training time

[youngAt19]

hi, thank you for sharing your codes.
I try to train densecl on imagenet with 4*2080ti, but it takes about 13 days to finish training, which is much longer time than yours.
Did you use imagenet in lmdb format so that training time is shorter?

[CoinCheung]

I used 16 tesla T4 gpu amount 2 nodes, and it takes me about 4 days to train it. From my observation of gpu occupance, the data loading is not the bottleneck of training speed.

[youngAt19]

Thanks for your reply. I thought it just takes you 53 hours to train the model from here

[youngAt19]

Thanks for your reply. I thought it just takes you 53 hours to train the model from here

[CoinCheung]

This is mocov2 time from the original repo of mocov2.

Training denseCl is time-consuming, I recommand you to use the pretrained model if you are using resnet50/101 backbone. It will save you time.