search

junfu1115 / DANet

Dual Attention Network for Scene Segmentation (CVPR2019)
MIT License
2.41k stars 483 forks source link

What kind of GPU is it in this paper? #58

Open angleboy8 opened 5 years ago

angleboy8 commented 5 years ago

This paper didn't introduce GPU setting specifically. Because attention map costs extensively, I want to know the memory of GPU. Thank you very much.

junfu1115 commented 5 years ago

@angleboy8 4XP40; Resnet101+PAM+CAM needs about 3.1G memory of GPU in testing phase and about 14.5G in training phase.

gaowu9595 commented 5 years ago

@junfu1115 Hi, Do you mean that the modle needs 14.5G per P40 with 8 batch-size in training phase?...I only have 2 GPUs (11G per GPU). And how does that affect the accuracy if I set the batch-size with 4 ? How should I set the batch-size, base size and crop size to get the results as close as yours with limited memory ?

shuangyumelody commented 4 years ago

@angleboy8 4XP40; Resnet101+PAM+CAM needs about 3.1G memory of GPU in testing phase and about 14.5G in training phase.

I want to know the input image size, Thank you very much!