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
103 lines
2.7 KiB
YAML
103 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}_td${denoising_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: walker2d-medium-v2
|
|
obs_dim: 17
|
|
action_dim: 6
|
|
denoising_steps: 20
|
|
cond_steps: 1
|
|
horizon_steps: 1
|
|
act_steps: 1
|
|
|
|
env:
|
|
n_envs: 40
|
|
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: ${oc.env:DPPO_WANDB_ENTITY}
|
|
project: rlpd-gym-${env_name}-finetune
|
|
run: ${now:%H-%M-%S}_${name}
|
|
|
|
train:
|
|
n_train_itr: 1000
|
|
n_critic_warmup_itr: 5
|
|
n_steps: 2000
|
|
gamma: 0.99
|
|
actor_lr: 1e-4
|
|
actor_weight_decay: 0
|
|
actor_lr_scheduler:
|
|
first_cycle_steps: 1000
|
|
warmup_steps: 10
|
|
min_lr: 1e-4
|
|
critic_lr: 1e-3
|
|
critic_weight_decay: 0
|
|
critic_lr_scheduler:
|
|
first_cycle_steps: 1000
|
|
warmup_steps: 10
|
|
min_lr: 1e-3
|
|
save_model_freq: 100
|
|
val_freq: 10
|
|
render:
|
|
freq: 1
|
|
num: 0
|
|
# RLPD specific
|
|
batch_size: 512
|
|
entropy_temperature: 1.0 # alpha in RLPD paper
|
|
target_ema_rate: 0.005 # rho in RLPD paper
|
|
scale_reward_factor: 1.0 # multiply reward by this amount for more stable value estimation
|
|
replay_ratio: 64 # number of batches to sample for each learning update
|
|
buffer_size: 1000000
|
|
|
|
model:
|
|
_target_: model.rl.gaussian_rlpd.RLPD_Gaussian
|
|
randn_clip_value: 3
|
|
actor:
|
|
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
|
mlp_dims: [512, 512, 512]
|
|
activation_type: ReLU
|
|
residual_style: True
|
|
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
|
horizon_steps: ${horizon_steps}
|
|
action_dim: ${action_dim}
|
|
critic:
|
|
_target_: model.common.critic.CriticObsAct
|
|
action_dim: ${action_dim}
|
|
action_steps: ${act_steps}
|
|
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
|
mlp_dims: [256, 256, 256]
|
|
activation_type: Mish
|
|
residual_style: True
|
|
use_layernorm: True
|
|
horizon_steps: ${horizon_steps}
|
|
device: ${device}
|
|
n_critics: 2 # Ensemble size for critic models
|
|
|
|
offline_dataset:
|
|
_target_: agent.dataset.sequence.StitchedSequenceQLearningDataset
|
|
dataset_path: ${offline_dataset_path}
|
|
horizon_steps: ${horizon_steps}
|
|
cond_steps: ${cond_steps}
|
|
device: ${device} |