def test_rl_track_setting_is_correct():
setting = SettingProxy(
IncrementalRLSetting,
"rl_track",
)
assert setting.nb_tasks == 8
assert setting.dataset == "MetaMonsterKong-v0"
assert setting.observation_space == NamedTupleSpace(
x=Image(0, 1, (3, 64, 64), dtype=np.float32),
task_labels=Sparse(spaces.Discrete(8)),
)
assert setting.action_space == spaces.Discrete(6)
# TODO: The reward range of the MetaMonsterKongEnv is (0, 50), which seems wrong.
# This isn't really a big deal though.
# assert setting.reward_space == spaces.Box(0, 100, shape=(), dtype=np.float32)
assert setting.steps_per_task == 200_000
assert setting.test_steps_per_task == 10_000
assert setting.known_task_boundaries_at_train_time is True
assert setting.known_task_boundaries_at_test_time is False
assert setting.monitor_training_performance is True
assert setting.train_transforms == [
Transforms.to_tensor, Transforms.three_channels
]
assert setting.val_transforms == [
Transforms.to_tensor, Transforms.three_channels
]
assert setting.test_transforms == [
Transforms.to_tensor, Transforms.three_channels
]
train_env = setting.train_dataloader()
assert train_env.observation_space == NamedTupleSpace(
x=Image(0, 1, (3, 64, 64), dtype=np.float32),
task_labels=spaces.Discrete(8),
)
assert train_env.reset() in train_env.observation_space
valid_env = setting.val_dataloader()
assert valid_env.observation_space == NamedTupleSpace(
x=Image(0, 1, (3, 64, 64), dtype=np.float32),
task_labels=spaces.Discrete(8),
)
# IDEA: Prevent submissions from calling the test_dataloader method or accessing the
# test_env / test_dataset property?
with pytest.raises(RuntimeError):
test_env = setting.test_dataloader()
test_env.reset()
with pytest.raises(RuntimeError):
test_env = setting.test_env
test_env.reset()
def test_sl_track_setting_is_correct():
setting = SettingProxy(
ClassIncrementalSetting,
"sl_track",
)
assert setting.nb_tasks == 12
assert setting.dataset == "synbols"
assert setting.observation_space == NamedTupleSpace(
x=Image(0, 1, (3, 32, 32), dtype=np.float32),
task_labels=spaces.Discrete(12),
)
assert setting.n_classes_per_task == 4
assert setting.action_space == spaces.Discrete(48)
assert setting.reward_space == spaces.Discrete(48)
assert setting.known_task_boundaries_at_train_time is True
assert setting.known_task_boundaries_at_test_time is False
assert setting.monitor_training_performance is True
assert setting.train_transforms == [
Transforms.to_tensor, Transforms.three_channels
]
assert setting.val_transforms == [
Transforms.to_tensor, Transforms.three_channels
]
assert setting.test_transforms == [
Transforms.to_tensor, Transforms.three_channels
]
def test_preserves_dtype_of_namedtuple_space():
input_space = NamedTupleSpace(
x=spaces.Box(0, 1, [32, 123, 123, 3]),
task_labels=spaces.MultiDiscrete([5 for _ in range(32)]),
dtype=Foo,
)
output_space = add_tensor_support(input_space)
assert output_space.dtype is input_space.dtype
def test_multitask_setting_test_env():
setting = MultiTaskSetting(dataset="mnist")
assert setting.phases == 1
assert setting.nb_tasks == 5
assert setting.observation_space == NamedTupleSpace(
x=Image(0.0, 1.0, (3, 28, 28), np.float32), task_labels=Discrete(5)
)
assert setting.action_space == Discrete(10)
# FIXME: Wait, actually, this test environment, will it be shuffled, or not?
with setting.test_dataloader(batch_size=32, num_workers=0) as test_env:
check_is_multitask_env(test_env, has_rewards=False)
def test_dqn_monsterkong_adds_channel_first_transform():
method = DQNMethod()
setting = IncrementalRLSetting(
dataset="monsterkong",
nb_tasks=2,
steps_per_task=1_000,
test_steps_per_task=1_000,
)
assert setting.max_steps == 2_000
assert setting.test_steps == 2_000
assert setting.nb_tasks == 2
assert setting.observation_space == NamedTupleSpace(
spaces={
"x": Image(0, 1, shape=(3, 64, 64), dtype=np.float32),
"task_labels": Sparse(spaces.Discrete(2)),
},
dtype=setting.Observations,
)
assert setting.action_space == spaces.Discrete(
6) # monsterkong has 6 actions.
# (Before the method gets to change the Setting):
# By default the setting gives the same shape of obs as the underlying env.
for env_method in [
setting.train_dataloader,
setting.val_dataloader,
setting.test_dataloader,
]:
print(f"Testing method {env_method.__name__}")
with env_method() as env:
reset_obs = env.reset()
# TODO: Fix this so the 'x' space actually gets tensor support.
# assert reset_obs in env.observation_space
assert reset_obs.numpy() in env.observation_space
assert reset_obs.x.shape == (3, 64, 64)
# Let the Method configure itself on the Setting:
method.configure(setting)
# (After the method gets to change the Setting):
for env_method in [
setting.train_dataloader,
setting.val_dataloader,
setting.test_dataloader,
]:
with env_method() as env:
reset_obs = env.reset()
# Fix this numpy bug.
assert reset_obs.numpy() in env.observation_space
assert reset_obs.x.shape == (64, 64, 3)
def test_domain_incremental_mnist_setup():
setting = DomainIncrementalSetting(
dataset="mnist",
increment=2,
)
setting.prepare_data(data_dir="data")
setting.setup()
assert setting.observation_space == NamedTupleSpace(
x=Image(0.0, 1.0, (3, 28, 28), np.float32), task_labels=Discrete(5))
for i in range(setting.nb_tasks):
setting.current_task_id = i
batch_size = 5
train_loader = setting.train_dataloader(batch_size=batch_size)
for j, (observations,
rewards) in enumerate(itertools.islice(train_loader, 100)):
x = observations.x
t = observations.task_labels
y = rewards.y
print(i, j, y, t)
assert x.shape == (batch_size, 3, 28, 28)
assert ((0 <= y) & (y < setting.n_classes_per_task)).all()
assert all(t == i)
x = x.permute(0, 2, 3, 1)[0]
assert x.shape == (28, 28, 3)
reward = train_loader.send([4 for _ in range(batch_size)])
# TODO: Why are we fine with getting `None` as the reward here? Is it
# because we're somehow setting it to be ``
assert reward is None
train_loader.close()
test_loader = setting.test_dataloader(batch_size=batch_size)
for j, (observations,
rewards) in enumerate(itertools.islice(test_loader, 100)):
assert rewards is None
x = observations.x
t = observations.task_labels
assert t is None
assert x.shape == (batch_size, 3, 28, 28)
x = x.permute(0, 2, 3, 1)[0]
assert x.shape == (28, 28, 3)
rewards = test_loader.send([4 for _ in range(batch_size)])
assert rewards is not None
y = rewards.y
assert ((0 <= y) & (y < setting.n_classes_per_task)).all()
def test_multitask_setting():
setting = MultiTaskSetting(dataset="mnist")
assert setting.phases == 1
assert setting.nb_tasks == 5
assert setting.observation_space == NamedTupleSpace(
x=Image(0.0, 1.0, (3, 28, 28), np.float32), task_labels=Discrete(5)
)
assert setting.action_space == Discrete(10)
with setting.train_dataloader(batch_size=32, num_workers=0) as train_env:
check_is_multitask_env(train_env, has_rewards=True)
with setting.val_dataloader(batch_size=32, num_workers=0) as val_env:
check_is_multitask_env(val_env, has_rewards=True)