training time consume about persformer_3dlane HOT 9 CLOSED

Hi, it tasks about half an hour on 8 v100 GPUs. So one hour for one epoch on 4 3090 GPUs is reasonable.

from persformer_3dlane.

Thanks for your reply!
By the way, if I use 4 3090 with batch size 8 per GPU, do I need to train 100 epoch to reach F1=0.47 as stated in this repo?
I've trained 20 epochs with 4 3090, I didn't change learning rate, the F1 is around 43 (20 epoch), and didn't get better after epoch 13.
I'm wondering if the training config needed to be modified?

BW

from persformer_3dlane.

Hi, refer to previous experiment, it will achieve the best score after about 40~50 epochs. You can set the ‘nepochs’ to 50.

from persformer_3dlane.

Thanks for your reply.
I trained 50 epochs on 4 3090, the F-measure is 0.454
I also trained 70 epochs on 8 3090, the F-measure is 0.4447
I didn't change any hyper-params.
I'm wondering if there's any thing wrong with my training.

BW

from persformer_3dlane.

Hi, for validating the performance , send your email to us and a best model would share with you.

from persformer_3dlane.

我在3090TI上训练了整整4天100轮，按照代码中默认的参数，但是最终的分数是0.66102999，我也想要最好的模型，可以给我也发一份吗😄. 我的邮箱是[email protected]

from persformer_3dlane.

@MyEyeAndEar i have sent it.

from persformer_3dlane.

Thank you so much！

from persformer_3dlane.

I trained 50 epochs on 4 3090, and the F-measure is 0.4672，Could you please send a best model to me for the further research？
My email is [email protected]

from persformer_3dlane.