Online finetuning runs (sometimes crash because of nans)

This commit is contained in:
Cadene
2024-02-16 15:13:24 +00:00
parent 228c045674
commit c202c2b3c2
5 changed files with 165 additions and 110 deletions

View File

@@ -80,7 +80,7 @@ expectile: 0.9
A_scaling: 3.0
# offline->online
offline_steps: ${train_steps}/2
offline_steps: 25000 # ${train_steps}/2
pretrained_model_path: ""
balanced_sampling: true
demo_schedule: 0.5