def train(self):
file = open('log_reward', 'w')
""" train agent"""
mainWorker = Worker.remote(self.env, self.hp)
worker = ActorPool([
Worker.remote(gym.make(self.hp.env_name), self.hp)
for i in range(8)
])
print(worker)
for step in range(self.hp.nb_steps):
#generate random pertrutbation
deltas = self.policy.sample_deltas()
positive_rewards = worker.map(
lambda a, v: a.explore.remote(*v),
[(self.normalizer, self.policy, 'positive', deltas[i])
for i in range(self.hp.nb_directions)])
positive_rewards = [i for i in positive_rewards]
negative_rewards = worker.map(
lambda a, v: a.explore.remote(*v),
[(self.normalizer, self.policy, 'negative', deltas[i])
for i in range(self.hp.nb_directions)])
negative_rewards = [i for i in negative_rewards]
# gathering the rewards
all_rewards = np.array(positive_rewards + negative_rewards)
#get the standard deviation of all rewards
sigma_r = all_rewards.std()
#soring the rewards to generate rollouts for updating weight
scores = {
k: max(r_pos, r_neg)
for k, (r_pos, r_neg
) in enumerate(zip(positive_rewards, negative_rewards))
}
order = sorted(scores.keys(),
key=lambda x: scores[x],
reverse=True)[:self.hp.nb_best_directions]
rollouts = [(positive_rewards[k], negative_rewards[k], deltas[k])
for k in order]
#update the policy with new weight
self.policy.update(rollouts, sigma_r)
#print result
reward_evaluation = ray.get(
mainWorker.explore.remote(self.normalizer, self.policy))
print("Step ", step, "=> Reward: ", reward_evaluation)
file.write(str(reward_evaluation) + '\n')
return self.policy.tetha
def test_map_gh23107(init):
sleep_time = 40
# Reference - https://github.com/ray-project/ray/issues/23107
@ray.remote
class DummyActor:
async def identity(self, s):
if s == 6:
await asyncio.sleep(sleep_time)
return s, time.time()
def func(a, v):
return a.identity.remote(v)
map_values = [1, 2, 3, 4, 5]
pool_map = ActorPool([DummyActor.remote() for i in range(2)])
pool_map.submit(func, 6)
start_time = time.time()
gen = pool_map.map(func, map_values)
assert all(elem[0] in [1, 2, 3, 4, 5] for elem in list(gen))
assert all(
abs(elem[1] - start_time) < sleep_time in [1, 2, 3, 4, 5]
for elem in list(gen))
pool_map_unordered = ActorPool([DummyActor.remote() for i in range(2)])
pool_map_unordered.submit(func, 6)
start_time = time.time()
gen = pool_map_unordered.map_unordered(func, map_values)
assert all(elem[0] in [1, 2, 3, 4, 5] for elem in list(gen))
assert all(
abs(elem[1] - start_time) < sleep_time in [1, 2, 3, 4, 5]
for elem in list(gen))
def extract_flame(fps):
files = list(DATASET_DIR.glob(f"*/*/video_{fps}fps.mp4"))
for i, video_file in enumerate(
tqdm(files, desc="Extracting flame parameters", leave=False)):
flame_h5_file = video_file.parent / f"flame_{fps}fps.h5"
if flame_h5_file.exists():
continue
flame_dir = video_file.parent / f"flame_{fps}fps"
gender = get_gender(video_file.parent.parent.name,
video_file.parent.name)
template_path = BASE_DIR / CONFIG["flame"][f"model_path_{gender}"]
# with open(template_model_fname, "rb") as f:
# template = pickle.load(f, encoding="latin1")
ringnet_file = video_file.parent / f"ringnet_{fps}fps.h5"
openface_file = video_file.parent / f"openface_{fps}fps.csv"
neutral_mesh_faces = Mesh(filename=str(video_file.parent /
"neutral_mesh.ply")).f
f = h5py.File(ringnet_file, "r")["flame_params"]
pool = ActorPool([
FrameOptimizer.remote(neutral_mesh_faces, template_path)
for _ in range(8)
])
openface_data = list(csv.reader(openface_file.open()))[1:]
data = f["pose"], f["shape"], f["expression"], openface_data
flame_dir.mkdir(parents=True, exist_ok=True)
runners = []
for i, (pose, shape, expression, openface) in enumerate(zip(*data), 1):
flame_file = flame_dir / f"{i:06}.npy"
if flame_file.exists():
continue
# Get 68 facial landmarks
landmarks = [float(x) for x in openface[299:435]]
# reshape the landmarks so that they are 2x51 (cut of the jaw (17 landmarks))
target_2d_lmks = np.array(landmarks).reshape(2, -1).T[17:]
runners.append(
(pose, shape, expression, target_2d_lmks, flame_file))
for file_name, flame_params in tqdm(
pool.map(lambda a, v: a.fit_lmk2d_v2.remote(*v), runners),
total=len(runners),
leave=False,
):
np.save(file_name, flame_params)
np_files = list(flame_dir.glob("*.npy"))
assert len(np_files) == len(openface_data)
results = defaultdict(list)
for file in flame_dir.glob("*.npy"):
for key, value in np.load(file, allow_pickle=True).item().items():
results[key].append(value)
with h5py.File(flame_h5_file, "w") as f:
for key, value in results.items():
f.create_dataset(key, data=np.vstack(value))
def test_map_gh23107(init):
# Reference - https://github.com/ray-project/ray/issues/23107
@ray.remote
class DummyActor:
async def identity(self, s):
return s
def func(a, v):
return a.identity.remote(v)
map_values = [1, 2, 3, 4, 5]
pool_map = ActorPool([DummyActor.remote() for i in range(2)])
pool_map.submit(func, 6)
gen = pool_map.map(func, map_values)
assert list(gen) == [1, 2, 3, 4, 5]
pool_map_unordered = ActorPool([DummyActor.remote() for i in range(2)])
pool_map_unordered.submit(func, 6)
gen = pool_map_unordered.map(func, map_values)
assert all(elem in [1, 2, 3, 4, 5] for elem in list(gen))
def test_map(init):
@ray.remote
class MyActor:
def __init__(self):
pass
def f(self, x):
return x + 1
def double(self, x):
return 2 * x
actors = [MyActor.remote() for _ in range(4)]
pool = ActorPool(actors)
index = 0
for v in pool.map(lambda a, v: a.double.remote(v), range(5)):
assert v == 2 * index
index += 1
def test_map_unordered(init):
@ray.remote
class MyActor:
def __init__(self):
pass
def f(self, x):
return x + 1
def double(self, x):
return 2 * x
actors = [MyActor.remote() for _ in range(4)]
pool = ActorPool(actors)
total = []
for v in pool.map(lambda a, v: a.double.remote(v), range(5)):
total += [v]
assert all(elem in [0, 2, 4, 6, 8] for elem in total)
import ray
from ray.util import ActorPool
@ray.remote
class Actor:
def double(self, n):
return n * 2
a1, a2 = Actor.remote(), Actor.remote()
pool = ActorPool([a1, a2])
# pool.map(..) returns a Python generator object ActorPool.map
gen = pool.map(lambda a, v: a.double.remote(v), [1, 2, 3, 4])
print(list(gen))
# [2, 4, 6, 8]
@ray.remote
class PoolActor:
def __init__(self):
self.id = str(uuid4())
def computation(self, num):
print(f'Actor with id {self.id} waiting for {num} sec')
for x in range(num):
sleep(1)
print(f'Actor with id {self.id} slept for {x} sec')
return num
# Create actors and add them to the pool
a1, a2, a3 = PoolActor.remote(), PoolActor.remote(), PoolActor.remote()
pool = ActorPool([a1, a2, a3])
print(list(pool.map(lambda a, v: a.computation.remote(v), [3, 4, 5, 4])))
pool.submit(lambda a, v: a.computation.remote(v), 3)
pool.submit(lambda a, v: a.computation.remote(v), 4)
pool.submit(lambda a, v: a.computation.remote(v), 5)
pool.submit(lambda a, v: a.computation.remote(v), 4)
print(pool.get_next())
print(pool.get_next())
print(pool.get_next())
print(pool.get_next())
class RayRunner(RunnerBase):
def __init__(self, *, random_seed: int = 0, volumes_dir: str = None, scratch_dir: str = None,
store_results=False, n_workers=None, blocklist=()) -> None:
if not ray.is_initialized():
ray.init()
super().__init__(random_seed=random_seed, volumes_dir=volumes_dir, scratch_dir=scratch_dir)
self.data_handler = DataHandler.remote()
self.ray_executor = RayExecutor.remote(random_seed=random_seed,
volumes_dir=volumes_dir, scratch_dir=scratch_dir,
store_results=store_results,blocklist=blocklist)
if n_workers is None:
n_workers = multiprocessing.cpu_count()
self.actor_pool = ActorPool([
RayExecutor.remote(random_seed=random_seed, volumes_dir=volumes_dir,
scratch_dir=scratch_dir, store_results=store_results,
blocklist=blocklist) for _ in range(n_workers)]
)
# Wait for primitives to be load on the workers
# time.sleep(len(d3m_index.search()) * 0.15)
def stop_ray(self):
ray.shutdown()
def get_request(self, request_id: str):
return ray.get(ray.ObjectID(binascii.unhexlify(request_id)))
def fit_pipeline_request(self, problem_description: Problem, pipeline: Pipeline,
input_data: typing.Sequence[ContainerType], *, timeout: float = None,
expose_outputs: bool = False) -> str:
request_id = self.data_handler.add_data.remote(input_data)
input_data_id = ray.get(request_id)
request_id = self.ray_executor.fit_pipeline.remote(self.data_handler, problem_description, pipeline, input_data_id,
timeout=timeout, expose_outputs=expose_outputs)
return request_id.hex()
def produce_pipeline_request(self, fitted_pipeline_id: str, input_data: typing.Sequence[ContainerType], *,
timeout: float = None, expose_outputs: bool = False) -> str:
request_id = self.data_handler.add_data.remote(input_data)
input_data_id = ray.get(request_id)
request_id = self.ray_executor.produce_pipeline.remote(self.data_handler, fitted_pipeline_id, input_data_id, timeout=timeout,
expose_outputs=expose_outputs)
return request_id.hex()
def evaluate_pipeline_request(
self, problem_description: Problem, pipeline: Pipeline,
input_data: typing.Sequence[ContainerType], *, metrics: typing.Sequence[typing.Dict],
data_preparation_pipeline: Pipeline = None, scoring_pipeline: Pipeline = None,
data_preparation_params: typing.Dict[str, str] = None, scoring_params: typing.Dict[str, str] = None,
timeout: float = None
) -> str:
request_id = self.data_handler.add_data.remote(input_data)
input_data_id = ray.get(request_id)
request_id = self.ray_executor.evaluate_pipeline.remote(
self.data_handler, problem_description, pipeline, input_data_id, metrics=metrics,
data_preparation_pipeline=data_preparation_pipeline, scoring_pipeline=scoring_pipeline,
data_preparation_params=data_preparation_params, scoring_params=scoring_params, timeout=timeout
)
return request_id.hex()
def fitted_pipeline_id_exists(self, fitted_pipeline_id):
request_id = self.ray_executor.fitted_pipeline_id_exists.remote(fitted_pipeline_id)
return ray.get(request_id)
def evaluate_pipelines(
self, problem_description: Problem, pipelines: typing.Sequence[Pipeline],
input_data: typing.Sequence[ContainerType], *, metrics: typing.Sequence[typing.Dict],
data_preparation_pipeline: Pipeline = None, scoring_pipeline: Pipeline = None,
data_preparation_params: typing.Dict[str, str] = None, scoring_params: typing.Dict[str, str] = None,
timeout: float = None
) -> typing.Sequence[PipelineResult]:
request_id = self.data_handler.add_data.remote(input_data)
input_data_id = ray.get(request_id)
args = []
for pipeline in pipelines:
args.append({
'data_handler': self.data_handler, 'problem_description': problem_description, 'pipeline': pipeline,
'input_data_id': input_data_id, 'metrics': metrics, 'data_preparation_pipeline': data_preparation_pipeline,
'scoring_pipeline': scoring_pipeline,'data_preparation_params': data_preparation_params,
'scoring_params': scoring_params,'timeout': timeout
})
return self.actor_pool.map(lambda actor, arg: actor.evaluate_pipeline.remote(**arg), args)
