def run(config, seed, device):
set_global_seeds(seed)
logdir = Path(config['log.dir']) / str(config['ID']) / str(seed)
env = make_env(config, seed)
env = VecMonitor(env)
if config['env.standardize_obs']:
env = VecStandardizeObservation(env, clip=5.)
if config['env.standardize_reward']:
env = VecStandardizeReward(env, clip=10., gamma=config['agent.gamma'])
agent = Agent(config, env, device)
runner = EpisodeRunner(reset_on_call=False)
engine = Engine(config, agent=agent, env=env, runner=runner)
train_logs = []
for i in count():
if agent.total_timestep >= config['train.timestep']:
break
train_logger = engine.train(i)
train_logs.append(train_logger.logs)
if i == 0 or (i+1) % config['log.freq'] == 0:
train_logger.dump(keys=None, index=0, indent=0, border='-'*50)
if i == 0 or (i+1) % config['checkpoint.freq'] == 0:
agent.checkpoint(logdir, i + 1)
agent.checkpoint(logdir, i + 1)
pickle_dump(obj=train_logs, f=logdir/'train_logs', ext='.pkl')
return None
def checkpoint(self, logdir, num_iter):
self.save(logdir / f'agent_{num_iter}.pth')
obs_env = get_wrapper(self.env, 'VecStandardizeObservation')
if obs_env is not None:
pickle_dump(obj=(obs_env.mean, obs_env.var),
f=logdir / f'obs_moments_{num_iter}',
ext='.pth')
def run(config, seed, device, logdir):
set_global_seeds(seed)
env = make_env(config, seed, 'train')
if config['use_lstm']:
agent = LSTMAgent(config, env, device)
else:
agent = Agent(config, env, device)
runner = StepRunner(reset_on_call=False)
engine = Engine(config, agent=agent, env=env, runner=runner)
train_logs = []
checkpoint_count = 0
for i in count():
if agent.total_timestep >= config['train.timestep']:
break
train_logger = engine.train(i)
train_logs.append(train_logger.logs)
if i == 0 or (i + 1) % config['log.freq'] == 0:
train_logger.dump(keys=None, index=0, indent=0, border='-' * 50)
if agent.total_timestep >= int(config['train.timestep'] *
(checkpoint_count /
(config['checkpoint.num'] - 1))):
agent.checkpoint(logdir, i + 1)
checkpoint_count += 1
pickle_dump(obj=train_logs, f=logdir / 'train_logs', ext='.pkl')
return None
def checkpoint(self, logdir, num_iter):
self.save(logdir / f'agent_{num_iter}.pth')
if self.config['env.normalize_obs']:
moments = (self.env.obs_moments.mean, self.env.obs_moments.var)
pickle_dump(obj=moments,
f=logdir / f'obs_moments_{num_iter}',
ext='.pth')
def run(config, seed, device, logdir):
set_global_seeds(seed)
print('Initializing...')
agent = Agent(config, make_env(config, seed), device)
es = CMAES([config['train.mu0']]*agent.num_params, config['train.std0'],
{'popsize': config['train.popsize'],
'seed': seed})
train_logs = []
checkpoint_count = 0
with ProcessPoolExecutor(max_workers=config['train.popsize'], initializer=initializer, initargs=(config, seed, device)) as executor:
print('Finish initialization. Training starts...')
for generation in range(config['train.generations']):
start_time = time.perf_counter()
solutions = es.ask()
out = list(executor.map(fitness, solutions, chunksize=2))
Rs, Hs = zip(*out)
es.tell(solutions, [-R for R in Rs])
logger = Logger()
logger('generation', generation+1)
logger('num_seconds', round(time.perf_counter() - start_time, 1))
logger('Returns', describe(Rs, axis=-1, repr_indent=1, repr_prefix='\n'))
logger('Horizons', describe(Hs, axis=-1, repr_indent=1, repr_prefix='\n'))
logger('fbest', es.result.fbest)
train_logs.append(logger.logs)
if generation == 0 or (generation+1)%config['log.freq'] == 0:
logger.dump(keys=None, index=0, indent=0, border='-'*50)
if (generation+1) >= int(config['train.generations']*(checkpoint_count/(config['checkpoint.num'] - 1))):
agent.from_vec(tensorify(es.result.xbest, 'cpu'))
agent.checkpoint(logdir, generation+1)
checkpoint_count += 1
pickle_dump(obj=train_logs, f=logdir/'train_logs', ext='.pkl')
return None
def run(config, seed, device, logdir):
set_global_seeds(seed)
env = make_env(config, seed, 'train')
eval_env = make_env(config, seed, 'eval')
random_agent = RandomAgent(config, env, device)
if config['agent.use_td3']:
agent = TD3Agent(config, env, device)
else:
agent = DDPGAgent(config, env, device)
runner = EpisodeRunner()
replay = ReplayBuffer(env, config['replay.capacity'], device)
engine = Engine(config,
agent=agent,
random_agent=random_agent,
env=env,
eval_env=eval_env,
runner=runner,
replay=replay,
logdir=logdir)
train_logs, eval_logs = engine.train()
pickle_dump(obj=train_logs, f=logdir / 'train_logs', ext='.pkl')
pickle_dump(obj=eval_logs, f=logdir / 'eval_logs', ext='.pkl')
return None
def run(config, seed, device):
set_global_seeds(seed)
logdir = Path(config['log.dir']) / str(config['ID']) / str(seed)
train_loader, test_loader = make_dataset(config)
if config['nn.type'] == 'VAE':
model = VAE(config, device)
elif config['nn.type'] == 'ConvVAE':
model = ConvVAE(config, device)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
engine = Engine(config,
model=model,
optimizer=optimizer,
train_loader=train_loader,
test_loader=test_loader)
train_logs = []
eval_logs = []
for epoch in range(config['train.num_epoch']):
train_logger = engine.train(epoch, logdir=logdir)
train_logs.append(train_logger.logs)
eval_logger = engine.eval(epoch, logdir=logdir)
eval_logs.append(eval_logger.logs)
pickle_dump(obj=train_logs, f=logdir / 'train_logs', ext='.pkl')
pickle_dump(obj=eval_logs, f=logdir / 'eval_logs', ext='.pkl')
return None
def save_running_average(self, f):
r"""Save the running averages for observation and reward in a dictionary by pickling.
It saves the mean and standard deviation for observation running average and the standard deviation
for reward running average. A dictionary with keys 'obs_avg' and 'r_avg' will be created. Each key
contains sub-keys ['mu', 'sigma'].
Args:
f (str): saving path
"""
out = self.running_averages
pickle_dump(obj=out, f=f, ext='.pkl')
def save_configs(self, f, method='pickle'):
r"""Save the list of configurations returned from :meth:`make_configs`.
Args:
f (str): file path
method (str): the method to save the list of configuration. Either 'pickle' or 'yaml'
"""
methods = ['pickle', 'yaml']
assert method in methods, f'expected {methods}, got {method}'
if method == 'pickle':
pickle_dump(obj=self.configs, f=f, ext='.pkl')
elif method == 'yaml':
yaml_dump(obj=self.configs, f=f, ext='.yml')
def run(config, seed, device, logdir):
set_global_seeds(seed)
torch.set_num_threads(1) # VERY IMPORTANT TO AVOID GETTING STUCK
print('Initializing...')
agent = Agent(config, make_env(config, seed, 'eval'), device)
es = OpenAIES(
[config['train.mu0']] * agent.num_params, config['train.std0'], {
'popsize': config['train.popsize'],
'seed': seed,
'sigma_scheduler_args': config['train.sigma_scheduler_args'],
'lr': config['train.lr'],
'lr_decay': config['train.lr_decay'],
'min_lr': config['train.min_lr'],
'antithetic': config['train.antithetic'],
'rank_transform': config['train.rank_transform']
})
train_logs = []
checkpoint_count = 0
with Pool(processes=config['train.popsize'] //
config['train.worker_chunksize']) as pool:
print('Finish initialization. Training starts...')
for generation in range(config['train.generations']):
t0 = time.perf_counter()
solutions = es.ask()
data = [(config, seed, device, solution) for solution in solutions]
out = pool.map(CloudpickleWrapper(fitness),
data,
chunksize=config['train.worker_chunksize'])
Rs, Hs = zip(*out)
es.tell(solutions, [-R for R in Rs])
logger = Logger()
logger('generation', generation + 1)
logger('num_seconds', round(time.perf_counter() - t0, 1))
logger('Returns',
describe(Rs, axis=-1, repr_indent=1, repr_prefix='\n'))
logger('Horizons',
describe(Hs, axis=-1, repr_indent=1, repr_prefix='\n'))
logger('fbest', es.result.fbest)
train_logs.append(logger.logs)
if generation == 0 or (generation + 1) % config['log.freq'] == 0:
logger.dump(keys=None, index=0, indent=0, border='-' * 50)
if (generation + 1) >= int(config['train.generations'] *
(checkpoint_count /
(config['checkpoint.num'] - 1))):
agent.from_vec(tensorify(es.result.xbest, 'cpu'))
agent.checkpoint(logdir, generation + 1)
checkpoint_count += 1
pickle_dump(obj=train_logs, f=logdir / 'train_logs', ext='.pkl')
return None
def run(config, seed, device, logdir):
set_global_seeds(seed)
env = make_env(config, seed)
env = VecMonitor(env)
env = VecStepInfo(env)
eval_env = make_env(config, seed)
eval_env = VecMonitor(eval_env)
agent = Agent(config, env, device)
replay = ReplayBuffer(env, config['replay.capacity'], device)
engine = Engine(config, agent=agent, env=env, eval_env=eval_env, replay=replay, logdir=logdir)
train_logs, eval_logs = engine.train()
pickle_dump(obj=train_logs, f=logdir/'train_logs', ext='.pkl')
pickle_dump(obj=eval_logs, f=logdir/'eval_logs', ext='.pkl')
return None
def process_es_result(self, result):
best_f_val = result['best_f_val']
best_return = -best_f_val
self.logger('generation', self.generation + 1)
self.logger('best_return', best_return)
if self.generation == 0 or (self.generation +
1) % self.config['log.interval'] == 0:
print('-' * 50)
self.logger.dump(keys=None, index=-1, indent=0)
print('-' * 50)
# Save the loggings and final parameters
if (self.generation + 1) == self.config['train.num_iteration']:
pickle_dump(obj=self.logger.logs,
f=self.logdir / 'result',
ext='.pkl')
np.save(self.logdir / 'trained_param', result['best_param'])
def test_pickle_yaml():
a = {'one': 1, 'two': [2, 3]}
b = {'three': 3, 'four': [4, 5]}
c = [a, b]
def check(x):
assert isinstance(x, list)
assert len(x) == 2
assert all([isinstance(i, dict) for i in x])
assert list(x[0].keys()) == ['one', 'two']
assert list(x[1].keys()) == ['three', 'four']
assert list(x[0].values()) == [1, [2, 3]]
assert list(x[1].values()) == [3, [4, 5]]
pickle_dump(c, '.tmp_pickle')
check(pickle_load('.tmp_pickle.pkl'))
os.unlink('.tmp_pickle.pkl')
yaml_dump(c, '.tmp_yaml')
check(yaml_load('.tmp_yaml.yml'))
os.unlink('.tmp_yaml.yml')
def learner(config, logdir, agent, engine, queue):
torch.set_num_threads(1) # VERY IMPORTANT TO AVOID GETTING STUCK
train_logs = []
checkpoint_count = 0
n = 0
while agent.total_timestep < config['train.timestep']:
D = []
while len(D) < config['train.batch_size']:
while queue.empty():
time.sleep(0.01)
D.append(queue.get_nowait())
train_logger = engine.train(n, D=D)
train_logs.append(train_logger.logs)
if n == 0 or (n + 1) % config['log.freq'] == 0:
train_logger.dump(keys=None, index=0, indent=0, border='-' * 50)
if agent.total_timestep >= int(config['train.timestep'] *
(checkpoint_count /
(config['checkpoint.num'] - 1))):
agent.checkpoint(logdir, n + 1)
checkpoint_count += 1
n += 1
pickle_dump(obj=train_logs, f=logdir / 'train_logs', ext='.pkl')
def algorithm(config, seed, device):
logdir = Path(config['log.dir']) / str(config['ID']) / str(seed)
seeder = Seeder(seed)
seeds = seeder(size=config['env.count'])
env_constructors = []
for seed in seeds:
env_constructors.append(partial(CraftingEnv, seed))
env = VecStandardize(SerialVecEnv(env_constructors),
clip_reward=100.0)
env_spec = EnvSpec(env)
agent = Agent(config, env_spec, device)
runner = RollingSegmentRunner(config, agent, env)
engine = Engine(agent, runner, env)
for i in range(config['train.iter']):
training_result = engine.train(i)
print(f'Training iteration {i} complete.')
if i % config['log.interval'] == 0:
logs = engine.log_train(training_result)
pickle_dump(obj=logs, f=logdir / f'iter_{i}_train_logs', ext='.pkl')
torch.save(engine.agent.policy.state_dict(),
logdir / 'trained_params')
def run_experiment(run, config, seeds, num_worker):
r"""A convenient function to launch a parallelized experiment (Master-Worker).
.. note::
It automatically creates all subfolders for logging the experiment. The topmost
folder is indicated by the logging directory specified in the configuration.
Then all subfolders for each configuration are created with the name of their ID.
Finally, under each configuration subfolder, a set subfolders are created for each
random seed (the random seed as folder name). Intuitively, an experiment could have
following directory structure::
- logs
- 0 # ID number
- 123 # random seed
- 345
- 567
- 1
- 123
- 345
- 567
- 2
- 123
- 345
- 567
- 3
- 123
- 345
- 567
- 4
- 123
- 345
- 567
Args:
run (function): an algorithm function to train on.
config (Config): a :class:`Config` object defining all configuration settings
seeds (list): a list of random seeds
num_worker (int): number of workers
"""
experiment = ExperimentMaster(ExperimentWorker, num_worker, run, config,
seeds)
log_path = Path(experiment.configs[0]['log.dir'])
if not log_path.exists():
log_path.mkdir(parents=True)
else:
msg = f"Logging directory '{log_path.absolute()}' already existed, do you want to clean it ?"
answer = ask_yes_or_no(msg)
if answer:
rmtree(log_path)
log_path.mkdir(parents=True)
else: # back up
old_log_path = log_path.with_name('old_' + log_path.name)
log_path.rename(old_log_path)
log_path.mkdir(parents=True)
print(
f"The old logging directory is renamed to '{old_log_path.absolute()}'. "
)
input('Please, press Enter to continue\n>>> ')
# save source files
source_path = Path(log_path / 'source_files/')
source_path.mkdir(parents=True)
[
copyfile(s, source_path / s.name)
for s in Path(inspect.getsourcefile(run)).parent.glob('*.py')
]
# Create subfolders for each ID and subsubfolders for each random seed
for config in experiment.configs:
ID = config['ID']
for seed in experiment.seeds:
p = log_path / f'{ID}' / f'{seed}'
p.mkdir(parents=True)
yaml_dump(obj=config, f=log_path / f'{ID}' / 'config', ext='.yml')
pickle_dump(experiment.configs, log_path / 'configs', ext='.pkl')
# Run experiment in parallel
results = experiment()
return results
def __call__(self, config, seed, device):
set_global_seeds(seed)
logdir = Path(config['log.dir']) / str(config['ID']) / str(seed)
if config['env.time_aware_obs']:
kwargs = {'extra_wrapper': [TimeAwareObservation]}
else:
kwargs = {}
env = make_vec_env(SerialVecEnv,
make_gym_env,
config['env.id'],
config['train.N'],
seed,
monitor=True,
**kwargs)
if config['eval.independent']:
eval_env = make_vec_env(SerialVecEnv, make_gym_env,
config['env.id'], config['eval.N'], seed)
if config['env.clip_action']:
env = VecClipAction(env)
if config['eval.independent']:
eval_env = VecClipAction(eval_env)
if config[
'env.standardize']: # running averages of observation and reward
env = VecStandardize(
venv=env,
use_obs=True,
use_reward=False, # A2C specific
clip_obs=10.,
clip_reward=10.,
gamma=0.99,
eps=1e-8)
env_spec = EnvSpec(env)
agent = Agent(config, env_spec, device)
runner = RollingSegmentRunner(config, agent, env)
if config['eval.independent']:
engine = Engine(agent, runner, config, eval_env=eval_env)
else:
engine = Engine(agent, runner, config)
train_logs = []
eval_logs = []
for i in count():
if 'train.iter' in config and i >= config[
'train.iter']: # enough iterations
break
elif 'train.timestep' in config and agent.total_T >= config[
'train.timestep']: # enough timesteps
break
train_output = engine.train(i)
if i == 0 or (i + 1) % config['log.interval'] == 0:
train_log = engine.log_train(train_output)
train_logs.append(train_log)
if config['eval.independent']:
with torch.no_grad(): # disable grad, save memory
eval_output = engine.eval(n=i)
eval_log = engine.log_eval(eval_output)
eval_logs.append(eval_log)
pickle_dump(obj=train_logs, f=logdir / 'train_logs', ext='.pkl')
pickle_dump(obj=eval_logs, f=logdir / 'eval_logs', ext='.pkl')
return None
def run_experiment(run,
config,
seeds,
log_dir,
max_workers,
chunksize=1,
use_gpu=False,
gpu_ids=None):
r"""A convenient function to parallelize the experiment (master-worker pipeline).
It is implemented by using `concurrent.futures.ProcessPoolExecutor`
It automatically creates all subfolders for each pair of configuration and random seed
to store the loggings of the experiment. The root folder is given by the user.
Then all subfolders for each configuration are created with the name of their job IDs.
Under each configuration subfolder, a set subfolders are created for each
random seed (the random seed as folder name). Intuitively, an experiment could have
following directory structure::
- logs
- 0 # ID number
- 123 # random seed
- 345
- 567
- 1
- 123
- 345
- 567
- 2
- 123
- 345
- 567
- 3
- 123
- 345
- 567
- 4
- 123
- 345
- 567
Args:
run (function): a function that defines an algorithm, it must take the
arguments `(config, seed, device, logdir)`
config (Config): a :class:`Config` object defining all configuration settings
seeds (list): a list of random seeds
log_dir (str): a string to indicate the path to store loggings.
max_workers (int): argument for ProcessPoolExecutor. if `None`, then all experiments run serially.
chunksize (int): argument for Executor.map()
use_gpu (bool): if `True`, then use CUDA. Otherwise, use CPU.
gpu_ids (list): if `None`, then use all available GPUs. Otherwise, only use the
GPU device defined in the list.
"""
configs = config.make_configs()
# create logging dir
log_path = Path(log_dir)
if not log_path.exists():
log_path.mkdir(parents=True)
else:
msg = f"Logging directory '{log_path.absolute()}' already existed, do you want to clean it ?"
answer = ask_yes_or_no(msg)
if answer:
rmtree(log_path)
log_path.mkdir(parents=True)
else: # back up
old_log_path = log_path.with_name('old_' + log_path.name)
log_path.rename(old_log_path)
log_path.mkdir(parents=True)
print(
f"The old logging directory is renamed to '{old_log_path.absolute()}'. "
)
input('Please, press Enter to continue\n>>> ')
# save source files
source_path = Path(log_path / 'source_files/')
source_path.mkdir(parents=True)
[
copyfile(s, source_path / s.name)
for s in Path(inspect.getsourcefile(run)).parent.glob('*.py')
]
# Create subfolders for each ID and subsubfolders for each random seed
for config in configs:
ID = config['ID']
for seed in seeds:
p = log_path / f'{ID}' / f'{seed}'
p.mkdir(parents=True)
yaml_dump(obj=config, f=log_path / f'{ID}' / 'config', ext='.yml')
pickle_dump(configs, log_path / 'configs', ext='.pkl')
# Create unique id for each job
jobs = list(enumerate(product(configs, seeds)))
def _run(job):
job_id, (config, seed) = job
# VERY IMPORTANT TO AVOID GETTING STUCK, oversubscription
# see following links
# https://github.com/pytorch/pytorch/issues/19163
# https://software.intel.com/en-us/intel-threading-building-blocks-openmp-or-native-threads
torch.set_num_threads(1)
if use_gpu:
num_gpu = torch.cuda.device_count()
if gpu_ids is None: # use all GPUs
device_id = job_id % num_gpu
else:
assert all([i >= 0 and i < num_gpu for i in gpu_ids])
device_id = gpu_ids[job_id % len(gpu_ids)]
torch.cuda.set_device(device_id)
device = torch.device(f'cuda:{device_id}')
else:
device = torch.device('cpu')
print(
f'@ Experiment: ID: {config["ID"]} ({len(configs)}), Seed: {seed}, Device: {device}, Job: {job_id} ({len(jobs)}), PID: {os.getpid()}'
)
print('#' * 50)
[print(f'# {key}: {value}') for key, value in config.items()]
print('#' * 50)
logdir = log_path / f'{config["ID"]}' / f'{seed}'
result = run(config, seed, device, logdir)
# Release all un-freed GPU memory
if use_gpu:
torch.cuda.empty_cache()
return result
if max_workers is None:
results = [_run(job) for job in jobs]
else:
with ProcessPoolExecutor(
max_workers=min(max_workers, len(jobs))) as executor:
results = list(
executor.map(CloudpickleWrapper(_run),
jobs,
chunksize=chunksize))
print(
color_str(
f'\nExperiment finished. Loggings are stored in {log_path.absolute()}. ',
'cyan',
bold=True))
return results
def evaluator(config, logdir, seed, make_env, learner_agent):
torch.set_num_threads(1) # VERY IMPORTANT TO AVOID GETTING STUCK
eval_logs = []
env = make_env(config, seed, 'train')
agent = Agent(config, env, torch.device('cpu'))
runner = EpisodeRunner(reset_on_call=True)
evaluated_steps = config['eval.freq']
while learner_agent.total_timestep < config['train.timestep']:
if learner_agent.total_timestep < evaluated_steps:
time.sleep(1.0)
else:
t0 = time.perf_counter()
agent.load_state_dict(
learner_agent.state_dict()) # copy to CPU by default
with torch.no_grad():
D = []
for _ in range(config['eval.num_episode']):
D += runner(agent, env, env.spec.max_episode_steps)
logger = Logger()
logger('num_seconds', round(time.perf_counter() - t0, 1))
logger('num_trajectories', len(D))
logger('num_timesteps', sum([len(traj) for traj in D]))
logger('accumulated_trained_timesteps',
learner_agent.total_timestep)
infos = [
info
for info in chain.from_iterable([traj.infos for traj in D])
if 'episode' in info
]
online_returns = [info['episode']['return'] for info in infos]
online_horizons = [info['episode']['horizon'] for info in infos]
logger(
'online_return',
describe(online_returns,
axis=-1,
repr_indent=1,
repr_prefix='\n'))
logger(
'online_horizon',
describe(online_horizons,
axis=-1,
repr_indent=1,
repr_prefix='\n'))
monitor_env = get_wrapper(env, 'VecMonitor')
logger(
'running_return',
describe(monitor_env.return_queue,
axis=-1,
repr_indent=1,
repr_prefix='\n'))
logger(
'running_horizon',
describe(monitor_env.horizon_queue,
axis=-1,
repr_indent=1,
repr_prefix='\n'))
logger.dump(keys=None,
index=0,
indent=0,
border=color_str('+' * 50, color='green'))
eval_logs.append(logger.logs)
evaluated_steps += config['eval.freq']
pickle_dump(obj=eval_logs, f=logdir / 'eval_logs', ext='.pkl')