search

koyeongmin / PINet

MIT License
207 stars 56 forks source link

Trying to train on other dataset, but the result is not good #12

Open FCInter opened 4 years ago

FCInter commented 4 years ago

I manage to reimplement similar test result on Tusimple dataset using your code. Thank you very much!

But when I'm trying to train on my own dataset, the result is not very good. The output lane by the model has zigzag shapes which look very strange. I load the pretrained .pkl checkpoint you provided.

I tested learning rate 1e-4, 2e-4, batchsize 16, 32. The result does not change much.

Is there any other parameter that I can finetune?

Or is there something wrong with my implementation.

I did not change much to the code. Only add some lines to the preprocessing code, to format my dataset to the Tusimple-style, and I have double-checked that these changes did not cause anything wrong.

koyeongmin commented 4 years ago

The provided weight file was trained on the tuSimple dataset only, and this dataset is not general because it consists of only sunny day and highway images. So the file can not be good pretrained model. I suggest to try to train from scratch. I am also trying to train the model on other dataset, Culane. It takes more time than tuSimple.

SoulmateB commented 4 years ago

hello, do you train the model on CULane? how is the result？ I train on CULane，the result is not good. Is there any other parameter that I can finetune?

koyeongmin commented 4 years ago

I am also trying to train the model on CULane. When I get the result, I will update about it. Thank you!

lucasjinreal commented 4 years ago

@koyeongmin How's the result on CULane?

koyeongmin commented 4 years ago

https://github.com/koyeongmin/PINet_new This repository includes CULane result! The concept is same, but networks and training parts are fixed a little bit.