def test_iql():
logger.reset()
# make tests small by mutating variant
iql.variant["algo_kwargs"]["start_epoch"] = -2
iql.variant["algo_kwargs"]["num_epochs"] = 2
iql.variant["algo_kwargs"]["batch_size"] = 2
iql.variant["algo_kwargs"]["num_eval_steps_per_epoch"] = 2
iql.variant["algo_kwargs"]["num_expl_steps_per_train_loop"] = 2
iql.variant["algo_kwargs"]["num_trains_per_train_loop"] = 100
iql.variant["algo_kwargs"]["min_num_steps_before_training"] = 2
iql.variant["qf_kwargs"] = dict(hidden_sizes=[2, 2])
iql.variant["seed"] = 25580
iql.main()
reference_csv = "tests/regression/iql/halfcheetah_online_progress.csv"
output_csv = os.path.join(logger.get_snapshot_dir(), "progress.csv")
print("comparing reference %s against output %s" %
(reference_csv, output_csv))
output = csv_util.get_exp(output_csv)
reference = csv_util.get_exp(reference_csv)
keys = [
"epoch",
"expl/num steps total",
"expl/Average Returns",
"trainer/Q1 Predictions Mean",
]
csv_util.check_equal(reference, output, keys)
def test_sac_online():
logger.reset()
# make tests small by mutating variant
sac.variant["algorithm_kwargs"]["num_epochs"] = 2
sac.variant["algorithm_kwargs"]["batch_size"] = 2
sac.variant["algorithm_kwargs"]["num_trains_per_train_loop"] = 100
sac.variant["qf_kwargs"] = dict(hidden_sizes=[2, 2])
sac.variant["policy_kwargs"] = dict(hidden_sizes=[2, 2])
sac.variant["seed"] = 25580
sac.main()
reference_csv = "tests/regression/simplegym/test_sac_progress.csv"
output_csv = os.path.join(logger.get_snapshot_dir(), "progress.csv")
print("comparing reference %s against output %s" % (reference_csv, output_csv))
output = csv_util.get_exp(output_csv)
reference = csv_util.get_exp(reference_csv)
keys = ["epoch", "expl/num steps total", "eval/Average Returns", "trainer/Q1 Predictions Mean", ]
csv_util.check_equal(reference, output, keys)
def skewfit_full_experiment_chester(args):
variant, log_dir, exp_prefix = args
base_log_dir = log_dir
logger.reset()
print("log dir is: ", base_log_dir)
script_name = main.__file__
seed = variant['seed']
actual_log_dir = setup_logger(
exp_prefix=exp_prefix,
variant=variant,
exp_id=0,
seed=seed,
snapshot_mode='gap_and_last',
snapshot_gap=25,
log_dir=base_log_dir,
script_name=script_name,
)
variant = variant['variant']
if torch.cuda.is_available():
use_gpu = True
else:
use_gpu = False
set_gpu_mode(use_gpu)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
np.random.seed(seed)
variant['skewfit_variant']['save_vae_data'] = True
full_experiment_variant_preprocess(variant)
train_vae_and_update_variant(variant)
skewfit_experiment(variant['skewfit_variant'])
def reset_execution_environment():
"""
Call this between calls to separate experiments.
:return:
"""
logger.reset()
