release
This commit is contained in:
allenzren
@@ -0,0 +1,121 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/lamp/lamp_low_dim_pre_diffusion_mlp_ta8_td100/2024-07-23_01-28-20/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: lamp
|
||||
randomness: low
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,123 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_unet_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/lamp/lamp_low_dim_pre_diffusion_unet_ta16_td100/2024-07-04_02-16-48/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 16
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: lamp
|
||||
randomness: low
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
groupnorm_eps: 1e-4
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,109 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/lamp/lamp_low_dim_pre_gaussian_mlp_ta8/2024-06-28_16-26-51/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: lamp
|
||||
randomness: low
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_ppo.PPO_Gaussian
|
||||
clip_ploss_coef: 0.01
|
||||
randn_clip_value: 3
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims:
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
fixed_std: 0.04
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
cond_dim: ${obs_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,121 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/lamp/lamp_med_dim_pre_diffusion_mlp_ta8_td100/2024-07-23_01-28-20/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: lamp
|
||||
randomness: med
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,122 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_unet_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/lamp/lamp_med_dim_pre_diffusion_unet_ta16_td100/2024-07-04_02-17-21/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 16
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: lamp
|
||||
randomness: med
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,109 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/lamp/lamp_med_dim_pre_gaussian_mlp_ta8/2024-06-28_16-26-56/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: lamp
|
||||
randomness: med
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_ppo.PPO_Gaussian
|
||||
clip_ploss_coef: 0.01
|
||||
randn_clip_value: 3
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims:
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
fixed_std: 0.04
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
cond_dim: ${obs_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,121 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/one_leg/one_leg_low_dim_pre_diffusion_mlp_ta8_td100/2024-07-22_20-01-16/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 700
|
||||
best_reward_threshold_for_success: 1
|
||||
specific:
|
||||
headless: true
|
||||
furniture: one_leg
|
||||
randomness: low
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,123 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_unet_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/one_leg/one_leg_low_dim_pre_diffusion_unet_ta16_td100/2024-07-03_22-23-38/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 16
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 700
|
||||
best_reward_threshold_for_success: 1
|
||||
specific:
|
||||
headless: true
|
||||
furniture: one_leg
|
||||
randomness: low
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
groupnorm_eps: 1e-4
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,109 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/one_leg/one_leg_low_dim_pre_gaussian_mlp_ta8/2024-06-26_23-43-02/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 700
|
||||
best_reward_threshold_for_success: 1
|
||||
specific:
|
||||
headless: true
|
||||
furniture: one_leg
|
||||
randomness: low
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_ppo.PPO_Gaussian
|
||||
clip_ploss_coef: 0.01
|
||||
randn_clip_value: 3
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims:
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
fixed_std: 0.04
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
cond_dim: ${obs_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,121 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/one_leg/one_leg_med_dim_pre_diffusion_mlp_ta8_td100/2024-07-23_01-28-11/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 700
|
||||
best_reward_threshold_for_success: 1
|
||||
specific:
|
||||
headless: true
|
||||
furniture: one_leg
|
||||
randomness: med
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,123 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_unet_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/one_leg/one_leg_med_dim_pre_diffusion_unet_ta16_td100/2024-07-04_02-16-16/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 16
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 700
|
||||
best_reward_threshold_for_success: 1
|
||||
specific:
|
||||
headless: true
|
||||
furniture: one_leg
|
||||
randomness: med
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
groupnorm_eps: 1e-4
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,109 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/one_leg/one_leg_med_dim_pre_gaussian_mlp_ta8/2024-06-28_16-27-02/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 700
|
||||
best_reward_threshold_for_success: 1
|
||||
specific:
|
||||
headless: true
|
||||
furniture: one_leg
|
||||
randomness: med
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_ppo.PPO_Gaussian
|
||||
clip_ploss_coef: 0.01
|
||||
randn_clip_value: 3
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims:
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
fixed_std: 0.04
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
cond_dim: ${obs_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,121 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/round_table/round_table_low_dim_pre_diffusion_mlp_ta8_td100/2024-07-23_01-28-26/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: round_table
|
||||
randomness: low
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,123 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_unet_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/round_table/round_table_low_dim_pre_diffusion_unet_ta16_td100/2024-07-04_02-19-48/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 16
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: round_table
|
||||
randomness: low
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
groupnorm_eps: 1e-4
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,109 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/round_table/round_table_low_dim_pre_gaussian_mlp_ta8/2024-06-28_16-26-51/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: round_table
|
||||
randomness: low
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_ppo.PPO_Gaussian
|
||||
clip_ploss_coef: 0.01
|
||||
randn_clip_value: 3
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims:
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
fixed_std: 0.04
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,121 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/round_table/round_table_med_dim_pre_diffusion_mlp_ta8_td100/2024-07-23_01-28-29/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: round_table
|
||||
randomness: med
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,122 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent
|
||||
|
||||
name: ${env_name}_ft_diffusion_unet_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/round_table/round_table_med_dim_pre_diffusion_unet_ta16_td100/2024-07-04_02-20-21/checkpoint/state_8000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
ft_denoising_steps: 5
|
||||
cond_steps: 1
|
||||
horizon_steps: 16
|
||||
act_steps: 8
|
||||
use_ddim: True
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: round_table
|
||||
randomness: med
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion_ppo.PPODiffusion
|
||||
# HP to tune
|
||||
gamma_denoising: 0.9
|
||||
clip_ploss_coef: 0.001
|
||||
clip_ploss_coef_base: 0.001
|
||||
clip_ploss_coef_rate: 3
|
||||
randn_clip_value: 3
|
||||
min_sampling_denoising_std: 0.04
|
||||
#
|
||||
use_ddim: ${use_ddim}
|
||||
ddim_steps: ${ft_denoising_steps}
|
||||
learn_eta: False
|
||||
eta:
|
||||
base_eta: 1
|
||||
input_dim: ${obs_dim}
|
||||
mlp_dims: [256, 256]
|
||||
action_dim: ${action_dim}
|
||||
min_eta: 0.1
|
||||
max_eta: 1.0
|
||||
_target_: model.diffusion.eta.EtaFixed
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
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 +1,109 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.finetune.train_ppo_gaussian_agent.TrainPPOGaussianAgent
|
||||
|
||||
name: ${env_name}_ft_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/round_table/round_table_med_dim_pre_gaussian_mlp_ta8/2024-06-28_16-26-38/checkpoint/state_3000.pt
|
||||
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
cond_steps: 1
|
||||
horizon_steps: 8
|
||||
act_steps: 8
|
||||
|
||||
env:
|
||||
n_envs: 1000
|
||||
name: ${env_name}
|
||||
env_type: furniture
|
||||
max_episode_steps: 1000
|
||||
best_reward_threshold_for_success: 2
|
||||
specific:
|
||||
headless: true
|
||||
furniture: round_table
|
||||
randomness: med
|
||||
normalization_path: ${normalization_path}
|
||||
act_steps: ${act_steps}
|
||||
sparse_reward: True
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_train_itr: 1000
|
||||
n_critic_warmup_itr: 1
|
||||
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
|
||||
gamma: 0.999
|
||||
actor_lr: 1e-5
|
||||
actor_weight_decay: 0
|
||||
actor_lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 10
|
||||
min_lr: 1e-6
|
||||
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: 50
|
||||
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: 8800
|
||||
update_epochs: 5
|
||||
vf_coef: 0.5
|
||||
target_kl: 1
|
||||
|
||||
model:
|
||||
_target_: model.rl.gaussian_ppo.PPO_Gaussian
|
||||
clip_ploss_coef: 0.01
|
||||
randn_clip_value: 3
|
||||
network_path: ${base_policy_path}
|
||||
actor:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims:
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
- 1024
|
||||
activation_type: ReLU
|
||||
residual_style: True
|
||||
fixed_std: 0.04
|
||||
learn_fixed_std: True
|
||||
std_min: 0.01
|
||||
std_max: 0.2
|
||||
cond_dim: ${obs_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
critic:
|
||||
_target_: model.common.critic.CriticObs
|
||||
obs_dim: ${obs_dim}
|
||||
mlp_dims: [512, 512, 512]
|
||||
activation_type: Mish
|
||||
residual_style: True
|
||||
horizon_steps: ${horizon_steps}
|
||||
cond_steps: ${cond_steps}
|
||||
device: ${device}
|
||||
@@ -0,0 +1,72 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: lamp
|
||||
randomness: low
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,74 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
|
||||
|
||||
name: ${env}_pre_diffusion_unet_ta${horizon_steps}_td${denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: lamp
|
||||
randomness: low
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 16
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
groupnorm_eps: 1e-4 # not important
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,64 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_gaussian_agent.TrainGaussianAgent
|
||||
|
||||
name: ${env}_pre_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: lamp
|
||||
randomness: low
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 3000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.common.gaussian.GaussianModel
|
||||
network:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
use_layernorm: False # worse with layernorm
|
||||
activation_type: ReLU # worse with Mish
|
||||
residual_style: True
|
||||
fixed_std: 0.1
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_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 +1,72 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: lamp
|
||||
randomness: med
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,73 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
|
||||
|
||||
name: ${env}_pre_diffusion_unet_ta${horizon_steps}_td${denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: lamp
|
||||
randomness: med
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 16
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,64 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_gaussian_agent.TrainGaussianAgent
|
||||
|
||||
name: ${env}_pre_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: lamp
|
||||
randomness: med
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 3000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.common.gaussian.GaussianModel
|
||||
network:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
use_layernorm: False # worse with layernorm
|
||||
activation_type: ReLU # worse with Mish
|
||||
residual_style: True
|
||||
fixed_std: 0.1
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_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 +1,72 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: one_leg
|
||||
randomness: low
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,74 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
|
||||
|
||||
name: ${env}_pre_diffusion_unet_ta${horizon_steps}_td${denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: one_leg
|
||||
randomness: low
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 16
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
groupnorm_eps: 1e-4 # not important
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,64 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_gaussian_agent.TrainGaussianAgent
|
||||
|
||||
name: ${env}_pre_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: one_leg
|
||||
randomness: low
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 3000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.common.gaussian.GaussianModel
|
||||
network:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
use_layernorm: False # worse with layernorm
|
||||
activation_type: ReLU # worse with Mish
|
||||
residual_style: True
|
||||
fixed_std: 0.1
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_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 +1,72 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: one_leg
|
||||
randomness: med
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,73 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
|
||||
|
||||
name: ${env}_pre_diffusion_unet_ta${horizon_steps}_td${denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: one_leg
|
||||
randomness: med
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 16
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,64 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_gaussian_agent.TrainGaussianAgent
|
||||
|
||||
name: ${env}_pre_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: one_leg
|
||||
randomness: med
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 58
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 10000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.common.gaussian.GaussianModel
|
||||
network:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
use_layernorm: False # worse with layernorm
|
||||
activation_type: ReLU # worse with Mish
|
||||
residual_style: True
|
||||
fixed_std: 0.1
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_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 +1,72 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: round_table
|
||||
randomness: low
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,73 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
|
||||
|
||||
name: ${env}_pre_diffusion_unet_ta${horizon_steps}_td${denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: round_table
|
||||
randomness: low
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 16
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,64 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_gaussian_agent.TrainGaussianAgent
|
||||
|
||||
name: ${env}_pre_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: round_table
|
||||
randomness: low
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 3000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.common.gaussian.GaussianModel
|
||||
network:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
use_layernorm: False # worse with layernorm
|
||||
activation_type: ReLU # worse with Mish
|
||||
residual_style: True
|
||||
fixed_std: 0.1
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_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 +1,72 @@
|
||||
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: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: round_table
|
||||
randomness: med
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
|
||||
time_dim: 32
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
cond_mlp_dims: [512, 64]
|
||||
use_layernorm: True # needed for larger MLP
|
||||
residual_style: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,73 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent
|
||||
|
||||
name: ${env}_pre_diffusion_unet_ta${horizon_steps}_td${denoising_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: round_table
|
||||
randomness: med
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
denoising_steps: 100
|
||||
horizon_steps: 16
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 8000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.diffusion.diffusion.DiffusionModel
|
||||
predict_epsilon: True
|
||||
denoised_clip_value: 1.0
|
||||
network:
|
||||
_target_: model.diffusion.unet.Unet1D
|
||||
diffusion_step_embed_dim: 16
|
||||
dim: 64
|
||||
dim_mults: [1, 2, 4]
|
||||
kernel_size: 5
|
||||
n_groups: 8
|
||||
smaller_encoder: False
|
||||
cond_predict_scale: True
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_steps}
|
||||
obs_dim: ${obs_dim}
|
||||
action_dim: ${action_dim}
|
||||
transition_dim: ${transition_dim}
|
||||
denoising_steps: ${denoising_steps}
|
||||
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 +1,64 @@
|
||||
defaults:
|
||||
- _self_
|
||||
hydra:
|
||||
run:
|
||||
dir: ${logdir}
|
||||
_target_: agent.pretrain.train_gaussian_agent.TrainGaussianAgent
|
||||
|
||||
name: ${env}_pre_gaussian_mlp_ta${horizon_steps}
|
||||
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
|
||||
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl
|
||||
|
||||
seed: 42
|
||||
device: cuda:0
|
||||
task: round_table
|
||||
randomness: med
|
||||
env: ${task}_${randomness}_dim
|
||||
obs_dim: 44
|
||||
action_dim: 10
|
||||
transition_dim: ${action_dim}
|
||||
horizon_steps: 8
|
||||
cond_steps: 1
|
||||
|
||||
wandb:
|
||||
entity: ${oc.env:DPPO_WANDB_ENTITY}
|
||||
project: furniture-${task}-${randomness}-pretrain
|
||||
run: ${now:%H-%M-%S}_${name}
|
||||
|
||||
train:
|
||||
n_epochs: 3000
|
||||
batch_size: 256
|
||||
learning_rate: 1e-4
|
||||
weight_decay: 1e-6
|
||||
lr_scheduler:
|
||||
first_cycle_steps: 10000
|
||||
warmup_steps: 100
|
||||
min_lr: 1e-5
|
||||
epoch_start_ema: 20
|
||||
update_ema_freq: 10
|
||||
save_model_freq: 1000
|
||||
|
||||
model:
|
||||
_target_: model.common.gaussian.GaussianModel
|
||||
network:
|
||||
_target_: model.common.mlp_gaussian.Gaussian_MLP
|
||||
mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]
|
||||
use_layernorm: False # worse with layernorm
|
||||
activation_type: ReLU # worse with Mish
|
||||
residual_style: True
|
||||
fixed_std: 0.1
|
||||
cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
|
||||
horizon_steps: ${horizon_steps}
|
||||
transition_dim: ${transition_dim}
|
||||
horizon_steps: ${horizon_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