search

NVlabs / OmniDrive

235 stars 12 forks source link

poor detection accuracy #17

Open troylhy1991 opened 2 months ago

troylhy1991 commented 2 months ago

the planning evaluation is close to what the paper reported after training for 6 epochs. however, the detection scores are very low, is this expected? ;(

exiawsh commented 1 month ago

Yes. Detection requires long time to converge. 12 epoch can obtain a relative good results. 24 epochs for full convergence.

troylhy1991 commented 1 month ago

many thx! turns out i was using default lr 1e-4 while my batchsize was only half. but it kind of introduces another confusing part to me: the planning sub task might be too easy to overfit ...... ;(