How can I keep the accuracy of my custom vision AI once exported as a TensorFlow lite model?

Question

Hello! I've created a custom vision project to analyze images of pottery sherds. I have three tags, with each tag associated with about 325 photos. I made an Android application on Android Studios that uses the exported TensorFlow lite model. When tested, the trained model works well on the custom vision website, but the accuracy is significantly worse on the Android app. I am using the same testing images. I used the metadata properties file provided when exporting to match my image preprocessing method as precisely as possible. I'm unsure which direction I should take my troubleshooting next. Should I try to remake the application to take a model.pb file and lose out on performance? Should I make a Python app that converts a model.pb file into a model.tflite file without Quantization? Any input would be greatly appreciated.

Answer 1

Hi Adam Cameron

Ideally, we should quantize the model to optimize model weights to run on edge devices and retain accuracy close to original model

Please check the documentation on Quantization aware training and post training quantization .

Also, Adopt preprocessing techniques before inputting data.

Hope it helps address your issue.

Please don’t forget to Accept Answer and Yes for "was this answer helpful" wherever the information provided helps you, this can be beneficial to other community members.

Thank you