Models Performs Poorly in production environment

Adesoji Alu asked 9 months ago

Good morning sir, the model I used is about 96% accurate on training accuracy and 98% on validation accuracy, but out of 10 pictures I tested with using FAST API, only 1 was correct, pls how do I correct this? secondly, how do I make an endpoint where all I need to do is upload images, train the model, and download the model in any supported format like savedmodel, json or tensorflowjs. e.g (https://teachablemachine.withgoogle.com/train) without returning to jupyter notebook for retraining everytime?

2 Answers
Best Answer
Chris Staff answered 9 months ago

Regarding your first question: by any chance, in production, do you feed different samples than from your training and validation datasets? Is preprocessing in your production pipeline done in exactly the same way? It sounds like a distribution issue.
Regarding your second question: this is too complex to answer in much detail. One way would be to create a JavaScript web application or a Streamlit app that allows you to upload files, then save them to Google Cloud Storage, S3 or a local disk attached to a web server, then retrain the model every few hours or days — and add some save_model stuff too, for e.g. TF Lite.

Adesoji answered 9 months ago

Thanks, sir. will do just that

Your Answer

16 + 19 =