run_dqn_atari.py 文件源码

def atari_model(img_in, num_actions, scope, reuse=False):
    # as described in https://storage.googleapis.com/deepmind-data/assets/papers/DeepMindNature14236Paper.pdf
    with tf.variable_scope(scope, reuse=reuse):
        out = img_in
        with tf.variable_scope("convnet"):
            # original architecture
            out = layers.convolution2d(out, num_outputs=32, kernel_size=8, stride=4, activation_fn=tf.nn.relu)
            out = layers.convolution2d(out, num_outputs=64, kernel_size=4, stride=2, activation_fn=tf.nn.relu)
            out = layers.convolution2d(out, num_outputs=64, kernel_size=3, stride=1, activation_fn=tf.nn.relu)
        out = layers.flatten(out)
        with tf.variable_scope("action_value"):
            out = layers.fully_connected(out, num_outputs=512,         activation_fn=tf.nn.relu)
            out = layers.fully_connected(out, num_outputs=num_actions, activation_fn=None)

        return out