search

TreB1eN / InsightFace_Pytorch

Pytorch0.4.1 codes for InsightFace
MIT License
1.73k stars 423 forks source link

GPU memory required to train r100 #5

Closed JingyuanHu closed 5 years ago

JingyuanHu commented 5 years ago

Thank you for your great work! However, I met a problem about limitation of GPU memory. Can you give me some suggestion about the minimum GPU memory required to train r100 ArcFace by your code?

TreB1eN commented 5 years ago

you mean resnet100? you can easily change the batchsize in the configuration if you face a memory issue, I hava 100 batchsize in resnet100 training

JingyuanHu commented 5 years ago

Thank you for your reply. Yes, I mean resnet100. My GPU is one Titan xp. I can change batchsize to 64 when backbone is resnet50. I want to try resnet100 and larger batchsize for better results. So I want to know the minimum GPU memory required. Maybe I can change a better to GPU?

TreB1eN commented 5 years ago

简单一点，如果你想要用大一点的batchsize,就直接把设置改了跑起来，然后用nvidia-smi 命令查看显存占用，多试几次就知道最大能放多少batchsize了，我一般是不会去算一个batch要占用多少显存的。提醒你batchsize不是越大越好的，而且改大了batchsize, learning rate也要调大。

JingyuanHu commented 5 years ago

谢谢

cvtower commented 5 years ago

简单一点，如果你想要用大一点的batchsize,就直接把设置改了跑起来，然后用nvidia-smi 命令查看显存占用，多试几次就知道最大能放多少batchsize了，我一般是不会去算一个batch要占用多少显存的。提醒你batchsize不是越大越好的，而且改大了batchsize, learning rate也要调大。

So what is the corresponding paras for the accuracy data mentioned in this repo? init lr:e-2 or e-3? paper: e-1 batchsize: deponds(default value is 200 for e-3?) paper:512 epoch: default value:8 Do you have any suggestion for 1080ti platform?