search

vikolss / DACS

Code from the paper "DACS: Domain Adaptation via Cross-domain Mixed Sampling"
MIT License
123 stars 19 forks source link

The parameters of the experiment of the cowMix mask #8

Closed manmanjun closed 3 years ago

manmanjun commented 3 years ago

In the /utils/transformmasks.py, how are the parameters of sigma, p set in the functions generate_cow_mask?

vikolss commented 3 years ago

Hi, sorry for late response. The parameters used are the same as the ones used in the paper ClassMix: Segmentation-Based Data Augmentation for Semi-Supervised Learning, namely p = 0.5 and sigma taken randomly in the range [8,32]. See this repo for details on the generation and implementation. The reason for using these particular values is partly because they were used in this existing work, and partly because they worked well empirically.