inference speed extremely slow about birefnet HOT 7 CLOSED

@razvypp
have you done it sucessfully ?
if yes, how to do?

from birefnet.

Hi, could you check your texts and correct their formats?

from birefnet.

About the comparison with u-2-net, I don't think that is a problem. It's almost impossible to bring much improvement on accuracy with same number of parameters.

Statistics about BiRefNet with different backbones can be referred to this issue.

You can also find some other issues where we talked about increasing the speed of inference (ONNX, FP16, ...). By now, there haven't been very good methods for it. You can wait for the version with swin_v1_tiny, which could be 4 times faster than the official one.

from birefnet.

Will the v1_tiny have the same performance on general datasets?

I failed to convert to onnx i get crash, is there a guide for this?

from birefnet.

Of course not, that's a trade-off.
Sorry, as for the issues I mentioned above, I currently have no time for this kind of thing.

from birefnet.

The well-trained BiRefNet with the swin_v1_tiny backbone has been uploaded to my Google Drive. Check the stuff in README for access to the weights, performance, predicted maps, and training log in the corresponding folder (exp-xxx). The performance is a bit lower than the official version, but still good (HCE↓: 1152 -> 1182 on DIS-VD). Feel free to download and use them.

from birefnet.

By the way, check the update in inference.py. Set the torch.set_float32_matmul_precision to 'high' can increase the FPS on A100 from 5 to 12 with ~0 performance downgrade (Because I set it to 'high' during training).

from birefnet.