def __init__(
self,
experiment_name,
use_gpu,
trainer_args
):
"""Train MTSAC with metaworld_experiments environment.
Args:
experiment_name: expeirment name to be used for logging and checkpointing
use_wandb: boolean, defines whether or not to log to wandb
use_gpu: boolean, defines whether or not to use to GPU for training
trainer_args: named tuple with args given by config
"""
# Define log and checkpoint dir
self.checkpoint_dir = os.path.join(
trainer_args.log_dir,
f"{experiment_name}-{trainer_args.project_id}"
)
print(f"Checkpoint dir: {self.checkpoint_dir}")
self.state_path = os.path.join(self.checkpoint_dir, "experiment_state.p")
self.env_state_path = os.path.join(self.checkpoint_dir, "env_state.p")
self.config_path = os.path.join(self.checkpoint_dir, "config.json")
self.experiment_name = experiment_name
# Only define viz_save_path if required to save visualizations local
self.viz_save_path = None
if trainer_args.save_visualizations_local:
self.viz_save_path = os.path.join(self.checkpoint_dir, "viz")
# Check if loading from existing experiment
self.loading_from_existing = os.path.exists(self.checkpoint_dir)
os.makedirs(self.checkpoint_dir, exist_ok=True)
# Save arguments for later retrieval
self.init_config(trainer_args)
num_tasks = trainer_args.num_tasks
# TODO: do we have to fix which GPU to use? run distributed across multiGPUs
if use_gpu:
set_gpu_mode(True, 0)
if trainer_args.seed is not None:
deterministic.set_seed(trainer_args.seed)
# Note: different classes whether it uses 10 or 50 tasks. Why?
mt_env = (
metaworld.MT10(seed=trainer_args.env_seed) if num_tasks <= 10
else metaworld.MT50(seed=trainer_args.env_seed)
)
train_task_sampler = MetaWorldTaskSampler(
mt_env, "train", add_env_onehot=True
)
# TODO: add some clarifying comments of why these asserts are required
assert num_tasks % 10 == 0, "Number of tasks have to divisible by 10"
assert num_tasks <= 500, "Number of tasks should be less or equal 500"
# TODO: do we have guarantees that in case seed is set, the tasks being sampled
# are the same?
mt_train_envs = train_task_sampler.sample(num_tasks)
env = mt_train_envs[0]()
if trainer_args.params_seed is not None:
torch.manual_seed(trainer_args.params_seed)
policy = create_policy_net(env_spec=env.spec, net_params=trainer_args)
qf1 = create_qf_net(env_spec=env.spec, net_params=trainer_args)
qf2 = create_qf_net(env_spec=env.spec, net_params=trainer_args)
if trainer_args.params_seed is not None:
calculate_mean_param("policy", policy)
calculate_mean_param("qf1", qf1)
calculate_mean_param("qf2", qf2)
if trainer_args.override_weight_initialization:
logging.warn("Overriding dendritic layer weight initialization")
self.override_weight_initialization([policy, qf1, qf2])
replay_buffer = PathBuffer(
capacity_in_transitions=trainer_args.num_buffer_transitions
)
max_episode_length = env.spec.max_episode_length
self.env_steps_per_epoch = int(max_episode_length * num_tasks)
self.num_epochs = trainer_args.timesteps // self.env_steps_per_epoch
sampler = RaySampler(
agent=policy,
envs=mt_train_envs,
max_episode_length=max_episode_length,
cpus_per_worker=trainer_args.cpus_per_worker,
gpus_per_worker=trainer_args.gpus_per_worker,
workers_per_env=trainer_args.workers_per_env,
seed=trainer_args.seed,
)
self._algo = CustomMTSAC(
env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
replay_buffer=replay_buffer,
sampler=sampler,
train_task_sampler=train_task_sampler,
gradient_steps_per_itr=int(
max_episode_length * trainer_args.num_grad_steps_scale
),
task_update_frequency=trainer_args.task_update_frequency,
num_tasks=num_tasks,
min_buffer_size=max_episode_length * num_tasks,
target_update_tau=trainer_args.target_update_tau,
discount=trainer_args.discount,
buffer_batch_size=trainer_args.buffer_batch_size,
policy_lr=trainer_args.policy_lr,
qf_lr=trainer_args.qf_lr,
reward_scale=trainer_args.reward_scale,
num_evaluation_episodes=trainer_args.eval_episodes,
fp16=trainer_args.fp16 if use_gpu else False,
log_per_task=trainer_args.log_per_task,
share_train_eval_env=trainer_args.share_train_eval_env
)
# Override with loaded networks if existing experiment
self.current_epoch = 0
if self.loading_from_existing:
self.load_experiment_state()
# Move all networks within the model on device
self._algo.to()
def mtppo_metaworld_mt10(ctxt, experiment_name, config_pth, seed, n_workers, n_tasks, use_wandb,
wandb_username, use_gpu):
"""Set up environment and algorithm and run the task.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
epochs (int): Number of training epochs.
batch_size (int): Number of environment steps in one batch.
n_workers (int): The number of workers the sampler should use.
n_tasks (int): Number of tasks to use. Should be a multiple of 10.
"""
params = get_params(config_pth)
set_seed(seed)
mt10 = metaworld.MT10()
train_task_sampler = MetaWorldTaskSampler(mt10,
'train',
lambda env, _: normalize(env),
add_env_onehot=True)
if use_wandb == 'True':
use_wandb = True
wandb.init(
name=experiment_name,
entity=wandb_username,
project="mt10",
group="Baselines{}".format("mt10"),
reinit=True,
config=params,
)
else:
use_wandb = False
assert n_tasks % 10 == 0
assert n_tasks <= 500
envs = [env_up() for env_up in train_task_sampler.sample(n_tasks)]
env = envs[0]
policy = create_policy_net(env_spec=env.spec,
net_params=params["net"]
)
value_function = create_vf_net(env_spec=env.spec,
net_params=params["net"]
)
sampler = RaySampler(agents=policy,
envs=env,
max_episode_length=env.spec.max_episode_length,
n_workers=n_workers,
worker_class=DefaultWorker)
gpu_training = True if use_gpu else False
algo = CustomMTPPO(env_spec=env.spec,
policy=policy,
value_function=value_function,
sampler=sampler,
train_task_sampler=train_task_sampler,
num_tasks=n_tasks,
task_update_frequency=params["training"]["task_update_frequency"],
num_eval_eps=params["general_setting"]["eval_episodes"],
policy_lr=params["training"]["policy_lr"],
vf_lr=params["training"]["vf_lr"],
ppo_eps=params["training"]["ppo_eps"],
minibatch_size=params["training"]["minibatch_size"],
ppo_epochs=params["training"]["ppo_epochs"],
num_train_per_epoch=params["training"]["num_train_per_epoch"],
discount=params["general_setting"]["discount"],
gae_lambda=params["training"]["gae_lambda"],
center_adv=False,
wandb_logging=use_wandb,
eval_freq=params["general_setting"]["eval_freq"],
stop_entropy_gradient=True,
entropy_method='max',
gpu_training=gpu_training
)
trainer = Trainer(ctxt)
trainer.setup(algo, env)
trainer.train(n_epochs=params["training"]["epochs"],
batch_size=params["training"]["batch_episodes_per_task"],
plot=False)
def mtsac_metaworld_mt50(
ctxt=None, *, config_pth, seed, timesteps, use_wandb, wandb_project_name, gpu
):
"""Train MTSAC with MT50 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
_gpu (int): The ID of the gpu to be used (used on multi-gpu machines).
num_tasks (int): Number of tasks to use. Should be a multiple of 10.
timesteps (int): Number of timesteps to run.
"""
"""Train MTSAC with metaworld_experiments environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
_gpu (int): The ID of the gpu to be used (used on multi-gpu machines).
timesteps (int): Number of timesteps to run.
"""
print(f"Initiation took {time() - t0:.2f} secs")
# Get experiment parameters (e.g. hyperparameters) and save the json file
params = get_params(config_pth)
with open(ctxt.snapshot_dir + "/params.json", "w") as json_file:
json.dump(params, json_file)
if use_wandb == "True":
use_wandb = True
wandb.init(
name=params["experiment_name"],
project=wandb_project_name,
group="Baselines{}".format("mt50"),
reinit=True,
config=params,
)
else:
use_wandb = False
num_tasks = 50
timesteps = timesteps
deterministic.set_seed(seed)
trainer = CustomTrainer(ctxt)
mt10 = metaworld.MT50()
train_task_sampler = MetaWorldTaskSampler(mt10, "train", add_env_onehot=True)
assert num_tasks % 10 == 0, "Number of tasks have to divisible by 10"
assert num_tasks <= 500, "Number of tasks should be less or equal 500"
mt50_train_envs = train_task_sampler.sample(num_tasks)
env = mt50_train_envs[0]()
params["net"]["policy_min_std"] = np.exp(params["net"]["policy_min_log_std"])
params["net"]["policy_max_std"] = np.exp(params["net"]["policy_max_log_std"])
policy = create_policy_net(env_spec=env.spec, net_params=params["net"])
qf1 = create_qf_net(env_spec=env.spec, net_params=params["net"])
qf2 = create_qf_net(env_spec=env.spec, net_params=params["net"])
replay_buffer = PathBuffer(
capacity_in_transitions=int(params["general_setting"]["num_buffer_transitions"])
)
max_episode_length = env.spec.max_episode_length
# Note: are the episode length the same among all tasks?
sampler = RaySampler(
agents=policy,
envs=mt50_train_envs,
max_episode_length=max_episode_length,
# 1 sampler worker for each environment
n_workers=num_tasks,
worker_class=DefaultWorker
)
test_sampler = RaySampler(
agents=policy,
envs=mt50_train_envs,
max_episode_length=max_episode_length,
# 1 sampler worker for each environment
n_workers=num_tasks,
worker_class=EvalWorker
)
# Number of transitions before a set of gradient updates
steps_between_updates = int(max_episode_length * num_tasks)
# epoch: 1 cycle of data collection + gradient updates
epochs = timesteps // steps_between_updates
mtsac = CustomMTSAC(
env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
replay_buffer=replay_buffer,
sampler=sampler,
train_task_sampler=train_task_sampler,
test_sampler=test_sampler,
gradient_steps_per_itr=int(max_episode_length * params["training"]["num_grad_steps_scale"]),
num_tasks=num_tasks,
min_buffer_size=max_episode_length * num_tasks,
target_update_tau=params["training"]["target_update_tau"],
discount=params["general_setting"]["discount"],
buffer_batch_size=params["training"]["buffer_batch_size"],
policy_lr=params["training"]["policy_lr"],
qf_lr=params["training"]["qf_lr"],
reward_scale=params["training"]["reward_scale"],
num_evaluation_episodes=params["general_setting"]["eval_episodes"],
task_update_frequency=params["training"]["task_update_frequency"],
wandb_logging=use_wandb,
evaluation_frequency=params["general_setting"]["evaluation_frequency"]
)
if gpu is not None:
set_gpu_mode(True, gpu)
mtsac.to()
trainer.setup(algo=mtsac, env=mt50_train_envs)
trainer.train(n_epochs=epochs, batch_size=steps_between_updates)
def mtsac_metaworld_mt10(ctxt=None,
*,
experiment_name,
config_pth,
seed,
use_wandb,
gpu):
"""Train MTSAC with MT10 environment.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by Trainer to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
_gpu (int): The ID of the gpu to be used (used on multi-gpu machines).
timesteps (int): Number of timesteps to run.
"""
print(f"Initiation took {time()-t0:.2f} secs")
device = torch.device("cuda") if gpu else torch.device("cpu")
print(f"Using GPU: {gpu}, Device: {device}")
# maybe overring other things - this is required, why?
if gpu:
set_gpu_mode(True)
else:
set_gpu_mode(False)
# Get experiment parameters (e.g. hyperparameters) and save the json file
params = get_params(config_pth)
with open(ctxt.snapshot_dir + "/params.json", "w") as json_file:
json.dump(params, json_file)
if use_wandb == "True":
use_wandb = True
wandb.init(
name=experiment_name,
project="mt10_debug",
group="Baselines{}".format("mt10"),
reinit=True,
config=params,
)
else:
use_wandb = False
num_tasks = params["net"]["num_tasks"]
timesteps = 15000000
deterministic.set_seed(seed)
trainer = Trainer(ctxt)
# Note: different classes whether it uses 10 or 50 tasks. Why?
if num_tasks <= 10:
mt_env = metaworld.MT10()
else:
mt_env = metaworld.MT50()
train_task_sampler = MetaWorldTaskSampler(mt_env,
"train",
add_env_onehot=True)
assert num_tasks % 10 == 0, "Number of tasks have to divisible by 10"
assert num_tasks <= 500, "Number of tasks should be less or equal 500"
mt_train_envs = train_task_sampler.sample(num_tasks)
env = mt_train_envs[0]()
params["net"]["policy_min_std"] = np.exp(
params["net"]["policy_min_log_std"])
params["net"]["policy_max_std"] = np.exp(
params["net"]["policy_max_log_std"])
policy = create_policy_net(env_spec=env.spec, net_params=params["net"])
print("Created policy")
qf1 = create_qf_net(env_spec=env.spec, net_params=params["net"])
qf2 = create_qf_net(env_spec=env.spec, net_params=params["net"])
print("Created value functions")
replay_buffer = PathBuffer(capacity_in_transitions=int(
params["general_setting"]["num_buffer_transitions"]))
max_episode_length = env.spec.max_episode_length
# Note: are the episode length the same among all tasks?
sampler = RaySampler(
agents=policy,
envs=mt_train_envs,
max_episode_length=max_episode_length,
cpus_per_worker=params["sampler"]["cpus_per_worker"],
gpus_per_worker=params["sampler"]["gpus_per_worker"],
seed=None, # set to get_seed() to make it deterministic
)
# will probably still need the sampler
test_sampler = sampler
# test_sampler = RaySampler(
# agents=policy,
# envs=mt_train_envs,
# max_episode_length=max_episode_length,
# # 1 sampler worker for each environment
# n_workers=num_tasks,
# worker_class=EvalWorker
# )
# Note: difference between sampler and test sampler is only the worker
# difference is one line in EvalWorker, uses average: a = agent_info["mean"]
# can we create a unified worker that cointais both rules?
# Number of transitions before a set of gradient updates
# Note: should we use avg episode length, if they are not same for all tasks?
batch_size = int(max_episode_length * num_tasks)
# TODO: this whole block seems unnecessary, it is not doing anything.
# Number of times policy is evaluated (also the # of epochs)
num_evaluation_points = timesteps // batch_size
epochs = timesteps // batch_size
# number of times new batch of samples + gradient updates are done per epoch
epoch_cycles = epochs // num_evaluation_points # this will always be equal to 1
epochs = epochs // epoch_cycles
mtsac = CustomMTSAC(
env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
replay_buffer=replay_buffer,
sampler=sampler,
train_task_sampler=train_task_sampler,
test_sampler=test_sampler,
gradient_steps_per_itr=1,
num_tasks=num_tasks,
steps_per_epoch=epoch_cycles,
min_buffer_size=max_episode_length * num_tasks,
target_update_tau=params["training"]["target_update_tau"],
discount=params["general_setting"]["discount"],
buffer_batch_size=params["training"]["buffer_batch_size"],
policy_lr=params["training"]["policy_lr"],
qf_lr=params["training"]["qf_lr"],
reward_scale=params["training"]["reward_scale"],
num_evaluation_episodes=params["general_setting"]["eval_episodes"],
task_update_frequency=params["training"]["task_update_frequency"],
wandb_logging=use_wandb,
evaluation_frequency=params["general_setting"]["evaluation_frequency"],
)
print("Created algo")
mtsac.to(device=device)
print("Moved networks to device")
trainer.setup(algo=mtsac, env=mt_train_envs)
print("Setup trainer")
trainer.train(n_epochs=epochs, batch_size=batch_size)