state_representation.py

class AtariNet(object):
    # ...
    # ...
    def _build(self):
        # ...
        # ...
        # convolutional layers for minimap features
        self.minimap_conv1 = tf.layers.conv2d(
            inputs=self.minimap_processed,
            filters=16,
            kernel_size=[8, 8],
            strides=[4, 4],
            padding="SAME",
            name="minimap_conv1")

        self.minimap_activation1 = tf.nn.relu(
            self.minimap_conv1,
            name="minimap_activation1")

        self.minimap_conv2 = tf.layers.conv2d(
            inputs=self.minimap_activation1,
            filters=32,
            kernel_size=[4, 4],
            strides=[2, 2],
            padding="SAME",
            name="minimap_conv2")

        self.minimap_activation2 = tf.nn.relu(
            self.minimap_conv2,
            name="minimap_activation2")

        # linear layer for non-spatial features (tanh activation)
        self.flat_linear = tf.layers.dense(
            inputs=self.flat_processed,
            units=64,
            activation=tf.tanh,
            name="flat_linear")

        # flatten and concatenate
        self.screen_flat = tf.layers.flatten(
            self.screen_activation2,
            name="screen_flat")

        self.minimap_flat = tf.layers.flatten(
            self.minimap_activation2,
            name="minimap_flat")

        self.concat = tf.concat(
            values=[self.screen_flat, self.minimap_flat, self.flat_linear],
            axis=1,
            name="concat")

        # linear layer with ReLU activation
        self.state_representation = tf.layers.dense(
            inputs=self.concat,
            units=256,
            activation=tf.nn.relu,
            name="state_representation")
        # ...
        # ...