search

lhoyer / DAFormer

[CVPR22] Official Implementation of DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation
Other
466 stars 91 forks source link

Global understanding of the `forward_train` function in dacs.py #75

Open adriengoleb opened 1 year ago

adriengoleb commented 1 year ago

Hello,

I'd like to understand a little better how this function works in the mmseg/models/uda/dacs.py file.

During training, subpots are displayed and filed in the folder work_dirs/local_basic/class_mix_debug.

How are these subplots updated? I've noticed that they don't change during training and remain identical for each image. What's the main purpose of these subplots ?

Finally, what does the "FDist Mask" plot correspond to ? How can I interpret it ?

Thanks for your feedback