f59154b6dd8116bf1d5e0757ffdab19fb93199ca,models/official/detection/modeling/architecture/heads.py,RetinanetHead,class_net,#RetinanetHead#Any#Any#Any#,333

Before Change

        self._num_classes * self._anchors_per_location,
        kernel_size=(3, 3),
        bias_initializer=tf.constant_initializer(-np.log((1 - 0.01) / 0.01)),
        kernel_initializer=tf.random_normal_initializer(stddev=1e-5),
        padding="same",
        name="class-predict")
    return classes

After Change

            tf.layers.separable_conv2d, depth_multiplier=1)
      else:
        conv2d_op = functools.partial(
            tf.layers.conv2d, kernel_initializer=tf.random_normal_initializer(
                stddev=0.01))
      features = conv2d_op(
          features,
          self._num_filters,
          kernel_size=(3, 3),
          bias_initializer=tf.zeros_initializer(),
          activation=None,
          padding="same",
          name="class-"+str(i))
      // The convolution layers in the class net are shared among all levels, but
      // each level has its batch normlization to capture the statistical
      // difference among different levels.
      features = self._batch_norm_relu(features, is_training=is_training,
                                       name="class-%d-%d"%(i, level),)
    if self._use_separable_conv:
      conv2d_op = functools.partial(
          tf.layers.separable_conv2d, depth_multiplier=1)
    else:
      conv2d_op = functools.partial(
          tf.layers.conv2d, kernel_initializer=tf.random_normal_initializer(
              stddev=1e-5))
    classes = conv2d_op(
        features,
        self._num_classes * self._anchors_per_location,
        kernel_size=(3, 3),

Instances