e0842e71dc
* v0.5 (#9) * update idql configs * update awr configs * update dipo configs * update qsm configs * update dqm configs * update project version to 0.5.0
109 lines
2.7 KiB
YAML
109 lines
2.7 KiB
YAML
defaults:
|
|
- _self_
|
|
hydra:
|
|
run:
|
|
dir: ${logdir}
|
|
_target_: agent.finetune.train_rlpd_agent.TrainRLPDAgent
|
|
|
|
name: ${env_name}_rlpd_mlp_ta${horizon_steps}
|
|
logdir: ${oc.env:DPPO_LOG_DIR}/gym-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
|
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
|
offline_dataset_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/train.npz
|
|
|
|
seed: 42
|
|
device: cuda:0
|
|
env_name: hopper-medium-v2
|
|
obs_dim: 11
|
|
action_dim: 3
|
|
cond_steps: 1
|
|
horizon_steps: 1
|
|
act_steps: 1
|
|
|
|
env:
|
|
n_envs: 1
|
|
name: ${env_name}
|
|
max_episode_steps: 1000
|
|
reset_at_iteration: False
|
|
save_video: False
|
|
best_reward_threshold_for_success: 3
|
|
wrappers:
|
|
mujoco_locomotion_lowdim:
|
|
normalization_path: ${normalization_path}
|
|
multi_step:
|
|
n_obs_steps: ${cond_steps}
|
|
n_action_steps: ${act_steps}
|
|
max_episode_steps: ${env.max_episode_steps}
|
|
reset_within_step: True
|
|
|
|
wandb:
|
|
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
|
project: rlpd-${env_name}
|
|
run: ${now:%H-%M-%S}_${name}
|
|
|
|
train:
|
|
n_train_itr: 250000
|
|
n_steps: 1
|
|
gamma: 0.99
|
|
actor_lr: 3e-4
|
|
actor_weight_decay: 0
|
|
actor_lr_scheduler:
|
|
first_cycle_steps: 1000
|
|
warmup_steps: 10
|
|
min_lr: 3e-4
|
|
critic_lr: 3e-4
|
|
critic_weight_decay: 0
|
|
critic_lr_scheduler:
|
|
first_cycle_steps: 1000
|
|
warmup_steps: 10
|
|
min_lr: 3e-4
|
|
save_model_freq: 50000
|
|
val_freq: 5000
|
|
render:
|
|
freq: 1
|
|
num: 0
|
|
log_freq: 200
|
|
# RLPD specific
|
|
batch_size: 256
|
|
target_ema_rate: 0.005
|
|
scale_reward_factor: 1
|
|
critic_num_update: 20
|
|
buffer_size: 1000000
|
|
n_eval_episode: 10
|
|
n_explore_steps: 5000
|
|
target_entropy: ${eval:'- ${action_dim} * ${act_steps}'}
|
|
init_temperature: 1
|
|
|
|
model:
|
|
_target_: model.rl.gaussian_rlpd.RLPD_Gaussian
|
|
randn_clip_value: 10
|
|
tanh_output: True # squash after sampling
|
|
backup_entropy: True
|
|
n_critics: 10 # Ensemble size for critic models
|
|
actor:
|
|
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
|
mlp_dims: [256, 256]
|
|
activation_type: ReLU
|
|
tanh_output: False # squash after sampling instead
|
|
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
|
horizon_steps: ${horizon_steps}
|
|
|
|
std_max: 7.3891
|
|
std_min: 0.0067
|
|
critic:
|
|
_target_: model.common.critic.CriticObsAct
|
|
mlp_dims: [256, 256]
|
|
activation_type: ReLU
|
|
use_layernorm: True
|
|
double_q: False # use ensemble
|
|
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
|
action_dim: ${action_dim}
|
|
action_steps: ${act_steps}
|
|
horizon_steps: ${horizon_steps}
|
|
device: ${device}
|
|
|
|
offline_dataset:
|
|
_target_: agent.dataset.sequence.StitchedSequenceQLearningDataset
|
|
dataset_path: ${offline_dataset_path}
|
|
horizon_steps: ${horizon_steps}
|
|
cond_steps: ${cond_steps}
|
|
device: ${device} |