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(_):
batch_size = 64 # used in qfunc and runner.
env = environment_impl.GymEnvironment(gym.make('CartPole-v0'))
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,
)
runner = runner_impl.ExperienceReplayRunner(
experience_capacity=100000, experience_sample_batch_size=batch_size)
# Train 500 episodes.
logging.ENV.debug_verbosity = 3
policy = policy_impl.GreedyPolicyWithRandomness(epsilon=0.1)
runner.Run(env=env, brain=qfunc, policy=policy, num_of_episodes=500)
# Test for 100 episodes.
logging.ENV.debug_verbosity = 4
policy = policy_impl.GreedyPolicy()
runner.Run(env=env, brain=qfunc, policy=policy, num_of_episodes=100)
# Demo with video.
env.TurnOnRendering(should_render=True, fps=24)
# env.StartRecording(video_filename='demo.mp4') # uncomment to record video.
# First 5 runs with random actions:
runner.Run(env=env,
brain=qfunc_impl.RandomQFunction(env.GetActionSpaceSize()),
policy=policy,
num_of_episodes=5)
# Then 10 runs with trained qfunc:
runner.Run(env=env, brain=qfunc, policy=policy, num_of_episodes=10)
def setUp(self) -> None:
self.env = environment_impl.SingleStateEnvironment(
action_space_size=1, step_limit=10)
self.qfunc = qfunc_impl.MemoizationQFunction(
action_space_size=3,
discount_factor=0.9,
learning_rate=0.9)
self.policy = policy_impl.GreedyPolicy()
def test_GreedyPolicy_choosesOptimalAction(self):
mock_qfunc = mock.MagicMock()
mock_qfunc.GetValues.return_value = numpy.array([[0.3, 0.7]])
policy = policy_impl.GreedyPolicy()
self.assertArrayEq(
numpy.array([[0, 1]]),
policy.Decide(env=environment_impl.SingleStateEnvironment(
action_space_size=2, step_limit=10),
brain=mock_qfunc,
state=numpy.array([[0]]),
episode_idx=0,
num_of_episodes=500))
def Demo(_):
env = screen_learning.ScreenGymEnvironment(gym.make('SpaceInvaders-v0'))
brain = a3c_impl.A3C(model=a3c_impl.CreateModel(
state_shape=env.GetStateShape(),
action_space_size=env.GetActionSpaceSize(),
hidden_layer_sizes=(12, ),
))
brain.Load('saved_models/a3c_invader.weights')
policy = policy_impl.GreedyPolicy()
env.StartRecording('a3c_invader.mp4')
runner = runner_impl.SimpleRunner()
runner.Run(env=env, brain=brain, policy=policy, num_of_episodes=10)
env.StopRecording()
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 test_GreedyPolicy(self):
# Tests that it can run; quality if not important for this test.
self.runner.Run(
env=self.env, qfunc=self.qfunc,
policy=policy_impl.GreedyPolicy(), num_of_episodes=1)