论文中提到batch_size=48,使用两张TitanXp，是否能够完成相应训练，并复现论文中指标

[lai-pf]

您好，请问论文中提到batch_size=48,使用两张TitanXp完成所有模型的训练
但在repo中您提到seedFormer使用两张V100,batch_size=48完成训练。
请问两张TitanXp或者单张24G3090是否能够完成相应训练，并复现论文中指标，因为V100单卡32G显存,TitanXp单卡12G显存，我不清楚哪个描述能够作为复现实验的基本显存需求。
希望得到您的回复，万分感谢

[hrzhou2]

We use two V100 gpus in the original implementation. You could change the batch size for your own settings.

[lai-pf]

I have tried it today. I used two 24G P40GPUs, it seems that I can only run in batch_size 32. And It cost near half hour an epoch,that means 400epochs need 8days? can you tell me the epoch time on your device or all training time.
I really like your code style, Simple and elegant. But it seems that I can't follow this work with my devices for it's long training time.

[hrzhou2]

It takes about 3 days on my device with the original settings.

Changing the batch_size to 32 often requires other modification on lr or decay. You can train it for fewer epochs since batch_size is smaller or you can use the pretrained model.

[Nevak1016]

@hrzhou2 Hi, I'm also experimenting with this method as a sota recently, but is it really possible to experiment with batchsize=48 using 2×V100? At present, I use 2×3090 and it prompts "CUDA out of memory".