🎉 Hi everyone! I introduced Test Time Augmentation (TTA) to the model testing tool. After testing with a slight trade-off in speed (from instant to a 1 sec delay), it is now 90% accurate in predicting any handwritten Amharic character.

Ladies and Gentlemen, I, am now successful in creating the first sub-10MB, microcontroller-friendly, CPU-first handwritten Geez OCR that is both cost-effective and accurate! 😊

I'm flying with joy right now! 🎉🎉🎉🎉

#AI #MachineLearning #Amharic #Geez #Tigrinya #OCR #DeepLearning #YAP_C_OCR

am preparing an exe file for everyone to use this testing tool

😊 I cant stress this enough isolated room is a must while testing your robot 🙄 am even inclined to add demo day in that list 😁

A gentle reminder that there is no perfect/predefined age to take the holy communion. Repent and claim your eternal life.

#Geez_OCR #Update I just finished preparing the standalone file and since it have to pack the entire tensorflow, opencv, PIL, and numpy library 😅 it is big above 700MB if you machine is a window give it a try

and DM me about your experience here is my gift after 3 sleepless night

i believe among my next steep is

1, to write a paper on this specially I have to document the new efficiency gain and accuracy with this project

2, prepare a web version so any one from any machine can try and test the model even from phone

3, expand this project to word OCR with similar objective of CPU first, Fast and Accurate model with limited resource

4, log the crazy tricks i used in this project here

Note: this should work fine even if your pc have 4GB ram only 👆 in the previous log

you can press the red X button to tell it when it predicted wrong and it will record it for future use if you export and share the sql database of your handwriting to share just press the green sql button and it will give you sql dataset of what you have tested so far and you can email me that.

Good Night 🫶

The size how can i reduce it to below 70 mb exe

700MB is big when i close my eye i keep wondering why olama setup is smaller than my standalone while rung llm but mine is dealing with 10mb model and yet it's more than 700 as stand alone am 100% shore that am missing some thing

I have identified few bugs

¤1 the drawing area is wider and as it find larger screen its expanding and adjusting but the model expect equal dimension image which will reduce accuracy

¤2 to open the app itäs taking 2 min in core i5, 16GB RAM, NVMe ssd Machine i can only imagine how much more minutes it will take in other machines so i need to work on that

[ lowkey it is related to the exe size and all the library it will unload to start working but it must have many unwanted dev packages i should remove in prod ]

¤3 as the app is launched if the user become impatience and keep clicking it will keep creating several duplicate process consuming more res so i have to add background process management

https://www.linkedin.com/in/yared-kassa/

i just pass 2.3k connection in LinkedIn 🫆