OCR Using Microsoft's Florence-2 Vision Model on Free Google Colab

In this video, I demonstrate how to implement Microsoft's recently released Florence-2 novel Foundational Vision Model on a free Google Colab workspace using a T4 GPU. I use Optical Character Recognition (OCR) as the primary use case to showcase the model's capabilities.
You'll learn:
1. An introduction to the Florence-2 Vision Model
2. Loading and configuring the Florence-2
3. Implementing OCR task with this advanced model
4. Evaluating the performance and results of OCR using Florence-2 Vision Model.
Code Link - colab.research...
Florence-2 Model - huggingface.co...
#florence2 #vision #multimodal #multimodalai #llm #microsoftai #googlecolab #ocr #machinelearning #ai #tutorial #freeresources #attention #objectdetection #segmentation

[@Steven_249]

wow... you are super smart..... especially when you change the code for OCR REGION....! Amazing !!!

[@theailearner1857]

Glad it helped!

[@kushaldulani]

Yes really, No one does that on RUclips, rest of all teach only basics. Thanks bro

[@jinanlionbridge4521]

Thanks for sharing! very useful

[@vishalranjan2429]

i want to intergate this in an android app , how to do it ?

[@sudabadri7051]

Good video

[@hegalzhang1457]

Great work, very useful, did you release code？

[@theailearner1857]

Glad it helped, I have provided the code link in the description.

[@despo13]

Thanks

[@trinityblood5622]

Any luck on Finetuning the OCR part with custom dataset other than English?

[@theailearner1857]

Haven't tried yet, but will try to make a video on finetuning.

[@seanthibert5961]

Any luck with making use of the raw OCR results? I find it picks up more than the ocr_with_region

[@ai_enthusiastic_]

How much RAM does it need to run on a CPU?

[@theailearner1857]

In full precision, it would need approximately 10-11 GB of RAM for inference. If you are not able run it on CPU, you can try with quantized model.

[@NimeshV-nf6uz]

Can I run this on cpu ?

[@theailearner1857]

Yes you can. Change the "device_map" argument to "cpu". And also make sure to not move input tensors to "cuda".

[@NimeshV-nf6uz]

@@theailearner1857 thanks 🤜🤛