Thanks! Great explanation. In practice, I've noticed Donut works fairly OK if samples are too similar to the ones in the training set (say CORD). In a typical prod scenario, one would normally try to maximize the generalization capacity to avoid maintaining multiple ML endpoints customly trained for a given family of documents. In this regard, I think Donut is still not market friendly. Still, a great addition, nonetheless. Thanks for the video!
Hey, thanks for appreciating the efforts. Yeah, agreed that you can’t fully rely on out-of-the-box checkpoints for serving in prod. Having said that, fine tuning on expected distribution set would surely help. Thanks!
Interesting work from the researchers. 'Prompt' has totally changed the way how models are being built.
True. :)
Thanks! Great explanation. In practice, I've noticed Donut works fairly OK if samples are too similar to the ones in the training set (say CORD). In a typical prod scenario, one would normally try to maximize the generalization capacity to avoid maintaining multiple ML endpoints customly trained for a given family of documents. In this regard, I think Donut is still not market friendly. Still, a great addition, nonetheless. Thanks for the video!
Hey, thanks for appreciating the efforts. Yeah, agreed that you can’t fully rely on out-of-the-box checkpoints for serving in prod. Having said that, fine tuning on expected distribution set would surely help. Thanks!
What solution is market friendly in your opinion?
Thanks for an amazing video! It was really usefull.
Glad you enjoyed Diego. Thanks :)
Great explanation. 👍
Thank you Akarsh :)
Thanks. Can donut be used for text region detection such as caption, oage number, serial number and classifiying them?
You could train it to specifically extract required entries from pdf. Making it simulate like region specific extracts
How do i call the model? There is no pipeline mentioned.
Thanks for the detailed explanation. Does it support prediction of arabic text?
I don’t think so. You can check the HF website or paper. But can be done if pretrained.
great explanation
can u please create video on layoutlmv3 and patch embedding
Thanks Shiv. Sure, will try to do them. 👍
Hi, thanks for the video. Is it possible to get confidence scores for the predictions and respective bounding boxes with this donut model?
Yeah, it should be returning it already I guess.
@@TechVizTheDataScienceGuy if possible, could you please guide me to some resource/ reference for it. I am unable to get it.
are you able to get score and boxes?