About train code

[Asthestarsfalll]

When training, how did the miou and accuracy calculate? On train dataset or validate dataset? I think it's calculated on val dataset due to https://github.com/RolandGao/RegSeg/blob/main/train.py#L238. I trained the base regseg model with config cityscapes_trainval_1000epochs.yam on Cityscapes and got the unbelievable results.

[Asthestarsfalll]

Well, I find the max_eval which might be the reason of results.

[RolandGao]

In the yaml config, the following two fields show the data during training and during validation:

train_split: trainval val_split: val

The trainval config trains on train+val and is meant for submitting to the test server. During training, it evaluates on the val set (which is seen during traing) only as a sanity check, so that's why the accuracy is so high.

The cityscapes_1000epochs.yaml config trains on train only and evaluates on val, so this might suit your needs better. Pay attention to the fields train_crop_size and loss_type. cityscapes_trainval_1000epochs.yaml uses a larger crop size and a better loss function than cityscapes_1000epochs.yaml at the cost of longer training time, so you can copy the two fields over if you can afford longer the training time.

[RolandGao]

Btw, it has nothing to do with max_eval. There are 500 images in the validation set and max_eval is set to 600, so no issues there

[Asthestarsfalll]

Thank you very much! I misunderstood the trainval as evaluating model during training and the train as not evaluating model.

[Asthestarsfalll]

@RolandGao
Hi, I wonder what is class_uniform_sampling, I can't find it on google and github. Could you please give the source of it, or paper. Thank you a lot !

[RolandGao]

It's first introduced by Nvidia in this paper https://arxiv.org/abs/1812.01593. The code is from here https://github.com/NVIDIA/semantic-segmentation.

[Asthestarsfalll]

@RolandGao Hi, I have send a email about some research of RegSeg to you, Did you received that?