search

donghao51 / SimMMDG

[NeurIPS 2023] SimMMDG: A Simple and Effective Framework for Multi-modal Domain Generalization
49 stars 1 forks source link

About parameters #7

Closed lz19991122 closed 3 days ago

lz19991122 commented 5 days ago

Hi, why different parameters for different cross domains?

python train_video_flow_audio_EPIC_SimMMDG.py D2 D3 to D1 --trans_hidden_num 1024 D1 D2 to D3 --alpha_trans 1.0

The same for MOOSA python train_video_flow_audio_EPIC_MOOSA.py D2 D3 to D1 --mask_ratio 0.7 --entropy_min_weight 0.001 D1 D3 to D2 --mask_ratio 0.7 --entropy_min_weight 0.1 D1 D2 to D3 --mask_ratio 0.3 --entropy_min_weight 0.1

donghao51 commented 4 days ago

Hi, we find that the default parameters work well in most cases. However, due to huge domain differences, in some cases we finetune the parameters to achieve better results.