Spritaro / condinst_tensorrt

MIT License
7 stars 1 forks source link

Great project. Will the SparseInst algorithm continue to be updated, it seems incomplete? #2

Closed Byronnar closed 1 year ago

Byronnar commented 1 year ago

Great project. Will the SparseInst algorithm continue to be updated, it seems incomplete?

Spritaro commented 1 year ago

Unfortunately, the model turned out to be quite difficult to train. It required large batch size (>64), didn't really work well with gradient accumulation, was unstable with half precision... I'd recommend to use the official implementation.

Byronnar commented 1 year ago

Thanks, I will try.

Byronnar commented 1 year ago

Unfortunately, the model turned out to be quite difficult to train. It required large batch size (>64), didn't really work well with gradient accumulation, was unstable with half precision... I'd recommend to use the official implementation.

I also want to ask whether the reproduction of CondInst can achieve the accuracy of the paper? I see that this open source warehouse only wrote about 20 mAPs. Is it because the training time is not enough?

Spritaro commented 1 year ago

Tweaking training parameters like batch size or data augmentation might increase accuracy. But this implementaiton of CondInst is probably too simple to reproduce the official results. For the sake of simplicity, I used CenterNet without deformable conv layers to detect instances, instead of FCOS as in the CondInst paper.

Byronnar commented 1 year ago

OK,Thank you.