Continuous actions using split MLP#
[7]:
%load_ext autoreload
%autoreload 2
The autoreload extension is already loaded. To reload it, use:
%reload_ext autoreload
[8]:
import warnings
warnings.filterwarnings('ignore')
You can skip training and instead load the last trained policy by changing the flag below.
[9]:
from navground.learning.utils.jupyter import skip_if, run_if
training = True
MASAC#
[10]:
from navground.learning import DefaultObservationConfig, DefaultStateConfig
from navground.learning.config.control_action_with_comm import ControlActionWithCommConfig
from navground.learning.examples.pad import get_env, comm, marker, PadReward
from navground.learning.scenarios.pad import render_kwargs
name = "DistributedCommFloatWithRewardSharing"
comm_size = 1
kwargs = render_kwargs(comm=True, low=-1, high=1,
color_low=(1, 0, 0), color_high=(0, 1, 0))
action = ControlActionWithCommConfig(use_acceleration_action=True, max_acceleration=1,
fix_orientation=True, comm_size=comm_size)
observation = DefaultObservationConfig(flat=False, include_velocity=True, include_target_direction=False)
state = DefaultStateConfig(include_y=False, include_position=True, include_velocity=True)
sensors = [comm(size=comm_size, binarize=False), marker()]
train_env = get_env(action=action, observation=observation, state=state,
sensors=sensors, start_in_opposite_sides=False,
reward=PadReward(neighbor_weight=0.5), include_success=False)
test_env = get_env(action=action, observation=observation, sensors=sensors, state=state,
start_in_opposite_sides=True,
include_success=True, init_success=False, intermediate_success=True,
render_mode='rgb_array', render_kwargs=kwargs)
[11]:
%%skip_if $training
import pathlib, os
from navground.learning.utils.benchmarl import NavgroundExperiment
path = max(pathlib.Path(f'logs/{name}/MASAC-SplitModel').glob('*/checkpoints/*.pt'), key=os.path.getmtime)
experiment = NavgroundExperiment.reload_from_file(str(path))
experiment.folder_name
[12]:
%%run_if $training
import os
from benchmarl.algorithms import MasacConfig
from benchmarl.models.mlp import MlpConfig
from navground.learning.utils.benchmarl.split_mlp import SplitMlpConfig
from benchmarl.experiment import ExperimentConfig
from navground.learning.utils.benchmarl import NavgroundExperiment, ExportPolicyCallback
algorithm = MasacConfig.get_from_yaml()
mlp = MlpConfig.get_from_yaml()
mlps_specs = [(1, None, {'num_cells': [256, 256]}), (1, ['pad/x', 'ego_velocity'], {'num_cells': [256, 256]})]
model = SplitMlpConfig(num_cells=mlp.num_cells, layer_class=mlp.layer_class,
activation_class=mlp.activation_class, activation_kwargs=mlp.activation_kwargs,
norm_class=mlp.norm_class, norm_kwargs=mlp.norm_kwargs,
mlps_specs=mlps_specs)
config = ExperimentConfig.get_from_yaml()
config.loggers = [ 'csv']
config.render = True
config.evaluation_interval = 6000
config.evaluation_episodes = 30
config.lr = 3e-4
config.max_n_iters = 1
config.checkpoint_at_end = True
config.save_folder = f'logs/{name}/MASAC-SplitModel'
os.makedirs(config.save_folder, exist_ok=True)
experiment_masac = NavgroundExperiment(
env=train_env,
eval_env=test_env,
config=config,
model_config=model,
critic_model_config=mlp,
algorithm_config=algorithm,
seed=0,
callbacks=[ExportPolicyCallback()]
)
experiment_masac.folder_name
[12]:
PosixPath('logs/DistributedCommFloatWithRewardSharing/MASAC-SplitModel/masac_navground_splitmlp__44bff93e_25_05_21-16_09_42')
[13]:
experiment_masac.policy[0][0][0].mlps
[13]:
[MultiAgentMLP(
MLP(
(0): Linear(in_features=3, out_features=256, bias=True)
(1): Tanh()
(2): Linear(in_features=256, out_features=256, bias=True)
(3): Tanh()
(4): Linear(in_features=256, out_features=2, bias=True)
),
n_agents=2,
share_params=True,
centralized=False,
agent_dim=-2),
MultiAgentMLP(
MLP(
(0): Linear(in_features=2, out_features=256, bias=True)
(1): Tanh()
(2): Linear(in_features=256, out_features=256, bias=True)
(3): Tanh()
(4): Linear(in_features=256, out_features=2, bias=True)
),
n_agents=2,
share_params=True,
centralized=False,
agent_dim=-2)]
[14]:
%%run_if $training
experiment_masac.run_for(iterations=40)
mean return = -6.207028388977051: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 40/40 [18:26<00:00, 27.66s/it]
[15]:
experiment_masac.plot_eval_logs(reward_low=-200, reward_high=0,
success=True, figsize=(8, 3))
[17]:
from stable_baselines3.common.evaluation import evaluate_policy
from navground.learning.parallel_env import make_vec_from_penv
test_venv = make_vec_from_penv(test_env)
policy_masac = experiment_masac.load_policy(name='best_policy')
evaluate_policy(policy_masac, test_venv, n_eval_episodes=30)
[17]:
(-30.328231702248257, 30.494501651731493)
[18]:
from navground.learning.evaluation.video import display_episode_video
display_episode_video(test_env, policy=policy_masac, factor=4, select=2, of=31, **kwargs)
[18]:
[20]:
from navground.learning.utils.plot import plot_policy
plot_policy(policy_masac,
variable={'pad/x': (-1, 1), 'neighbor/comm': (-1, 1)},
fix={'ego_velocity': 0.1},
actions={0: 'acceleration', 1: 'message'}, width=5, height=3)
Alternate Action and Communication training#
[21]:
%%skip_if $training
import pathlib, os
from navground.learning.utils.benchmarl import NavgroundExperiment
path = max(pathlib.Path(f'logs/{name}/MASAC-AltModel').glob('*/checkpoints/*.pt'), key=os.path.getmtime)
experiment = NavgroundExperiment.reload_from_file(str(path))
experiment.folder_name
[22]:
%%run_if $training
from navground.learning.utils.benchmarl.callbacks import AlternateActorCallback
learning_rates = [(2, 3e-4), (2, 1e-3)]
alt_cb = AlternateActorCallback(loss='loss_actor',
learning_rates=learning_rates)
config = ExperimentConfig.get_from_yaml()
config.loggers = [ 'csv']
config.render = True
config.evaluation_interval = 6000
config.evaluation_episodes = 30
config.lr = 3e-4
config.max_n_iters = 1
config.checkpoint_at_end = True
config.save_folder = f'logs/{name}/MASAC-AltModel'
os.makedirs(config.save_folder, exist_ok=True)
experiment_masac_alt = NavgroundExperiment(
env=train_env,
eval_env=test_env,
config=config,
model_config=model,
critic_model_config=mlp,
algorithm_config=algorithm,
seed=0,
callbacks=[ExportPolicyCallback(), alt_cb]
)
experiment_masac_alt.folder_name
[22]:
PosixPath('logs/DistributedCommFloatWithRewardSharing/MASAC-AltModel/masac_navground_splitmlp__013da278_25_05_21-16_30_29')
[23]:
%%run_if $training
experiment_masac_alt.run_for(iterations=40)
mean return = -5.425860404968262: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 40/40 [20:17<00:00, 30.43s/it]
[24]:
experiment_masac_alt.plot_eval_logs(reward_low=-200, reward_high=0,
success=True, figsize=(8, 3))
[29]:
policy_masac_alt = experiment_masac_alt.load_policy(name='best_policy')
evaluate_policy(policy_masac_alt, test_venv, n_eval_episodes=30)
[29]:
(-11.591048258543015, 13.943050052673952)
[32]:
display_episode_video(test_env, policy=policy_masac_alt, factor=4, select=15, of=31, **kwargs)
[32]:
[31]:
plot_policy(policy_masac_alt,
variable={'pad/x': (-1, 1), 'neighbor/comm': (-1, 1)},
fix={'ego_velocity': 0.1},
actions={0: 'acceleration', 1: 'message'}, width=5, height=3)
