def _prepare_reanalyze_data(self, replay_buffer: ReplayBuffer, env_ids,
positions, n1, n2):
"""
Get the n1 + n2 steps of experience indicated by ``positions`` and return
as the first n1 as ``exp1`` and the next n2 steps as ``exp2``.
"""
batch_size = env_ids.shape[0]
n = n1 + n2
flat_env_ids = env_ids.expand_as(positions).reshape(-1)
flat_positions = positions.reshape(-1)
exp = replay_buffer.get_field(None, flat_env_ids, flat_positions)
if self._data_transformer_ctor is not None:
if self._data_transformer is None:
observation_spec = dist_utils.extract_spec(exp.observation)
self._data_transformer = create_data_transformer(
self._data_transformer_ctor, observation_spec)
# DataTransformer assumes the shape of exp is [B, T, ...]
# It also needs exp.batch_info and exp.replay_buffer.
exp = alf.nest.map_structure(lambda x: x.unsqueeze(1), exp)
exp = exp._replace(batch_info=BatchInfo(flat_env_ids,
flat_positions),
replay_buffer=replay_buffer)
exp = self._data_transformer.transform_experience(exp)
exp = exp._replace(batch_info=(), replay_buffer=())
exp = alf.nest.map_structure(lambda x: x.squeeze(1), exp)
def _split1(x):
shape = x.shape[1:]
x = x.reshape(batch_size, n, *shape)
return x[:, :n1, ...].reshape(batch_size * n1, *shape)
def _split2(x):
shape = x.shape[1:]
x = x.reshape(batch_size, n, *shape)
return x[:, n1:, ...].reshape(batch_size * n2, *shape)
with alf.device(self._device):
exp = convert_device(exp)
exp1 = alf.nest.map_structure(_split1, exp)
exp2 = alf.nest.map_structure(_split2, exp)
return exp1, exp2
def test_frame_stacker(self, stack_axis=0):
data_spec = DataItem(step_type=alf.TensorSpec((), dtype=torch.int32),
observation=dict(scalar=alf.TensorSpec(()),
vector=alf.TensorSpec((7, )),
matrix=alf.TensorSpec((5, 6)),
tensor=alf.TensorSpec(
(2, 3, 4))))
replay_buffer = ReplayBuffer(data_spec=data_spec,
num_environments=2,
max_length=1024,
num_earliest_frames_ignored=2)
frame_stacker = FrameStacker(
data_spec.observation,
stack_size=3,
stack_axis=stack_axis,
fields=['scalar', 'vector', 'matrix', 'tensor'])
new_spec = frame_stacker.transformed_observation_spec
self.assertEqual(new_spec['scalar'].shape, (3, ))
self.assertEqual(new_spec['vector'].shape, (21, ))
if stack_axis == -1:
self.assertEqual(new_spec['matrix'].shape, (5, 18))
self.assertEqual(new_spec['tensor'].shape, (2, 3, 12))
elif stack_axis == 0:
self.assertEqual(new_spec['matrix'].shape, (15, 6))
self.assertEqual(new_spec['tensor'].shape, (6, 3, 4))
def _step_type(t, period):
if t % period == 0:
return StepType.FIRST
if t % period == period - 1:
return StepType.LAST
return StepType.MID
observation = alf.nest.map_structure(
lambda spec: spec.randn((1000, 2)), data_spec.observation)
state = common.zero_tensor_from_nested_spec(frame_stacker.state_spec,
2)
def _get_stacked_data(t, b):
if stack_axis == -1:
return dict(scalar=observation['scalar'][t, b],
vector=observation['vector'][t, b].reshape(-1),
matrix=observation['matrix'][t, b].transpose(
0, 1).reshape(5, 18),
tensor=observation['tensor'][t, b].permute(
1, 2, 0, 3).reshape(2, 3, 12))
elif stack_axis == 0:
return dict(scalar=observation['scalar'][t, b],
vector=observation['vector'][t, b].reshape(-1),
matrix=observation['matrix'][t, b].reshape(15, 6),
tensor=observation['tensor'][t,
b].reshape(6, 3, 4))
def _check_equal(stacked, expected, b):
self.assertEqual(stacked['scalar'][b], expected['scalar'])
self.assertEqual(stacked['vector'][b], expected['vector'])
self.assertEqual(stacked['matrix'][b], expected['matrix'])
self.assertEqual(stacked['tensor'][b], expected['tensor'])
for t in range(1000):
batch = DataItem(
step_type=torch.tensor([_step_type(t, 17),
_step_type(t, 22)]),
observation=alf.nest.map_structure(lambda x: x[t],
observation))
replay_buffer.add_batch(batch)
timestep, state = frame_stacker.transform_timestep(batch, state)
if t == 0:
for b in (0, 1):
expected = _get_stacked_data([0, 0, 0], b)
_check_equal(timestep.observation, expected, b)
if t == 1:
for b in (0, 1):
expected = _get_stacked_data([0, 0, 1], b)
_check_equal(timestep.observation, expected, b)
if t == 2:
for b in (0, 1):
expected = _get_stacked_data([0, 1, 2], b)
_check_equal(timestep.observation, expected, b)
if t == 16:
for b in (0, 1):
expected = _get_stacked_data([14, 15, 16], b)
_check_equal(timestep.observation, expected, b)
if t == 17:
for b, t in ((0, [17, 17, 17]), (1, [15, 16, 17])):
expected = _get_stacked_data(t, b)
_check_equal(timestep.observation, expected, b)
if t == 18:
for b, t in ((0, [17, 17, 18]), (1, [16, 17, 18])):
expected = _get_stacked_data(t, b)
_check_equal(timestep.observation, expected, b)
if t == 22:
for b, t in ((0, [20, 21, 22]), (1, [22, 22, 22])):
expected = _get_stacked_data(t, b)
_check_equal(timestep.observation, expected, b)
batch_info = BatchInfo(env_ids=torch.tensor([0, 1, 0, 1],
dtype=torch.int64),
positions=torch.tensor([0, 1, 18, 22],
dtype=torch.int64))
# [4, 2, ...]
experience = replay_buffer.get_field(
'', batch_info.env_ids.unsqueeze(-1),
batch_info.positions.unsqueeze(-1) + torch.arange(2))
experience = experience._replace(batch_info=batch_info,
replay_buffer=replay_buffer)
experience = frame_stacker.transform_experience(experience)
expected = _get_stacked_data([0, 0, 0], 0)
_check_equal(experience.observation, expected, (0, 0))
expected = _get_stacked_data([0, 0, 1], 0)
_check_equal(experience.observation, expected, (0, 1))
expected = _get_stacked_data([0, 0, 1], 1)
_check_equal(experience.observation, expected, (1, 0))
expected = _get_stacked_data([0, 1, 2], 1)
_check_equal(experience.observation, expected, (1, 1))
expected = _get_stacked_data([17, 17, 18], 0)
_check_equal(experience.observation, expected, (2, 0))
expected = _get_stacked_data([17, 18, 19], 0)
_check_equal(experience.observation, expected, (2, 1))
expected = _get_stacked_data([22, 22, 22], 1)
_check_equal(experience.observation, expected, (3, 0))
expected = _get_stacked_data([22, 22, 23], 1)
_check_equal(experience.observation, expected, (3, 1))
def _test_preprocess_experience(self, train_reward_function, td_steps,
reanalyze_ratio, expected):
"""
The following summarizes how the data is generated:
.. code-block:: python
# position: 01234567890123
step_type0 = 'FMMMLFMMLFMMMM'
step_type1 = 'FMMMMMLFMMMMLF'
scale = 1. for current model
2. for target model
observation = [position] * 3
reward = position if train_reward_function and td_steps!=-1
else position * (step_type == LAST)
value = 0.5 * position * scale
action_probs = scale * [position, position+1, position] for env 0
scale * [position+1, position, position] for env 1
action = 1 for env 0
0 for env 1
"""
reanalyze_td_steps = 2
num_unroll_steps = 4
batch_size = 2
obs_dim = 3
observation_spec = alf.TensorSpec([obs_dim])
action_spec = alf.BoundedTensorSpec((),
minimum=0,
maximum=1,
dtype=torch.int32)
reward_spec = alf.TensorSpec(())
time_step_spec = ds.time_step_spec(observation_spec, action_spec,
reward_spec)
global _mcts_model_id
_mcts_model_id = 0
muzero = MuzeroAlgorithm(observation_spec,
action_spec,
model_ctor=_create_mcts_model,
mcts_algorithm_ctor=MockMCTSAlgorithm,
num_unroll_steps=num_unroll_steps,
td_steps=td_steps,
train_game_over_function=True,
train_reward_function=train_reward_function,
reanalyze_ratio=reanalyze_ratio,
reanalyze_td_steps=reanalyze_td_steps,
data_transformer_ctor=partial(FrameStacker,
stack_size=2))
data_transformer = FrameStacker(observation_spec, stack_size=2)
time_step = common.zero_tensor_from_nested_spec(
time_step_spec, batch_size)
dt_state = common.zero_tensor_from_nested_spec(
data_transformer.state_spec, batch_size)
state = muzero.get_initial_predict_state(batch_size)
transformed_time_step, dt_state = data_transformer.transform_timestep(
time_step, dt_state)
alg_step = muzero.rollout_step(transformed_time_step, state)
alg_step_spec = dist_utils.extract_spec(alg_step)
experience = ds.make_experience(time_step, alg_step, state)
experience_spec = ds.make_experience(time_step_spec, alg_step_spec,
muzero.train_state_spec)
replay_buffer = ReplayBuffer(data_spec=experience_spec,
num_environments=batch_size,
max_length=16,
keep_episodic_info=True)
# 01234567890123
step_type0 = 'FMMMLFMMLFMMMM'
step_type1 = 'FMMMMMLFMMMMLF'
dt_state = common.zero_tensor_from_nested_spec(
data_transformer.state_spec, batch_size)
for i in range(len(step_type0)):
step_type = [step_type0[i], step_type1[i]]
step_type = [
ds.StepType.MID if c == 'M' else
(ds.StepType.FIRST if c == 'F' else ds.StepType.LAST)
for c in step_type
]
step_type = torch.tensor(step_type, dtype=torch.int32)
reward = reward = torch.full([batch_size], float(i))
if not train_reward_function or td_steps == -1:
reward = reward * (step_type == ds.StepType.LAST).to(
torch.float32)
time_step = time_step._replace(
discount=(step_type != ds.StepType.LAST).to(torch.float32),
step_type=step_type,
observation=torch.tensor([[i, i + 1, i], [i + 1, i, i]],
dtype=torch.float32),
reward=reward,
env_id=torch.arange(batch_size, dtype=torch.int32))
transformed_time_step, dt_state = data_transformer.transform_timestep(
time_step, dt_state)
alg_step = muzero.rollout_step(transformed_time_step, state)
experience = ds.make_experience(time_step, alg_step, state)
replay_buffer.add_batch(experience)
state = alg_step.state
env_ids = torch.tensor([0] * 14 + [1] * 14, dtype=torch.int64)
positions = torch.tensor(list(range(14)) + list(range(14)),
dtype=torch.int64)
experience = replay_buffer.get_field(None,
env_ids.unsqueeze(-1).cpu(),
positions.unsqueeze(-1).cpu())
experience = experience._replace(replay_buffer=replay_buffer,
batch_info=BatchInfo(
env_ids=env_ids,
positions=positions),
rollout_info_field='rollout_info')
processed_experience = muzero.preprocess_experience(experience)
import pprint
pprint.pprint(processed_experience.rollout_info)
alf.nest.map_structure(lambda x, y: self.assertEqual(x, y),
processed_experience.rollout_info, expected)
def test_preprocess_experience(self):
"""
The following summarizes how the data is generated:
.. code-block:: python
# position: 01234567890123
step_type0 = 'FMMMLFMMLFMMMM'
step_type1 = 'FMMMMMLFMMMMLF'
reward = position if train_reward_function and td_steps!=-1
else position * (step_type == LAST)
action = t + 1 for env 0
t for env 1
"""
num_unroll_steps = 4
batch_size = 2
obs_dim = 3
observation_spec = alf.TensorSpec([obs_dim])
action_spec = alf.BoundedTensorSpec((1, ),
minimum=0,
maximum=1,
dtype=torch.float32)
reward_spec = alf.TensorSpec(())
time_step_spec = ds.time_step_spec(observation_spec, action_spec,
reward_spec)
repr_learner = PredictiveRepresentationLearner(
observation_spec,
action_spec,
num_unroll_steps=num_unroll_steps,
decoder_ctor=partial(SimpleDecoder,
target_field='reward',
decoder_net_ctor=partial(
EncodingNetwork, fc_layer_params=(4, ))),
encoding_net_ctor=LSTMEncodingNetwork,
dynamics_net_ctor=LSTMEncodingNetwork)
time_step = common.zero_tensor_from_nested_spec(
time_step_spec, batch_size)
state = repr_learner.get_initial_predict_state(batch_size)
alg_step = repr_learner.rollout_step(time_step, state)
alg_step = alg_step._replace(output=torch.tensor([[1.], [0.]]))
alg_step_spec = dist_utils.extract_spec(alg_step)
experience = ds.make_experience(time_step, alg_step, state)
experience_spec = ds.make_experience(time_step_spec, alg_step_spec,
repr_learner.train_state_spec)
replay_buffer = ReplayBuffer(data_spec=experience_spec,
num_environments=batch_size,
max_length=16,
keep_episodic_info=True)
# 01234567890123
step_type0 = 'FMMMLFMMLFMMMM'
step_type1 = 'FMMMMMLFMMMMLF'
for i in range(len(step_type0)):
step_type = [step_type0[i], step_type1[i]]
step_type = [
ds.StepType.MID if c == 'M' else
(ds.StepType.FIRST if c == 'F' else ds.StepType.LAST)
for c in step_type
]
step_type = torch.tensor(step_type, dtype=torch.int32)
reward = reward = torch.full([batch_size], float(i))
time_step = time_step._replace(
discount=(step_type != ds.StepType.LAST).to(torch.float32),
step_type=step_type,
observation=torch.tensor([[i, i + 1, i], [i + 1, i, i]],
dtype=torch.float32),
reward=reward,
env_id=torch.arange(batch_size, dtype=torch.int32))
alg_step = repr_learner.rollout_step(time_step, state)
alg_step = alg_step._replace(output=i + torch.tensor([[1.], [0.]]))
experience = ds.make_experience(time_step, alg_step, state)
replay_buffer.add_batch(experience)
state = alg_step.state
env_ids = torch.tensor([0] * 14 + [1] * 14, dtype=torch.int64)
positions = torch.tensor(list(range(14)) + list(range(14)),
dtype=torch.int64)
experience = replay_buffer.get_field(None,
env_ids.unsqueeze(-1).cpu(),
positions.unsqueeze(-1).cpu())
experience = experience._replace(replay_buffer=replay_buffer,
batch_info=BatchInfo(
env_ids=env_ids,
positions=positions),
rollout_info_field='rollout_info')
processed_experience = repr_learner.preprocess_experience(experience)
pprint.pprint(processed_experience.rollout_info)
# yapf: disable
expected = PredictiveRepresentationLearnerInfo(
action=torch.tensor(
[[[ 1., 2., 3., 4., 5.]],
[[ 2., 3., 4., 5., 5.]],
[[ 3., 4., 5., 5., 5.]],
[[ 4., 5., 5., 5., 5.]],
[[ 5., 5., 5., 5., 5.]],
[[ 6., 7., 8., 9., 9.]],
[[ 7., 8., 9., 9., 9.]],
[[ 8., 9., 9., 9., 9.]],
[[ 9., 9., 9., 9., 9.]],
[[10., 11., 12., 13., 14.]],
[[11., 12., 13., 14., 14.]],
[[12., 13., 14., 14., 14.]],
[[13., 14., 14., 14., 14.]],
[[14., 14., 14., 14., 14.]],
[[ 0., 1., 2., 3., 4.]],
[[ 1., 2., 3., 4., 5.]],
[[ 2., 3., 4., 5., 6.]],
[[ 3., 4., 5., 6., 6.]],
[[ 4., 5., 6., 6., 6.]],
[[ 5., 6., 6., 6., 6.]],
[[ 6., 6., 6., 6., 6.]],
[[ 7., 8., 9., 10., 11.]],
[[ 8., 9., 10., 11., 12.]],
[[ 9., 10., 11., 12., 12.]],
[[10., 11., 12., 12., 12.]],
[[11., 12., 12., 12., 12.]],
[[12., 12., 12., 12., 12.]],
[[13., 13., 13., 13., 13.]]]).unsqueeze(-1),
mask=torch.tensor(
[[[ True, True, True, True, True]],
[[ True, True, True, True, False]],
[[ True, True, True, False, False]],
[[ True, True, False, False, False]],
[[ True, False, False, False, False]],
[[ True, True, True, True, False]],
[[ True, True, True, False, False]],
[[ True, True, False, False, False]],
[[ True, False, False, False, False]],
[[ True, True, True, True, True]],
[[ True, True, True, True, False]],
[[ True, True, True, False, False]],
[[ True, True, False, False, False]],
[[ True, False, False, False, False]],
[[ True, True, True, True, True]],
[[ True, True, True, True, True]],
[[ True, True, True, True, True]],
[[ True, True, True, True, False]],
[[ True, True, True, False, False]],
[[ True, True, False, False, False]],
[[ True, False, False, False, False]],
[[ True, True, True, True, True]],
[[ True, True, True, True, True]],
[[ True, True, True, True, False]],
[[ True, True, True, False, False]],
[[ True, True, False, False, False]],
[[ True, False, False, False, False]],
[[ True, False, False, False, False]]]),
target=torch.tensor(
[[[ 0., 1., 2., 3., 4.]],
[[ 1., 2., 3., 4., 4.]],
[[ 2., 3., 4., 4., 4.]],
[[ 3., 4., 4., 4., 4.]],
[[ 4., 4., 4., 4., 4.]],
[[ 5., 6., 7., 8., 8.]],
[[ 6., 7., 8., 8., 8.]],
[[ 7., 8., 8., 8., 8.]],
[[ 8., 8., 8., 8., 8.]],
[[ 9., 10., 11., 12., 13.]],
[[10., 11., 12., 13., 13.]],
[[11., 12., 13., 13., 13.]],
[[12., 13., 13., 13., 13.]],
[[13., 13., 13., 13., 13.]],
[[ 0., 1., 2., 3., 4.]],
[[ 1., 2., 3., 4., 5.]],
[[ 2., 3., 4., 5., 6.]],
[[ 3., 4., 5., 6., 6.]],
[[ 4., 5., 6., 6., 6.]],
[[ 5., 6., 6., 6., 6.]],
[[ 6., 6., 6., 6., 6.]],
[[ 7., 8., 9., 10., 11.]],
[[ 8., 9., 10., 11., 12.]],
[[ 9., 10., 11., 12., 12.]],
[[10., 11., 12., 12., 12.]],
[[11., 12., 12., 12., 12.]],
[[12., 12., 12., 12., 12.]],
[[13., 13., 13., 13., 13.]]]))
# yapf: enable
alf.nest.map_structure(lambda x, y: self.assertEqual(x, y),
processed_experience.rollout_info, expected)
def _reanalyze1(self, replay_buffer: ReplayBuffer, env_ids, positions,
mcts_state_field):
"""Reanalyze one batch.
This means:
1. Re-plan the policy using MCTS for n1 = 1 + num_unroll_steps to get fresh policy
and value target.
2. Caluclate the value for following n2 = reanalyze_td_steps so that we have value
for a total of 1 + num_unroll_steps + reanalyze_td_steps.
3. Use these values and rewards from replay buffer to caculate n2-step
bootstraped value target for the first n1 steps.
In order to do 1 and 2, we need to get the observations for n1 + n2 steps
and processs them using data_transformer.
"""
batch_size = env_ids.shape[0]
n1 = self._num_unroll_steps + 1
n2 = self._reanalyze_td_steps
env_ids, positions = self._next_n_positions(
replay_buffer, env_ids, positions, self._num_unroll_steps + n2)
# [B, n1]
positions1 = positions[:, :n1]
game_overs = replay_buffer.get_field('discount', env_ids,
positions1) == 0.
steps_to_episode_end = replay_buffer.steps_to_episode_end(
positions1, env_ids)
bootstrap_n = steps_to_episode_end.clamp(max=n2)
exp1, exp2 = self._prepare_reanalyze_data(replay_buffer, env_ids,
positions, n1, n2)
bootstrap_position = positions1 + bootstrap_n
discount = replay_buffer.get_field('discount', env_ids,
bootstrap_position)
sum_reward = self._sum_discounted_reward(replay_buffer, env_ids,
positions1,
bootstrap_position, n2)
if not self._train_reward_function:
rewards = self._get_reward(replay_buffer, env_ids,
bootstrap_position)
with alf.device(self._device):
bootstrap_n = convert_device(bootstrap_n)
discount = convert_device(discount)
sum_reward = convert_device(sum_reward)
game_overs = convert_device(game_overs)
# 1. Reanalyze the first n1 steps to get both the updated value and policy
self._mcts.set_model(self._target_model)
mcts_step = self._mcts.predict_step(
exp1, alf.nest.get_field(exp1, mcts_state_field))
self._mcts.set_model(self._model)
candidate_actions = ()
if not _is_empty(mcts_step.info.candidate_actions):
candidate_actions = mcts_step.info.candidate_actions
candidate_actions = candidate_actions.reshape(
batch_size, n1, *candidate_actions.shape[1:])
candidate_action_policy = mcts_step.info.candidate_action_policy
candidate_action_policy = candidate_action_policy.reshape(
batch_size, n1, *candidate_action_policy.shape[1:])
values1 = mcts_step.info.value.reshape(batch_size, n1)
# 2. Calulate the value of the next n2 steps so that n2-step return
# can be computed.
model_output = self._target_model.initial_inference(
exp2.observation)
values2 = model_output.value.reshape(batch_size, n2)
# 3. Calculate n2-step return
values = torch.cat([values1, values2], dim=1)
# [B, n1]
bootstrap_pos = torch.arange(n1).unsqueeze(0) + bootstrap_n
values = values[torch.arange(batch_size).unsqueeze(-1),
bootstrap_pos]
values = values * discount * (self._discount**bootstrap_n.to(
torch.float32))
values = values + sum_reward
if not self._train_reward_function:
# For this condition, we need to set the value at and after the
# last step to be the last reward.
values = torch.where(game_overs, convert_device(rewards),
values)
return candidate_actions, candidate_action_policy, values