def main():
args = get_args()
init_logging(args.log_level)
get_logger().info("Running with args {}".format(args))
ptitle("Master: {}".format(
"Training" if args.test_date is None else "Testing"))
cfg, srcs = load_config(args)
if args.test_date is None:
OnPolicyRunner(
config=cfg,
output_dir=args.output_dir,
loaded_config_src_files=srcs,
seed=args.seed,
mode="train",
deterministic_cudnn=args.deterministic_cudnn,
deterministic_agents=args.deterministic_agents,
extra_tag=args.extra_tag,
disable_tensorboard=args.disable_tensorboard,
disable_config_saving=args.disable_config_saving,
).start_train(
checkpoint=args.checkpoint,
restart_pipeline=args.restart_pipeline,
max_sampler_processes_per_worker=args.
max_sampler_processes_per_worker,
)
else:
OnPolicyRunner(
config=cfg,
output_dir=args.output_dir,
loaded_config_src_files=srcs,
seed=args.seed,
mode="test",
deterministic_cudnn=args.deterministic_cudnn,
deterministic_agents=args.deterministic_agents,
extra_tag=args.extra_tag,
disable_tensorboard=args.disable_tensorboard,
disable_config_saving=args.disable_config_saving,
).start_test(
experiment_date=args.test_date,
checkpoint=args.checkpoint,
skip_checkpoints=args.skip_checkpoints,
max_sampler_processes_per_worker=args.
max_sampler_processes_per_worker,
)
def test_ppo_trains(self, tmpdir):
cfg = PPOBabyAIGoToObjExperimentConfig()
output_dir = tmpdir.mkdir("experiment_output")
train_runner = OnPolicyRunner(
config=cfg,
output_dir=output_dir,
loaded_config_src_files=None,
seed=1,
mode="train",
deterministic_cudnn=True,
)
start_time_str = train_runner.start_train(
max_sampler_processes_per_worker=1)
test_runner = OnPolicyRunner(
config=cfg,
output_dir=output_dir,
loaded_config_src_files=None,
seed=1,
mode="test",
deterministic_cudnn=True,
)
test_results = test_runner.start_test(
experiment_date=start_time_str,
skip_checkpoints=1,
max_sampler_processes_per_worker=1,
)
assert (len(test_results) == 1
), f"Too many or too few test results ({test_results})"
tr = test_results[0]
assert (tr["training_steps"] == round(
math.ceil(cfg.TOTAL_RL_TRAIN_STEPS /
(cfg.ROLLOUT_STEPS * cfg.NUM_TRAIN_SAMPLERS))) *
cfg.ROLLOUT_STEPS *
cfg.NUM_TRAIN_SAMPLERS), "Incorrect number of training steps"
assert len(tr["tasks"]
) == cfg.NUM_TEST_TASKS, "Incorrect number of test tasks"
assert tr["success"] == sum(
task["success"] for task in tr["tasks"]) / len(
tr["tasks"]), "Success counts don't seem to match"
assert (
tr["success"] > 0.95
), "PPO did not seem to converge for the go_to_obj task (success {}).".format(
tr["success"])