def main(_):
batch_size = 64 # used in qfunc and runner.
env = environment_impl.GymEnvironment(gym.make('MountainCar-v0'))
env.SetGymEnvMaxEpisodeSteps(400)
qfunc = qfunc_impl.DQN(
model=qfunc_impl.CreateModel(
state_shape=env.GetStateShape(),
action_space_size=env.GetActionSpaceSize(),
hidden_layer_sizes=(64, )),
training_batch_size=batch_size,
discount_factor=0.99,
)
qfunc.Load('saved_models/mountaincar_shape_64_rmsprop_gamma_099.weights')
policy = policy_impl.GreedyPolicy()
runner = runner_impl.NoOpRunner()
env.TurnOnRendering(should_render=True, fps=24)
logging.ENV.debug_verbosity = 9
env.StartRecording(video_filename='mountaincar_demo.mp4')
# First 5 runs with random actions:
rand_qfunc = qfunc_impl.RandomQFunction(env.GetActionSpaceSize())
runner.Run(env=env, brain=rand_qfunc, policy=policy, num_of_episodes=5)
# Then 10 runs with trained qfunc:
runner.Run(env=env, brain=qfunc, policy=policy, num_of_episodes=10)
env.StopRecording()
def main(_):
env = environment_impl.GymEnvironment(gym.make('Seaquest-v0'))
env.TurnOnRendering(should_render=True, fps=24)
qfunc = qfunc_impl.RandomValueQFunction(
action_space_size=env.GetActionSpaceSize())
policy = policy_impl.GreedyPolicyWithRandomness(epsilon=1.0)
runner = runner_impl.NoOpRunner()
runner.Run(env, qfunc, policy, num_of_episodes=10)
def Demo(self,
num_of_episodes: int = 10,
save_video_to: t.Text = 'demo.mp4'):
"""Starts a demo run.
Args:
num_of_episodes: number of runs to demo.
save_video_to: saves the demo video for the run to a file of this
name. It must ends with mp4.
"""
self.env.TurnOnRendering(should_render=True, fps=24)
self.env.StartRecording(video_filename=save_video_to)
runner_impl.NoOpRunner().Run(env=self.env,
brain=self.qfunc,
policy=policy_impl.GreedyPolicy(),
num_of_episodes=num_of_episodes)
self.env.StopRecording()
self.env.TurnOnRendering(should_render=False)
def main(_):
batch_size = 64 # used in qfunc and runner.
env = environment_impl.GymEnvironment(gym.make('Acrobot-v1'))
qfunc = qfunc_impl.DQN(
model=qfunc_impl.CreateModel(
state_shape=env.GetStateShape(),
action_space_size=env.GetActionSpaceSize(),
hidden_layer_sizes=(20, 20, 20)),
training_batch_size=batch_size,
discount_factor=0.99,
)
qfunc.LoadModel(
'saved_models/acrobot_v1_shape_20-20-20_rmsprop_gamma_0.99.model')
policy = policy_impl.GreedyPolicy()
runner = runner_impl.NoOpRunner()
env.TurnOnRendering(should_render=True, fps=10)
logging.ENV.debug_verbosity = 9
runner.Run(env=env, qfunc=qfunc, policy=policy, num_of_episodes=10)
def setUp(self) -> None:
self.env = mock.MagicMock()
self.qfunc = mock.MagicMock()
self.policy = mock.MagicMock()
self.runner = runner_impl.NoOpRunner()