mittagessen

Sorry from my side on not following up on this either. I ran a couple of experiments myself in December to see which component actually causes the loss of accuracy...

Yes, if it isn't too much trouble for you I'd appreciate you updating the pull request for current `main`. If not I can do it by mid-March. It is actually...

Thanks. There's one small thing. The binary datasets (`lib/arrow_dataset.py` and `ArrowIPCRecognitionDataset` in `lib/dataset/recognition.py`) also need the legacy polygon switch. I'll have a look at the pretrain tests.

The test could be just a new flag in the metadata (as it is right now) but it needs to check internally that when loading multiple arrow files the values...

My apologies. Yes, that makes sense. I'll look through everything and merge in the second half of the week.

On 24/03/06 02:48AM, Robin Champenois wrote: > Okay, I can make those changes, or wait for you to look through the code and comment precisely the changes you'd like to...

Sorry for the delay. I've merged everything and it seem to work as intended. I'll write a release note and then tag a new release.

> If we prove that it is really equivalent (PIL.ImageDraw.Draw is different by only a few pixels, and the new PiecewiseAffineTransform has been shown to change the output by https://github.com/scikit-image/scikit-image/pull/6963#issuecomment-1872077806...

0.04 is below 5% change for a good model so I'd deem that acceptable given the speedup. I'm a bit busy the next couple of days. If you could create...

On 24/04/04 12:46PM, anutkk wrote: > @mittagessen Sure thing, no problem. On the main branch or the v5.0 branch? main is better. I'm squishing small bugs I've missed in the...