`minADE` During Training

Question

`minADE` During Training

Closed this issue 2 years ago · 5 comments

hi, @stepankonev!
I have trained the model for about 40 epochs, using this configuration. An evaluation of minADE on the validation set of waymo motion is made , and the picture below is the outcome. Is the change of this metrics normal? Other configurations like setting trainable_cov = false have also been tried, and get analogous outcome on the change of minADE.
Thanks.

Answer 1 · 2022-08-05T16:00:32.000Z

Hello! Could you clarify what you mean by change please? Basically speaking I mot sure that I have the track of minADE while training.

Answer 2 · 2022-09-03T10:52:30.000Z

Hi @MichaelYu781, Could you explain what does the y-axis represent in your plot? I just split the training folder of the waymo motion dataset into (90%-10%) train-dev sets. But when I evaluate the single decoder model on the dev set with about 100, 000 training steps, the minADE is still above 16+ meters.

Answer 3 · 2022-09-05T12:54:03.000Z

Hi, @Jerrrrykun. I follows the definition https://waymo.com/open/challenges/2022/motion-prediction/ here.

Answer 4 · 2022-09-07T02:13:38.000Z

Hi, @Jerrrrykun. I follows the definition https://waymo.com/open/challenges/2022/motion-prediction/ here.

Thanks for your reply. I am using the same definition in my own implementation, too. I have solved it. My problem lies in the normalization part.

Answer 5 · 2022-11-18T07:19:42.000Z

hi, @Jerrrrykun. May you further explain the solution of the large minADE problem? I have trained the model with more than 120 epochs, but the minADE result is even worse than what you previously got.