Closed de-code closed 2 years ago
Thanks for bringing this up -- yeah, we do have plans to extend to more models. And in the latest updates, we've added models based on EfficientDet, which I've found is 90% faster on CPUs, has smaller model weights, and is much easier to install. Would you like to have a try -- lp.AutoLayoutModel("lp://efficientdet/PubLayNet")
?
PS: this requires you to upgrade lp to the latest version: pip install -U layoutparser[effdet]
.
That is perfect thank you. I should have tried to explore the available models more. It is indeed much faster and usable (got around 220 ms). I haven't seen any noticeable reduction in performance for my current use-case (detecting figures, although just looked at one document so far).
AutoLayoutModel("lp://efficientdet/PubLayNet") This can be used for lp.Detectron2LayoutModel also ? I have trained a custom model I am trying to increase speed for predicting layout detection ? Need some help it would be glad someone who knows better can help .
Motivation
(Apologies if this is not the right place to ask questions) The
faster_rcnn_R_50_FPN_3x
(PubLayNet) seems to be quite slow on a CPU. Locally it's around 3 seconds per image. In Google Colab it's more than 6 seconds. (It's around 350 ms with a GPU though). Something that would make this work on a CPU at a more reasonable speed could make it more "accessible". It would also make the download of the model, and PyTorch itself smaller. I was wondering whether you have any plans to train a smaller model on one of the related datasets?Related resources Something like YOLOv5s perhaps?
Additional context n/a