the mIoU in ade20k is not good like in the ade20k benchmark

[daixiaolei623]

thank you for your great work.
However, the mIoU i test with the maskformer_swin_large_IN21k_384_bs16_160k_res640.yaml and model is only 0.4022,
but your mIoU in ade20k benchmark is 0.4967, which is much better.
Could you please tell me why?
thank you.

[daixiaolei623]

I used the model_final_45388b.pkl and maskformer_swin_large_IN21k_384_bs16_160k_res640.yaml

[bowenc0221]

We follow Swin Transformer setting to train our model in both train and val set, but that won't give you +9 mIoU difference. There might be something wrong when you save predictions.

[daixiaolei623]

@bowenc0221
thank you.
i will not use this result to ade20k benchmark in public, and my mIoU is hidden in benchmark, so you don't have to worry about the leakage of this result.
i just want to know how the maskfromer can get the 0.4967 mIoU in ade20k benchmark, because it is a great progress.
Do you mean that the model_final_45388b.pkl is not the best model?
or There might be something wrong when you save predictions? but i did save the predictions (sem_seg in your code),Is this wrong?

[bowenc0221]

Do you mean that the model_final_45388b.pkl is not the best model?

model_final_45388b.pkl is the best model trained on train set, and the model we used to evaluate on the ADE20K benchmark is trained on train + val set following Swin Transformer.

but i did save the predictions (sem_seg in your code),Is this wrong?

You didn't use multi-scale inference.

[daixiaolei623]

@bowenc0221
thank you.
i am a student of Jiangsu university about CS, i just want to learn this excellent model.
Could you please share the corresponding model and multi-scale inference to let me learn this. i do not understand how to do this.
thank you.

[daixiaolei623]

@bowenc0221
my email is asertu@126.com