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

change default ft_denoising_steps in eval configs to 0 (assume evaluating pre-trained models)

Browse Source
This commit is contained in:
allenzren 2025-02-04 11:48:59 -05:00 committed by Allen Z. Ren
parent fc42865c77
commit 9032d02eae
28 changed files with 55 additions and 55 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
cfg/d3il/eval/avoid_m1/eval_diffusion_mlp.yaml
View File

@ -19,7 +19,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 25
render_num: 40
@ -48,7 +48,7 @@ env:
reset_within_step: False
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/furniture/eval/lamp_low/eval_diffusion_mlp.yaml
View File

@ -21,7 +21,7 @@ horizon_steps: 8
act_steps: 8
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
render_num: 0
@ -42,7 +42,7 @@ env:
sparse_reward: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/furniture/eval/lamp_low/eval_diffusion_unet.yaml
View File

@ -21,7 +21,7 @@ horizon_steps: 16
act_steps: 8
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
render_num: 0
@ -42,7 +42,7 @@ env:
sparse_reward: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/furniture/eval/one_leg_low/eval_diffusion_mlp.yaml
View File

@ -21,7 +21,7 @@ horizon_steps: 8
act_steps: 8
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
render_num: 0
@ -42,7 +42,7 @@ env:
sparse_reward: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/furniture/eval/one_leg_low/eval_diffusion_unet.yaml
View File

@ -21,7 +21,7 @@ horizon_steps: 16
act_steps: 8
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
render_num: 0
@ -42,7 +42,7 @@ env:
sparse_reward: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/furniture/eval/round_table_low/eval_diffusion_mlp.yaml
View File

@ -21,7 +21,7 @@ horizon_steps: 8
act_steps: 8
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
render_num: 0
@ -42,7 +42,7 @@ env:
sparse_reward: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/furniture/eval/round_table_low/eval_diffusion_unet.yaml
View File

@ -21,7 +21,7 @@ horizon_steps: 16
act_steps: 8
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
render_num: 0
@ -42,7 +42,7 @@ env:
sparse_reward: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/gym/eval/halfcheetah-v2/eval_diffusion_mlp.yaml
View File

@ -19,7 +19,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 250 # 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
@ -41,7 +41,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/gym/eval/hopper-v2/eval_diffusion_mlp.yaml
View File

@ -19,7 +19,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 250 # 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
@ -41,7 +41,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/gym/eval/kitchen-v0/eval_diffusion_mlp.yaml
View File

@ -19,7 +19,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 70
render_num: 0
@ -41,7 +41,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/gym/eval/walker2d-v2/eval_diffusion_mlp.yaml
View File

@ -19,7 +19,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 250 # 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
@ -41,7 +41,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/can/eval_diffusion_mlp.yaml
View File

@ -20,7 +20,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 300 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -45,7 +45,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/can/eval_diffusion_mlp_img.yaml
View File

@ -23,7 +23,7 @@ horizon_steps: 4
act_steps: 4
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: 300 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -59,7 +59,7 @@ shape_meta:
shape: [7]
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/can/eval_diffusion_unet.yaml
View File

@ -20,7 +20,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 75 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -45,7 +45,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/can/eval_diffusion_unet_img.yaml
View File

@ -23,7 +23,7 @@ horizon_steps: 4
act_steps: 4
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: 300 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -59,7 +59,7 @@ shape_meta:
shape: [7]
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/lift/eval_diffusion_mlp.yaml
View File

@ -20,7 +20,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 300 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -45,7 +45,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/lift/eval_diffusion_mlp_img.yaml
View File

@ -23,7 +23,7 @@ horizon_steps: 4
act_steps: 4
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: 300 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -59,7 +59,7 @@ shape_meta:
shape: [7]
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/lift/eval_diffusion_unet.yaml
View File

@ -20,7 +20,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 75 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -45,7 +45,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/lift/eval_diffusion_unet_img.yaml
View File

@ -23,7 +23,7 @@ horizon_steps: 4
act_steps: 4
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: 300 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -59,7 +59,7 @@ shape_meta:
shape: [7]
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/square/eval_diffusion_mlp.yaml
View File

@ -20,7 +20,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 400 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -45,7 +45,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/square/eval_diffusion_mlp_img.yaml
View File

@ -23,7 +23,7 @@ horizon_steps: 4
act_steps: 4
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: 400 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -59,7 +59,7 @@ shape_meta:
shape: [7]
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/square/eval_diffusion_unet.yaml
View File

@ -20,7 +20,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 4
act_steps: 4
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 100 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -45,7 +45,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/square/eval_diffusion_unet_img.yaml
View File

@ -23,7 +23,7 @@ horizon_steps: 4
act_steps: 4
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: 400 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -59,7 +59,7 @@ shape_meta:
shape: [7]
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/transport/eval_diffusion_mlp.yaml
View File

@ -20,7 +20,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 8
act_steps: 8
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 400 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -48,7 +48,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/transport/eval_diffusion_mlp_img.yaml
View File

@ -23,7 +23,7 @@ horizon_steps: 8
act_steps: 8
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: 200 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -63,7 +63,7 @@ shape_meta:
shape: [14]
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/transport/eval_diffusion_unet.yaml
View File

@ -20,7 +20,7 @@ denoising_steps: 20
cond_steps: 1
horizon_steps: 16
act_steps: 8
ft_denoising_steps: 10
ft_denoising_steps: 0
n_steps: 100 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -48,7 +48,7 @@ env:
reset_within_step: True
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

4
cfg/robomimic/eval/transport/eval_diffusion_unet_img.yaml
View File

@ -23,7 +23,7 @@ horizon_steps: 16
act_steps: 8
use_ddim: True
ddim_steps: 5
ft_denoising_steps: 5
ft_denoising_steps: 0
n_steps: 400 # each episode takes max_episode_steps / act_steps steps
render_num: 0
@ -63,7 +63,7 @@ shape_meta:
shape: [14]
model:
_target_: model.diffusion.diffusion_eval_ft.DiffusionEvalFT
_target_: model.diffusion.diffusion_eval_ft.DiffusionEval
ft_denoising_steps: ${ft_denoising_steps}
predict_epsilon: True
denoised_clip_value: 1.0

2
model/diffusion/diffusion_eval.py
View File

@ -20,8 +20,8 @@ class DiffusionEval(DiffusionModel):
def __init__(
self,
network_path,
ft_denoising_steps, # if running pre-trained model (not fine-tuned), set it to zero; if running fine-tuned model, need to specify the correct number of denoising steps fine-tuned, so that here it knows which model (base or ft) to use for each denoising step
use_ddim=False,
ft_denoising_steps=0, # if running fine-tuned model, need to specify the correct number of denoising steps fine-tuned, so that here it knows which model (base or ft) to use for each denoising step
**kwargs,
):
# do not let base class load model