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 11:51:47 -05:00
committed by Allen Z. Ren
parent fc42865c77
commit 9032d02eae
28 changed files with 55 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files
cfg/gym/eval/walker2d-v2/eval_diffusion_mlp.yaml
+2 -2
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