Deep convolutional network to predict 23 classes of skin diseases found in http://www.dermnet.com/dermatology-pictures-skin-disease-pictures from raw images. We make use of a pretrained ResNet152 since many visual semantics like edges or object shapes should be transferrable from Imagenet. We append 2 fully connected layers to fine-tune for our use case.
This project is an entry to API World Hackathon 2017.
Run the setup.sh script to download raw images and ResNet152 embeddings. The script will also download a trained model (this is a 2-fully-connected network stacked on top of ResNet152). Unzipping the files should produce the following folders: train/, test/, train_emb/, test_emb/. trained_models includes a trained version of the fine-tune net with 0.53 percent accuracy on held-out test data.
To install all the libraries needed, run pip install -r requirements.txt.
To start the Flask app, do python run.py. There is a single POST route (/predict) that takes a JSON from key image to a base64 encoded image. It will return a class and a score.