def test_assemble_curriculum_config():
file_contents = """
behavior1:
curriculum:
foo: 5
behavior2:
curriculum:
foo: 6
"""
trainer_config = _load_config(file_contents)
curriculum_config = assemble_curriculum_config(trainer_config)
assert curriculum_config == {
"behavior1": {
"foo": 5
},
"behavior2": {
"foo": 6
}
}
# Check that nothing is returned if no curriculum.
file_contents = """
behavior1:
foo: 3
behavior2:
foo: 4
"""
trainer_config = _load_config(file_contents)
curriculum_config = assemble_curriculum_config(trainer_config)
assert curriculum_config == {}
# Check that method doesn't break if 1st level entity isn't a dict.
# Note: this is a malformed configuration.
file_contents = """
behavior1: 3
behavior2: 4
"""
trainer_config = _load_config(file_contents)
curriculum_config = assemble_curriculum_config(trainer_config)
assert curriculum_config == {}
def run_training(run_seed: int, options: RunOptions) -> None:
"""
Launches training session.
:param options: parsed command line arguments
:param run_seed: Random seed used for training.
:param run_options: Command line arguments for training.
"""
with hierarchical_timer("run_training.setup"):
base_path = "results"
write_path = os.path.join(base_path, options.run_id)
maybe_init_path = (
os.path.join(base_path, options.run_id) if options.initialize_from else None
)
run_logs_dir = os.path.join(write_path, "run_logs")
port = options.base_port
# Check if directory exists
handle_existing_directories(
write_path, options.resume, options.force, maybe_init_path
)
# Make run logs directory
os.makedirs(run_logs_dir, exist_ok=True)
# Configure CSV, Tensorboard Writers and StatsReporter
# We assume reward and episode length are needed in the CSV.
csv_writer = CSVWriter(
write_path,
required_fields=[
"Environment/Cumulative Reward",
"Environment/Episode Length",
],
)
tb_writer = TensorboardWriter(write_path, clear_past_data=not options.resume)
gauge_write = GaugeWriter()
console_writer = ConsoleWriter()
StatsReporter.add_writer(tb_writer)
StatsReporter.add_writer(csv_writer)
StatsReporter.add_writer(gauge_write)
StatsReporter.add_writer(console_writer)
if options.env_path is None:
port = UnityEnvironment.DEFAULT_EDITOR_PORT
env_factory = create_environment_factory(
options.env_path,
options.no_graphics,
run_seed,
port,
options.env_args,
os.path.abspath(run_logs_dir), # Unity environment requires absolute path
)
engine_config = EngineConfig(
width=options.width,
height=options.height,
quality_level=options.quality_level,
time_scale=options.time_scale,
target_frame_rate=options.target_frame_rate,
capture_frame_rate=options.capture_frame_rate,
)
env_manager = SubprocessEnvManager(env_factory, engine_config, options.num_envs)
curriculum_config = assemble_curriculum_config(options.behaviors)
maybe_meta_curriculum = try_create_meta_curriculum(
curriculum_config, env_manager, options.lesson
)
sampler_manager, resampling_interval = create_sampler_manager(
options.parameter_randomization, run_seed
)
trainer_factory = TrainerFactory(
options.behaviors,
options.run_id,
write_path,
options.keep_checkpoints,
not options.inference,
options.resume,
run_seed,
maybe_init_path,
maybe_meta_curriculum,
options.multi_gpu,
)
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
write_path,
options.run_id,
options.save_freq,
maybe_meta_curriculum,
not options.inference,
run_seed,
sampler_manager,
resampling_interval,
)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
write_run_options(write_path, options)
write_timing_tree(run_logs_dir)