search

voldemortX / pytorch-auto-drive

PytorchAutoDrive: Segmentation models (ERFNet, ENet, DeepLab, FCN...) and Lane detection models (SCNN, RESA, LSTR, LaneATT, BézierLaneNet...) based on PyTorch with fast training, visualization, benchmarking & deployment help
BSD 3-Clause "New" or "Revised" License
837 stars 137 forks source link

How can I modify Input size of model? #115

Closed daeunni closed 1 year ago

daeunni commented 1 year ago

Hi, thank you for your impressive work ! :D

btw, I'd like to unify model's input sizes of TUSimple and CULane dataset . and I thought it can if I modify below part in the cfg file.

https://github.com/voldemortX/pytorch-auto-drive/blob/2b0d5ec5f7536c9d2d2b6d8498718a8fca2ab276/configs/lane_detection/baseline/erfnet_culane.py#L27

however when I revise the input size in config file, it raise some error like that. RuntimeError: input and target batch or spatial sizes don't match: target [20, 288, 800], input [20, 5, 360, 640]

How can I solve that? Is there anyone who know how to unify TUSimple & CULane model's input size? Best,

voldemortX commented 1 year ago

@daeunni I suppose you want to change 288 to 360? You'll have to change all 288 related things, including data processing (L5-6) and the flatten size (L65) for segmentation based methods.

voldemortX commented 1 year ago

train & test settings as well (L27 L50)

daeunni commented 1 year ago

Yes! I want to CULane model's current default input size input_size=(288, 800) to input_size=(360, 640) (same as TUSimple model's)

Okay thank you for your quick reply. Is there any problem for model or performance after changing input size ?

voldemortX commented 1 year ago

Okay thank you for your quick reply. Is there any problem for model or performance after changing input size ?

That wasn't tested properly, though I suspect the influence should be small. The main problem comes with the change of aspect ratio.

daeunni commented 1 year ago

Oh thank you for your quick reply :))

however, if I change CULane's input size (288, 800) to (360, 640), I think the aspect ratio is changed 1 : 2.77(=800/288) to 1 : 1.77(=640/360). Is it okay? :0

I'm new to Lane Detection, so if my thought was wrong about aspect ratio, plz tell me.

voldemortX commented 1 year ago

Yes the difference between original image (590/1640) & training size (360/640) could bring unnatural distortions to the image contents. Ideally, the aspect ratio should remain similar to original.

daeunni commented 1 year ago

Oh okay I understand :) thank you for your kind reply! I'll ask another questions if I can't solve this problem :D