573c32dd3af062a57399e3f8e1b5be7ac0a3f375,frcnn/rpn.py,RPN,__init__,#RPN#,20

Before Change

        // According to Faster RCNN paper we need to initialize layers with
        // "from a zero-mean Gaussian distribution with standard deviation 0.01"
        // we use the truncated version (redraws when more than 2 std from mean.)
        self._initializer = tf.contrib.layers.variance_scaling_initializer(
            factor=1., uniform=True, mode="FAN_AVG"
        )

        // We could use normal relu without any problems.
        self._rpn_activation = tf.nn.relu6

After Change

        // According to Faster RCNN paper we need to initialize layers with
        // "from a zero-mean Gaussian distribution with standard deviation 0.0
        self._initializer = tf.random_normal_initializer(mean=0.0, stddev=0.01)

        // We could use normal relu without any problems.
        self._rpn_activation = tf.nn.relu6

Instances