The BEiT experiment on the ADE20K dataset cannot be completely replicated.

microsoft / unilm

Large-scale Self-supervised Pre-training Across Tasks, Languages, and Modalities

https://aka.ms/GeneralAI

MIT License

19.63k stars 2.51k forks source link

The BEiT experiment on the ADE20K dataset cannot be completely replicated. #585

Open linfangjian01 opened 2 years ago

linfangjian01 commented 2 years ago

Hi, the results of my repeated training on ADE20K using the BEiT-base pretrain model were 53.2 and 53.3 respectively. 0.3-0.4 lower than the official result (53.2 vs. 53.6). We are using the official config entirely.

happycoding1996 commented 2 years ago

Same situation. Beit-Base I only obtain 53.0, and Beit-Large 55.9, following the official implementation of MMSegmentation. It is hard to reproduce the results mentioned in the paper.