dense net training from scratch accuracy keeps dropping

[cxy7452]

Hi,

This is great, thanks for sharing the code!

I was using your code with processed TuSimple data trying to train dense net model from scratch according to your training procedure and config. While the training loss keeps decreasing, the accuracy also keeps decreasing (from around 0.5 at the beginning down to now close to 0.03 at the 500th epoch). I am using the same learning rate setup as your suggested setup. Any thoughts on why this is happening?

[MaybeShewill-CV]

@cxy7452 Since I have not find a suitable pre-trained weights on dense net yet all my experiments was done using the vgg structure as the base cnn net

[yiyichun]

Hello, @cxy7452 : How do you train? I have a problem with nan .(https://github.com/MaybeShewill-CV/lanenet-lane-detection/issues/4) Do you need to read the json file during training? Or just need train.txt 、 val.txt、binary segmentation label file 、 instance segmentation label file and the original image ? thank you :)

[cxy7452]

@yiyichun Like @MaybeShewill-CV mentioned, I just downloaded TuSimple dataset, and processed the dataset into the same format that's required for this repo, which is just train.txt, val.txt, binary segmentation label image, instance segmentation label image, and original image. Make sure your label images are correctly done and the paths in your text files are correct as well.

[yiyichun]

Hi , @cxy7452: Thank you for your reply. My training data is as follows: original image:

gt_image_binary:

gt_image_instance:

The path of the above three is recorded in train.txt Is this correct? thank you :)

[MaybeShewill-CV]

@yiyichun The label image looks correct for now. You'd better check the label follow the readme.md file

[yiyichun]

Hi, @MaybeShewill-CV , Does the image file need png file? Is the jpg image file available?

[MaybeShewill-CV]

@yiyichun jpg file is also accepted. But you have to make sure the pixel value of the label image is right.

[stubbornstubborn]

Hi,@yiyichun , How is your gt_image_binary and gt_iamge_instance generated?

[yiyichun]

Hi, @ding-hai-tao , I am modifying data_provider\lanenet_hnet_data_processor.py to get these 2 files. However, I am not sure if this is correct, maybe @MaybeShewill-CV can have correct instructions.

[south-ocean]

Hello, I also meet the problem that dense net training from scratch accuracy keeps dropping , can you share some experience? @cxy7452 @MaybeShewill-CV @yiyichun @ding-hai-tao

[MaybeShewill-CV]

@south-ocean You'd better use pretrained dense net weights params as initialized weights.

[yiyichun]

@south-ocean , I have also encountered the same problem. Try to reduce the learning rate.

[south-ocean]

Your solution didn't resolve my problem, can you leave me a Contact？ Thank you! @yiyichun @MaybeShewill-CV @cxy7452

[MaybeShewill-CV]

@south-ocean You can try this if the solution can resolve your problem. #33 @mukaman84