About the inference result

[hsdjkfnsfc]

Thanks for your amazing work. However, I faced a problem when I ran the eva.py.

the output looks like a hard segmentation, not a soft matting. Here is my result.

[hsdjkfnsfc]

The result predicted by the model train on my dataset, and I found your model is the matting result, I want to know why.

[Windaway]

This training code is best executed under Pytorch 1.x, otherwise you may need to modify the Norm Layer order of AEAL. In addition, it is best to provide your training log.

[hsdjkfnsfc]

I didn't save the log, the terminal has the print like this, I didn't find any useful info about the bug.

[Windaway]

I'm not sure what your training data looks like (maybe that's a problem), but your loss seems to be smaller than the loss on the Adobe dataset?

[hsdjkfnsfc]

here is an example of my data.

When I fintune on your model, the output looks normal, which is a matting result. but when I random initialize the model param and train on my data, the output looks like segmentation.
The log that I showed before was fintune your model on my data.

[Windaway]

Maybe some hyper-parameters should be tuned.

[hsdjkfnsfc]

It's a pytorch problem! the result was normal when I trained based on torch1.x.

[Windaway]

You may use the pytorch 2.0 branch for training https://github.com/aipixel/AEMatter/tree/Pytorch2.0.