search

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
525 stars 105 forks source link

dataset details #14

Closed pawopawo closed 4 years ago

pawopawo commented 4 years ago

Why did you search directly in the train set and did not sample the val set from the cityscapes train set? Will the searched arch not overfit in the cityscapes val set?

chenwydj commented 4 years ago

Hi @pawopawo !

What did you mean by "sample the val set from the cityscapes train set"?

We only leveraged the training set for the search, without using the validation set. We followed the same setting in DARTS.

pawopawo commented 4 years ago

image darts sample the val set from the train set.

chenwydj commented 4 years ago

DARTS split the training set half and half for optimizing w and alpha: https://github.com/quark0/darts/blob/master/cnn/train_search.py#L91

Same as what we did: https://github.com/TAMU-VITA/FasterSeg/blob/master/search/train_search.py#L108