HidupSehat Machine Learning Model Development

In our app, there are 2 deep learning models and 1 machine learning model. Each model will be explained in each folder.

1. Our Object Detection model trained with a pre-trained model SSD MobileNet V2 with FPN-Lite Feature Extractor 640x640, shared box predictor and focal loss, trained on COCO 2017 dataset with training images scaled to 640x640. We set 40000 steps 64 batch on 7400+ images from 36 food classes.

2. Yoga Pose Detection we will use the MoveNet Lightning model, extract the keypoints and calculate the similarity between the user and the yoga image with Euclidean distance on 8 different angles.

3. Recommendation based on user inputs in Diary feature on our app. From there we can recommend the feeds content, and calculate the words similarity using Cosine Similarity and the TF-IDF method.

4. Our Data Scraping on food nutrients sourced from FatSecret and feeds content are from HaloDoc using BeautifulSoup

Library, Pre-Trained Model, & Method Used

• TensorFlow
• TensorFlow SSD MobileNet v2 fpnlite 640x640
• TensorFlow Movenet Lightning
• Euclidean distance
• Cosine Similarity
• NLTK
• Numpy
• Pandas
• Regex
• TF-IDF
• Scikit-learn
• BeautifulSoup
• Matplotlib
• difPy
• PIL
• xml