5bae78c35afe28b014d78b9fe3b78c7f37785091,examples/atari_dqn/convnet.py,ConvNet,_build,#ConvNet#Any#,88

Before Change

    def _build(self, convnet_pars):
        with tf.variable_scope(None, default_name=self._name):
            self._scope_name = tf.get_default_graph().get_name_scope()
            self._x = tf.placeholder(tf.float32,
                                     shape=[None,
                                            convnet_pars["height"],
                                            convnet_pars["width"],
                                            convnet_pars["history_length"]],
                                     name="input")
            hidden_1 = tf.layers.conv2d(
                self._x, 32, 8, 4, activation=tf.nn.relu,
                kernel_initializer=tf.glorot_uniform_initializer(),
                bias_initializer=tf.glorot_uniform_initializer(),
                name="hidden_1"
            )
            hidden_2 = tf.layers.conv2d(
                hidden_1, 64, 4, 2, activation=tf.nn.relu,
                kernel_initializer=tf.glorot_uniform_initializer(),
                bias_initializer=tf.glorot_uniform_initializer(),
                name="hidden_2"
            )
            hidden_3 = tf.layers.conv2d(
                hidden_2, 64, 3, 1, activation=tf.nn.relu,
                kernel_initializer=tf.glorot_uniform_initializer(),
                bias_initializer=tf.glorot_uniform_initializer(),
                name="hidden_3"
            )
            flatten = tf.reshape(hidden_3, [-1, 7 * 7 * 64], name="flatten")
            self._features = tf.layers.dense(
                flatten, 512, activation=tf.nn.relu,
                kernel_initializer=tf.glorot_uniform_initializer(),
                bias_initializer=tf.glorot_uniform_initializer(),
                name="features"
            )
            self.q = tf.layers.dense(
                self._features, convnet_pars["n_actions"],
                kernel_initializer=tf.glorot_uniform_initializer(),
                bias_initializer=tf.glorot_uniform_initializer(),
                name="q"
            )

            self._target_q = tf.placeholder("float32", [None], name="target_q")
            self._action = tf.placeholder("uint8", [None], name="action")

            action_one_hot = tf.one_hot(self._action,
                                        convnet_pars["n_actions"],
                                        name="action_one_hot")
            self._q_acted = tf.reduce_sum(self.q * action_one_hot,
                                          axis=1,
                                          name="q_acted")

            loss = tf.losses.huber_loss(self._target_q, self._q_acted)
            tf.summary.scalar("huber_loss", loss)
            tf.summary.scalar("average_q", tf.reduce_mean(self.q))
            self._merged = tf.summary.merge(
                tf.get_collection(tf.GraphKeys.SUMMARIES,
                                  scope=self._scope_name)
            )

After Change

    def _build(self, convnet_pars):
        with tf.variable_scope(None, default_name=self._name):
            self._scope_name = tf.get_default_graph().get_name_scope() + "/"
            self._x = tf.placeholder(tf.float32,
                                     shape=[None,
                                            convnet_pars["height"],
                                            convnet_pars["width"],
                                            convnet_pars["history_length"]],
                                     name="input")
            hidden_1 = tf.layers.conv2d(
                self._x, 32, 8, 4, activation=tf.nn.relu,
                kernel_initializer=tf.glorot_uniform_initializer(),
                bias_initializer=tf.glorot_uniform_initializer(),
                name="hidden_1"
            )
            hidden_2 = tf.layers.conv2d(
                hidden_1, 64, 4, 2, activation=tf.nn.relu,
                kernel_initializer=tf.glorot_uniform_initializer(),
                bias_initializer=tf.glorot_uniform_initializer(),
                name="hidden_2"
            )
            hidden_3 = tf.layers.conv2d(
                hidden_2, 64, 3, 1, activation=tf.nn.relu,
                kernel_initializer=tf.glorot_uniform_initializer(),
                bias_initializer=tf.glorot_uniform_initializer(),
                name="hidden_3"
            )
            flatten = tf.reshape(hidden_3, [-1, 7 * 7 * 64], name="flatten")
            self._features = tf.layers.dense(
                flatten, 512, activation=tf.nn.relu,
                kernel_initializer=tf.glorot_uniform_initializer(),
                bias_initializer=tf.glorot_uniform_initializer(),
                name="features"
            )
            self.q = tf.layers.dense(
                self._features, convnet_pars["n_actions"],
                kernel_initializer=tf.glorot_uniform_initializer(),
                bias_initializer=tf.glorot_uniform_initializer(),
                name="q"
            )

            self._target_q = tf.placeholder("float32", [None], name="target_q")
            self._action = tf.placeholder("uint8", [None], name="action")

            action_one_hot = tf.one_hot(self._action,
                                        convnet_pars["n_actions"],
                                        name="action_one_hot")
            self._q_acted = tf.reduce_sum(self.q * action_one_hot,
                                          axis=1,
                                          name="q_acted")

            loss = tf.losses.huber_loss(self._target_q, self._q_acted)
            tf.summary.scalar("huber_loss", loss)
            tf.summary.scalar("average_q", tf.reduce_mean(self.q))
            self._merged = tf.summary.merge(
                tf.get_collection(tf.GraphKeys.SUMMARIES,
                                  scope=self._scope_name)
            )

Instances