Using Google's AutoML and tfjs, build a machine learning model quickly and iteratively without tons of time for training jobs to run. AutoML is a great tool for testing your data and getting a sense of accuracy, and it uses its own internal model to determine the best training methods -- DNN, CNN, GAN, etc. It's a bit of a black box, but the finished models are accessible via a URL endpoint. Tensorflow's lightweight tfjs can also run models on the client side. We tested with posenet, caffe, tiny-yolo, and others to get additional context before sending an image to AutoML for classification. We also used these client-side models and processing to create our data set because we were using faces, but the same process could be used with object detection and webcam feeds.
To upload that data in the necessary format to Google Storage, we used some python scripts (but this could easily be done alongside the ML processing itself).