xingangpan / ocda-driving-example Goto Github PK

Hi. Thanks for your interesting paper!
My question is how to obtain the label as 'xxx_train_id.png' in the C-Driving. In the original BDD100k dataset, I find only segmentation labels for image-10k are provided, without weather information. As for image-100k with weather information, there is no segmentation label. https://doc.bdd100k.com/download.html#semantic-segmentation
In your C-Driving dataset, there are subsets (different weather) of images from image-100k with segmentation labels.
Could you please tell me how to get the segmentation labels for these images?
Thanks!

Hello! Thank you for you great work!

I found a question that the pretrained model is resnet101, but the used model is vgg16_bn.

This question is bothering me and I look forward to your reply!

Thanks for your code. The paper claims that OCDA-Driving uses dynamic transferable embedding and curriculum training in addition to adversarial adaptation, but I cannot find the corresponding code.

compute_iou.py:18: RuntimeWarning: invalid value encountered in true_divide
return np.diag(hist) / (hist.sum(1) + hist.sum(0) - np.diag(hist))
i wanna know why this error occor when i test the training module?Thank you!

When I use the normal BN instead of SyncBN, that is, I set "--model DeeplabVGG".

The gradient and loss explode in a few iterations.

iter =        0/   40000, loss_seg1 = 2.930 loss_seg2 = 0.000 loss_adv1 = 0.691, loss_adv2 = 0.000 loss_D1 = 0.693 loss_D2 = 0.000
iter =        1/   40000, loss_seg1 = 52.802 loss_seg2 = 0.000 loss_adv1 = 0.925, loss_adv2 = 0.000 loss_D1 = 0.714 loss_D2 = 0.000
iter =        2/   40000, loss_seg1 = 643.913 loss_seg2 = 0.000 loss_adv1 = 0.000, loss_adv2 = 0.000 loss_D1 = 15.293 loss_D2 = 0.000
iter =        3/   40000, loss_seg1 = 26799679471616.000 loss_seg2 = 0.000 loss_adv1 = 5.030, loss_adv2 = 0.000 loss_D1 = 2.718 loss_D2 = 0.000

do you know how to solve it?

Thanks!

thank you for your submission, but i can't get a good training result, the pretrained model can't exceed a good result too