def setUp(self):
defaultSettings = SettingsManager.getDefaultSettings()
defaultSettings.setProperty('noiseStd', 1.0)
defaultSettings.setProperty('initialStateDistributionMinRange',
np.asarray([np.pi - np.pi, -2]))
defaultSettings.setProperty('initialStateDistributionMaxRange',
np.asarray([np.pi + np.pi, 2]))
defaultSettings.setProperty('initialStateDistributionType', 'Uniform')
defaultSettings.setProperty('dt', 0.025)
defaultSettings.setProperty('initSigmaActions', 1.0)
defaultSettings.setProperty('initialStateDistributionMinRange',
np.asarray([np.pi - np.pi, -2]))
defaultSettings.setProperty('initialStateDistributionMaxRange',
np.asarray([np.pi + np.pi, 2]))
self.sampler = EpisodeWithStepsSampler()
self.episodeManager = self.sampler.getEpisodeDataManager()
self.stepManager = self.episodeManager.subDataManager
self.pendulum = Pendulum(self.episodeManager)
self.sampler.stepSampler.setIsActiveSampler(
StepBasedEpisodeTerminationSampler(self.episodeManager,
'bla',
numTimeSteps=40))
initialStateSampler = InitialStateSamplerStandard(self.episodeManager)
dummyActionAndReward = DummyActionAndReward(
self.episodeManager.subDataManager, 1)
self.sampler.setTransitionFunction(self.pendulum.getExpectedNextState)
self.sampler.setInitStateSampler(initialStateSampler.sampleInitState)
self.sampler.setActionPolicy(dummyActionAndReward.sampleAction)
self.sampler.setRewardFunction(dummyActionAndReward.sampleReward)
self.sampler.setReturnFunction(dummyActionAndReward.sampleReward)
class testStepSampler(unittest.TestCase):
def setUp(self):
settings = SettingsManager.getDefaultSettings()
settings.setProperty('numTimeSteps', 40)
self.dataManager = DataUtil.createTestManagerSteps()
self.stepSamplerEpisodes = EpisodeWithStepsSampler(
self.dataManager, 'episodes', 'steps')
#Todo: get function for data manager
def tearDown(self):
pass
def testSampling(self):
environment = TestEnvironment(self.dataManager)
policy = TestPolicy(self.dataManager)
reward = TestReward(self.dataManager)
self.stepSamplerEpisodes.setInitStateSampler(environment.initState)
self.stepSamplerEpisodes.setTransitionFunction(environment)
self.stepSamplerEpisodes.setActionPolicy(policy)
self.stepSamplerEpisodes.setRewardFunction(reward)
data = self.dataManager.createDataObject([10, 100])
data[Ellipsis] >> self.stepSamplerEpisodes
states = data.getDataEntry('states', 1)
actions = data.getDataEntry('actions', 2)
rewards = data.getDataEntry('rewards', 3)
statesTarget = np.array(range(1, 41))
self.assertTrue(
(abs(states.transpose() - statesTarget) < 0.00001).all())
self.assertTrue((abs(states * 2 - actions) < 0.00001).all())
self.assertTrue((abs(states * 2 - rewards) < 0.00001).all())
class Test(unittest.TestCase):
def setUp(self):
self.sampler = EpisodeWithStepsSampler()
self.dataManager = self.sampler.getEpisodeDataManager()
self.tf = TransitionFunction(self.dataManager, 2, 1)
# Test that sampler and data managers are set up correctly
def test_init(self):
stepManager = self.sampler.getEpisodeDataManager().subDataManager
self.assertIn('states', stepManager.dataEntries)
self.assertIn('actions', stepManager.dataEntries)
# self.assertIn('nextStates', stepManager.dataEntries)
stepSampler = self.sampler.stepSampler
def setUp(self):
self.sampler = EpisodeWithStepsSampler()
self.episodeManager = self.sampler.getEpisodeDataManager()
double_link = DoubleLink(self.episodeManager)
self.sampler.stepSampler.setIsActiveSampler(
StepBasedEpisodeTerminationSampler(self.episodeManager, 'steps',
40))
initialStateSampler = InitialStateSamplerStandard(self.episodeManager)
dummyActionAndReward = DummyActionAndReward(
self.episodeManager.subDataManager, 2, True)
self.sampler.setTransitionFunction(double_link.getExpectedNextState)
self.sampler.setInitStateSampler(initialStateSampler.sampleInitState)
self.sampler.setActionPolicy(dummyActionAndReward.sampleAction)
self.sampler.setRewardFunction(dummyActionAndReward.sampleReward)
self.sampler.setReturnFunction(dummyActionAndReward.sampleReward)
def setUp(self):
defaultSettings = SettingsManager.getDefaultSettings()
defaultSettings.setProperty('numTimeSteps', 20)
self.sampler = EpisodeWithStepsSampler(samplerNameSteps='steps')
self.sampler.numSamples = 3
self.dataManager = self.sampler.getEpisodeDataManager()
self.sampler.stepSampler.setTerminationFunction(
NumStepsTerminationFunction(self.dataManager,
None,
numTimeSteps=20))
environment = MountainCar(self.dataManager)
self.sampler.setContextSampler(environment.sampleContext)
self.sampler.setActionPolicy(environment.sampleAction)
self.sampler.setTransitionFunction(environment.transitionFunction)
self.sampler.setRewardFunction(environment.sampleReward)
self.sampler.setInitStateSampler(environment.sampleInitState)
@Mapping.MappingMethod()
def getAction(self, states):
return states * 2 * np.ones(2)
class TestReward(Mapping):
def __init__(self, dataManager):
Mapping.__init__(self, dataManager, ['states', 'actions'], ['rewards'])
@Mapping.MappingMethod()
def getReward(self, states, actions):
return states[:, 0:1] * 2
sampler = EpisodeWithStepsSampler()
dataManager = sampler.getEpisodeDataManager()
environment = TestEnvironment(dataManager)
policy = TestPolicy(dataManager)
reward = TestReward(dataManager)
sampler.setInitStateSampler(environment.initState)
sampler.setTransitionFunction(environment)
sampler.setActionPolicy(policy)
sampler.setRewardFunction(reward)
data = dataManager.createDataObject([1, 20])
data[...] >> sampler >> data
class Test(unittest.TestCase):
# Test that sampler and data managers are set up correctly
def setUp(self):
defaultSettings = SettingsManager.getDefaultSettings()
defaultSettings.setProperty('numTimeSteps', 20)
self.sampler = EpisodeWithStepsSampler(samplerNameSteps='steps')
self.sampler.numSamples = 3
self.dataManager = self.sampler.getEpisodeDataManager()
self.sampler.stepSampler.setTerminationFunction(
NumStepsTerminationFunction(self.dataManager,
None,
numTimeSteps=20))
environment = MountainCar(self.dataManager)
self.sampler.setContextSampler(environment.sampleContext)
self.sampler.setActionPolicy(environment.sampleAction)
self.sampler.setTransitionFunction(environment.transitionFunction)
self.sampler.setRewardFunction(environment.sampleReward)
self.sampler.setInitStateSampler(environment.sampleInitState)
def test_sampleData(self):
self.sampler.setParallelSampling(True)
newData = self.dataManager.createDataObject([3, 20])
self.sampler >> newData[...]
print('States:', newData[1, :].states)
print('NextStates:', newData[1, :].nextStates)
def setUp(self):
settings = SettingsManager.getDefaultSettings()
settings.setProperty('numTimeSteps', 40)
self.dataManager = DataUtil.createTestManagerSteps()
self.stepSamplerEpisodes = EpisodeWithStepsSampler(
self.dataManager, 'episodes', 'steps')
def setUp(self):
self.sampler = EpisodeWithStepsSampler()
self.dataManager = self.sampler.getEpisodeDataManager()
self.tf = TransitionFunction(self.dataManager, 2, 1)
class Test(unittest.TestCase):
def setUp(self):
defaultSettings = SettingsManager.getDefaultSettings()
defaultSettings.setProperty('noiseStd', 1.0)
defaultSettings.setProperty('initialStateDistributionMinRange',
np.asarray([np.pi - np.pi, -2]))
defaultSettings.setProperty('initialStateDistributionMaxRange',
np.asarray([np.pi + np.pi, 2]))
defaultSettings.setProperty('initialStateDistributionType', 'Uniform')
defaultSettings.setProperty('dt', 0.025)
defaultSettings.setProperty('initSigmaActions', 1.0)
defaultSettings.setProperty('initialStateDistributionMinRange',
np.asarray([np.pi - np.pi, -2]))
defaultSettings.setProperty('initialStateDistributionMaxRange',
np.asarray([np.pi + np.pi, 2]))
self.sampler = EpisodeWithStepsSampler()
self.episodeManager = self.sampler.getEpisodeDataManager()
self.stepManager = self.episodeManager.subDataManager
self.pendulum = Pendulum(self.episodeManager)
self.sampler.stepSampler.setIsActiveSampler(
StepBasedEpisodeTerminationSampler(self.episodeManager,
'bla',
numTimeSteps=40))
initialStateSampler = InitialStateSamplerStandard(self.episodeManager)
dummyActionAndReward = DummyActionAndReward(
self.episodeManager.subDataManager, 1)
self.sampler.setTransitionFunction(self.pendulum.getExpectedNextState)
self.sampler.setInitStateSampler(initialStateSampler.sampleInitState)
self.sampler.setActionPolicy(dummyActionAndReward.sampleAction)
self.sampler.setRewardFunction(dummyActionAndReward.sampleReward)
self.sampler.setReturnFunction(dummyActionAndReward.sampleReward)
def testGenerating(self):
data = self.episodeManager.getDataObject(10)
self.sampler.numSamples = 100
self.sampler.setParallelSampling(True)
data >> self.sampler
self.assertEqual(data[:, 1].states.shape, (100, 2))
self.assertEqual(data[1, :].states.shape, (40, 2))
class Test(unittest.TestCase):
def setUp(self):
self.sampler = EpisodeWithStepsSampler()
self.episodeManager = self.sampler.getEpisodeDataManager()
double_link = DoubleLink(self.episodeManager)
self.sampler.stepSampler.setIsActiveSampler(
StepBasedEpisodeTerminationSampler(self.episodeManager, 'steps',
40))
initialStateSampler = InitialStateSamplerStandard(self.episodeManager)
dummyActionAndReward = DummyActionAndReward(
self.episodeManager.subDataManager, 2, True)
self.sampler.setTransitionFunction(double_link.getExpectedNextState)
self.sampler.setInitStateSampler(initialStateSampler.sampleInitState)
self.sampler.setActionPolicy(dummyActionAndReward.sampleAction)
self.sampler.setRewardFunction(dummyActionAndReward.sampleReward)
self.sampler.setReturnFunction(dummyActionAndReward.sampleReward)
def testGenerating(self):
data = self.episodeManager.getDataObject(10)
self.sampler.numSamples = 100
self.sampler.setParallelSampling(True)
data[...] >> self.sampler
self.assertEqual(data[:, 1].states.shape, (100, 4))
self.assertEqual(data[1, :].states.shape, (40, 4))
defaultSettings.setProperty('dt', 0.025)
defaultSettings.setProperty('initSigmaActions', 1e-10)
# ... to sample the initial states ...
defaultSettings.setProperty('initialStateDistributionType', 'Uniform')
defaultSettings.setProperty('numTimeSteps', steps_per_epoch)
# ... and for the image creating preprocessor. Note that the same settings object is used here which is then
# distributed by the setting manager to all its clients. (The Pendulum, the InitalStateSampler as well as the
# Preprocessor going to be setting clients
img_size = 48 # pixels, height and width. (currently only squared images possible)
defaultSettings.setProperty('imgSize', img_size)
defaultSettings.setProperty('lineWidth', 3) # again pixel
# Next we can create a "Sampler". It will later sample the values from the desired environment
sampler = EpisodeWithStepsSampler()
# We also need the data managers for the episode and step layer - we get them from the sampler
episodeManager = sampler.getEpisodeDataManager()
# Finally, we can create our environment. Try exchanging the DoubleLink with the QuadLink or the Pendulum...
n_link_pendulum = Pendulum(episodeManager)
# ... make sure you change the number of joints too
number_of_joints = 1
# We can still add further settings, e.g. for the initialStateSampler created next
defaultSettings.setProperty(
'initialStateDistributionMinRange',
np.tile(np.asarray([np.pi - np.pi, -2]), [number_of_joints]))
defaultSettings.setProperty(
'initialStateDistributionMaxRange',