Python tensorflow.contrib.slim 模块，max_pool2d() 实例源码

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

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

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

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

def vgg_arg_scope(weight_decay=0.0005):
  """Defines the VGG arg scope.

  Args:
    weight_decay: The l2 regularization coefficient.

  Returns:
    An arg_scope.
  """
  with slim.arg_scope([slim.conv2d, slim.fully_connected],
                      activation_fn=tf.nn.relu,
                      weights_regularizer=slim.l2_regularizer(weight_decay),
                      biases_initializer=tf.zeros_initializer()):
      with slim.arg_scope([slim.conv2d], padding='SAME') as arg_sc:
          with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
              return arg_sc

def _extra_conv_arg_scope_with_bn(weight_decay=0.00001,
                     activation_fn=None,
                     batch_norm_decay=0.997,
                     batch_norm_epsilon=1e-5,
                     batch_norm_scale=True):

  batch_norm_params = {
      'decay': batch_norm_decay,
      'epsilon': batch_norm_epsilon,
      'scale': batch_norm_scale,
      'updates_collections': tf.GraphKeys.UPDATE_OPS,
  }

  with slim.arg_scope(
      [slim.conv2d],
      weights_regularizer=slim.l2_regularizer(weight_decay),
      weights_initializer=slim.variance_scaling_initializer(),
      activation_fn=tf.nn.relu,
      normalizer_fn=slim.batch_norm,
      normalizer_params=batch_norm_params):
    with slim.arg_scope([slim.batch_norm], **batch_norm_params):
      with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
        return arg_sc

def encoder(self, x):
        with tf.variable_scope('encoder'):
            net = resnet_utils.conv2d_same(x, 64, 7, stride=2, scope='conv1')
            net = tf.pad(net, [[0, 0], [1, 1], [1, 1], [0, 0]])
            x = slim.max_pool2d(net, [3, 3], stride=2, padding='VALID', scope='pool1')
            x_features_all, _ = resnet_v1.resnet_v1(x,
                                                self._blocks_encoder,
                                                global_pool=False,
                                                include_root_block=False,
                                                scope=self._resnet_scope)
            x_features_all = tf.reduce_mean(x_features_all, axis=[1, 2])
            x_features_labeled, x_features_unlabeled = tf.split(x_features_all, 2)

        x_features_tiled = tf.tile(x_features_unlabeled, [self._num_classes, 1])  # (100, 256) --> (2100, 256)
        x_features = tf.concat([x_features_labeled, x_features_tiled], 0) # (2100, 256) --> (2200, 256)
        return x_features

def encoder(self, x):
        with tf.variable_scope('encoder'):
            net = resnet_utils.conv2d_same(x, 64, 7, stride=2, scope='conv1')
            net = tf.pad(net, [[0, 0], [1, 1], [1, 1], [0, 0]])
            x = slim.max_pool2d(net, [3, 3], stride=2, padding='VALID', scope='pool1')
            x_features_all, _ = resnet_v1.resnet_v1(x,
                                                self._blocks_encoder,
                                                global_pool=False,
                                                include_root_block=False,
                                                scope=self._resnet_scope)
            x_features_all = tf.reduce_mean(x_features_all, axis=[1, 2])
            x_features_labeled, x_features_unlabeled = tf.split(x_features_all, 2)

        x_features_tiled = tf.tile(x_features_unlabeled, [self._num_classes, 1])  # (100, 256) --> (2100, 256)
        x_features = tf.concat([x_features_labeled, x_features_tiled], 0) # (2100, 256) --> (2200, 256)
        return x_features

def squeezenet(inputs,
               num_classes=1000,
               is_training=True,
               keep_prob=0.5,
               spatial_squeeze=True,
               scope='squeeze'):
    """
    squeezenetv1.1
    """
    with tf.name_scope(scope, 'squeeze', [inputs]) as sc:
        end_points_collection = sc + '_end_points'
        # Collect outputs for conv2d, fully_connected and max_pool2d.
        with slim.arg_scope([slim.conv2d, slim.max_pool2d,
                             slim.avg_pool2d, fire_module],
                            outputs_collections=end_points_collection):
            nets = squeezenet_inference(inputs, is_training, keep_prob)
            nets = slim.conv2d(nets, num_classes, [1, 1],
                               activation_fn=None,
                               normalizer_fn=None,
                               scope='logits')
            end_points = slim.utils.convert_collection_to_dict(end_points_collection)
            if spatial_squeeze:
                nets = tf.squeeze(nets, [1, 2], name='logits/squeezed')
            return nets, end_points

def densenet_inference(inputs, is_training, keep_prob, growth_rate, reduction):

    first_output_fea = growth_rate * 2

    nets = slim.conv2d(inputs, first_output_fea,
                       [5, 5], scope='conv0')
    nets = slim.max_pool2d(nets, [3, 3], padding='SAME', scope='pool0')  # 56*48*64

    nets = densenet_block(nets, 6, growth_rate, True,
                          'block1', is_training, keep_prob)
    nets = transition_block(nets, reduction, 'trans1', is_training, keep_prob)  # 28*24*256

    nets = densenet_block(nets, 12, growth_rate, True,
                          'block2', is_training, keep_prob)
    nets = transition_block(nets, reduction, 'trans2', is_training, keep_prob)  # 14*12*640

    nets = densenet_block(nets, 24, growth_rate, True,
                          'block3', is_training, keep_prob)
    nets = transition_block(nets, reduction, 'trans3', is_training, keep_prob)  # 7*6*1408

    nets = densenet_block(nets, 16, growth_rate, True,
                          'block4', is_training, keep_prob)  # 7*6*1920
    nets = slim.avg_pool2d(nets, [7, 6], scope='pool4')  # 1*1*1920
    return nets

def _root_block(input,
                initial_conv_filters,
                weight_decay=5e-4,
                ksize=(7,7),
                is_pool=True):
  ''' Adds an initial conv block, with batch norm and relu for the DPN
  Args:
      input: input tensor
      initial_conv_filters: number of filters for initial conv block
      weight_decay: weight decay factor
  Returns: a keras tensor
  '''
  x = slim.conv2d(input,
                  initial_conv_filters,
                  ksize,
                  padding='SAME',
                  stride=(1, 1),
                  weights_regularizer=slim.l2_regularizer(weight_decay),
                  weights_initializer=tf.contrib.layers.xavier_initializer(),
                  biases_initializer=None)
  x = slim.batch_norm(x)
  x = tf.nn.relu(x)
  if is_pool:
    x = slim.max_pool2d(x, (3, 3), stride=(2, 2), padding='SAME')
  return x

def reduction_a(net, k, l, m, n):
    with tf.variable_scope('Branch_0'):
        tower_conv = slim.conv2d(net, n, 3, stride=2, padding='VALID',
                                 scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_1'):
        tower_conv1_0 = slim.conv2d(net, k, 1, scope='Conv2d_0a_1x1')
        tower_conv1_1 = slim.conv2d(tower_conv1_0, l, 3,
                                    scope='Conv2d_0b_3x3')
        tower_conv1_2 = slim.conv2d(tower_conv1_1, m, 3,
                                    stride=2, padding='VALID',
                                    scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_2'):
        tower_pool = slim.max_pool2d(net, 3, stride=2, padding='VALID',
                                     scope='MaxPool_1a_3x3')
    net = tf.concat([tower_conv, tower_conv1_2, tower_pool], 3)
    return net

def reduction_b(net):
    with tf.variable_scope('Branch_0'):
        tower_conv = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        tower_conv_1 = slim.conv2d(tower_conv, 384, 3, stride=2,
                                   padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_1'):
        tower_conv1 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        tower_conv1_1 = slim.conv2d(tower_conv1, 256, 3, stride=2,
                                    padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_2'):
        tower_conv2 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        tower_conv2_1 = slim.conv2d(tower_conv2, 256, 3,
                                    scope='Conv2d_0b_3x3')
        tower_conv2_2 = slim.conv2d(tower_conv2_1, 256, 3, stride=2,
                                    padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_3'):
        tower_pool = slim.max_pool2d(net, 3, stride=2, padding='VALID',
                                     scope='MaxPool_1a_3x3')
    net = tf.concat([tower_conv_1, tower_conv1_1,
                        tower_conv2_2, tower_pool], 3)
    return net

def _extra_conv_arg_scope_with_bn(weight_decay=0.00001,
                     activation_fn=None,
                     batch_norm_decay=0.997,
                     batch_norm_epsilon=1e-5,
                     batch_norm_scale=True):

  batch_norm_params = {
      'decay': batch_norm_decay,
      'epsilon': batch_norm_epsilon,
      'scale': batch_norm_scale,
      'updates_collections': tf.GraphKeys.UPDATE_OPS_EXTRA,
  }

  with slim.arg_scope(
      [slim.conv2d],
      weights_regularizer=slim.l2_regularizer(weight_decay),
      weights_initializer=slim.variance_scaling_initializer(),
      activation_fn=tf.nn.relu,
      normalizer_fn=slim.batch_norm,
      normalizer_params=batch_norm_params):
    with slim.arg_scope([slim.batch_norm], **batch_norm_params):
      with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
        return arg_sc

def subsample(inputs, factor, scope=None):
  """Subsamples the input along the spatial dimensions.

  Args:
    inputs: A `Tensor` of size [batch, height_in, width_in, channels].
    factor: The subsampling factor.
    scope: Optional variable_scope.

  Returns:
    output: A `Tensor` of size [batch, height_out, width_out, channels] with the
      input, either intact (if factor == 1) or subsampled (if factor > 1).
  """
  if factor == 1:
    return inputs
  else:
    return slim.max_pool2d(inputs, [1, 1], stride=factor, scope=scope)

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

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

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

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

def _image_to_head(self, is_training, reuse=None):
    with tf.variable_scope(self._scope, self._scope, reuse=reuse):
      net = slim.repeat(self._image, 2, slim.conv2d, 64, [3, 3],
                          trainable=False, scope='conv1')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool1')
      net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
                        trainable=False, scope='conv2')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool2')
      net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3],
                        trainable=is_training, scope='conv3')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool3')
      net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
                        trainable=is_training, scope='conv4')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool4')
      net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
                        trainable=is_training, scope='conv5')

    self._act_summaries.append(net)
    self._layers['head'] = net

    return net

def create_network(self, input, trainable):
        if trainable:
            wr = slim.l2_regularizer(self.regularization)
        else:
            wr = None

        # the input is stack of black and white frames.
        # put the stack in the place of channel (last in tf)
        input_t = tf.transpose(input, [0, 2, 3, 1])

        net = slim.conv2d(input_t, 8, (7, 7), data_format="NHWC",
            activation_fn=tf.nn.relu, stride=3, weights_regularizer=wr, trainable=trainable)
        net = slim.max_pool2d(net, 2, 2)
        net = slim.conv2d(net, 16, (3, 3), data_format="NHWC",
            activation_fn=tf.nn.relu, weights_regularizer=wr, trainable=trainable)
        net = slim.max_pool2d(net, 2, 2)
        net = slim.flatten(net)
        net = slim.fully_connected(net, 256, activation_fn=tf.nn.relu,
            weights_regularizer=wr, trainable=trainable)
        q_state_action_values = slim.fully_connected(net, self.dim_actions,
            activation_fn=None, weights_regularizer=wr, trainable=trainable)

        return q_state_action_values

def __arg_scope(self, weight_decay=0.0005, data_format='NHWC'):
        """Defines the VGG arg scope.

        Args:
          weight_decay: The l2 regularization coefficient.

        Returns:
          An arg_scope.
        """
        with slim.arg_scope([slim.conv2d, slim.fully_connected],
                            activation_fn=tf.nn.relu,
                            weights_regularizer=slim.l2_regularizer(weight_decay),
                            weights_initializer=tf.contrib.layers.xavier_initializer(),
                            biases_initializer=tf.zeros_initializer()):
            with slim.arg_scope([slim.conv2d, slim.max_pool2d],
                                padding='SAME',
                                data_format=data_format):
                with slim.arg_scope([custom_layers.pad2d,
                                     custom_layers.l2_normalization,
                                     custom_layers.channel_to_last],
                                    data_format=data_format) as sc:
                    return sc

def cnn_layers(inputs, scope, end_points_collection, dropout_keep_prob=0.8, is_training=True):
  with slim.arg_scope([slim.conv2d, slim.fully_connected, slim.max_pool2d],
                      outputs_collections=[end_points_collection]):
    with slim.arg_scope([slim.conv2d],
                        normalizer_fn=slim.batch_norm, normalizer_params={'is_training': is_training},
                        activation_fn=leaky_relu):
      net = slim.conv2d(inputs, 32, [3, 3], scope='conv1')
      net = slim.max_pool2d(net, [2, 2], 2, scope='pool1')
      net = slim.conv2d(net, 64, [3, 3], scope='conv2')
      net = slim.max_pool2d(net, [2, 2], 2, scope='pool2')
      net = slim.conv2d(net, 128, [3, 3], scope='conv3')
      net = slim.conv2d(net, 64, [1, 1], scope='conv4')
      box_net = net = slim.conv2d(net, 128, [3, 3], scope='conv5')

      net = slim.max_pool2d(net, [2, 2], 2, scope='pool5')
      net = slim.conv2d(net, 256, [3, 3], scope='conv6')
      net = slim.conv2d(net, 128, [1, 1], scope='conv7')
      net = slim.conv2d(net, 256, [3, 3], scope='conv8')

      box_net = _reorg(box_net, 2)
      net = tf.concat([box_net, net], 3)
      net = slim.conv2d(net, 256, [3, 3], scope='conv9')
      net = slim.conv2d(net, 75, [1, 1], activation_fn=None, scope='conv10')

  return net, end_points_collection

def cnn_layers(inputs, scope, end_points_collection, dropout_keep_prob=0.8, is_training=True):
  with slim.arg_scope([slim.conv2d, slim.fully_connected, slim.max_pool2d],
                      outputs_collections=[end_points_collection]):
    net = slim.conv2d(inputs, 48, [5, 5], scope='conv1')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool1')
    net = slim.conv2d(net, 64, [5, 5], scope='conv2')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool2')
    net = slim.conv2d(net, 128, [5, 5], scope='conv3')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool3')
    net = slim.conv2d(net, 160, [5, 5], scope='conv4')
    net = slim.conv2d(net, 192, [5, 5], scope='conv5')
    net = slim.conv2d(net, 192, [5, 5], scope='conv6')
    net = slim.conv2d(net, 192, [5, 5], scope='conv7')
    net = slim.flatten(net)

    # By removing the fc layer, we'll get much smaller model with almost the same performance
    # net = slim.fully_connected(net, 3072, scope='fc8')

  return net, end_points_collection

def cnn_layers(inputs, scope, end_points_collection, dropout_keep_prob=0.8, is_training=True):
  with slim.arg_scope([slim.conv2d, slim.fully_connected, slim.max_pool2d],
                      outputs_collections=[end_points_collection]):
    net = slim.conv2d(inputs, 32, [5, 5], scope='conv1')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool1')
    net = slim.conv2d(net, 64, [5, 5], scope='conv2')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool2')
    net = slim.conv2d(net, 64, [5, 5], scope='conv3')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool3')
    net = slim.conv2d(net, 64, [5, 5], scope='conv4')
    net = slim.conv2d(net, 64, [5, 5], scope='conv5')
    net = slim.conv2d(net, 64, [5, 5], scope='conv6')
    net = slim.conv2d(net, 64, [5, 5], scope='conv7')
    net = slim.flatten(net)

    net = slim.fully_connected(net, 256, scope='fc3')

  return net, end_points_collection

def cnn_layers(inputs, scope, end_points_collection, dropout_keep_prob=0.8, is_training=True):
  with slim.arg_scope([slim.conv2d, slim.fully_connected, slim.max_pool2d],
                      outputs_collections=[end_points_collection]):
    net = slim.conv2d(inputs, 32, [5, 5], scope='conv1')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool1')
    net = slim.conv2d(net, 64, [5, 5], scope='conv2')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool2')
    net = slim.conv2d(net, 64, [5, 5], scope='conv3')
    net = slim.max_pool2d(net, [2, 2], 2, scope='pool3')
    net = slim.conv2d(net, 64, [5, 5], scope='conv4')
    net = slim.conv2d(net, 64, [5, 5], scope='conv5')
    net = slim.conv2d(net, 64, [5, 5], scope='conv6')
    net = slim.conv2d(net, 64, [5, 5], scope='conv7')
    net = slim.flatten(net)

    net = slim.fully_connected(net, 128, scope='fc3')

  return net, end_points_collection

def reduction_a(net, k, l, m, n):
    with tf.variable_scope('Branch_0'):
        tower_conv = slim.conv2d(net, n, 3, stride=2, padding='VALID',
                                 scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_1'):
        tower_conv1_0 = slim.conv2d(net, k, 1, scope='Conv2d_0a_1x1')
        tower_conv1_1 = slim.conv2d(tower_conv1_0, l, 3,
                                    scope='Conv2d_0b_3x3')
        tower_conv1_2 = slim.conv2d(tower_conv1_1, m, 3,
                                    stride=2, padding='VALID',
                                    scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_2'):
        tower_pool = slim.max_pool2d(net, 3, stride=2, padding='VALID',
                                     scope='MaxPool_1a_3x3')
    net = tf.concat([tower_conv, tower_conv1_2, tower_pool], 3)
    return net

def reduction_b(net):
    with tf.variable_scope('Branch_0'):
        tower_conv = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        tower_conv_1 = slim.conv2d(tower_conv, 384, 3, stride=2,
                                   padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_1'):
        tower_conv1 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        tower_conv1_1 = slim.conv2d(tower_conv1, 256, 3, stride=2,
                                    padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_2'):
        tower_conv2 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        tower_conv2_1 = slim.conv2d(tower_conv2, 256, 3,
                                    scope='Conv2d_0b_3x3')
        tower_conv2_2 = slim.conv2d(tower_conv2_1, 256, 3, stride=2,
                                    padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_3'):
        tower_pool = slim.max_pool2d(net, 3, stride=2, padding='VALID',
                                     scope='MaxPool_1a_3x3')
    net = tf.concat([tower_conv_1, tower_conv1_1,
                        tower_conv2_2, tower_pool], 3)
    return net

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

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

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

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

def _image_to_head(self, is_training, reuse=False):
    with tf.variable_scope(self._scope, self._scope, reuse=reuse):
      net = slim.repeat(self._image, 2, slim.conv2d, 64, [3, 3],
                          trainable=False, scope='conv1')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool1')
      net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
                        trainable=False, scope='conv2')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool2')
      net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3],
                        trainable=is_training, scope='conv3')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool3')
      net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
                        trainable=is_training, scope='conv4')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool4')
      net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
                        trainable=is_training, scope='conv5')

    self._act_summaries.append(net)
    self._layers['head'] = net

    return net

def vgg16(inputs, num_classes, batch_size):
    with slim.arg_scope([slim.conv2d, slim.fully_connected],
                        activation_fn=tf.nn.relu,
                        weights_initializer=tf.truncated_normal_initializer(0.0, 0.01)):
        net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], padding="SAME", scope='conv1')
        net = slim.max_pool2d(net, [2, 2], scope='pool1')
        net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], padding="SAME", scope='conv2')
        net = slim.max_pool2d(net, [2, 2], scope='pool2')
        net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], padding="SAME", scope='conv3')
        net = slim.max_pool2d(net, [2, 2], scope='pool3')
        net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], padding="SAME", scope='conv4')
        net = slim.max_pool2d(net, [2, 2], scope='pool4')
        net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], padding="SAME", scope='conv5')
        net = slim.max_pool2d(net, [2, 2], scope='pool5')
        net = tf.reshape(net, (batch_size, 7 * 7 * 512))
        net = slim.fully_connected(net, 4096, scope='fc6')
        net = slim.dropout(net, 0.5, scope='dropout6')
        net = slim.fully_connected(net, 4096, scope='fc7')
        net = slim.dropout(net, 0.5, scope='dropout7')
        net = slim.fully_connected(net, 1000, activation_fn=None, scope='fc8')
    return net

def create_model(self, model_input, num_classes=2, l2_penalty=1e-8, **unused_params):
    net = slim.conv2d(model_input, 64, [3, 3], scope='conv1_1')
    # net = slim.conv2d(net, 64, [3, 3], scope='conv1_2')
    net = slim.max_pool2d(net, [2, 2], scope='pool1')

    # net = slim.conv2d(net, 128, [3, 3], scope='conv2_1')
    # net = slim.conv2d(net, 128, [3, 3], scope='conv2_2')
    # net = slim.max_pool2d(net, [2, 2], scope='pool2')

    # net = slim.conv2d(net, 258, [3, 3], scope='conv3_1')
    # net = slim.conv2d(net, 258, [3, 3], scope='conv3_2')
    # net = slim.max_pool2d(net, [2, 2], scope='pool3')

    net = slim.flatten(net)
    output = slim.fully_connected(
        net, num_classes - 1, activation_fn=tf.nn.sigmoid,
        weights_regularizer=slim.l2_regularizer(l2_penalty))
    return {"predictions": output}

def _extra_conv_arg_scope_with_bn(weight_decay=0.00001,
                     activation_fn=None,
                     batch_norm_decay=0.997,
                     batch_norm_epsilon=1e-5,
                     batch_norm_scale=True):

  batch_norm_params = {
      'decay': batch_norm_decay,
      'epsilon': batch_norm_epsilon,
      'scale': batch_norm_scale,
      'updates_collections': tf.GraphKeys.UPDATE_OPS,
  }

  with slim.arg_scope(
      [slim.conv2d],
      weights_regularizer=slim.l2_regularizer(weight_decay),
      weights_initializer=slim.variance_scaling_initializer(),
      activation_fn=tf.nn.relu,
      normalizer_fn=slim.batch_norm,
      normalizer_params=batch_norm_params):
    with slim.arg_scope([slim.batch_norm], **batch_norm_params):
      with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
        return arg_sc

def subsample(inputs, factor, scope=None):
  """Subsamples the input along the spatial dimensions.

  Args:
    inputs: A `Tensor` of size [batch, height_in, width_in, channels].
    factor: The subsampling factor.
    scope: Optional variable_scope.

  Returns:
    output: A `Tensor` of size [batch, height_out, width_out, channels] with the
      input, either intact (if factor == 1) or subsampled (if factor > 1).
  """
  if factor == 1:
    return inputs
  else:
    return slim.max_pool2d(inputs, [1, 1], stride=factor, scope=scope)

def max_pool_with_argmax(net, stride):
  """
  Tensorflow default implementation does not provide gradient operation on max_pool_with_argmax
  Therefore, we use max_pool_with_argmax to extract mask and
  plain max_pool for, eeem... max_pooling.
  """
  with tf.name_scope('MaxPoolArgMax'):
    _, mask = tf.nn.max_pool_with_argmax(
      net,
      ksize=[1, stride, stride, 1],
      strides=[1, stride, stride, 1],
      padding='SAME')
    mask = tf.stop_gradient(mask)
    net = slim.max_pool2d(net, kernel_size=[stride, stride],  stride=FLAGS.pool_size)
    return net, mask


# Thank you, @https://github.com/Pepslee

def reduction_a(net, k, l, m, n):
    with tf.variable_scope('Branch_0'):
        tower_conv = slim.conv2d(net, n, 3, stride=2, padding='VALID',
                                 scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_1'):
        tower_conv1_0 = slim.conv2d(net, k, 1, scope='Conv2d_0a_1x1')
        tower_conv1_1 = slim.conv2d(tower_conv1_0, l, 3,
                                    scope='Conv2d_0b_3x3')
        tower_conv1_2 = slim.conv2d(tower_conv1_1, m, 3,
                                    stride=2, padding='VALID',
                                    scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_2'):
        tower_pool = slim.max_pool2d(net, 3, stride=2, padding='VALID',
                                     scope='MaxPool_1a_3x3')
    net = tf.concat([tower_conv, tower_conv1_2, tower_pool], 3)
    return net

def reduction_b(net):
    with tf.variable_scope('Branch_0'):
        tower_conv = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        tower_conv_1 = slim.conv2d(tower_conv, 384, 3, stride=2,
                                   padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_1'):
        tower_conv1 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        tower_conv1_1 = slim.conv2d(tower_conv1, 256, 3, stride=2,
                                    padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_2'):
        tower_conv2 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
        tower_conv2_1 = slim.conv2d(tower_conv2, 256, 3,
                                    scope='Conv2d_0b_3x3')
        tower_conv2_2 = slim.conv2d(tower_conv2_1, 256, 3, stride=2,
                                    padding='VALID', scope='Conv2d_1a_3x3')
    with tf.variable_scope('Branch_3'):
        tower_pool = slim.max_pool2d(net, 3, stride=2, padding='VALID',
                                     scope='MaxPool_1a_3x3')
    net = tf.concat([tower_conv_1, tower_conv1_1,
                     tower_conv2_2, tower_pool], 3)
    return net

def _build_network(self):

        with slim.arg_scope([slim.conv2d],
                        activation_fn=tf.nn.relu,
                        weights_regularizer=slim.l2_regularizer(self.weight_decay),
                        weights_initializer= self.weights_initializer,
                        biases_initializer = self.biases_initializer):
            with slim.arg_scope([slim.conv2d, slim.max_pool2d],
                                padding='SAME',
                                data_format = self.data_format):
                with tf.variable_scope(self.basenet_type):
                    basenet, end_points = net_factory.get_basenet(self.basenet_type, self.inputs);

                with tf.variable_scope('extra_layers'):
                    self.net, self.end_points = self._add_extra_layers(basenet, end_points);

                with tf.variable_scope('seglink_layers'):
                    self._add_seglink_layers();

def _extra_conv_arg_scope_with_bn(weight_decay=0.00001,
                     activation_fn=None,
                     batch_norm_decay=0.997,
                     batch_norm_epsilon=1e-5,
                     batch_norm_scale=True):

  batch_norm_params = {
      'decay': batch_norm_decay,
      'epsilon': batch_norm_epsilon,
      'scale': batch_norm_scale,
      'updates_collections': tf.GraphKeys.UPDATE_OPS_EXTRA,
  }

  with slim.arg_scope(
      [slim.conv2d],
      weights_regularizer=slim.l2_regularizer(weight_decay),
      weights_initializer=slim.variance_scaling_initializer(),
      activation_fn=tf.nn.relu,
      normalizer_fn=slim.batch_norm,
      normalizer_params=batch_norm_params):
    with slim.arg_scope([slim.batch_norm], **batch_norm_params):
      with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
        return arg_sc

def subsample(inputs, factor, scope=None):
  """Subsamples the input along the spatial dimensions.

  Args:
    inputs: A `Tensor` of size [batch, height_in, width_in, channels].
    factor: The subsampling factor.
    scope: Optional variable_scope.

  Returns:
    output: A `Tensor` of size [batch, height_out, width_out, channels] with the
      input, either intact (if factor == 1) or subsampled (if factor > 1).
  """
  if factor == 1:
    return inputs
  else:
    return slim.max_pool2d(inputs, [1, 1], stride=factor, scope=scope)

def _build_base(self):
    with tf.variable_scope(self._scope, self._scope):
      net = resnet_utils.conv2d_same(self._image, 64, 7, stride=2, scope='conv1')
      net = tf.pad(net, [[0, 0], [1, 1], [1, 1], [0, 0]])
      net = slim.max_pool2d(net, [3, 3], stride=2, padding='VALID', scope='pool1')

    return net

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

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

def _image_to_head(self, is_training, reuse=False):
    with tf.variable_scope(self._scope, self._scope, reuse=reuse):

      # [VGG16] conv1
      # input shape : 224 * 224 * 3
      # output shape : 112 * 112 * 64
      net = slim.repeat(self._image, 2, slim.conv2d, 64, [3, 3],
                          trainable=False, scope='conv1')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool1')

      # [VGG16] conv2
      # input shape : 112 * 112 * 64
      # output shape : 56 * 56 * 128
      net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
                        trainable=False, scope='conv2')
      net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool2')

      # [Hand Detection] REMOVE net = slim.max_pool2d(net, [2, 2], padding='SAME', scope='pool3') 
      # [Hand Detection] conv3
      # input shape : 56 * 56 * 128
      # output shape : 56 * 56 * 256
      net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3],
                        trainable=is_training, scope='conv3')
      to_be_normalized_1 = net

      # [Hand Detection] conv4
      # input shape : 56 * 56 * 256
      # output shape : 56 * 56 * 256
      net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
                        trainable=is_training, scope='conv4')
      to_be_normalized_2 = net 

      # [Hand Detection] conv5
      # input shape : 56 * 56 * 256
      # output shape : 56 * 56 * 256
      net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
                        trainable=is_training, scope='conv5')
      to_be_normalized_3 = net

      return to_be_normalized_1, to_be_normalized_2, to_be_normalized_3