def test_saveLoad(self):
a3c = a3c_impl.A3C(model=a3c_impl.CreateModel(
state_shape=(3, ),
action_space_size=2,
hidden_layer_sizes=(3, ),
), )
tmp_file = tempfile.NamedTemporaryFile().name
s = numpy.array([[1, 2, 3]])
for _ in range(10):
a3c.UpdateFromTransitions([
base.Transition(s=s,
a=numpy.array([[1, 0]]),
r=1.0,
sp=numpy.array([[4, 5, 6]])),
])
a3c.Save(tmp_file)
saved_values = a3c.GetValues(s)
a3c = a3c_impl.A3C(model=a3c_impl.CreateModel(
state_shape=(3, ),
action_space_size=2,
hidden_layer_sizes=(3, ),
), )
a3c.Load(tmp_file)
numpy_util.TestUtil.AssertArrayEqual(saved_values, a3c.GetValues(s))
def Train(_):
env = environment_impl.GymEnvironment(gym.make('CartPole-v0'))
async_env_runners = [] # type: t.List[async_runner_impl.AsyncEnvRunner]
for _ in range(10):
async_env_runners.append(
async_runner_impl.AsyncEnvRunner(
env=environment_impl.GymEnvironment(gym.make('CartPole-v0')),
runner=runner_impl.NStepExperienceRunner(n_step_return=10),
))
brain = async_runner_impl.AsyncBrain(
a3c_impl.A3C(model=a3c_impl.CreateModel(
state_shape=env.GetStateShape(),
action_space_size=env.GetActionSpaceSize(),
hidden_layer_sizes=(12, ),
)))
policy = policy_impl.PolicyWithDecreasingRandomness(
base_policy=policy_impl.PiWeightedPolicy(),
initial_epsilon=0.2,
final_epsilon=0.05,
decay_by_half_after_num_of_episodes=500,
)
runner = async_runner_impl.ParallelRunner(async_env_runners)
runner.AddCallback(
async_runner_impl.AsyncRunnerExtension(
runner_extension_impl.ProgressTracer(
report_every_num_of_episodes=100)))
runner.AddCallback(
async_runner_impl.AsyncRunnerExtension(
runner_extension_impl.ModelSaver(
save_filepath='saved_models/a3c_cartpole_12.weights',
use_averaged_value_over_num_of_episodes=30)))
runner.Run(brain=brain, policy=policy, num_of_episodes=2000)
def test_convergence(self):
a3c = a3c_impl.A3C(
model=a3c_impl.CreateModel(
state_shape=(3, ),
action_space_size=2,
hidden_layer_sizes=(3, ),
),
# optimizer=a3c_impl.CreateDefaultOptimizer(learning_rate=0.05),
)
s = numpy.array([[1, 2, 3]])
a1 = numpy.array([[1, 0]])
a2 = numpy.array([[0, 1]])
for _ in range(10):
# Needs to train for both actions as one step, otherwise it shows some
# "staggering" effect.
a3c.UpdateFromTransitions([
base.Transition(s=s, a=a1, r=1.0, sp=None),
])
a3c.UpdateFromTransitions([
base.Transition(s=s, a=a2, r=-1.0, sp=s),
])
logging.printf('%s', a3c.GetValues(s))
old_value_a1 = a3c.GetActionValues(a3c.GetValues(s), a1)
# Trains for one step, for both actions.
a3c.UpdateFromTransitions([
base.Transition(s=s, a=a1, r=1.0, sp=None),
])
a3c.UpdateFromTransitions([
base.Transition(s=s, a=a2, r=-1.0, sp=s),
])
self.assertGreaterEqual(a3c.GetActionValues(a3c.GetValues(s), a1),
old_value_a1)
def __init__(self):
self._env = environment_impl.GymEnvironment(
gym_env=gym.make('CartPole-v0'))
self._brain = a3c_impl.A3C(model=a3c_impl.CreateModel(
state_shape=self._env.GetStateShape(),
action_space_size=self._env.GetActionSpaceSize(),
hidden_layer_sizes=(12, ),
))
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 main(_):
running_environment.ForceCpuForTheRun()
env = environment_impl.GymEnvironment(gym.make('CartPole-v0'))
brain = a3c_impl.A3C(model=a3c_impl.CreateModel(
state_shape=env.GetStateShape(),
action_space_size=env.GetActionSpaceSize(),
hidden_layer_sizes=(12, ),
))
policy = policy_impl.PolicyWithDecreasingRandomness(
base_policy=policy_impl.PiWeightedPolicy(),
initial_epsilon=0.4,
final_epsilon=0.05,
decay_by_half_after_num_of_episodes=500,
)
runner = a3c_impl.NStepExperienceRunner()
# runner = runner_impl.SimpleRunner()
runner.AddCallback(
runner_extension_impl.ProgressTracer(report_every_num_of_episodes=100))
runner.Run(env=env, brain=brain, policy=policy, num_of_episodes=1200)