lots of convolutional layers, to obtain pool5
  if modalities == ['rgb']:
    with tf.name_scope('pre_rgb'):
      x = (images_reshaped + 128.) / 255. # Convert to brightness between 0 and 1.
      if data_augment.relight and is_training:
        x = tf_utils.distort_image(x, fast_mode=data_augment.relight_fast)
      x = (x-0.5)*2.0
    scope_name = encoder
  elif modalities == ['depth']:
    with tf.name_scope('pre_d'):
      d_image = images_reshaped
      x = 2*(d_image[...,0] - 80.0)/100.0
      y = d_image[...,1]
      d_image = tf.concat([tf.expand_dims(x, -1), tf.expand_dims(y, -1)], 3)
      x = d_image
    scope_name = 'd_'+encoder

  resnet_is_training = is_training and (not freeze_conv)
  with slim.arg_scope(resnet_v2.resnet_utils.resnet_arg_scope(resnet_is_training)):
    fn = getattr(tf_utils, encoder)
    x, end_points = fn(x, num_classes=None, global_pool=False,
                       output_stride=None, reuse=None,
                       scope=scope_name)
  vars_ = slim.get_variables_to_restore()

  conv_feat = x
  return conv_feat, vars_ 

def build_inceptionv3_graph(images, endpoint, is_training, checkpoint,
                            reuse=False):
  """Builds an InceptionV3 model graph.

  Args:
    images: A 4-D float32 `Tensor` of batch images.
    endpoint: String, name of the InceptionV3 endpoint.
    is_training: Boolean, whether or not to build a training or inference graph.
    checkpoint: String, path to the pretrained model checkpoint.
    reuse: Boolean, whether or not we are reusing the embedder.
  Returns:
    inception_output: `Tensor` holding the InceptionV3 output.
    inception_variables: List of inception variables.
    init_fn: Function to initialize the weights (if not reusing, then None).
  """
  with slim.arg_scope(inception.inception_v3_arg_scope()):
    _, endpoints = inception.inception_v3(
        images, num_classes=1001, is_training=is_training)
    inception_output = endpoints[endpoint]
    inception_variables = slim.get_variables_to_restore()
    inception_variables = [
        i for i in inception_variables if 'global_step' not in i.name]
    if is_training and not reuse:
      init_saver = tf.train.Saver(inception_variables)
      def init_fn(scaffold, sess):
        del scaffold
        init_saver.restore(sess, checkpoint)
    else:
      init_fn = None
    return inception_output, inception_variables, init_fn 

def variables_to_restore(scope=None, strip_scope=False):
  """Returns a list of variables to restore for the specified list of methods.

  It is supposed that variable name starts with the method's scope (a prefix
  returned by _method_scope function).

  Args:
    methods_names: a list of names of configurable methods.
    strip_scope: if True will return variable names without method's scope.
      If methods_names is None will return names unchanged.
    model_scope: a scope for a whole model.

  Returns:
    a dictionary mapping variable names to variables for restore.
  """
  if scope:
    variable_map = {}
    method_variables = slim.get_variables_to_restore(include=[scope])
    for var in method_variables:
      if strip_scope:
        var_name = var.op.name[len(scope) + 1:]
      else:
        var_name = var.op.name
      variable_map[var_name] = var

    return variable_map
  else:
    return {v.op.name: v for v in slim.get_variables_to_restore()} 

def get_repr_from_image(images_reshaped, modalities, data_augment, encoder,
                        freeze_conv, wt_decay, is_training):
  # Pass image through lots of convolutional layers, to obtain pool5
  if modalities == ['rgb']:
    with tf.name_scope('pre_rgb'):
      x = (images_reshaped + 128.) / 255. # Convert to brightness between 0 and 1.
      if data_augment.relight and is_training:
        x = tf_utils.distort_image(x, fast_mode=data_augment.relight_fast)
      x = (x-0.5)*2.0
    scope_name = encoder
  elif modalities == ['depth']:
    with tf.name_scope('pre_d'):
      d_image = images_reshaped
      x = 2*(d_image[...,0] - 80.0)/100.0
      y = d_image[...,1]
      d_image = tf.concat([tf.expand_dims(x, -1), tf.expand_dims(y, -1)], 3)
      x = d_image
    scope_name = 'd_'+encoder

  resnet_is_training = is_training and (not freeze_conv)
  with slim.arg_scope(resnet_v2.resnet_utils.resnet_arg_scope(resnet_is_training)):
    fn = getattr(tf_utils, encoder)
    x, end_points = fn(x, num_classes=None, global_pool=False,
                       output_stride=None, reuse=None,
                       scope=scope_name)
  vars_ = slim.get_variables_to_restore()

  conv_feat = x
  return conv_feat, vars_ 

def finalize(preprocess_fn, network_factory, checkpoint_path, image_shape,
             output_filename):
    """Finalize model, i.e., strip off training variables and only save model
    variables to checkpoint file.

    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.
    checkpoint_path : str
        The checkpoint file to load.
    image_shape : Tuple[int, int, int]
        Image shape (height, width, channels).
    output_filename : str
        The checkpoint file to write.

    """
    with tf.Session(graph=tf.Graph()) as session:
        input_var = tf.placeholder(tf.uint8, (None, ) + image_shape)
        image_var = tf.map_fn(
            lambda x: preprocess_fn(x, is_training=False),
            input_var, back_prop=False, dtype=tf.float32)
        network_factory(image_var)

        loader = tf.train.Saver(slim.get_variables_to_restore())
        loader.restore(session, checkpoint_path)

        saver = tf.train.Saver(slim.get_model_variables())
        saver.save(session, output_filename, global_step=None) 

def build_inceptionv3_graph(images, endpoint, is_training, checkpoint,
                            reuse=False):
  """Builds an InceptionV3 model graph.

  Args:
    images: A 4-D float32 `Tensor` of batch images.
    endpoint: String, name of the InceptionV3 endpoint.
    is_training: Boolean, whether or not to build a training or inference graph.
    checkpoint: String, path to the pretrained model checkpoint.
    reuse: Boolean, whether or not we are reusing the embedder.
  Returns:
    inception_output: `Tensor` holding the InceptionV3 output.
    inception_variables: List of inception variables.
    init_fn: Function to initialize the weights (if not reusing, then None).
  """
  with slim.arg_scope(inception.inception_v3_arg_scope()):
    _, endpoints = inception.inception_v3(
        images, num_classes=1001, is_training=is_training)
    inception_output = endpoints[endpoint]
    inception_variables = slim.get_variables_to_restore()
    inception_variables = [
        i for i in inception_variables if 'global_step' not in i.name]
    if is_training and not reuse:
      init_saver = tf.train.Saver(inception_variables)
      def init_fn(scaffold, sess):
        del scaffold
        init_saver.restore(sess, checkpoint)
    else:
      init_fn = None
    return inception_output, inception_variables, init_fn 

def variables_to_restore(scope=None, strip_scope=False):
  """Returns a list of variables to restore for the specified list of methods.

  It is supposed that variable name starts with the method's scope (a prefix
  returned by _method_scope function).

  Args:
    methods_names: a list of names of configurable methods.
    strip_scope: if True will return variable names without method's scope.
      If methods_names is None will return names unchanged.
    model_scope: a scope for a whole model.

  Returns:
    a dictionary mapping variable names to variables for restore.
  """
  if scope:
    variable_map = {}
    method_variables = slim.get_variables_to_restore(include=[scope])
    for var in method_variables:
      if strip_scope:
        var_name = var.op.name[len(scope) + 1:]
      else:
        var_name = var.op.name
      variable_map[var_name] = var

    return variable_map
  else:
    return {v.op.name: v for v in slim.get_variables_to_restore()}