release
This commit is contained in:
allenzren
@@ -0,0 +1,102 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/halfcheetah-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-04-42/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
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: 5000
|
||||
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
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,103 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/halfcheetah-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-04-42/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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
|
||||
eta: 0.0001
|
||||
action_gradient_steps: 10
|
||||
buffer_size: 400000
|
||||
batch_size: 5000
|
||||
replay_ratio: 64
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_dipo.DIPODiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,102 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/halfcheetah-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-04-42/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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
|
||||
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: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,111 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/halfcheetah-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-04-42/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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: 20 # how many samples to score during eval
|
||||
critic_tau: 0.001 # rate of target q network update
|
||||
use_expectile_exploration: True
|
||||
buffer_size: 5000
|
||||
batch_size: 512
|
||||
replay_ratio: 16
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_idql.IDQLDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,110 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ppo_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/gym-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/halfcheetah-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-04-42/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
ft_denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 40
|
||||
name: ${env_name}
|
||||
max_episode_steps: 1000
|
||||
reset_at_iteration: False
|
||||
save_video: False
|
||||
best_reward_threshold_for_success: 3 # success rate not relevant for gym tasks
|
||||
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}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
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: 5000
|
||||
update_epochs: 5
|
||||
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: 0.01
|
||||
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: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
cond_dim: ${obs_dim}
|
||||
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 +1,118 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_exact_diffusion_agent.TrainPPOExactDiffusionAgent
|
||||
|
||||
name: ${env_name}_ppo_exact_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/gym-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/halfcheetah-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-04-42/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
ft_denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune-exact
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
gamma: 0.99
|
||||
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: 5000
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo_exact.PPOExactDiffusion
|
||||
sde:
|
||||
_target_: model.diffusion.sde_lib.VPSDE # VP for DDPM
|
||||
N: ${denoising_steps}
|
||||
rtol: 1e-4
|
||||
atol: 1e-4
|
||||
sde_hutchinson_type: Rademacher
|
||||
sde_step_size: 0.05
|
||||
sde_method: euler
|
||||
sde_num_epsilon: 2
|
||||
sde_min_beta: 1e-10
|
||||
sde_probability_flow: True
|
||||
#
|
||||
gamma_denoising: 0.99
|
||||
clip_ploss_coef: 0.01
|
||||
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: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
cond_dim: ${obs_dim}
|
||||
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 +1,103 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/halfcheetah-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-04-42/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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 # rate of target q network update
|
||||
buffer_size: 5000
|
||||
batch_size: 256
|
||||
replay_ratio: 32
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_qsm.QSMDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,87 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/halfcheetah-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-04-42/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
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: 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
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
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}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,108 @@
|
||||
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}
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
ft_denoising_steps: 20
|
||||
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: 10000
|
||||
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.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}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,94 @@
|
||||
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}
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
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: 10000
|
||||
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: True
|
||||
cond_dim: ${obs_dim}
|
||||
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 +1,102 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/hopper-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-10-05/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
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: 5000
|
||||
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
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,103 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/hopper-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-10-05/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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
|
||||
eta: 0.0001
|
||||
action_gradient_steps: 10
|
||||
buffer_size: 400000
|
||||
batch_size: 5000
|
||||
replay_ratio: 64
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_dipo.DIPODiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,102 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/hopper-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-10-05/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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
|
||||
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: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,111 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/hopper-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-10-05/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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: 20 # how many samples to score during eval
|
||||
critic_tau: 0.001 # rate of target q network update
|
||||
use_expectile_exploration: True
|
||||
buffer_size: 5000
|
||||
batch_size: 512
|
||||
replay_ratio: 16
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_idql.IDQLDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,110 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ppo_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/gym-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/hopper-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-10-05/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
ft_denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 40
|
||||
name: ${env_name}
|
||||
max_episode_steps: 1000
|
||||
reset_at_iteration: False
|
||||
save_video: False
|
||||
best_reward_threshold_for_success: 3 # success rate not relevant for gym tasks
|
||||
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}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
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: 5000
|
||||
update_epochs: 5
|
||||
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: 0.01
|
||||
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: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
cond_dim: ${obs_dim}
|
||||
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 +1,117 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_exact_diffusion_agent.TrainPPOExactDiffusionAgent
|
||||
|
||||
name: ${env_name}_ppo_exact_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/gym-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/hopper-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-10-05/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
ft_denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune-exact
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
gamma: 0.99
|
||||
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: 5000
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo_exact.PPOExactDiffusion
|
||||
sde:
|
||||
_target_: model.diffusion.sde_lib.VPSDE # VP for DDPM
|
||||
N: ${denoising_steps}
|
||||
rtol: 1e-4
|
||||
atol: 1e-4
|
||||
sde_hutchinson_type: Rademacher
|
||||
sde_step_size: 0.05
|
||||
sde_method: euler
|
||||
sde_num_epsilon: 2
|
||||
sde_min_beta: 1e-10
|
||||
sde_probability_flow: True
|
||||
#
|
||||
gamma_denoising: 0.99
|
||||
clip_ploss_coef: 0.01
|
||||
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: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
residual_style: True
|
||||
cond_dim: ${obs_dim}
|
||||
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 +1,103 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/hopper-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-10-05/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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 # rate of target q network update
|
||||
buffer_size: 5000
|
||||
batch_size: 256
|
||||
replay_ratio: 32
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_qsm.QSMDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,87 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/hopper-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-10-05/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
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: 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
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
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}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,108 @@
|
||||
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}
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
ft_denoising_steps: 20
|
||||
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: 10000
|
||||
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.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}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,94 @@
|
||||
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}
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
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: 10000
|
||||
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: True
|
||||
cond_dim: ${obs_dim}
|
||||
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 +1,102 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/walker2d-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-06-12/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: walker2d-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
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: 5000
|
||||
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
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,103 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/walker2d-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-06-12/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: walker2d-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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
|
||||
eta: 0.0001
|
||||
action_gradient_steps: 10
|
||||
buffer_size: 400000
|
||||
batch_size: 5000
|
||||
replay_ratio: 64
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_dipo.DIPODiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,102 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/walker2d-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-06-12/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: walker2d-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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
|
||||
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: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,111 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/walker2d-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-06-12/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: walker2d-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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: 20 # how many samples to score during eval
|
||||
critic_tau: 0.001 # rate of target q network update
|
||||
use_expectile_exploration: True
|
||||
buffer_size: 5000
|
||||
batch_size: 512
|
||||
replay_ratio: 16
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_idql.IDQLDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,110 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ppo_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/gym-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/walker2d-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-06-12/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: walker2d-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
ft_denoising_steps: 10
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
env:
|
||||
n_envs: 40
|
||||
name: ${env_name}
|
||||
max_episode_steps: 1000
|
||||
reset_at_iteration: False
|
||||
save_video: False
|
||||
best_reward_threshold_for_success: 3 # success rate not relevant for gym tasks
|
||||
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}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
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: 5000
|
||||
update_epochs: 5
|
||||
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: 0.01
|
||||
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: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
cond_dim: ${obs_dim}
|
||||
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 +1,103 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/walker2d-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-06-12/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: walker2d-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 5
|
||||
n_steps: 500
|
||||
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 # rate of target q network update
|
||||
buffer_size: 5000
|
||||
batch_size: 256
|
||||
replay_ratio: 32
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_qsm.QSMDiffusion
|
||||
# Sampling HPs
|
||||
min_sampling_denoising_std: 0.10
|
||||
randn_clip_value: 3
|
||||
#
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,87 @@
|
||||
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}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/walker2d-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-06-12/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: walker2d-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
cond_steps: 1
|
||||
horizon_steps: 4
|
||||
act_steps: 4
|
||||
|
||||
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: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env_name}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 0
|
||||
n_steps: 500
|
||||
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: 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
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
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}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,108 @@
|
||||
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}
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: walker2d-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
ft_denoising_steps: 20
|
||||
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: 10000
|
||||
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.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}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
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 +1,94 @@
|
||||
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}
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
|
||||
|
||||
device: cuda:0
|
||||
env_name: walker2d-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
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: 10000
|
||||
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: True
|
||||
cond_dim: ${obs_dim}
|
||||
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 +1,71 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
|
||||
|
||||
name: ${env}_pre_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/gym/${env}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env: halfcheetah-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
horizon_steps: 4
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 3000
|
||||
batch_size: 128
|
||||
learning_rate: 1e-3
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 3000
|
||||
warmup_steps: 1
|
||||
min_lr: 1e-4
|
||||
epoch_start_ema: 10
|
||||
update_ema_freq: 5
|
||||
save_model_freq: 100
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
out_activation_type: Identity
|
||||
use_layernorm: False
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
|
||||
ema:
|
||||
decay: 0.995
|
||||
|
||||
train_dataset:
|
||||
_target_: agent.dataset.sequence.StitchedActionSequenceDataset
|
||||
dataset_path: ${train_dataset_path}
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,71 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
|
||||
|
||||
name: ${env}_pre_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/gym/${env}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env: hopper-medium-v2
|
||||
obs_dim: 11
|
||||
action_dim: 3
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
horizon_steps: 4
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 3000
|
||||
batch_size: 128
|
||||
learning_rate: 1e-3
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 3000
|
||||
warmup_steps: 1
|
||||
min_lr: 1e-4
|
||||
epoch_start_ema: 10
|
||||
update_ema_freq: 5
|
||||
save_model_freq: 100
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
out_activation_type: Identity
|
||||
use_layernorm: False
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
|
||||
ema:
|
||||
decay: 0.995
|
||||
|
||||
train_dataset:
|
||||
_target_: agent.dataset.sequence.StitchedActionSequenceDataset
|
||||
dataset_path: ${train_dataset_path}
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,71 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
|
||||
|
||||
name: ${env}_pre_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/gym/${env}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env: walker2d-medium-v2
|
||||
obs_dim: 17
|
||||
action_dim: 6
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 20
|
||||
horizon_steps: 4
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: gym-${env}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 3000
|
||||
batch_size: 128
|
||||
learning_rate: 1e-3
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 3000
|
||||
warmup_steps: 1
|
||||
min_lr: 1e-4
|
||||
epoch_start_ema: 10
|
||||
update_ema_freq: 5
|
||||
save_model_freq: 100
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
cond_dim: ${obs_dim}
|
||||
time_dim: 16
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: ReLU
|
||||
out_activation_type: Identity
|
||||
use_layernorm: False
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
|
||||
ema:
|
||||
decay: 0.995
|
||||
|
||||
train_dataset:
|
||||
_target_: agent.dataset.sequence.StitchedActionSequenceDataset
|
||||
dataset_path: ${train_dataset_path}
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
Reference in New Issue
Block a user