search

PJLab-ADG / DriveArena

DriveArena: A Closed-loop Generative Simulation Platform for Autonomous Driving
https://pjlab-adg.github.io/DriveArena/
Apache License 2.0
253 stars 18 forks source link

Train question #32

Closed nkuwangfeng closed 17 hours ago

nkuwangfeng commented 3 days ago

I saw in the paper that you trained 200k batches, does it have to be trained for such a long time to see the effect? I have trained 50K so far, and the result still looks very poor with an average loss of 0.12, which has not decreased for a long time. The effect of val is only visible in the outline, and the details are not very clear. I would like to ask if this is normal for you and if it is necessary to continue waiting?

nkuwangfeng commented 3 days ago

WorldDreamer

zzz5y commented 3 days ago

Hello may i ask you what device are you using for training! I want to train my model too.

nkuwangfeng commented 3 days ago

Hello may i ask you what device are you using for training! I want to train my model too.

A100 40G 8卡

zzz5y commented 3 days ago

Hello may i ask you what device are you using for training! I want to train my model too.

A100 40G 8卡

好的好的 谢谢！

zzz5y commented 2 days ago

Hello,May i ask you how long do you need for training a iteration?

jokester-zzz commented 1 day ago

The complete training process of 200k steps took approximately 2-3 days.

And after about 10k steps, we achieved the results shown in the figure. (The image below is what we generated.)

image

zzz5y commented 1 day ago

The complete training process of 200k steps took approximately 2-3 days.

And after about 10k steps, we achieved the results shown in the figure. (The image below is what we generated.)

image

OK, thanks you for reply!

jokester-zzz commented 17 hours ago

I will close this issue. Please feel free to reopen it if you have any further questions or concerns~