Miao Zheng comments

Results 80 comments of


Miao Zheng

ValueError: Input and output must have the same number of spatial dimensions,

The shape of the seg map must be [H, W], so the pixel value is 0 or 1 rather than [0,0,0] or [1,1,1]

How to train segmentation model with large dataset?

`workers_per_gpu` is the `num_workers` for pytorch dataloader, and I think `workers_per_gpu=16` is a little high for your CPU, and I suggest you set the `num_workers` lower

How to train segmentation model with large dataset?

I see, and suggest you try to set`workers_per_gpu=0` and `persistent_workers=False` to debug.

SETR Class Weights

You might try to double-check the Config of SETR, and there is no `class_weight` in model cfg. ref: https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/setr_mla.py

result per class returns nan

Would you mind sharing more information about your dataset and how you implement it?

result per class returns nan

Could you please check the num_class in your model configuration?

Binary segmentation may only output zeros

Thanks for your bug reporting. We will fix it asap. @xiexinch please have a look.

When I started testing the data with the model, the values of decode.loss_ce, loss, and aux.loss_ce of the model were always 0, and the values of decode.acc_seg, and aux.acc_seg were always 100.！！！！！

would you like to provide more information about your issue

When I started testing the data with the model, the values of decode.loss_ce, loss, and aux.loss_ce of the model were always 0, and the values of decode.acc_seg, and aux.acc_seg were always 100.！！！！！

I would like to suggest you double-check your dataset, and there might be something wrong with the ground truth.

Add SwinV2(UperNet) and FD-SwinV2 to model_zoo

We are planning these models indeed. However, it won't be finished in a short time due to the lack of developers. Collaborations are appreciated If you are interested in it.