Questions about the data stream

[Yulu-gan]

Thanks for your solid and excellent work! but I have some questions. Under the mmseg framework, where did you call mmseg / models / uda / dacs.py, and RSC as wll as the whole backbone training process are all in dacs.py? I'm looking forward to your reply.

[Yulu-gan]

I've found '_dacs_a999fdthings.py' under .../base/uda ， is this the config we use? but I find '_nameuda' in '_gta2cs_uda_warm_fdthings_rcs_croppl_a999_daformer_mitb5_s0.py' is **'dacs_a999_fd_things_rcs0.01cpl**'

[lhoyer]

Thank you for your interest in our work!

If you want to train DAFormer with the standalone config file, please have a look at: https://github.com/lhoyer/DAFormer/blob/master/configs/daformer/gta2cs_uda_warm_fdthings_rcs_croppl_a999_daformer_mitb5_s0.py

You can find the relevant config flags for the feature distance in line 21-23 and the config flags for rare class sampling in line 30-31.

The feature distance is implemented here: https://github.com/lhoyer/DAFormer/blob/master/mmseg/models/uda/dacs.py#L157

The rare class sampling is implemented here: https://github.com/lhoyer/DAFormer/blob/master/mmseg/datasets/uda_dataset.py

The file configs/base/uda/dacs_a999_fdthings.py is only used for automatic config file generation using experiments.py. So, you can ignore is when you use configs/daformer/gta2cs_uda_warm_fdthings_rcs_croppl_a999_daformer_mitb5_s0.py

[Yulu-gan]

Got it. Thank you for your patience.