def run_training(**kwargs):
from sipga.common.experiment_analysis import EnvironmentEvaluator
alg = kwargs.pop('alg')
env_id = kwargs.pop('env_id')
logger_kwargs = kwargs.pop('logger_kwargs')
log_dir = logger_kwargs['log_dir']
learn_fn = utils.get_learn_function(alg)
evaluator = EnvironmentEvaluator(log_dir=log_dir)
ac, env = learn_fn(env_id, logger_kwargs=logger_kwargs, **kwargs)
evaluator.eval(env=env, ac=ac, num_evaluations=32)
def check_alg(alg_name, env_id):
"""" Run one epoch update with algorithm."""
print(f'Run {alg_name}.')
defaults = U.get_defaults_kwargs(alg=alg_name, env_id=env_id)
defaults['epochs'] = 1
defaults['num_mini_batches'] = 4
defaults['steps_per_epoch'] = 1000
defaults['verbose'] = False
learn_fn = U.get_learn_function(alg_name)
defaults['logger_kwargs'] = setup_logger_kwargs(exp_name='unittest',
seed=None,
base_dir='/var/tmp/',
datestamp=True,
use_tensor_board=True,
verbose=False)
return learn_fn(env_id, **defaults)
def fit(self, epochs=None, env=None):
""" Train the model for a given number of epochs.
Parameters
----------
epochs: int
Number of epoch to train. If None, use the standard setting from the
defaults.py of the corresponding algorithm.
env: gym.Env
provide a custom environment for fitting the model, e.g. pass a
virtual environment (based on NN approximation)
Returns
-------
None
"""
assert self.compiled, 'Call model.compile() before model.fit()'
if self.multi_thread:
# start all tasks which are stored in the scheduler
self.scheduler.run()
else:
# single thread training
if epochs is None:
epochs = self.kwargs.pop('epochs')
else:
self.kwargs.pop('epochs') # pop to avoid double kwargs
# fit() can also take a custom env, e.g. a virtual environment
env_id = self.env_id if env is None else env
learn_func = utils.get_learn_function(self.alg)
ac, env = learn_func(
env_id=env_id,
logger_kwargs=self.logger_kwargs,
epochs=epochs,
**self.kwargs
)
self.model = ac
self.env = env
self.trained = True