release
This commit is contained in:
allenzren
@@ -0,0 +1,107 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_diffusion_mlp_ta4_td20/2024-07-10_01-46-16/checkpoint/state_8000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: false
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 300
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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: 1
|
||||
max_adv_weight: 100
|
||||
beta: 10
|
||||
buffer_size: 3000
|
||||
batch_size: 256
|
||||
replay_ratio: 64
|
||||
critic_update_ratio: 4
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_awr.AWRDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,108 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_diffusion_mlp_ta4_td20/2024-07-10_01-46-16/checkpoint/state_8000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: False
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 300
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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: 1
|
||||
eta: 0.0001
|
||||
action_gradient_steps: 10
|
||||
buffer_size: 400000
|
||||
batch_size: 5000
|
||||
replay_ratio: 64
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_dipo.DIPODiffusion
|
||||
# HP to tune
|
||||
min_sampling_denoising_std: 0.1
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,107 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_diffusion_mlp_ta4_td20/2024-07-10_01-46-16/checkpoint/state_8000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: false
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 300
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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: 1
|
||||
eta: 1.0
|
||||
buffer_size: 400000
|
||||
batch_size: 5000
|
||||
replay_ratio: 64
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_dql.DQLDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.1
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,116 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_diffusion_mlp_ta4_td20/2024-07-10_01-46-16/checkpoint/state_8000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: False
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 300
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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: 1
|
||||
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: 3000
|
||||
batch_size: 512
|
||||
replay_ratio: 16
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_idql.IDQLDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.1
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic_q:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
critic_v:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,115 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_diffusion_mlp_ta4_td20/2024-07-10_01-46-16/checkpoint/state_8000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
ft_denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: false
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 500
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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: 0.01
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.1
|
||||
min_logprob_denoising_std: 0.1
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
ft_denoising_steps: ${ft_denoising_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
cond_steps: ${cond_steps}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,164 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_img_agent.TrainPPOImgDiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_mlp_img_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_diffusion_mlp_img_ta4_td100/2024-07-30_22-27-34/checkpoint/state_4000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}-img.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}-img/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 9
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: False
|
||||
use_image_obs: True
|
||||
wrappers:
|
||||
robomimic_image:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos']
|
||||
image_keys: ['agentview_image']
|
||||
shape_meta: ${shape_meta}
|
||||
multi_step:
|
||||
n_obs_steps: ${cond_steps}
|
||||
n_action_steps: ${act_steps}
|
||||
max_episode_steps: ${env.max_episode_steps}
|
||||
reset_within_step: True
|
||||
|
||||
shape_meta:
|
||||
obs:
|
||||
rgb:
|
||||
shape: [3, 96, 96]
|
||||
state:
|
||||
shape: [9]
|
||||
action:
|
||||
shape: [7]
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 500
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
augment: True
|
||||
grad_accumulate: 20
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 500
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
critic_lr: 1e-3
|
||||
critic_weight_decay: 0
|
||||
critic_lr_scheduler:
|
||||
first_cycle_steps: 500
|
||||
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: 500
|
||||
logprob_batch_size: 1000
|
||||
update_epochs: 10
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.01
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.1
|
||||
min_logprob_denoising_std: 0.1
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.VisionDiffusionMLP
|
||||
backbone:
|
||||
_target_: model.common.vit.VitEncoder
|
||||
obs_shape: ${shape_meta.obs.rgb.shape}
|
||||
cfg:
|
||||
patch_size: 8
|
||||
depth: 1
|
||||
embed_dim: 128
|
||||
num_heads: 4
|
||||
embed_style: embed2
|
||||
embed_norm: 0
|
||||
augment: False
|
||||
spatial_emb: 128
|
||||
time_dim: 32
|
||||
mlp_dims: [768, 768, 768]
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.ViTCritic
|
||||
spatial_emb: 128
|
||||
augment: False
|
||||
backbone:
|
||||
_target_: model.common.vit.VitEncoder
|
||||
obs_shape: ${shape_meta.obs.rgb.shape}
|
||||
cfg:
|
||||
patch_size: 8
|
||||
depth: 1
|
||||
embed_dim: 128
|
||||
num_heads: 4
|
||||
embed_style: embed2
|
||||
embed_norm: 0
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
ft_denoising_steps: ${ft_denoising_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
cond_steps: ${cond_steps}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,117 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_unet_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_diffusion_unet_ta4_td20/2024-06-29_02-48-45/checkpoint/state_8000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
ft_denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: false
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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: 0.01
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.1
|
||||
min_logprob_denoising_std: 0.1
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
ft_denoising_steps: ${ft_denoising_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
cond_steps: ${cond_steps}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,103 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_gaussian_mlp_ta4/2024-06-28_15-02-32/checkpoint/state_5000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: false
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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.rl.gaussian_ppo.PPO_Gaussian
|
||||
clip_ploss_coef: 0.01
|
||||
randn_clip_value: 3
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [1024, 1024, 1024]
|
||||
residual_style: True
|
||||
fixed_std: 0.1
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,141 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_img_agent.TrainPPOImgGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gaussian_mlp_img_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_gaussian_mlp_img_ta4/2024-07-30_18-44-32/checkpoint/state_4000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}-img.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}-img/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 9
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: False
|
||||
use_image_obs: True
|
||||
wrappers:
|
||||
robomimic_image:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos']
|
||||
image_keys: ['agentview_image']
|
||||
shape_meta: ${shape_meta}
|
||||
multi_step:
|
||||
n_obs_steps: ${cond_steps}
|
||||
n_action_steps: ${act_steps}
|
||||
max_episode_steps: ${env.max_episode_steps}
|
||||
reset_within_step: True
|
||||
|
||||
shape_meta:
|
||||
obs:
|
||||
rgb:
|
||||
shape: [3, 96, 96]
|
||||
state:
|
||||
shape: [9]
|
||||
action:
|
||||
shape: [7]
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 500
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
augment: True
|
||||
grad_accumulate: 10
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 500
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
critic_lr: 1e-3
|
||||
critic_weight_decay: 0
|
||||
critic_lr_scheduler:
|
||||
first_cycle_steps: 500
|
||||
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
|
||||
logprob_batch_size: 1000
|
||||
update_epochs: 10
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_ppo.PPO_Gaussian
|
||||
clip_ploss_coef: 0.01
|
||||
randn_clip_value: 3
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_VisionMLP
|
||||
backbone:
|
||||
_target_: model.common.vit.VitEncoder
|
||||
obs_shape: ${shape_meta.obs.rgb.shape}
|
||||
cfg:
|
||||
patch_size: 8
|
||||
depth: 1
|
||||
embed_dim: 128
|
||||
num_heads: 4
|
||||
embed_style: embed2
|
||||
embed_norm: 0
|
||||
augment: False
|
||||
spatial_emb: 128
|
||||
mlp_dims: [768, 768, 768]
|
||||
residual_style: True
|
||||
fixed_std: 0.1
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.ViTCritic
|
||||
spatial_emb: 128
|
||||
augment: False
|
||||
backbone:
|
||||
_target_: model.common.vit.VitEncoder
|
||||
obs_shape: ${shape_meta.obs.rgb.shape}
|
||||
cfg:
|
||||
patch_size: 8
|
||||
depth: 1
|
||||
embed_dim: 128
|
||||
num_heads: 4
|
||||
embed_style: embed2
|
||||
embed_norm: 0
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,104 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gaussian_transformer_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_gaussian_transformer_ta4/2024-06-28_15-02-39/checkpoint/state_5000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: false
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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.rl.gaussian_ppo.PPO_Gaussian
|
||||
clip_ploss_coef: 0.01
|
||||
randn_clip_value: 3
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.transformer.Gaussian_Transformer
|
||||
transformer_embed_dim: 160
|
||||
transformer_num_heads: 4
|
||||
transformer_num_layers: 4
|
||||
fixed_std: 0.1
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,104 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gmm_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_gmm_mlp_ta4/2024-06-28_15-03-08/checkpoint/state_5000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
num_modes: 5
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: false
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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.rl.gmm_ppo.PPO_GMM
|
||||
clip_ploss_coef: 0.01
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.mlp_gmm.GMM_MLP
|
||||
mlp_dims: [1024, 1024, 1024]
|
||||
residual_style: True
|
||||
fixed_std: 0.1
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
num_modes: ${num_modes}
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,105 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gmm_transformer_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_gmm_transformer_ta4/2024-06-28_15-03-15/checkpoint/state_5000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
num_modes: 5
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: false
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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.rl.gmm_ppo.PPO_GMM
|
||||
clip_ploss_coef: 0.01
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.transformer.GMM_Transformer
|
||||
transformer_embed_dim: 160
|
||||
transformer_num_heads: 4
|
||||
transformer_num_layers: 4
|
||||
fixed_std: 0.1
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
num_modes: ${num_modes}
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,108 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_diffusion_mlp_ta4_td20/2024-07-10_01-46-16/checkpoint/state_8000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: False
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 300
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
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: 1
|
||||
q_grad_coeff: 50
|
||||
critic_tau: 0.005 # rate of target q network update
|
||||
buffer_size: 3000
|
||||
batch_size: 256
|
||||
replay_ratio: 32
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_qsm.QSMDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.1
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObsAct
|
||||
action_dim: ${action_dim}
|
||||
action_steps: ${act_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256, 256]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,92 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square/square_pre_diffusion_mlp_ta4_td20/2024-07-10_01-46-16/checkpoint/state_8000.pt
|
||||
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: square
|
||||
obs_dim: 23
|
||||
action_dim: 7
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 50
|
||||
name: ${env_name}
|
||||
best_reward_threshold_for_success: 1
|
||||
max_episode_steps: 400
|
||||
save_video: false
|
||||
wrappers:
|
||||
robomimic_lowdim:
|
||||
normalization_path: ${normalization_path}
|
||||
low_dim_keys: ['robot0_eef_pos',
|
||||
'robot0_eef_quat',
|
||||
'robot0_gripper_qpos',
|
||||
'object'] # same order of preprocessed observations
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: robomimic-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 300
|
||||
n_critic_warmup_itr: 2
|
||||
n_steps: 400
|
||||
gamma: 0.999
|
||||
lr: 1e-5
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 1000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-5
|
||||
save_model_freq: 100
|
||||
val_freq: 10
|
||||
render:
|
||||
freq: 1
|
||||
num: 0
|
||||
# RWR specific
|
||||
max_reward_weight: 100
|
||||
beta: 10
|
||||
batch_size: 1000
|
||||
update_epochs: 16
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_rwr.RWRDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.1
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
Reference in New Issue
Block a user