http://latte-art-classifier.herokuapp.com/
This app uses a tf.keras convolutional neural network trained on images of latte art with a single-node dense layer to estimate the grade of an image of coffee (binary classification).
The app is built on the plotly-dash framework. The model is loaded in and used in app callbacks to carry out the classification when an image is uploaded.
Given that the single-node linear-activation output of the neural network produces a continuous value, the distribution of predictions generated from the entire data-set can be used to determine a grading scheme. In other words, through quantile analysis of the distribution of the predictions of the testing and training sets, each output could be discretised into qualitatively-determined bins corresponding to the their respective grades (A+ to F).