what cfgs used in training on CULane dataset?

harryhan618 / SCNN_Pytorch

Pytorch implementation of "Spatial As Deep: Spatial CNN for Traffic Scene Understanding"

MIT License

250 stars 68 forks source link

what cfgs used in training on CULane dataset? #46

Closed zlxu2011 closed 4 years ago

zlxu2011 commented 4 years ago

Hi, Recently, when I test with model you provided on culane dataset, I found that the model has trained at least 13 epochs, but according cfg file from experiments/exp10, it will only training maximum 11 epochs (L212), so can you tell me what cfgs you used in training on culane dataset? Besides, the lr (16e-2) is also a little bit large and will cause loss to be Nan.

harryhan618 commented 4 years ago

The number of epoch doesn't really matter that much. I remember around 10k iterations would be fine. Also, I trained with 8 GPUs. So maybe you can set lr to 1e-2 or so.

zlxu2011 commented 4 years ago

OK, I got it! Thank you for your reply.