largs = dict(learner_args)
largs['name'] = f"{learner_args['name']}_{name}"
learner = rl.PolicyGradient(env=env,
model=model,
optimizer=optim,
memory_updater=memory_updater,
crit=crit,
action_selector=sp.PolicyGradientActionSelection(),
callbacks=[],
**largs
)
return learner
root = 'tests/experiment'
experiment = Experiment(root=root)
params = experiment.get_params()
params['factory_args']['learner_args']['dump_path'] = root
params['factory_args']['core'] = Core(**params['core_args'])
with with_experiment(experiment=experiment, overwrite=params['overwrite']):
learner = Ensemble(model_class=Model,
trainer_factory=factory,
trainer_args=params['factory_args'],
n_model=params['n_learner'],
callbacks=[])
learner.fit(**params['fit'])
print('done training')
# raise NotImplemented
def run(root, path_script):
experiment = Experiment(root=root)
factory = experiment.get_factory()
params = experiment.get_params()
# callbacks_factory = experiment.get_factory(source_function='get_ens_callbacks')
params['factory_args']['learner_args']['dump_path'] = root
# params['factory_args']['dump_path'] = root
Model = experiment.get_model_class()
experiment.document_script(path_script, overwrite=params['overwrite'])
env = MultiInstanceGym(**params['env_args'])
params['factory_args']['model_args'][
'in_nodes'] = env.observation_space.shape[0]
params['factory_args']['model_args']['out_nodes'] = env.action_space.n
params['factory_args']['env'] = env
dqn_player = get_player(key=params['player_type'])
selection_strategy = get_selection_strategy(
params['selection_strategy'], params.get('selection_args', {}))
with with_experiment(experiment=experiment, overwrite=params['overwrite']):
memory = get_memory(params['memory_type'], params['memory_args'])
params['factory_args']['learner_args']['memory'] = memory
learner = DQNEnsemble(
model_class=Model,
trainer_factory=factory,
memory=memory,
env=env,
player=dqn_player,
selection_strategy=selection_strategy,
trainer_args=params['factory_args'],
n_model=params['n_learner'],
dump_path=root,
callbacks=[
rcb.EpisodeUpdater(**params.get('memory_update', {}),
frequency=5),
rcb.UncertaintyUpdater(hat=hat.EntropyHat()),
cb.Checkpointer(frequency=10),
rcb.EnvironmentEvaluator(
env=env,
n_evaluations=10,
action_selector=sp.GreedyValueSelection(
post_pipeline=[hat.EnsembleHat()]),
metrics={
'det_val_reward_mean': np.mean,
'deter_val_reward_std': np.std
},
frequency=10,
epoch_name='det_val_epoch'),
rcb.EnvironmentEvaluator(
env=env,
n_evaluations=10,
action_selector=get_selection_strategy(
params['eval_selection_strategy'],
params.get('selection_args', {})),
metrics={
'prob_val_reward_mean': np.mean,
'prob_val_reward_std': np.std
},
frequency=10,
epoch_name='prob_val_epoch'),
rcb.EnsembleRewardPlotter(frequency=10,
metrics={
'det_val_reward_mean':
'det_val_epoch',
'prob_val_reward_mean':
'prob_val_epoch',
})
])
learner.fit(**params['fit'])
def run(root, path_script):
print(root, path_script)
experiment = Experiment(root=root)
factory = experiment.get_factory()
params = experiment.get_params()
params['factory_args']['learner_args']['dump_path'] = root
Model = experiment.get_model_class()
experiment.document_script(path_script, overwrite=params['overwrite'])
# env = MultiInstanceGym(**params['factory_args']['env_args'])
env = TorchGym(params['factory_args']['env_args']['env_name'])
params['factory_args']['model_args'][
'in_nodes'] = env.observation_space.shape[0]
params['factory_args']['model_args']['out_nodes'] = env.action_space.n
dqn_player = DQNPlayer()
# selection_strategy = sp.AdaptiveQActionSelectionEntropy(warm_up=0,
# post_pipeline=[EnsembleHatStd()])
# selection_strategy = sp.QActionSelection(post_pipeline=[EnsembleHat()])
selection_strategy = sp.EpsilonGreedyActionSelection(
action_space=[0, 1, 2, 3], post_pipeline=[EnsembleHat()])
with with_experiment(experiment=experiment, overwrite=params['overwrite']):
memory = Memory(**params["factory_args"]['memory_args'])
params['factory_args']['learner_args']['memory'] = memory
learner = DQNEnsemble(
model_class=Model,
trainer_factory=factory,
memory=memory,
env=env,
player=dqn_player,
selection_strategy=selection_strategy,
trainer_args=params['factory_args'],
n_model=params['n_learner'],
callbacks=[
rcb.EpisodeUpdater(**params.get('memory_update', {})),
cb.Checkpointer(frequency=1),
# rcb.UncertaintyUpdater(),
rcb.EnvironmentEvaluator(
env=TorchGym(
params['factory_args']['env_args']['env_name']),
n_evaluations=10,
action_selector=sp.GreedyValueSelection(
post_pipeline=[EnsembleHat()]),
metrics={
'det_val_reward_mean': np.mean,
'deter_val_reward_std': np.std
},
frequency=1,
epoch_name='det_val_epoch'),
rcb.EnvironmentEvaluator(env=TorchGym(
params['factory_args']['env_args']['env_name']),
n_evaluations=10,
action_selector=selection_strategy,
metrics={
'prob_val_reward_mean': np.mean,
'prob_val_reward_std': np.std
},
frequency=1,
epoch_name='prob_val_epoch'),
rcb.EnsembleRewardPlotter(
metrics={
'det_val_reward_mean': 'det_val_epoch',
'prob_val_reward_mean': 'prob_val_epoch',
}),
])
# learner.load_checkpoint(path=f'{root}/checkpoint', tag='checkpoint')
learner.fit(**params['fit'])
from scipy.stats import entropy
import torch
from pymatch.ReinforcementLearning.memory import PriorityMemory
from pymatch.DeepLearning.hat import EnsembleHatStd, EnsembleHat
import pymatch.DeepLearning.callback as cb
import pymatch.ReinforcementLearning.callback as rcb
import pymatch.ReinforcementLearning.selection_policy as sp
from pymatch.ReinforcementLearning.torch_gym import TorchGym
from pymatch.utils.experiment import Experiment, with_experiment
from pymatch.utils.functional import interactive_python_mode, one_hot_encoding
from pymatch.DeepLearning.ensemble import DQNEnsemble
from pymatch.ReinforcementLearning.player import DQNPlayerCertainty
root = 'tests/test_experiment/exp_1'
experiment = Experiment(root=root)
factory = experiment.get_factory()
params = experiment.get_params()
params['factory_args']['learner_args']['dump_path'] = root
Model = experiment.get_model_class()
env = TorchGym(**params['factory_args']['env_args'])
params['factory_args']['model_args'][
'in_nodes'] = env.env.observation_space.shape[0]
params['factory_args']['model_args']['out_nodes'] = env.action_space.n
dqn_player = DQNPlayerCertainty()
selection_strategy = sp.AdaptiveQActionSelection(
temperature=params['temp'],
min_length=10,
post_pipeline=[EnsembleHatStd()])
import pymatch.ReinforcementLearning.callback as cb
from pymatch.ReinforcementLearning.selection_policy import GreedyValueSelection
from pymatch.ReinforcementLearning.torch_gym import TorchGym, CartPole
from pymatch.utils.experiment import Experiment
from pymatch.utils.functional import interactive_python_mode
from pymatch.DeepLearning.ensemble import Ensemble
if interactive_python_mode():
path_scipt = 'tests/boosting/train_boosting.py'
root = 'tests/boosting/exp'
else:
path_scipt = sys.argv[0]
root = sys.argv[1]
experiment = Experiment(root=root)
factory = experiment.get_factory()
params = experiment.get_params()
params['factory_args']['learner_args']['dump_path'] = root
Model = experiment.get_model_class()
core = experiment.get_model_class(source_file='core', source_class='Core')(**params['core_args'])
params['factory_args']['core'] = core
learner = Ensemble(model_class=Model,
trainer_factory=factory,
trainer_args=params['factory_args'],
n_model=params['n_learner'],
callbacks=[cb.EnvironmentEvaluator(
env=CartPole(**params['factory_args']['env_args']),
n_evaluations=10,