accury problem ade20k

[zxmyyds]

I trained ADE20k, but why the accuracy makes such a difference

[CoinCheung]

What modifications did you made? And which dataset did you use? Would you please tell me how did you reach this result?

[zxmyyds]

@CoinCheung I did not make any modification and trained ade20k 2016 according to the instructions. I also wonder why some training shows 0. Have you encountered such a situation

[CoinCheung]

Then tell me what is your environment is like, and the steps following which I can achieve your result?

[zxmyyds]

I just did these steps

[CoinCheung]

Maybe you should adjust your batch size to match the configuration of what I used. I used 8 gpus, but you only used one.

If your gpu memory cannot afford such batch size, you should try to reduce learning rate.

The training parameters I provided is not carefully tuned(because authors did not have such experiments in the paper), maybe you can have a round of grid search to find the best configuration.

[zxmyyds]

Thanks I used the city training set and its accuracy is normal and I need to try it again

[CoinCheung]

Maybe you have achieved the results that you see correct, so I am closing this.

[SCO]

Thanks I used the city training set and its accuracy is normal and I need to try it again

I met the same problem on one single GPU without any modification, did you solve the problem? thank you