Getting lower mIOU than reported in paper

VITA-Group / FasterSeg

[ICLR 2020] "FasterSeg: Searching for Faster Real-time Semantic Segmentation" by Wuyang Chen, Xinyu Gong, Xianming Liu, Qian Zhang, Yuan Li, Zhangyang Wang

MIT License

526 stars 107 forks source link

Getting lower mIOU than reported in paper #32

Closed soyebn closed 4 years ago

soyebn commented 4 years ago

One of the best works for semantic seg with NAS. Thank You for sharing train as well as search code.

I downloaded student's weights from the location mentioned in the readme https://drive.google.com/file/d/1O56HnA0ug2M3K4SR3_AUzIs0wegy9BX6/view?usp=sharing and ran train.py to check the validation accuracy but I got 70.5%. But your paper suggest it should generate 73.1%. Am I doing something wrong?

Thanks.

chenwydj commented 4 years ago

Hi @soyebn!

Thank you for your interest in our work!

73.1% is the best achievable validation performance during the training. The provided checkpoint is from the last epoch. It should give you similar results with our code.