def run_trial_test(spec_file, spec_name=False):
spec = spec_util.get(spec_file, spec_name)
spec = util.override_test_spec(spec)
info_space = InfoSpace()
info_space.tick('trial')
trial = Trial(spec, info_space)
trial_data = trial.run()
assert isinstance(trial_data, pd.DataFrame)
def run_benchmark(spec, const):
benchmark_specs = benchmarker.generate_specs(spec, const)
logger.info('Running benchmark')
for spec_name, benchmark_spec in benchmark_specs.items():
# run only if not already exist; benchmark mode only
if not any(spec_name in filename for filename in os.listdir('data')):
info_space = InfoSpace()
info_space.tick('experiment')
Experiment(benchmark_spec, info_space).run()
def run_trial_test(spec_file, spec_name=False, distributed=False):
spec = spec_util.get(spec_file, spec_name)
spec = util.override_test_spec(spec)
info_space = InfoSpace()
info_space.tick('trial')
if distributed:
spec['meta']['distributed'] = True
if os.environ.get('CI') != 'true': # CI has not enough CPU
spec['meta']['max_session'] = 2
trial = Trial(spec, info_space)
trial_data = trial.run()
assert isinstance(trial_data, pd.DataFrame)
def prepath_to_info_space(prepath):
'''Create info_space from prepath such that it returns the same prepath with spec'''
from slm_lab.experiment.monitor import InfoSpace
_, _, _, _, experiment_ts, ckpt = prepath_split(prepath)
trial_index, session_index = prepath_to_idxs(prepath)
# create info_space for prepath
info_space = InfoSpace()
info_space.experiment_ts = experiment_ts
info_space.ckpt = ckpt
info_space.set('experiment', 0)
info_space.set('trial', trial_index)
info_space.set('session', session_index)
return info_space
def plot_session_from_file(session_df_filepath):
'''
Method to plot session from its session_df file
@example
from slm_lab.experiment import analysis
filepath = 'data/reinforce_cartpole_2018_01_22_211751/reinforce_cartpole_t0_s0_session_df.csv'
analysis.plot_session_from_file(filepath)
'''
from slm_lab.experiment.monitor import InfoSpace
spec_name = util.prepath_to_spec_name(session_df_filepath)
session_spec = {'name': spec_name}
session_df = util.read(session_df_filepath,
header=[0, 1, 2, 3],
index_col=0,
dtype=np.float32)
session_data = util.session_df_to_data(session_df)
tn, sn = session_df_filepath.replace('_session_df.csv', '').split('_')[-2:]
info_space = InfoSpace()
info_space.set('experiment', 0)
info_space.set('trial', int(tn[1:]))
info_space.set('session', int(sn[1:]))
session_fig = plot_session(session_spec, info_space, session_data)
viz.save_image(
session_fig,
session_df_filepath.replace('_session_df.csv', '_session_graph.png'))
def run_by_mode(spec_file, spec_name, lab_mode):
logger.info(f'Running lab in mode: {lab_mode}')
spec = spec_util.get(spec_file, spec_name)
info_space = InfoSpace()
analysis.save_spec(spec, info_space, unit='experiment')
# '@' is reserved for 'enjoy@{prepath}'
os.environ['lab_mode'] = lab_mode.split('@')[0]
os.environ['PREPATH'] = util.get_prepath(spec, info_space)
reload(logger) # to set PREPATH properly
if lab_mode == 'search':
info_space.tick('experiment')
Experiment(spec, info_space).run()
elif lab_mode.startswith('train'):
if '@' in lab_mode:
prepath = lab_mode.split('@')[1]
spec, info_space = util.prepath_to_spec_info_space(prepath)
else:
info_space.tick('trial')
Trial(spec, info_space).run()
elif lab_mode.startswith('enjoy'):
prepath = lab_mode.split('@')[1]
spec, info_space = util.prepath_to_spec_info_space(prepath)
Session(spec, info_space).run()
elif lab_mode.startswith('enjoy'):
prepath = lab_mode.split('@')[1]
spec, info_space = util.prepath_to_spec_info_space(prepath)
Session(spec, info_space).run()
elif lab_mode == 'dev':
spec = util.override_dev_spec(spec)
info_space.tick('trial')
Trial(spec, info_space).run()
else:
logger.warn('lab_mode not recognized; must be one of `search, train, enjoy, benchmark, dev`.')
def run_by_mode(spec_file, spec_name, lab_mode):
logger.info(f'Running lab in mode: {lab_mode}')
spec = spec_util.get(spec_file, spec_name)
info_space = InfoSpace()
os.environ['PREPATH'] = util.get_prepath(spec, info_space)
reload(logger) # to set PREPATH properly
# expose to runtime, '@' is reserved for 'enjoy@{prepath}'
os.environ['lab_mode'] = lab_mode.split('@')[0]
if lab_mode == 'search':
info_space.tick('experiment')
Experiment(spec, info_space).run()
elif lab_mode == 'train':
info_space.tick('trial')
Trial(spec, info_space).run()
elif lab_mode.startswith('enjoy'):
prepath = lab_mode.split('@')[1]
spec, info_space = util.prepath_to_spec_info_space(prepath)
Session(spec, info_space).run()
elif lab_mode == 'generate_benchmark':
benchmarker.generate_specs(spec, const='agent')
elif lab_mode == 'benchmark':
# TODO allow changing const to env
run_benchmark(spec, const='agent')
elif lab_mode == 'dev':
spec = util.override_dev_spec(spec)
info_space.tick('trial')
Trial(spec, info_space).run()
else:
logger.warn(
'lab_mode not recognized; must be one of `search, train, enjoy, benchmark, dev`.'
)
def __init__(self, spec, info_space=InfoSpace()):
self.spec = spec
if info_space.get('trial') is None:
info_space.tick('trial')
self.info_space = info_space
self.coor, self.index = self.info_space.get_coor_idx(self)
self.session_data_dict = {}
self.data = None
def test_aeb_space(test_spec):
global aeb_space
if aeb_space is None:
aeb_space = AEBSpace(test_spec, InfoSpace())
env_space = EnvSpace(test_spec, aeb_space)
agent_space = AgentSpace(test_spec, aeb_space)
aeb_space.init_body_space()
return aeb_space
def __init__(self, spec, info_space=None):
info_space = info_space or InfoSpace()
init_thread_vars(spec, info_space, unit='trial')
self.spec = spec
self.info_space = info_space
self.coor, self.index = self.info_space.get_coor_idx(self)
self.session_data_dict = {}
self.data = None
analysis.save_spec(spec, info_space, unit='trial')
logger.info(f'Initialized trial {self.index}')
def mock_spec_info_space(predir, trial_index=None, session_index=None):
'''Helper for retro analysis to build mock info_space and spec'''
from slm_lab.experiment.monitor import InfoSpace
_, _, _, spec_name, experiment_ts, _ = util.prepath_split(predir)
info_space = InfoSpace()
info_space.experiment_ts = experiment_ts
info_space.set('experiment', 0)
if trial_index is None:
filepath = f'{predir}/{spec_name}_spec.json'
else:
info_space.set('trial', trial_index)
filepath = f'{predir}/{spec_name}_t{trial_index}_spec.json'
if session_index is not None:
info_space.set('session', session_index)
spec = util.read(filepath)
return spec, info_space
def test_prioritized_replay_memory(request):
memspec = spec_util.get('base.json', 'base_prioritized_replay_memory')
memspec = util.override_test_spec(memspec)
aeb_mem_space = AEBSpace(memspec, InfoSpace())
env_space = EnvSpace(memspec, aeb_mem_space)
aeb_mem_space.init_body_space()
agent_space = AgentSpace(memspec, aeb_mem_space)
agent = agent_space.agents[0]
body = agent.nanflat_body_a[0]
res = (body.memory, ) + request.param
return res
def __init__(self, spec, info_space=InfoSpace()):
self.spec = spec
spec['meta']['train_mode'] = True
if info_space.get('experiment') is None:
info_space.tick('experiment')
self.info_space = info_space
self.coor, self.index = self.info_space.get_coor_idx(self)
self.trial_data_dict = {}
self.data = None
SearchClass = getattr(search, 'RaySearch')
self.search = SearchClass(self)
def run_trial_test_dist(spec_file, spec_name=False):
spec = spec_util.get(spec_file, spec_name)
spec = spec_util.override_test_spec(spec)
info_space = InfoSpace()
info_space.tick('trial')
spec['meta']['distributed'] = True
spec['meta']['max_session'] = 2
trial = Trial(spec, info_space)
# manually run the logic to obtain global nets for testing to ensure global net gets updated
global_nets = trial.init_global_nets()
# only test first network
if ps.is_list(global_nets): # multiagent only test first
net = list(global_nets[0].values())[0]
else:
net = list(global_nets.values())[0]
session_datas = trial.parallelize_sessions(global_nets)
trial.session_data_dict = {data.index[0]: data for data in session_datas}
trial_data = analysis.analyze_trial(trial)
trial.close()
assert isinstance(trial_data, pd.DataFrame)
def __init__(self, spec, info_space=None):
info_space = info_space or InfoSpace()
init_thread_vars(spec, info_space, unit='experiment')
self.spec = spec
self.info_space = info_space
self.coor, self.index = self.info_space.get_coor_idx(self)
self.trial_data_dict = {}
self.data = None
SearchClass = getattr(search, spec['meta'].get('search'))
self.search = SearchClass(self)
analysis.save_spec(spec, info_space, unit='experiment')
logger.info(f'Initialized experiment {self.index}')
def mock_info_space_spec(predir, trial_index=None, session_index=None):
'''Helper for retro analysis to build mock info_space and spec'''
from slm_lab.experiment.monitor import InfoSpace
spec_name = util.prepath_to_spec_name(predir)
experiment_ts = util.prepath_to_experiment_ts(predir)
info_space = InfoSpace()
info_space.experiment_ts = experiment_ts
info_space.set('experiment', 0)
if trial_index is None:
filepath = f'{predir}/{spec_name}_spec.json'
else:
info_space.set('trial', trial_index)
filepath = f'{predir}/{spec_name}_t{trial_index}_spec.json'
if session_index is not None:
info_space.set('session', session_index)
spec = util.read(filepath)
return spec, info_space
def __init__(self, spec, info_space=InfoSpace()):
self.spec = spec
if info_space.get('session') is None:
info_space.tick('session')
self.info_space = info_space
self.coor, self.index = self.info_space.get_coor_idx(self)
# TODO option to set rand_seed. also set np random seed
self.torch_rand_seed = torch.initial_seed()
self.data = None
self.aeb_space = AEBSpace(self.spec, self.info_space)
self.env_space = EnvSpace(self.spec, self.aeb_space)
self.agent_space = AgentSpace(self.spec, self.aeb_space)
logger.info(util.self_desc(self))
self.aeb_space.init_body_space()
self.aeb_space.post_body_init()
def generic_algo_test(spec, algo_name):
'''Need new InfoSpace() per trial otherwise session id doesn't tick correctly'''
trial = Trial(spec, info_space=InfoSpace())
trial_data = trial.run()
folders = [x for x in os.listdir('data/') if x.startswith(algo_name)]
assert len(folders) == 1
path = 'data/' + folders[0]
sess_data = util.read(path + '/' + algo_name + '_t0_s0_session_df.csv')
rewards = sess_data['0.2'].replace("reward", -1).astype(float)
print(f'rewards: {rewards}')
maxr = rewards.max()
'''Delete test data folder and trial'''
shutil.rmtree(path)
del trial
return maxr
def run_new_mode(spec_file, spec_name, lab_mode):
'''Run to generate new data with `search, train, dev`'''
spec = spec_util.get(spec_file, spec_name)
info_space = InfoSpace()
analysis.save_spec(spec, info_space, unit='experiment') # first save the new spec
if lab_mode == 'search':
info_space.tick('experiment')
Experiment(spec, info_space).run()
elif lab_mode.startswith('train'):
info_space.tick('trial')
Trial(spec, info_space).run()
elif lab_mode == 'dev':
spec = spec_util.override_dev_spec(spec)
info_space.tick('trial')
Trial(spec, info_space).run()
else:
raise ValueError(f'Unrecognizable lab_mode not of {TRAIN_MODES}')
def __init__(self, spec, info_space=None):
info_space = info_space or InfoSpace()
init_thread_vars(spec, info_space, unit='session')
self.spec = spec
self.info_space = info_space
self.coor, self.index = self.info_space.get_coor_idx(self)
# TODO option to set rand_seed. also set np random seed
self.torch_rand_seed = torch.initial_seed()
self.data = None
self.aeb_space = AEBSpace(self.spec, self.info_space)
self.env_space = EnvSpace(self.spec, self.aeb_space)
self.agent_space = AgentSpace(self.spec, self.aeb_space)
logger.info(util.self_desc(self))
self.aeb_space.init_body_space()
self.aeb_space.post_body_init()
logger.info(f'Initialized session {self.index}')
def __init__(self, spec, info_space=None):
info_space = info_space or InfoSpace()
init_thread_vars(spec, info_space, unit='session')
self.spec = deepcopy(spec)
self.info_space = info_space
self.coor, self.index = self.info_space.get_coor_idx(self)
self.random_seed = 100 * (info_space.get('trial') or 0) + self.index
torch.cuda.manual_seed_all(self.random_seed)
torch.manual_seed(self.random_seed)
np.random.seed(self.random_seed)
self.data = None
self.aeb_space = AEBSpace(self.spec, self.info_space)
self.env_space = EnvSpace(self.spec, self.aeb_space)
self.agent_space = AgentSpace(self.spec, self.aeb_space)
logger.info(util.self_desc(self))
self.aeb_space.init_body_space()
self.aeb_space.post_body_init()
logger.info(f'Initialized session {self.index}')
def plot_session_from_file(session_df_filepath):
'''
Method to plot session from its session_df file
@example
from slm_lab.experiment import analysis
filepath = 'data/reinforce_cartpole_2018_01_22_211751/reinforce_cartpole_t0_s0_session_df.csv'
analysis.plot_session_from_file(filepath)
'''
from slm_lab.experiment.monitor import InfoSpace
spec_name = util.prepath_to_spec_name(session_df_filepath)
session_spec = {'name': spec_name}
session_df = util.read(session_df_filepath, header=[0, 1, 2, 3], index_col=0, dtype=np.float32)
session_data = util.session_df_to_data(session_df)
tn, sn = session_df_filepath.replace('_session_df.csv', '').split('_')[-2:]
info_space = InfoSpace()
info_space.set('experiment', 0)
info_space.set('trial', int(tn[1:]))
info_space.set('session', int(sn[1:]))
session_fig = plot_session(session_spec, info_space, session_data)
viz.save_image(session_fig, session_df_filepath.replace('_session_df.csv', '_session_graph.png'))
import os # NOTE increase if needed. Pytorch thread overusage https://github.com/pytorch/pytorch/issues/975 os.environ['OMP_NUM_THREADS'] = '1' from slm_lab.agent import Agent from slm_lab.env import OpenAIEnv from slm_lab.experiment import analysis from slm_lab.experiment.monitor import Body, InfoSpace, enable_aeb_space from slm_lab.lib import logger, util from slm_lab.spec import spec_util from irl_benchmark.rl.slm_session import Session spec = spec_util.get(spec_file='ppo.json', spec_name='ppo_mlp_shared_pendulum') info_space = InfoSpace() os.environ['PREPATH'] = util.get_prepath(spec, info_space) os.environ['lab_mode'] = 'training' spec['env'][0]['max_episode'] = 10 session = Session(spec, info_space) data, agent = session.run() print(data)
def test_info_space():
return InfoSpace()
