def test_session(test_spec):
spec_util.tick(test_spec, 'trial')
spec_util.tick(test_spec, 'session')
spec_util.save(test_spec, unit='trial')
session = Session(test_spec)
session_metrics = session.run()
assert isinstance(session_metrics, dict)
def test_experiment():
spec = spec_util.get('demo.json', 'dqn_cartpole')
spec_util.save(spec, unit='experiment')
spec = spec_util.override_spec(spec, 'test')
spec_util.tick(spec, 'experiment')
experiment_df = Experiment(spec).run()
assert isinstance(experiment_df, pd.DataFrame)
def test_trial_demo():
spec = spec_util.get('demo.json', 'dqn_cartpole')
spec_util.save(spec, unit='experiment')
spec = spec_util.override_spec(spec, 'test')
spec_util.tick(spec, 'trial')
trial_metrics = Trial(spec).run()
assert isinstance(trial_metrics, dict)
def train(self, specfile, specname):
"enjoy + optimizes agent + periodic eval"
logger.info(
f'Running lab mode:train with specfile:{specfile} specname:{specname}'
)
spec = spec_util.get(specfile, specname)
# FIXME Why does this need to be in env?
os.environ['lab_mode'] = 'train'
spec_util.save(spec) # first save the new spec
spec_util.tick(spec, 'trial')
Trial(spec).run()
def __init__(self, spec, global_nets=None):
self.spec = spec
self.index = self.spec['meta']['session']
util.set_random_seed(self.spec)
util.set_cuda_id(self.spec)
util.set_logger(self.spec, logger, 'session')
spec_util.save(spec, unit='session')
self.agent, self.env = make_agent_env(self.spec, global_nets)
with util.ctx_lab_mode('eval'): # env for eval
self.eval_env = make_env(self.spec)
logger.info(util.self_desc(self))
def test_demo_performance():
spec = spec_util.get('demo.json', 'dqn_cartpole')
spec_util.save(spec, unit='experiment')
for env_spec in spec['env']:
env_spec['max_frame'] = 2000
spec_util.tick(spec, 'trial')
trial = Trial(spec)
spec_util.tick(spec, 'session')
session = Session(spec)
session.run()
last_reward = session.agent.body.train_df.iloc[-1]['total_reward']
assert last_reward > 50, f'last_reward is too low: {last_reward}'
def dev(self, specfile, specname):
"train + limit the number of trials & sessions. Useful for iterative development."
logger.info(
f'Running lab mode:dev with specfile:{specfile} specname:{specname}'
)
spec = spec_util.get(specfile, specname)
# FIXME Why does this need to be in env?
os.environ['lab_mode'] = 'dev'
spec_util.save(spec) # first save the new spec
# spec = spec_util.override_dev_spec(spec)
spec['meta']['max_session'] = 1
spec['meta']['max_trial'] = 2
spec_util.tick(spec, 'trial')
Trial(spec).run()
def run_spec(spec, lab_mode):
'''Run a spec in lab_mode'''
os.environ['lab_mode'] = lab_mode # set lab_mode
spec = spec_util.override_spec(spec, lab_mode) # conditionally override spec
if lab_mode in TRAIN_MODES:
spec_util.save(spec) # first save the new spec
if lab_mode == 'search':
spec_util.tick(spec, 'experiment')
Experiment(spec).run()
else:
spec_util.tick(spec, 'trial')
Trial(spec).run()
elif lab_mode in EVAL_MODES:
Session(spec).run()
else:
raise ValueError(f'Unrecognizable lab_mode not of {TRAIN_MODES} or {EVAL_MODES}')
def __init__(self, spec):
self.spec = spec
self.index = self.spec['meta']['experiment']
util.set_logger(self.spec, logger, 'trial')
spec_util.save(spec, unit='experiment')
def test_trial(test_spec):
spec_util.tick(test_spec, 'trial')
spec_util.save(test_spec, unit='trial')
trial = Trial(test_spec)
trial_metrics = trial.run()
assert isinstance(trial_metrics, dict)
