Optimize model parameters according to DETR.

liuruijin17 / LSTR

This is an official repository of End-to-end Lane Shape Prediction with Transformers.

BSD 3-Clause "New" or "Revised" License

647 stars 130 forks source link

Optimize model parameters according to DETR. #20

Closed lunachy closed 3 years ago

lunachy commented 3 years ago

Do you make grid search about parameters of enc_layers、dec_layers? According to DETR, more encoder layers and decoder layers are beneficial. Maybe enc_layers and dec_layers set to 3 at least.

liuruijin17 commented 3 years ago

Yes, we also tried it, but the performance was not that significantly improved (or even decreased), probably due to overfitting. We actually set up 2 directly and found the overall effect to be good :joy:, so we didn't carefully tune the number of transformer layers.

lunachy commented 3 years ago

According to the paper, I guess you test this on tusimple dataset.

liuruijin17 commented 3 years ago

Yes, the TuSimple validation dataset.