Got a poor result for separate training on ADE20K

[mrwu-mac]

Thank you for your excellent work!

I try to train Painter on ADE20K with the same training params with joint training (but a total batch size = 4 * 1 * 32). But I got only 8. mIOU, is the batch size affecting the result?

[mrwu-mac]

And this is my result and result from your pretrained model. My result seems to non-smooth, is there any possible solution?