Miao Zheng

Of course, you can! I think to-do list is complete, but for pre-trained ckpts, we will train them with your code and then upload them. Thanks for your contribution. We...

Yes. SPynet is in our plan. BTW, would you like to participate in this work? You might create a pr for this algorithm or help to review our code?

Many thanks for your support! It means a lot to us.

Would you mind proving the implementation of your dataset and the documentation link you referred to?

I think you might print `kwargs` after `def __init__(self, **kwargs)` before `super` to check the `times` in `kwargs`

Do you use `RepeatDataet` in dataset config? https://github.com/open-mmlab/mmsegmentation/blob/4eaa8e69191cc293b64dafe47f1f88a7d468c93c/mmseg/datasets/dataset_wrappers.py#L176

just remove `times=40000` from your config ```python data = dict( samples_per_gpu=2, workers_per_gpu=2, train=dict( type=dataset_type, times=40000, ... ```

I suggest you try to save the checkpoints at different times, like 1k epoch 10k epoch 20k epoch, and check whether the model weights are abnormal.

There might be imbalance labels in your dataset, and some labels like `background` and `trans_lane` are in the large proportion in one image. I suggest you use class_weight in loss...

Please check the label of prediction is all 0, as there are `PALETTE` for this dataset, and label 0 is corresponding to [0, 0, 0] just black. https://github.com/open-mmlab/mmsegmentation/blob/4eaa8e69191cc293b64dafe47f1f88a7d468c93c/mmseg/datasets/isaid.py#L20-L29