search

speedinghzl / DSRG

Weakly-Supervised Semantic Segmentation Network with Deep Seeded Region Growing (CVPR 2018).
MIT License
250 stars 36 forks source link

ResNet Code #3

Closed DQDH closed 5 years ago

DQDH commented 6 years ago

Hi, @speedinghzl . Can you release the code for resnet101 result in the paper? Thanks.

speedinghzl commented 6 years ago

@hardBird123 Sorry about the late reply. I have no plan to release the ResNet code. Thanks. It is not difficult to achieve good performance with Resnet-101, in fact, overfitting is a key obstacle. I just added a dropout layer with ratio 0.5 before the last layer and adjusted the learning rate to avoid gradient explosion.

DQDH commented 5 years ago

@speedinghzl Thanks. I have another problem that when I running the code, it is divided two parts, one always in GPU0 ,another part in GPU2, can you tell how to modify the code？

speedinghzl commented 5 years ago

@hardBird123 You can add CUDA_VISIBLE_DEVICES=2 before the bash command in run.sh. At the same time, setting GPU=0 in run.sh.

DQDH commented 5 years ago

Thanks a lot!