def her_td3_experiment(variant):
import gym
import multiworld.envs.mujoco
import multiworld.envs.pygame
import railrl.samplers.rollout_functions as rf
import railrl.torch.pytorch_util as ptu
from railrl.exploration_strategies.base import (
PolicyWrappedWithExplorationStrategy)
from railrl.exploration_strategies.epsilon_greedy import EpsilonGreedy
from railrl.exploration_strategies.gaussian_strategy import GaussianStrategy
from railrl.exploration_strategies.ou_strategy import OUStrategy
from railrl.torch.grill.launcher import get_video_save_func
from railrl.demos.her_td3bc import HerTD3BC
from railrl.torch.networks import FlattenMlp, TanhMlpPolicy
from railrl.data_management.obs_dict_replay_buffer import (
ObsDictRelabelingBuffer)
if 'env_id' in variant:
env = gym.make(variant['env_id'])
else:
env = variant['env_class'](**variant['env_kwargs'])
observation_key = variant['observation_key']
desired_goal_key = variant['desired_goal_key']
variant['algo_kwargs']['her_kwargs']['observation_key'] = observation_key
variant['algo_kwargs']['her_kwargs']['desired_goal_key'] = desired_goal_key
if variant.get('normalize', False):
raise NotImplementedError()
achieved_goal_key = desired_goal_key.replace("desired", "achieved")
replay_buffer = ObsDictRelabelingBuffer(
env=env,
observation_key=observation_key,
desired_goal_key=desired_goal_key,
achieved_goal_key=achieved_goal_key,
**variant['replay_buffer_kwargs'])
demo_train_buffer = ObsDictRelabelingBuffer(
env=env,
observation_key=observation_key,
desired_goal_key=desired_goal_key,
achieved_goal_key=achieved_goal_key,
**variant['replay_buffer_kwargs'])
demo_test_buffer = ObsDictRelabelingBuffer(
env=env,
observation_key=observation_key,
desired_goal_key=desired_goal_key,
achieved_goal_key=achieved_goal_key,
**variant['replay_buffer_kwargs'])
obs_dim = env.observation_space.spaces['observation'].low.size
action_dim = env.action_space.low.size
goal_dim = env.observation_space.spaces['desired_goal'].low.size
exploration_type = variant['exploration_type']
if exploration_type == 'ou':
es = OUStrategy(action_space=env.action_space, **variant['es_kwargs'])
elif exploration_type == 'gaussian':
es = GaussianStrategy(
action_space=env.action_space,
**variant['es_kwargs'],
)
elif exploration_type == 'epsilon':
es = EpsilonGreedy(
action_space=env.action_space,
**variant['es_kwargs'],
)
else:
raise Exception("Invalid type: " + exploration_type)
qf1 = FlattenMlp(input_size=obs_dim + action_dim + goal_dim,
output_size=1,
**variant['qf_kwargs'])
qf2 = FlattenMlp(input_size=obs_dim + action_dim + goal_dim,
output_size=1,
**variant['qf_kwargs'])
policy = TanhMlpPolicy(input_size=obs_dim + goal_dim,
output_size=action_dim,
**variant['policy_kwargs'])
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
algorithm = HerTD3BC(env,
qf1=qf1,
qf2=qf2,
policy=policy,
exploration_policy=exploration_policy,
demo_train_buffer=demo_train_buffer,
demo_test_buffer=demo_test_buffer,
replay_buffer=replay_buffer,
demo_path=variant["demo_path"],
**variant['algo_kwargs'])
if variant.get("save_video", False):
rollout_function = rf.create_rollout_function(
rf.multitask_rollout,
max_path_length=algorithm.max_path_length,
observation_key=algorithm.observation_key,
desired_goal_key=algorithm.desired_goal_key,
)
video_func = get_video_save_func(
rollout_function,
env,
policy,
variant,
)
algorithm.post_epoch_funcs.append(video_func)
algorithm.to(ptu.device)
algorithm.train()
def grill_her_td3_experiment(variant):
import railrl.samplers.rollout_functions as rf
import railrl.torch.pytorch_util as ptu
from railrl.data_management.obs_dict_replay_buffer import \
ObsDictRelabelingBuffer
from railrl.exploration_strategies.base import (
PolicyWrappedWithExplorationStrategy
)
from railrl.demos.her_td3bc import HerTD3BC
from railrl.torch.networks import FlattenMlp, TanhMlpPolicy
grill_preprocess_variant(variant)
env = get_envs(variant)
es = get_exploration_strategy(variant, env)
observation_key = variant.get('observation_key', 'latent_observation')
desired_goal_key = variant.get('desired_goal_key', 'latent_desired_goal')
achieved_goal_key = desired_goal_key.replace("desired", "achieved")
obs_dim = (
env.observation_space.spaces[observation_key].low.size
+ env.observation_space.spaces[desired_goal_key].low.size
)
action_dim = env.action_space.low.size
qf1 = FlattenMlp(
input_size=obs_dim + action_dim,
output_size=1,
**variant['qf_kwargs']
)
qf2 = FlattenMlp(
input_size=obs_dim + action_dim,
output_size=1,
**variant['qf_kwargs']
)
policy = TanhMlpPolicy(
input_size=obs_dim,
output_size=action_dim,
**variant['policy_kwargs']
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
replay_buffer = ObsDictRelabelingBuffer(
env=env,
observation_key=observation_key,
desired_goal_key=desired_goal_key,
achieved_goal_key=achieved_goal_key,
**variant['replay_buffer_kwargs']
)
demo_train_buffer = ObsDictRelabelingBuffer(
env=env,
observation_key=observation_key,
desired_goal_key=desired_goal_key,
achieved_goal_key=achieved_goal_key,
**variant['replay_buffer_kwargs']
)
demo_test_buffer = ObsDictRelabelingBuffer(
env=env,
observation_key=observation_key,
desired_goal_key=desired_goal_key,
achieved_goal_key=achieved_goal_key,
**variant['replay_buffer_kwargs']
)
algo_kwargs = variant['algo_kwargs']
algo_kwargs['replay_buffer'] = replay_buffer
base_kwargs = algo_kwargs['base_kwargs']
base_kwargs['training_env'] = env
base_kwargs['render'] = variant["render"]
base_kwargs['render_during_eval'] = variant["render"]
her_kwargs = algo_kwargs['her_kwargs']
her_kwargs['observation_key'] = observation_key
her_kwargs['desired_goal_key'] = desired_goal_key
# algorithm = HerTd3(
# env,
# qf1=qf1,
# qf2=qf2,
# policy=policy,
# exploration_policy=exploration_policy,
# **variant['algo_kwargs']
# )
env.vae.to(ptu.device)
algorithm = HerTD3BC(
env,
qf1=qf1,
qf2=qf2,
policy=policy,
exploration_policy=exploration_policy,
demo_train_buffer=demo_train_buffer,
demo_test_buffer=demo_test_buffer,
demo_path=variant["demo_path"],
add_demo_latents=True,
**variant['algo_kwargs']
)
if variant.get("save_video", True):
rollout_function = rf.create_rollout_function(
rf.multitask_rollout,
max_path_length=algorithm.max_path_length,
observation_key=algorithm.observation_key,
desired_goal_key=algorithm.desired_goal_key,
)
video_func = get_video_save_func(
rollout_function,
env,
algorithm.eval_policy,
variant,
)
algorithm.post_epoch_funcs.append(video_func)
algorithm.to(ptu.device)
if not variant.get("do_state_exp", False):
env.vae.to(ptu.device)
algorithm.train()