v0.6 (#18)
* Sampling over both env and denoising steps in DPPO updates (#13) * sample one from each chain * full random sampling * Add Proficient Human (PH) Configs and Pipeline (#16) * fix missing cfg * add ph config * fix how terminated flags are added to buffer in ibrl * add ph config * offline calql for 1M gradient updates * bug fix: number of calql online gradient steps is the number of new transitions collected * add sample config for DPPO with ta=1 * Sampling over both env and denoising steps in DPPO updates (#13) * sample one from each chain * full random sampling * fix diffusion loss when predicting initial noise * fix dppo inds * fix typo * remove print statement --------- Co-authored-by: Justin M. Lidard <jlidard@neuronic.cs.princeton.edu> Co-authored-by: allenzren <allen.ren@princeton.edu> * update robomimic configs * better calql formulation * optimize calql and ibrl training * optimize data transfer in ppo agents * add kitchen configs * re-organize config folders, rerun calql and rlpd * add scratch gym locomotion configs * add kitchen installation dependencies * use truncated for termination in furniture env * update furniture and gym configs * update README and dependencies with kitchen * add url for new data and checkpoints * update demo RL configs * update batch sizes for furniture unet configs * raise error about dropout in residual mlp * fix observation bug in bc loss --------- Co-authored-by: Justin Lidard <60638575+jlidard@users.noreply.github.com> Co-authored-by: Justin M. Lidard <jlidard@neuronic.cs.princeton.edu>
This commit is contained in:
Allen Z. Ren
co-authored by
Justin M. Lidard
Justin Lidard
Justin M. Lidard
Justin Lidard
parent
7b10df690d
commit
dc8e0c9edc
@@ -0,0 +1,99 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_awr_diffusion_agent.TrainAWRDiffusionAgent
|
||||
|
||||
name: ${env_name}_awr_diffusion_mlp_ta${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
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 10
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-scratch
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 1000
|
||||
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
|
||||
# AWR specific
|
||||
scale_reward_factor: 0.01
|
||||
max_adv_weight: 100
|
||||
beta: 10
|
||||
buffer_size: 100000 # * n_envs
|
||||
batch_size: 256
|
||||
replay_ratio: 128
|
||||
critic_update_ratio: 4
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_awr.AWRDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,101 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_dipo_diffusion_agent.TrainDIPODiffusionAgent
|
||||
|
||||
name: ${env_name}_dipo_diffusion_mlp_ta${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
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 10
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-scratch
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 1000
|
||||
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
|
||||
# DIPO specific
|
||||
scale_reward_factor: 0.01
|
||||
target_ema_rate: 0.005
|
||||
buffer_size: 1000000
|
||||
action_lr: 0.0001
|
||||
action_gradient_steps: 10
|
||||
replay_ratio: 128
|
||||
batch_size: 256
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_dipo.DIPODiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,100 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_dql_diffusion_agent.TrainDQLDiffusionAgent
|
||||
|
||||
name: ${env_name}_dql_diffusion_mlp_ta${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
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 10
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-scratch
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 1000
|
||||
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
|
||||
# DQL specific
|
||||
scale_reward_factor: 0.01
|
||||
target_ema_rate: 0.005
|
||||
buffer_size: 1000000
|
||||
eta: 1.0
|
||||
replay_ratio: 128
|
||||
batch_size: 256
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_dql.DQLDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,108 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_idql_diffusion_agent.TrainIDQLDiffusionAgent
|
||||
|
||||
name: ${env_name}_idql_diffusion_mlp_ta${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
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 10
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-scratch
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 1000
|
||||
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
|
||||
# IDQL specific
|
||||
scale_reward_factor: 0.01
|
||||
eval_deterministic: True
|
||||
eval_sample_num: 10 # how many samples to score during eval
|
||||
critic_tau: 0.001 # rate of target q network update
|
||||
use_expectile_exploration: True
|
||||
buffer_size: 100000 # * n_envs
|
||||
replay_ratio: 128
|
||||
batch_size: 256
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_idql.IDQLDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
critic_q:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
critic_v:
|
||||
_target_: model.common.critic.CriticObs
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,106 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_nopre_ppo_diffusion_mlp_ta${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
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
denoising_steps: 10
|
||||
ft_denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 10
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-scratch
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 1000
|
||||
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
|
||||
# PPO specific
|
||||
reward_scale_running: True
|
||||
reward_scale_const: 1.0
|
||||
gae_lambda: 0.95
|
||||
batch_size: 10000
|
||||
update_epochs: 10
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.99
|
||||
clip_ploss_coef: 0.1
|
||||
clip_ploss_coef_base: 0.1
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.1
|
||||
min_logprob_denoising_std: 0.1
|
||||
#
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
ft_denoising_steps: ${ft_denoising_steps}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,93 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_nopre_ppo_gaussian_mlp_ta${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
|
||||
|
||||
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: 10
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-scratch
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 1000
|
||||
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
|
||||
# PPO specific
|
||||
reward_scale_running: True
|
||||
reward_scale_const: 1.0
|
||||
gae_lambda: 0.95
|
||||
batch_size: 1000
|
||||
update_epochs: 10
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_ppo.PPO_Gaussian
|
||||
clip_ploss_coef: 0.1
|
||||
randn_clip_value: 3
|
||||
#
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: False # with new logvar head
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,100 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_qsm_diffusion_agent.TrainQSMDiffusionAgent
|
||||
|
||||
name: ${env_name}_qsm_diffusion_mlp_ta${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
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 10
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-scratch
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 1000
|
||||
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
|
||||
# QSM specific
|
||||
scale_reward_factor: 0.01
|
||||
q_grad_coeff: 50
|
||||
critic_tau: 0.005
|
||||
buffer_size: 100000 # * n_envs
|
||||
replay_ratio: 128
|
||||
batch_size: 256
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_qsm.QSMDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,84 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_rwr_diffusion_agent.TrainRWRDiffusionAgent
|
||||
|
||||
name: ${env_name}_rwr_diffusion_mlp_ta${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
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 1
|
||||
act_steps: 1
|
||||
|
||||
env:
|
||||
n_envs: 10
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-scratch
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 1000
|
||||
gamma: 0.99
|
||||
lr: 1e-4
|
||||
weight_decay: 0
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-4
|
||||
save_model_freq: 100
|
||||
val_freq: 10
|
||||
render:
|
||||
freq: 1
|
||||
num: 0
|
||||
# RWR specific
|
||||
max_reward_weight: 100
|
||||
beta: 10
|
||||
batch_size: 256
|
||||
update_epochs: 128
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_rwr.RWRDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.1
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
action_dim: ${action_dim}
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
Reference in New Issue
Block a user