dodox/dppo
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

add evaluation agents and some example configs

Browse Source
This commit is contained in:
allenzren
2024-09-17 16:32:45 -04:00
parent bc52beca1e
commit c9f24ba0c3
13 changed files with 1120 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
cfg/furniture/eval/one_leg_low/eval_diffusion_mlp.yaml
+67
View File
@@ -0,0 +1,67 @@
defaults:
- _self_
hydra:
run:
dir: ${logdir}
_target_: agent.eval.eval_diffusion_agent.EvalDiffusionAgent
name: ${env_name}_eval_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-eval/${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
cond_steps: 1
horizon_steps: 8
act_steps: 8
use_ddim: True
ddim_steps: 5
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
render_num: 0
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}
obs_steps: ${cond_steps}
act_steps: ${act_steps}
sparse_reward: True
model:
_target_: model.diffusion.diffusion.DiffusionModel
predict_epsilon: True
denoised_clip_value: 1.0
randn_clip_value: 3
#
use_ddim: ${use_ddim}
ddim_steps: ${ddim_steps}
network_path: ${base_policy_path}
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}
denoising_steps: ${denoising_steps}
device: ${device}
cfg/gym/eval/hopper-v2/eval_diffusion_mlp.yaml
+62
View File
@@ -0,0 +1,62 @@
defaults:
- _self_
hydra:
run:
dir: ${logdir}
_target_: agent.eval.eval_diffusion_agent.EvalDiffusionAgent
name: ${env_name}_eval_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/gym-eval/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
base_policy_path: ${oc.env:DPPO_LOG_DIR}/gym-pretrain/hopper-medium-v2_pre_diffusion_mlp_ta4_td20/2024-06-12_23-10-05/checkpoint/state_3000.pt
normalization_path: ${oc.env:DPPO_DATA_DIR}/gym/${env_name}/normalization.npz
seed: 42
device: cuda:0
env_name: hopper-medium-v2
obs_dim: 11
action_dim: 3
transition_dim: ${action_dim}
denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
n_steps: 500 # each episode can take maximum (max_episode_steps / act_steps, =250 right now) steps but may finish earlier in gym. We only count episodes finished within n_steps for evaluation.
render_num: 0
env:
n_envs: 40
name: ${env_name}
max_episode_steps: 1000
reset_at_iteration: False
save_video: False
best_reward_threshold_for_success: 3 # success rate not relevant for gym tasks
wrappers:
mujoco_locomotion_lowdim:
normalization_path: ${normalization_path}
multi_step:
n_obs_steps: ${cond_steps}
n_action_steps: ${act_steps}
max_episode_steps: ${env.max_episode_steps}
reset_within_step: True
model:
_target_: model.diffusion.diffusion.DiffusionModel
predict_epsilon: True
denoised_clip_value: 1.0
#
network_path: ${base_policy_path}
network:
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
time_dim: 16
mlp_dims: [512, 512, 512]
activation_type: ReLU
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}
denoising_steps: ${denoising_steps}
device: ${device}
cfg/robomimic/eval/can/eval_diffusion_mlp.yaml
+66
View File
@@ -0,0 +1,66 @@
defaults:
- _self_
hydra:
run:
dir: ${logdir}
_target_: agent.eval.eval_diffusion_agent.EvalDiffusionAgent
name: ${env_name}_eval_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-eval/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/can/can_pre_diffusion_mlp_ta4_td20/2024-06-28_13-29-54/checkpoint/state_5000.pt # use 8000 for comparing policy parameterizations
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
seed: 42
device: cuda:0
env_name: can
obs_dim: 23
action_dim: 7
transition_dim: ${action_dim}
denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
n_steps: 300 # each episode takes max_episode_steps / act_steps steps
render_num: 0
env:
n_envs: 50
name: ${env_name}
best_reward_threshold_for_success: 1
max_episode_steps: 300
save_video: False
wrappers:
robomimic_lowdim:
normalization_path: ${normalization_path}
low_dim_keys: ['robot0_eef_pos',
'robot0_eef_quat',
'robot0_gripper_qpos',
'object'] # same order of preprocessed observations
multi_step:
n_obs_steps: ${cond_steps}
n_action_steps: ${act_steps}
max_episode_steps: ${env.max_episode_steps}
reset_within_step: True
model:
_target_: model.diffusion.diffusion.DiffusionModel
predict_epsilon: True
denoised_clip_value: 1.0
randn_clip_value: 3
#
network_path: ${base_policy_path}
network:
_target_: model.diffusion.mlp_diffusion.DiffusionMLP
time_dim: 16
mlp_dims: [512, 512, 512]
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}
denoising_steps: ${denoising_steps}
device: ${device}
cfg/robomimic/eval/can/eval_diffusion_mlp_img.yaml
+98
View File
@@ -0,0 +1,98 @@
defaults:
- _self_
hydra:
run:
dir: ${logdir}
_target_: agent.eval.eval_diffusion_img_agent.EvalImgDiffusionAgent
name: ${env_name}_eval_diffusion_mlp_img_ta${horizon_steps}_td${denoising_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-eval/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/can/can_pre_diffusion_mlp_img_ta4_td100/2024-07-30_22-23-55/checkpoint/state_5000.pt
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}-img.json
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}-img/normalization.npz
seed: 42
device: cuda:0
env_name: can
obs_dim: 9
action_dim: 7
transition_dim: ${action_dim}
denoising_steps: 100
cond_steps: 1
img_cond_steps: 1
horizon_steps: 4
act_steps: 4
use_ddim: True
ddim_steps: 5
n_steps: 300 # each episode takes max_episode_steps / act_steps steps
render_num: 0
env:
n_envs: 50
name: ${env_name}
best_reward_threshold_for_success: 1
max_episode_steps: 300
save_video: False
use_image_obs: True
wrappers:
robomimic_image:
normalization_path: ${normalization_path}
low_dim_keys: ['robot0_eef_pos',
'robot0_eef_quat',
'robot0_gripper_qpos']
image_keys: ['robot0_eye_in_hand_image']
shape_meta: ${shape_meta}
multi_step:
n_obs_steps: ${cond_steps}
n_action_steps: ${act_steps}
max_episode_steps: ${env.max_episode_steps}
reset_within_step: True
shape_meta:
obs:
rgb:
shape: [3, 96, 96]
state:
shape: [9]
action:
shape: [7]
model:
_target_: model.diffusion.diffusion.DiffusionModel
predict_epsilon: True
denoised_clip_value: 1.0
randn_clip_value: 3
#
use_ddim: ${use_ddim}
ddim_steps: ${ddim_steps}
network_path: ${base_policy_path}
network:
_target_: model.diffusion.mlp_diffusion.VisionDiffusionMLP
backbone:
_target_: model.common.vit.VitEncoder
obs_shape: ${shape_meta.obs.rgb.shape}
num_channel: ${eval:'3 * ${img_cond_steps}'} # each image patch is history concatenated
img_h: ${shape_meta.obs.rgb.shape[1]}
img_w: ${shape_meta.obs.rgb.shape[2]}
cfg:
patch_size: 8
depth: 1
embed_dim: 128
num_heads: 4
embed_style: embed2
embed_norm: 0
augment: False
spatial_emb: 128
time_dim: 32
mlp_dims: [512, 512, 512]
residual_style: True
img_cond_steps: ${img_cond_steps}
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}
denoising_steps: ${denoising_steps}
device: ${device}
cfg/robomimic/eval/can/eval_gaussian_mlp.yaml
+60
View File
@@ -0,0 +1,60 @@
defaults:
- _self_
hydra:
run:
dir: ${logdir}
_target_: agent.eval.eval_gaussian_agent.EvalGaussianAgent
name: ${env_name}_eval_gaussian_mlp_ta${horizon_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-eval/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/can/can_pre_gaussian_mlp_ta4/2024-06-28_13-31-00/checkpoint/state_5000.pt
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
seed: 42
device: cuda:0
env_name: can
obs_dim: 23
action_dim: 7
transition_dim: ${action_dim}
cond_steps: 1
horizon_steps: 4
act_steps: 4
n_steps: 300 # each episode takes max_episode_steps / act_steps steps
render_num: 0
env:
n_envs: 50
name: ${env_name}
best_reward_threshold_for_success: 1
max_episode_steps: 300
save_video: False
wrappers:
robomimic_lowdim:
normalization_path: ${normalization_path}
low_dim_keys: ['robot0_eef_pos',
'robot0_eef_quat',
'robot0_gripper_qpos',
'object'] # same order of preprocessed observations
multi_step:
n_obs_steps: ${cond_steps}
n_action_steps: ${act_steps}
max_episode_steps: ${env.max_episode_steps}
reset_within_step: True
model:
_target_: model.common.gaussian.GaussianModel
randn_clip_value: 3
#
network_path: ${base_policy_path}
network:
_target_: model.common.mlp_gaussian.Gaussian_MLP
mlp_dims: [512, 512, 512]
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}
cfg/robomimic/eval/can/eval_gaussian_mlp_img.yaml
+87
View File
@@ -0,0 +1,87 @@
defaults:
- _self_
hydra:
run:
dir: ${logdir}
_target_: agent.eval.eval_gaussian_img_agent.EvalImgGaussianAgent
name: ${env_name}_eval_gaussian_mlp_img_ta${horizon_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-eval/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/can/can_pre_gaussian_mlp_img_ta4/2024-07-28_21-54-40/checkpoint/state_1000.pt
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}-img.json
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}-img/normalization.npz
seed: 42
device: cuda:0
env_name: can
obs_dim: 9
action_dim: 7
transition_dim: ${action_dim}
cond_steps: 1
img_cond_steps: 1
horizon_steps: 4
act_steps: 4
n_steps: 300 # each episode takes max_episode_steps / act_steps steps
render_num: 0
env:
n_envs: 50
name: ${env_name}
best_reward_threshold_for_success: 1
max_episode_steps: 300
save_video: False
use_image_obs: True
wrappers:
robomimic_image:
normalization_path: ${normalization_path}
low_dim_keys: ['robot0_eef_pos',
'robot0_eef_quat',
'robot0_gripper_qpos']
image_keys: ['robot0_eye_in_hand_image']
shape_meta: ${shape_meta}
multi_step:
n_obs_steps: ${cond_steps}
n_action_steps: ${act_steps}
max_episode_steps: ${env.max_episode_steps}
reset_within_step: True
shape_meta:
obs:
rgb:
shape: [3, 96, 96]
state:
shape: [9]
action:
shape: [7]
model:
_target_: model.common.gaussian.GaussianModel
randn_clip_value: 3
#
network_path: ${base_policy_path}
network:
_target_: model.common.mlp_gaussian.Gaussian_VisionMLP
backbone:
_target_: model.common.vit.VitEncoder
obs_shape: ${shape_meta.obs.rgb.shape}
num_channel: ${eval:'3 * ${img_cond_steps}'} # each image patch is history concatenated
img_h: ${shape_meta.obs.rgb.shape[1]}
img_w: ${shape_meta.obs.rgb.shape[2]}
cfg:
patch_size: 8
depth: 1
embed_dim: 128
num_heads: 4
embed_style: embed2
embed_norm: 0
augment: False
spatial_emb: 128
mlp_dims: [512, 512, 512]
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}