def test_one_step(self):
exp_source = experience.ExperienceSource(self.envs,
DummyAgent(),
steps_count=1)
for exp in exp_source:
self.assertEqual(1, len(exp))
break
def test_two_steps(self):
exp_source = experience.ExperienceSource(self.env,
DummyAgent(),
steps_count=2)
for exp in exp_source:
self.assertEqual(2, len(exp))
break
def test_short_game(self):
env = gym.make('CartPole-v0')
exp_source = experience.ExperienceSource(env, DummyAgent(), steps_count=1)
for step, exp in enumerate(exp_source):
self.assertIsInstance(exp, tuple)
self.assertIsInstance(exp[0], experience.Experience)
if exp[0].done:
break
def test_one_step(self):
exp_source = experience.ExperienceSource(self.env, DummyAgent(), steps_count=1)
for exp in exp_source:
self.assertEqual(1, len(exp))
self.assertIsInstance(exp, tuple)
self.assertIsInstance(exp[0], experience.Experience)
self.assertAlmostEqual(exp[0].reward, -1.0)
self.assertFalse(exp[0].done)
break
def test_short_game(self):
env = gym.make('CartPole-v0')
exp_source = experience.ExperienceSource(env,
dummy_agent,
steps_count=1)
for step, exp in enumerate(exp_source):
self.assertIsInstance(exp, tuple)
self.assertIsInstance(exp[0], experience.Experience)
if len(exp) == 1:
self.assertTrue(exp[0].done)
break
def test_state(self):
actions_count = self.envs[0].action_space.n
my_agent = StatefulAgent(self.envs[0].action_space)
steps = 3
exp_source = experience.ExperienceSource(self.envs, my_agent, steps_count=steps)
for _, exp in zip(range(100), exp_source):
prev_act = None
for e in exp:
if prev_act is not None:
self.assertEqual(e.action, (prev_act+1) % actions_count)
prev_act = e.action
if len(exp) != steps:
self.assertTrue(exp[-1].done)
def setUpClass(cls):
env = gym.make("MountainCar-v0")
cls.source = experience.ExperienceSource(env, agent=DummyAgent())
input_shape=(1 if grayscale else 3, im_height, im_width))
if params.cuda_enabled:
model.cuda()
loss_fn = nn.MSELoss(size_average=False)
optimizer = optim.Adam(model.parameters(),
lr=run.getfloat("learning", "lr"))
action_selector = ActionSelectorEpsilonGreedy(epsilon=run.getfloat(
"defaults", "epsilon"),
params=params)
target_net = agent.TargetNet(model)
dqn_agent = agent.DQNAgent(dqn_model=model,
action_selector=action_selector)
exp_source = experience.ExperienceSource(env=env_pool,
agent=dqn_agent,
steps_count=run.getint(
"defaults", "n_steps"))
exp_replay = experience.ExperienceReplayBuffer(exp_source,
buffer_size=run.getint(
"exp_buffer", "size"))
use_target_dqn = run.getboolean("dqn", "target_dqn", fallback=False)
use_double_dqn = run.getboolean("dqn", "double_dqn", fallback=False)
if use_target_dqn:
target_model = target_net.target_model
else:
target_model = model
def batch_to_train(batch):
"""
