search

jeferson-sb / dogAI

Dog breed image recognition with Teachable Machine & Tensorflow.js
https://dogai.netlify.app
MIT License
18 stars 10 forks source link

Different classification results for own Teachable Machine model - in the TM UI and on this website #193

Open KonstantinaD opened 2 years ago

KonstantinaD commented 2 years ago

Hi Jeferson, Thank you for exposing this website, it helps me a lot with my project. I trained my own model in Teachable Machine and provided my model files in model folder of your app. The classification results I see on the TM UI (https://teachablemachine.withgoogle.com/train/image) and on your locally run website are often different - sometimes both in terms of class and confidence. Have you compared your website results with the TM UI ones and found differences? If yes, do you know why they might be there? Also, I trained my own image classification model using ml5.js library, saved the model and put the model files in the model folder of your app. Using your existing code, I test my model in this way. Do you know if this is a reliable way to test? Thank you in advance.

jeferson-sb commented 2 years ago

Hi @KonstantinaD how are ya?

yeah, those classifications' results may vary, for this project I'm using an older version of Tensorflow.js that is compatible with teachablemachine image plugin@teachablemachine/image which is being used to load the model and do the predictions, unfortunately that library hasn't been updated for a long time 😅

Not sure how this would work with ml5.js worth taking a lot of the docs, but generally speaking always test within the app and in order to improve the results you can make another dataset with more images and stuff. Hope that helps :)

KonstantinaD commented 2 years ago

Thank you, Jeferson. I'll try to change the code a bit using a later version if available. Again, a great interface which is my only testing tool right now :)

On Tue, 17 May 2022 at 3:32, Jeferson S. @.***> wrote:

Hi @KonstantinaD how are ya?

yeah, those classifications' results may vary, for this project I'm using an older version of Tensorflow.js that is compatible with teachablemachine image @.***/image which is being used to load the model and do the predictions, unfornatuely this library hasn't been updated for a long time 😅

Not sure how this would work with ml5.js worth taking a lot the docs, but generally speaking always test within the app and in order to improve the results you can make another dataset with more images and stuff. Hope that helps :)

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>