LeanAI

LeanAI is a machine learning powered web app which aids anyone in achieving their fitness goals. Given a few physique photos and basic metrics (weight, height) LeanAI will predict bf% and give a comprehensive diet plan to achieve desired physique goals.

Hugging Faces Repository:

https://huggingface.co/spaces/kalandjl/leanai-gradio/tree/main

Model Architecture

The core LeanAI model is a pytorch ensemble of 11 different image regression models:

- resnet50
- efficientnet_b3
- densenet121
- resnet34
- mobilenetv3_large_100 
- efficientnet_b0
- resnet50d                 
- densenet121              
- regnety_040               
- ghostnet_100                
- maxvit_tiny_rw_224            
- efficientformerv2_s0

All of these models were trained on public body fat estimation data and varying augmentation datablocks. Fastai's python library was used to streamline this process. See more in the /core folder.

Front end

Simple Next.JS frontend which access hugging faces repo api. Use bellow commands to run:

https://github.com/kalandjl/LeanAI
cd client

npm run dev

Notebooks

To access my notebooks which I used to train the models, run:

• Most notebooks will not work without ensemble.pt file, individual model.pkl files, or /images folder & coresponded .csv files

jupyter notebook

and navigate to the cloned git repo /core folder.

FastAPI server

Server may not run if python --version is greater than 3.10

• Server will not work without ensemble.pt file: either download your own or use my notebooks to produce one.

cd /server
python app.py