def make_embedding(tasks_to_accomplish, tasks_finished, gpu_idx, embedding_type_name ) -> List[List[Tuple]]:
GPU_NUM = gpu_idx # 원하는 GPU 번호 입력
device = torch.device(f'cuda:{GPU_NUM}' if torch.cuda.is_available() else 'cpu')
torch.cuda.set_device(device) # change allocation of current GPU
model = SentenceTransformer(embedding_type_name )
while True:
try:
if tasks_to_accomplish.empty():
raise Empty
task = tasks_to_accomplish.get_nowait()
except Empty:
break
else:
ret = []
for i in tqdm(task):
sentence_embeddings = tuple(model.encode([i])[0].tolist())
ret.append((sentence_embeddings, i))
print(' {} : embedding making finished'.format(current_process().name))
for data in ret:
print("{} put : {}".format(current_process().name, data[1]))
tasks_finished.put(data)
print(' {} : END pushing embed-sent {} data to finished_queue'.format(current_process().name, len(ret)))
time.sleep(0.5)
return True
def build_net(env, seeds):
torch.manual_seed(seeds[0])
net = Net(env.observation_space.shape, env.action_space.n)
logger.debug("in build_net,current_process: %s,seeds:%s",
mp.current_process(), seeds)
for seed in seeds[1:]:
print(mp.current_process(), seed)
net = mutate_net(net, seed, copy_net=False)
return net
def get_frames(path):
print(mp.current_process().name + ' is getting frames at ' + path )
frames = dict()
contents = os.listdir(path)
for i in range(len(contents)):
if not i % 50:
print(mp.current_process().name + ' is ' + '{:.2f}'.format((i*1.0)/len(contents) * 100) + '% done')
content = os.path.join(path, contents[i])
frames[i] = torch.Tensor(tfm.resize(skio.imread(content), (3,224,224)))
return frames
def my_func(x, d):
print(mp.current_process())
print('process id:', os.getpid())
print('1:', mp.current_process(), id(d), id(d['y']), id(d['arr']), d['y'], d['arr'][x+500])
d['y'] = 100
print('2:', mp.current_process(), id(d), id(d['y']), id(d['arr']), d['y'], d['arr'][x+500])
d['arr'][x:x+1000] = 10
print('3:', mp.current_process(), id(d), id(d['y']), id(d['arr']), d['y'], d['arr'][x+500])
for i in range(20):
y = np.sum(d['arr'])
return x*np.sum(d['arr'])
def worker_func(input_queue, output_queue, top_parent_cache, device="cpu"):
env = make_env()
while True:
try:
# parents:(parent_seed,net,child_seed)
parents = input_queue.get()
population = []
if parents is None:
break
#logger.debug("current_process: %s,parents:%s", mp.current_process(), parents)
#logger.debug("current_process: %s,top_parent_cache:%s", mp.current_process(), top_parent_cache)
for net_seeds in parents:
if len(net_seeds) > 1:
#logger.debug("current_process: %s,net_seeds[:-1]:%s,top_parent_cache: %s", mp.current_process(),
#net_seeds[0], top_parent_cache)
logger.debug("current_process:inside1,%s,net_seeds:%s",
mp.current_process(), net_seeds)
net = Net(env.observation_space.shape, env.action_space.n)
#net = net_seeds[1]
net.load_state_dict(net_seeds[1])
if net is not None:
net = mutate_net(net, net_seeds[-1], device).to(device)
else:
assert False
#net = build_net(env, net_seeds, device).to(device)
else:
net = build_net(env, net_seeds, device).to(device)
logger.debug("current_process:inside2,%s",
mp.current_process())
reward, steps = evaluate(env, net, device)
population.append((net, net_seeds, reward, steps))
#logger.debug("before, current_process: %s,seeds:%s", mp.current_process(), population)
population.sort(key=lambda p: p[2], reverse=True)
#logger.debug("output queue put, current_process: %s,population:%s", mp.current_process(), population[:][1])
for i in range(PARENTS_COUNT):
logger.debug("current_process:inside3,%s, population[i][0],%s",
mp.current_process(), population[i][0])
#top_parent_cache[population[i][1][-1]] = population[i][0].state_dict()
output_queue.put(
OutputItem(seeds=population[i][1],
net=population[i][0].state_dict(),
reward=population[i][2],
steps=population[i][3]))
except Exception as e:
logger.debug("comme here")
logger.Exception("Unexpected exception! %s", e)
def worker_func(input_queue, output_queue, device="cpu"):
#env = make_env()#gym.make("RoboschoolHalfCheetah-v1")
env_pool = [make_env()]
#cache = {}
# first generation -- just evaluate given single seeds
parents = input_queue.get()
for seed in parents:
net = build_net(env_pool[0], seed).to(device)
#net.zero_noise(batch_size=1)
reward, steps = evaluate(env_pool[0], net, device)
output_queue.put((seed, reward, steps))
while True:
parents = input_queue.get()
print(mp.current_process(), parents)
if parents is None:
break
parents.sort()
for parent_seeds, children_iter in itertools.groupby(
parents, key=lambda s: s[:-1]):
batch = list(children_iter)
children_seeds = [b[-1] for b in batch]
net = build_net(env_pool[0], parent_seeds).to(device)
#net.set_noise_seeds(children_seeds)
batch_size = len(children_seeds)
while len(env_pool) < batch_size:
env_pool.append(make_env())
rewards, steps = evaluate_batch(env_pool[:batch_size], net, device)
for seeds, reward, step in zip(batch, rewards, steps):
output_queue.put((seeds, reward, step))
def train_worker(config, thr_incr):
"""
Args:
config (dict): dictionary containing configuration details.
thr_incr (float): A threshold analysis is performed at the end of the training
using the trained model and the validation sub-dataset to find the optimal binarization
threshold. The specified value indicates the increment between 0 and 1 used during the
ROC analysis (e.g. 0.1). Flag: ``-t``, ``--thr-increment``
"""
current = mp.current_process()
# ID of process used to assign a GPU
ID = int(current.name[-1]) - 1
# Use GPU i from the array specified in the config file
config["gpu_ids"] = [config["gpu_ids"][ID]]
# Call ivado cmd_train
try:
# Save best validation score
config["command"] = "train"
best_training_dice, best_training_loss, best_validation_dice, best_validation_loss = \
ivado.run_command(config, thr_increment=thr_incr)
except Exception:
logging.exception('Got exception on main handler')
logging.info("Unexpected error:", sys.exc_info()[0])
raise
# Save config file in output path
config_copy = open(config["path_output"] + "/config_file.json", "w")
json.dump(config, config_copy, indent=4)
return config["path_output"], best_training_dice, best_training_loss, best_validation_dice, \
best_validation_loss
def kernel(act_net, train_queue, device, num_threads, env_seed):
ae = Autoencoder84(
1,
pretrained="./Design/Models/AE/Autoencoder84.dat").to(device).eval()
env = sc.StageCreator(RES, ae=ae, boundary=.8, mode="goal",
seed=env_seed) # individual environment
env.mode = "selfplay"
buffer = common.ExperienceBuffer(buffer_size=REPLAY_SIZE // num_threads,
device=device) # indiovidual exp replay
agent = common.Agent_SAC_HER(act_net,
env,
buffer,
GAMMA,
device=device,
unroll_steps=UNROLL)
thread_id = mp.current_process().name
exp_count = 0
while True:
exp_count = agent.play_episode(
random=False) # play same episode over and over again
mean_episode_reward, mean_step_count = agent.get_mean_reward_and_steps(
)
if len(buffer) < REPLAY_INITIAL // num_threads:
continue
batch = buffer.sample(BATCH_SIZE)
train_queue.put((batch, exp_count, mean_episode_reward,
mean_step_count, thread_id))
def __init__(self,
env,
actor: torch.nn.Module,
replay_buffer: SharedReplayBuffer,
scheduler: Scheduler,
argp,
logger: SummaryWriter = None):
self.env = env
self.actor = actor
self.replay_buffer = replay_buffer
self.scheduler = scheduler
self.num_trajectories = argp.num_trajectories
self.trajectory_length = argp.episode_length
self.schedule_switch = argp.schedule_switch
self.discount_factor = argp.discount_factor
self.discounts = torch.cumprod(
torch.ones([self.trajectory_length - 1]) * 0.99, dim=-1)
self.log_every = 10
self.logger = logger
if argp.num_workers > 1:
self.process_id = current_process()._identity[0] # process ID
else:
self.process_id = 1
def worker_func(input_queue, output_queue, device="cpu"):
env = make_env() #gym.make("RoboschoolHalfCheetah-v1")
cache = {}
while True:
parents = input_queue.get()
print(mp.current_process(), parents)
if parents is None:
break
new_cache = {}
for net_seeds in parents:
if len(net_seeds) > 1:
net = cache.get(net_seeds[:-1])
if net is not None:
net = mutate_net(net, net_seeds[-1])
else:
net = build_net(env, net_seeds).to(device)
else:
net = build_net(env, net_seeds).to(device)
# store{(seed,):net}
new_cache[net_seeds] = net
reward, steps = evaluate(env, net, device)
output_queue.put(
OutputItem(seeds=net_seeds, reward=reward, steps=steps))
cache.update(new_cache)
def train_worker(config, thr_incr):
current = mp.current_process()
# ID of process used to assign a GPU
ID = int(current.name[-1]) - 1
# Use GPU i from the array specified in the config file
config["gpu"] = config["gpu"][ID]
# Call ivado cmd_train
try:
# Save best validation score
best_training_dice, best_training_loss, best_validation_dice, best_validation_loss = \
ivado.run_command(config, thr_increment=thr_incr)
except:
logging.exception('Got exception on main handler')
print("Unexpected error:", sys.exc_info()[0])
raise
# Save config file in log directory
config_copy = open(config["log_directory"] + "/config_file.json", "w")
json.dump(config, config_copy, indent=4)
return config[
"log_directory"], best_training_dice, best_training_loss, best_validation_dice, best_validation_loss
def fun(datum):
global global_data
if global_data is None:
global_data = str(multiprocessing.current_process())
with multiprocessing.Pool(processes=2) as p:
result = list(p.imap_unordered(funfun, [datum + k for k in range(9)]))
return {'global_data': global_data, 'results': result}
def bert_worker(i, artist, track, lyrics, get_embedding_fn, agg_type,
tokenizer, models):
wid = int(current_process().name[-1]) - 1
print(f'[processor {wid}] i={i:<6} {artist:<20} {track:>30}')
embedding = get_embedding_fn(tokenizer, agg_type, models[wid], lyrics)
with open(f'wid_{wid}_done.txt', 'a') as f:
f.write(f'"{artist}","{track}","{embedding}"\n')
return (artist, track, embedding)
def run(self):
nStepBuffer = []
bufferState, bufferAction, bufferReward, bufferNextState = [], [], [], []
for self.epIdx in range(self.trainStep):
print("episode index:" + str(self.epIdx) + " from" +
current_process().name + "\n")
state = self.env.reset()
done = False
rewardSum = 0
stepCount = 0
# clear the nstep buffer
nStepBuffer.clear()
while not done:
episode = self.epsilon_by_episode(
self.globalEpisodeCount.value)
action = self.select_action(self.localNet, state, episode)
nextState, reward, done, info = self.env.step(action)
nStepBuffer.append((state, action, nextState, reward))
if len(nStepBuffer) > self.nStepForward:
R = sum([
nStepBuffer[i][3] * (self.gamma**i)
for i in range(self.nStepForward)
])
state, action, _, _ = nStepBuffer.pop(0)
bufferAction.append(action)
bufferState.append(state)
bufferReward.append(R)
bufferNextState.append(nextState)
state = nextState
rewardSum += reward
if self.totalStep % self.updateGlobalFrequency == 0 and len(
bufferAction
) > 0: # update global and assign to local net
# sync
self.update_net_and_sync(bufferAction, bufferState,
bufferReward, bufferNextState)
bufferAction.clear()
bufferState.clear()
bufferReward.clear()
bufferNextState.clear()
if done:
# print("done in step count: {}".format(stepCount))
# print("reward sum = " + str(rewardSum))
# done and print information
# pass
self.recordInfo(rewardSum, stepCount)
stepCount += 1
self.totalStep += 1
self.resultQueue.put(None)
def train_parallel(args):
if torch.cuda.is_available() and not args.cpu:
pid = mp.current_process().pid
torch.cuda.set_device(args.pid_to_cuda[pid])
args.devices = [args.pid_to_cuda[pid]]
args.checkpoint_path = args.checkpoint_path + f".{args.devices[0]}"
result = train(args)
return result
def _initializer(*args):
"""
Process initializer function that is called when mp.Pool is started.
:param args: arguments that are to be copied to the target process. This can be a tuple for convenience.
"""
global process_initial_data
process_id, process_initial_data = args
assert 'OMP_NUM_THREADS' in os.environ
torch.set_num_threads(int(os.environ['OMP_NUM_THREADS']))
# manually set process name
with process_id.get_lock():
mp.current_process().name = 'PoolWorker-' + str(process_id.value)
logger.info('pid=' + str(os.getpid()) + ' : ' +
mp.current_process().name)
process_id.value += 1
def run(self):
torch.set_num_threads(1)
bufferState, bufferAction, bufferReward, bufferNextState = [], [], [], []
for self.epIdx in range(self.trainStep):
print("episode index:" + str(self.epIdx) + " from" + current_process().name + "\n")
state = self.env.reset()
done = False
rewardSum = 0
# clear the nstep buffer
self.nStepBuffer.clear()
for stepCount in range(self.episodeLength):
epsilon = self.epsilon_by_episode(self.globalEpisodeCount.value)
action = self.select_action(self.localNet, state, epsilon)
nextState, reward, done, info = self.env.step(action)
if stepCount == 0:
print("at step 0: from " + current_process().name + "\n")
print(info)
if done:
nextState = None
self.update_net_and_sync(state, action, nextState, reward)
state = nextState
rewardSum += reward * pow(self.gamma, stepCount)
self.totalStep += 1
if done:
# print("done in step count: {}".format(stepCount))
# print("reward sum = " + str(rewardSum))
# done and print information
# pass
break
self.recordInfo(rewardSum, stepCount)
self.resultQueue.put(None)
def multi_inference(embedding_type_name, tasks_to_accomplish, tasks_finished,
gpu_idx, src_sentences, src_sentence_embeddings) -> bool:
'''
param
- embedding_type_name : str
- setnece transformer 에서 제공하는 embedding 종류
- tasks_to_accomplish : Queue
- sentence(str) 이 담긴 queue
- tasks_finished : Queue
- inference 결과 로그를 담는 queue
- src_sentences : List[str]
- source pool sentence List
- src_sentence_embeddgins
- source pool settence vector 들의 List
desc
- 해당하는 emedding_type_name로 sentence encoder를 로딩
- tasks_to_accomplish 에서 job(str)을 가져옴
- src_sentence_embeddings 를 보고 가장 유사한 embeddig을 찾고 매핑된 src_sentence를 구한다
'''
GPU_NUM = gpu_idx # 원하는 GPU 번호 입력
device = torch.device(
f'cuda:{GPU_NUM}' if torch.cuda.is_available() else 'cpu')
torch.cuda.set_device(device) # change allocation of current GPU
model = SentenceTransformer(embedding_type_name)
levin = CustomLevin()
while True:
try:
if tasks_to_accomplish.empty():
raise Empty
task = tasks_to_accomplish.get_nowait()
except Empty:
break
else:
ret = []
# ret : [(sentence_embeddings, concat_synom) ... ]
for i in tqdm(task):
# i -> 한 프로세서에 배분된 target sentence의 리스트
target = i
sbert_inference = predict(model, levin, target, src_sentences,
src_sentence_embeddings)
ret.append([i, sbert_inference])
for data in ret:
# target = data[0]
# inference = data[1]
# sbert_only_inference = data[2]
tasks_finished.put(data)
print(' {} : END pushing {} data to finished_queue'.format(
current_process().name, len(ret)))
time.sleep(0.5)
return True
def main(args):
if torch.cuda.is_available() and not args.cpu:
pid = mp.current_process().pid
torch.cuda.set_device(args.pid_to_cuda[pid])
set_random_seed(args.seed)
task = build_task(args)
result = task.train()
return result
def worker_func(input_queue, output_queue, device_w="cpu"):
new_env = make_env()
parent_list = []
for i in range(PARENTS_COUNT):
parent_list.append(i)
#parent_list = [0, 1]
logger.debug("in work_func,current_process: {0},parent_list:{1}".format(
mp.current_process(), parent_list))
while True:
get_item = input_queue.get()
parents_w = get_item[0]
pro_list = get_item[1]
batch_steps_w = 0
child = []
logger.debug(
"in worker_func, current_process: {0},parents[0][0]:{1},len of parents:{2},pro_list:{3}"
.format(mp.current_process(), parents_w[0]['fc.2.bias'],
len(parents_w), pro_list))
for _ in range(SEEDS_PER_WORKER):
#solve pro do not sum to 1
pro_list = np.array(pro_list)
pro_list = pro_list / sum(pro_list)
parent = np.random.choice(parent_list, p=pro_list)
#parent = rand_pick(parent_list, pro_list)
#parent = np.random.randint(PARENTS_COUNT)
child_seed = np.random.randint(MAX_SEED)
child_net = mutate_net(new_env, parents_w[parent],
child_seed).to(device_w)
reward, steps = evaluate(new_env, child_net, device_w)
batch_steps_w += steps
child.append((child_net.state_dict(), reward, steps))
child.sort(key=lambda p: p[1], reverse=True)
#logger.debug("middle, current_process: {0},child[0][1]:{1},child[0][2]:{2},len of "
# "child:{3}".format(mp.current_process(), child[0][1], child[0][2], len(child)))
for i in range(PARENTS_COUNT):
#output_queue.put(child[i])
output_queue.put(
OutputItem(child_net=child[i][0],
reward=child[i][1],
steps=batch_steps_w))
def get_batches(group, bsize=1): keys = group.keys() for i in range(0, len(keys), bsize): batch = torch.Tensor(np.zeros([bsize,3,224,224])) for j in range(bsize): if i+j < len(keys): key = keys[i+j] batch[j] = group[key] else: print(mp.current_process().name, i, j, i+j) yield batch
def kernel(alice, bob, train_queue, device, seed=None):
""" Execute by each single thread """
env = sc.StageCreator(RES, mode="selfplay", boundary=0, seed=seed)
agent = SelfplayAgent(alice, bob, env, device=device, gamma=GAMMA)
while True:
for _ in range(REPORT_INTERV):
train_queue.put(agent.selfplay_episode(random=True))
# Calculate rewards and send them to master for logging
a_data, b_data = agent.selfplay_episode(random=True)
mean_r_a, mean_r_b = agent.get_selfplay_rewards()
train_queue.put((a_data, b_data, mean_r_a, mean_r_b, mp.current_process().name))
def runEnv(env):
env.reset()
print("Hello, World! from " + current_process().name + "\n")
print('agentID')
print(id(env.agent))
print('MapID')
print(id(env.mapMat))
print('envID')
print(id(env))
for i in range(4):
action = random.randint(0, env.nbActions - 1)
state, reward, done, info = env.step(action)
print("Hello, World! from " + current_process().name + "\n")
print('step ' + str(i))
#print(state)
#print(reward)
#print(done)
print(info)
if done:
print("great!!!!!!!!!!!!!!!!!!!")
break
def kernel(alice, bob, train_queue, device):
""" Execute by each single thread """
env = sc.StageCreator(target=False)
env = sc.ScreenOutput(RES, env)
agent = SelfplayAgent(alice, bob, env, device=device, gamma=GAMMA)
while True:
for _ in range(10):
train_queue.put(agent.selfplay_episode())
a_data, b_data = agent.selfplay_episode()
mean_r_a, mean_r_b = agent.get_selfplay_rewards()
train_queue.put(
(a_data, b_data, mean_r_a, mean_r_b, mp.current_process().name))
def _init1(args):
trainer, keep_on_gpu, cuda_pool = args
if cuda_pool:
procid = (current_process()._identity or [1])[0] - 1
cuda_id = cuda_pool[procid]
else:
cuda_id = None
if (cuda_id is not None
and torch.device(trainer.device).type.startswith('cuda')):
trainer.device = torch.device(f'cuda:{cuda_id}')
trainer.init()
if not keep_on_gpu:
trainer.model = trainer.to(device='cpu')
return trainer
def worker_func(obj, f):
print(mp.current_process())
print('process id:', os.getpid())
time.sleep(2)
y = np.random.rand(int(100e6))
obj.donothing()
obj.calc(f)
obj.donothing()
print(id(obj), obj.x)
time.sleep(2)
obj.modify(f)
print(id(obj), obj.x)
obj.donothing()
time.sleep(2)
def get_logger() -> logging.Logger:
"""Get a `logging.Logger` to stderr. It can be called whenever we wish to
log some message. Messages can get mixed-up
(https://docs.python.org/3.6/library/multiprocessing.html#logging), but it
works well in most cases.
# Returns
logger: the `logging.Logger` object
"""
if _new_logger():
if mp.current_process().name == "MainProcess":
_new_logger(logging.DEBUG)
_set_log_formatter()
return _LOGGER
def test_multiProcess(self):
print("Hello, World! from " + current_process().name + "\n")
print(self.globalPolicyNet.state_dict())
for gp in self.globalPolicyNet.parameters():
gp.grad = torch.ones_like(gp)
#gp.grad.fill_(1)
self.globalOptimizer.step()
print('globalNetID:')
print(id(self.globalPolicyNet))
print('globalOptimizer:')
print(id(self.globalOptimizer))
print('localNetID:')
print(id(self.localNet))
def run(self) -> None:
self._proc_id = current_process()
ex = self._build_sacred_experiment()
@ex.automain
def main(_run, _config, _log):
# Load config and logger
config = self._set_seed(_config)
# run the framework
self.run_sacred_framework(_run, config, _log)
ex.run()
os._exit(os.EX_OK)
def __init__(self, param_server, shared_replay_buffer, parser_args):
self.device = "cuda:0" if torch.cuda.is_available() else "cpu"
self.param_server = param_server
if parser_args.num_worker > 1:
self.pid = current_process()._identity[0]
else:
self.pid = 1
self.logging = parser_args.logging
self.logger = None
if self.pid == 1 and self.logging: # only create one logger
self.logger = SummaryWriter()
self.shared_replay_buffer = shared_replay_buffer
self.env = gym.make("Swimmer-v2")
# self.env = gym.make("Pendulum-v0")
# self.env = gym.make("HalfCheetah-v2")
self.num_actions = self.env.action_space.shape[0]
self.num_obs = self.env.observation_space.shape[0]
self.actor = Actor(num_actions=self.num_actions,
num_obs=self.num_obs,
log_std_init=np.log(parser_args.init_std)).to(
self.device)
self.critic = Critic(num_actions=self.num_actions,
num_obs=self.num_obs).to(self.device)
self.target_actor = copy.deepcopy(self.actor).to(self.device)
self.target_actor.freeze_net()
self.target_critic = copy.deepcopy(self.critic).to(self.device)
self.target_critic.freeze_net()
self.actor_loss = ActorLoss(alpha=parser_args.entropy_reg)
self.critic_loss = Retrace(num_actions=self.num_actions)
self.num_trajectories = parser_args.num_trajectories
self.update_targnets_every = parser_args.update_targnets_every
self.learning_steps = parser_args.learning_steps
self.num_runs = parser_args.num_runs
self.render = parser_args.render
self.log_every = parser_args.log_interval
self.num_grads = parser_args.num_grads
self.grad_ctr = 0
self.cond = Condition()
