class TestSingleEnvExperiment(unittest.TestCase):
def setUp(self):
np.random.seed(0)
torch.manual_seed(0)
self.env = GymEnvironment('CartPole-v0')
self.env.seed(0)
self.experiment = None
def test_adds_label(self):
experiment = MockExperiment(dqn(), self.env, quiet=True)
self.assertEqual(experiment._writer.label, "_dqn_CartPole-v0")
def test_writes_training_returns_eps(self):
experiment = MockExperiment(dqn(), self.env, quiet=True)
experiment.train(episodes=3)
np.testing.assert_equal(
experiment._writer.data["evaluation/returns/episode"]["values"],
np.array([22.0, 20.0, 24.0]),
)
np.testing.assert_equal(
experiment._writer.data["evaluation/returns/episode"]["steps"],
np.array([1, 2, 3]),
)
def test_writes_test_returns(self):
experiment = MockExperiment(dqn(), self.env, quiet=True)
experiment.train(episodes=5)
returns = experiment.test(episodes=4)
expected_mean = 9.5
expected_std = 0.5
np.testing.assert_equal(np.mean(returns), expected_mean)
np.testing.assert_equal(
experiment._writer.data["evaluation/returns-test/mean"]["values"],
np.array([expected_mean]),
)
np.testing.assert_equal(
experiment._writer.data["evaluation/returns-test/std"]["values"],
np.array([expected_std]),
)
np.testing.assert_equal(
experiment._writer.data["evaluation/returns-test/mean"]["steps"],
np.array([95.]),
)
def test_writes_loss(self):
experiment = MockExperiment(dqn(),
self.env,
quiet=True,
write_loss=True)
self.assertTrue(experiment._writer.write_loss)
experiment = MockExperiment(dqn(),
self.env,
quiet=True,
write_loss=False)
self.assertFalse(experiment._writer.write_loss)
class TestParallelEnvExperiment(unittest.TestCase):
def setUp(self):
np.random.seed(0)
torch.manual_seed(0)
self.env = GymEnvironment('CartPole-v0')
self.env.seed(0)
self.experiment = MockExperiment(self.make_agent(), self.env, quiet=True)
for i, env in enumerate(self.experiment._envs):
env.seed(i)
def test_adds_default_label(self):
self.assertEqual(self.experiment._writer.label, "a2c_CartPole-v0")
def test_adds_custom_label(self):
env = GymEnvironment('CartPole-v0')
experiment = MockExperiment(self.make_agent(), env, name='a2c', quiet=True)
self.assertEqual(experiment._writer.label, "a2c_CartPole-v0")
def test_writes_training_returns_eps(self):
self.experiment.train(episodes=3)
np.testing.assert_equal(
self.experiment._writer.data["evaluation/returns/episode"]["steps"],
np.array([1, 2, 3]),
)
np.testing.assert_equal(
self.experiment._writer.data["evaluation/returns/episode"]["values"],
np.array([10., 11., 17.]),
)
def test_writes_test_returns(self):
self.experiment.train(episodes=5)
returns = self.experiment.test(episodes=4)
self.assertEqual(len(returns), 4)
np.testing.assert_equal(
self.experiment._writer.data["evaluation/returns-test/mean"]["values"],
np.array([np.mean(returns)]),
)
np.testing.assert_equal(
self.experiment._writer.data["evaluation/returns-test/std"]["values"],
np.array([np.std(returns)]),
)
def test_writes_loss(self):
experiment = MockExperiment(self.make_agent(), self.env, quiet=True, write_loss=True)
self.assertTrue(experiment._writer.write_loss)
experiment = MockExperiment(self.make_agent(), self.env, quiet=True, write_loss=False)
self.assertFalse(experiment._writer.write_loss)
def make_agent(self):
return a2c.device('cpu').env(self.env).build()
class TestExperiment(unittest.TestCase):
def setUp(self):
np.random.seed(0)
torch.manual_seed(0)
self.env = GymEnvironment('CartPole-v0')
self.env.seed(0)
self.experiment = None
def test_adds_label(self):
experiment = MockExperiment(dqn(), self.env, quiet=True, episodes=3)
self.assertEqual(experiment._writer.label, "_dqn_CartPole-v0")
def test_writes_returns_eps(self):
experiment = MockExperiment(dqn(), self.env, quiet=True, episodes=3)
np.testing.assert_equal(
experiment._writer.data["evaluation/returns/episode"]["values"],
np.array([14.0, 19.0, 26.0]),
)
np.testing.assert_equal(
experiment._writer.data["evaluation/returns/episode"]["steps"],
np.array([1, 2, 3]),
)
def test_writes_loss(self):
experiment = MockExperiment(dqn(),
self.env,
quiet=True,
write_loss=True,
episodes=3)
self.assertTrue(experiment._writer.write_loss)
experiment = MockExperiment(dqn(),
self.env,
quiet=True,
write_loss=False,
episodes=3)
self.assertFalse(experiment._writer.write_loss)
def test_runs_multi_env(self):
experiment = MockExperiment(a2c(n_envs=3),
self.env,
quiet=True,
episodes=3)
self.assertEqual(
len(experiment._writer.data["evaluation/returns/episode"]
["values"]), 3)