def test_console_writer(self):
# Test write_stats
with self.assertLogs("mlagents.trainers", level="INFO") as cm:
category = "category1"
console_writer = ConsoleWriter()
statssummary1 = StatsSummary(mean=1.0, std=1.0, num=1)
console_writer.write_stats(
category,
{
"Environment/Cumulative Reward": statssummary1,
"Is Training": statssummary1,
},
10,
)
statssummary2 = StatsSummary(mean=0.0, std=0.0, num=1)
console_writer.write_stats(
category,
{
"Environment/Cumulative Reward": statssummary1,
"Is Training": statssummary2,
},
10,
)
# Test hyperparameter writing
console_writer.add_property(
"category1", StatsPropertyType.HYPERPARAMETERS, {"example": 1.0}
)
self.assertIn(
"Mean Reward: 1.000. Std of Reward: 1.000. Training.", cm.output[0]
)
self.assertIn("Not Training.", cm.output[1])
self.assertIn("Hyperparameters for behavior name", cm.output[2])
self.assertIn("example:\t1.0", cm.output[2])
def test_selfplay_console_writer(self):
with self.assertLogs("mlagents.trainers", level="INFO") as cm:
category = "category1"
console_writer = ConsoleWriter()
console_writer.add_property(category, StatsPropertyType.SELF_PLAY,
True)
console_writer.add_property(category,
StatsPropertyType.SELF_PLAY_TEAM, 1)
statssummary1 = StatsSummary(mean=1.0, std=1.0, num=1)
console_writer.write_stats(
category,
{
"Environment/Cumulative Reward": statssummary1,
"Is Training": statssummary1,
"Self-play/ELO": statssummary1,
"Self-play/Mean Opponent ELO": statssummary1,
"Self-play/Std Opponent ELO": statssummary1,
},
10,
)
self.assertIn("Mean Reward: 1.000. Std of Reward: 1.000. Training.",
cm.output[0])
self.assertIn(
"category1 Team 1: ELO: 1.000. Mean Opponent ELO: 1.000. Std Opponent ELO: 1.000.",
cm.output[1],
)
def test_selfplay_console_writer(self):
with self.assertLogs("mlagents.trainers", level="INFO") as cm:
category = "category1"
console_writer = ConsoleWriter()
console_writer.add_property(category, StatsPropertyType.SELF_PLAY,
True)
statssummary1 = StatsSummary(
mean=1.0,
std=1.0,
num=1,
sum=1.0,
aggregation_method=StatsAggregationMethod.AVERAGE,
)
console_writer.write_stats(
category,
{
"Environment/Cumulative Reward": statssummary1,
"Is Training": statssummary1,
"Self-play/ELO": statssummary1,
},
10,
)
self.assertIn("Mean Reward: 1.000. Std of Reward: 1.000. Training.",
cm.output[0])
def get_default_stats_writers(run_options: RunOptions) -> List[StatsWriter]:
"""
The StatsWriters that mlagents-learn always uses:
* A TensorboardWriter to write information to TensorBoard
* A GaugeWriter to record our internal stats
* A ConsoleWriter to output to stdout.
"""
checkpoint_settings = run_options.checkpoint_settings
return [
TensorboardWriter(
checkpoint_settings.write_path,
clear_past_data=not checkpoint_settings.resume,
),
GaugeWriter(),
ConsoleWriter(),
]
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"):
model_path = f"./models/{options.run_id}"
maybe_init_path = (
f"./models/{options.initialize_from}" if options.initialize_from else None
)
summaries_dir = "./summaries"
port = options.base_port
# Configure CSV, Tensorboard Writers and StatsReporter
# We assume reward and episode length are needed in the CSV.
csv_writer = CSVWriter(
summaries_dir,
required_fields=[
"Environment/Cumulative Reward",
"Environment/Episode Length",
],
)
handle_existing_directories(
model_path, summaries_dir, options.resume, options.force, maybe_init_path
)
tb_writer = TensorboardWriter(summaries_dir, 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
)
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)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum_config, env_manager, options.lesson
)
sampler_manager, resampling_interval = create_sampler_manager(
options.sampler_config, run_seed
)
trainer_factory = TrainerFactory(
options.trainer_config,
summaries_dir,
options.run_id,
model_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,
model_path,
summaries_dir,
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_timing_tree(summaries_dir, options.run_id)
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.
"""
options.checkpoint_settings.run_id = "test8"
with hierarchical_timer("run_training.setup"):
checkpoint_settings = options.checkpoint_settings
env_settings = options.env_settings
engine_settings = options.engine_settings
base_path = "results"
write_path = os.path.join(base_path, checkpoint_settings.run_id)
maybe_init_path = (os.path.join(base_path,
checkpoint_settings.initialize_from)
if checkpoint_settings.initialize_from else None)
run_logs_dir = os.path.join(write_path, "run_logs")
port: Optional[int] = env_settings.base_port
# Check if directory exists
handle_existing_directories(
write_path,
checkpoint_settings.resume,
checkpoint_settings.force,
maybe_init_path,
)
# Make run logs directory
os.makedirs(run_logs_dir, exist_ok=True)
# Load any needed states
if checkpoint_settings.resume:
GlobalTrainingStatus.load_state(
os.path.join(run_logs_dir, "training_status.json"))
# 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 checkpoint_settings.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)
engine_config = EngineConfig(
width=engine_settings.width,
height=engine_settings.height,
quality_level=engine_settings.quality_level,
time_scale=engine_settings.time_scale,
target_frame_rate=engine_settings.target_frame_rate,
capture_frame_rate=engine_settings.capture_frame_rate,
)
if env_settings.env_path is None:
port = None
# Begin training
env_settings.env_path = "C:/Users/Sebastian/Desktop/RLUnity/Training/mFindTarget_new/RLProject.exe"
env_factory = create_environment_factory(
env_settings.env_path,
engine_settings.no_graphics,
run_seed,
port,
env_settings.env_args,
os.path.abspath(
run_logs_dir), # Unity environment requires absolute path
)
env_manager = SubprocessEnvManager(env_factory, engine_config,
env_settings.num_envs)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum, env_manager, restore=checkpoint_settings.resume)
sampler_manager, resampling_interval = create_sampler_manager(
options.parameter_randomization, run_seed)
max_steps = options.behaviors['Brain'].max_steps
options.behaviors['Brain'].max_steps = 10
trainer_factory = TrainerFactory(options,
write_path,
not checkpoint_settings.inference,
checkpoint_settings.resume,
run_seed,
maybe_init_path,
maybe_meta_curriculum,
False,
total_steps=0)
trainer_factory.trainer_config[
'Brain'].hyperparameters.learning_rate_schedule = ScheduleType.CONSTANT
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
write_path,
checkpoint_settings.run_id,
maybe_meta_curriculum,
not checkpoint_settings.inference,
run_seed,
sampler_manager,
resampling_interval,
)
try:
# Get inital weights
tc.init_weights(env_manager)
inital_weights = deepcopy(tc.weights)
finally:
env_manager.close()
write_run_options(write_path, options)
write_timing_tree(run_logs_dir)
write_training_status(run_logs_dir)
options.behaviors['Brain'].max_steps = max_steps
step = 0
counter = 0
max_meta_updates = 200
while counter < max_meta_updates:
sample = np.random.random_sample()
if (sample > 1):
print("Performing Meta-learning on Carry Object stage")
env_settings.env_path = "C:/Users/Sebastian/Desktop/RLUnity/Training/mCarryObject_new/RLProject.exe"
else:
print("Performing Meta-learning on Find Target stage")
env_settings.env_path = "C:/Users/Sebastian/Desktop/RLUnity/Training/mFindTarget_new/RLProject.exe"
env_factory = create_environment_factory(
env_settings.env_path,
engine_settings.no_graphics,
run_seed,
port,
env_settings.env_args,
os.path.abspath(
run_logs_dir), # Unity environment requires absolute path
)
env_manager = SubprocessEnvManager(env_factory, engine_config,
env_settings.num_envs)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum,
env_manager,
restore=checkpoint_settings.resume)
sampler_manager, resampling_interval = create_sampler_manager(
options.parameter_randomization, run_seed)
trainer_factory = TrainerFactory(options,
write_path,
not checkpoint_settings.inference,
checkpoint_settings.resume,
run_seed,
maybe_init_path,
maybe_meta_curriculum,
False,
total_steps=step)
trainer_factory.trainer_config[
'Brain'].hyperparameters.learning_rate_schedule = ScheduleType.CONSTANT
trainer_factory.trainer_config[
'Brain'].hyperparameters.learning_rate = 0.0005 * (
1 - counter / max_meta_updates)
trainer_factory.trainer_config[
'Brain'].hyperparameters.beta = 0.005 * (
1 - counter / max_meta_updates)
trainer_factory.trainer_config[
'Brain'].hyperparameters.epsilon = 0.2 * (
1 - counter / max_meta_updates)
print("Current lr: {}\nCurrent beta: {}\nCurrent epsilon: {}".format(
trainer_factory.trainer_config['Brain'].hyperparameters.
learning_rate,
trainer_factory.trainer_config['Brain'].hyperparameters.beta,
trainer_factory.trainer_config['Brain'].hyperparameters.epsilon))
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
write_path,
checkpoint_settings.run_id,
maybe_meta_curriculum,
not checkpoint_settings.inference,
run_seed,
sampler_manager,
resampling_interval,
)
try:
# Get inital weights
print("Start learning at step: " + str(step) + " meta_step: " +
str(counter))
print("Inital weights: " + str(inital_weights[8]))
weights_after_train = tc.start_learning(env_manager,
inital_weights)
print(tc.trainers['Brain'].optimizer)
# weights_after_train = tc.weights
# print("Trained weights: " + str(weights_after_train[8]))
step += options.behaviors['Brain'].max_steps
print("meta step:" + str(step))
# print(weights_after_train)
# equal = []
# for i, weight in enumerate(tc.weights):
# equal.append(np.array_equal(inital_weights[i], weights_after_train[i]))
# print(all(equal))
finally:
print(len(weights_after_train), len(inital_weights))
for i, weight in enumerate(weights_after_train):
inital_weights[i] = weights_after_train[i]
env_manager.close()
write_run_options(write_path, options)
write_timing_tree(run_logs_dir)
write_training_status(run_logs_dir)
counter += 1
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"):
checkpoint_settings = options.checkpoint_settings
env_settings = options.env_settings
engine_settings = options.engine_settings
base_path = "results"
write_path = os.path.join(base_path, checkpoint_settings.run_id)
maybe_init_path = (
os.path.join(base_path, checkpoint_settings.initialize_from)
if checkpoint_settings.initialize_from is not None
else None
)
run_logs_dir = os.path.join(write_path, "run_logs")
port: Optional[int] = env_settings.base_port
# Check if directory exists
validate_existing_directories(
write_path,
checkpoint_settings.resume,
checkpoint_settings.force,
maybe_init_path,
)
# Make run logs directory
os.makedirs(run_logs_dir, exist_ok=True)
# Load any needed states
if checkpoint_settings.resume:
GlobalTrainingStatus.load_state(
os.path.join(run_logs_dir, "training_status.json")
)
# Configure Tensorboard Writers and StatsReporter
tb_writer = TensorboardWriter(
write_path, clear_past_data=not checkpoint_settings.resume
)
gauge_write = GaugeWriter()
console_writer = ConsoleWriter()
StatsReporter.add_writer(tb_writer)
StatsReporter.add_writer(gauge_write)
StatsReporter.add_writer(console_writer)
if env_settings.env_path is None:
port = None
env_factory = create_environment_factory(
env_settings.env_path,
engine_settings.no_graphics,
run_seed,
port,
env_settings.env_args,
os.path.abspath(run_logs_dir), # Unity environment requires absolute path
)
engine_config = EngineConfig(
width=engine_settings.width,
height=engine_settings.height,
quality_level=engine_settings.quality_level,
time_scale=engine_settings.time_scale,
target_frame_rate=engine_settings.target_frame_rate,
capture_frame_rate=engine_settings.capture_frame_rate,
)
env_manager = SubprocessEnvManager(
env_factory, engine_config, env_settings.num_envs
)
env_parameter_manager = EnvironmentParameterManager(
options.environment_parameters, run_seed, restore=checkpoint_settings.resume
)
trainer_factory = TrainerFactory(
trainer_config=options.behaviors,
output_path=write_path,
train_model=not checkpoint_settings.inference,
load_model=checkpoint_settings.resume,
seed=run_seed,
param_manager=env_parameter_manager,
init_path=maybe_init_path,
multi_gpu=False,
)
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
write_path,
checkpoint_settings.run_id,
env_parameter_manager,
not checkpoint_settings.inference,
run_seed,
)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
write_run_options(write_path, options)
write_timing_tree(run_logs_dir)
write_training_status(run_logs_dir)
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"):
checkpoint_settings = options.checkpoint_settings
env_settings = options.env_settings
engine_settings = options.engine_settings
base_path = "results"
write_path = os.path.join(base_path, checkpoint_settings.run_id)
maybe_init_path = (
os.path.join(base_path, checkpoint_settings.initialize_from)
if checkpoint_settings.initialize_from
else None
)
run_logs_dir = os.path.join(write_path, "run_logs")
port: Optional[int] = env_settings.base_port
# Check if directory exists
handle_existing_directories(
write_path,
checkpoint_settings.resume,
checkpoint_settings.force,
maybe_init_path,
)
# Make run logs directory
os.makedirs(run_logs_dir, exist_ok=True)
# Load any needed states
if checkpoint_settings.resume:
GlobalTrainingStatus.load_state(
os.path.join(run_logs_dir, "training_status.json")
)
# 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 checkpoint_settings.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 env_settings.env_path is None:
port = None
env_factory = create_environment_factory(
env_settings.env_path,
engine_settings.no_graphics,
run_seed,
port,
env_settings.env_args,
os.path.abspath(run_logs_dir), # Unity environment requires absolute path
)
engine_config = EngineConfig(
width=engine_settings.width,
height=engine_settings.height,
quality_level=engine_settings.quality_level,
time_scale=engine_settings.time_scale,
target_frame_rate=engine_settings.target_frame_rate,
capture_frame_rate=engine_settings.capture_frame_rate,
)
env_manager = SubprocessEnvManager(
env_factory, engine_config, env_settings.num_envs
)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum, env_manager, restore=checkpoint_settings.resume
)
maybe_add_samplers(options.parameter_randomization, env_manager, run_seed)
trainer_factory = TrainerFactory(
options.behaviors,
write_path,
not checkpoint_settings.inference,
checkpoint_settings.resume,
run_seed,
maybe_init_path,
maybe_meta_curriculum,
False,
)
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
write_path,
checkpoint_settings.run_id,
maybe_meta_curriculum,
not checkpoint_settings.inference,
run_seed,
)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
write_run_options(write_path, options)
write_timing_tree(run_logs_dir)
write_training_status(run_logs_dir)