tensorflow/tfjs-models

tfjs-models is a Google-maintained collection of pre-trained machine learning models that have been ported to TensorFlow.js, the JavaScript version of the TensorFlow library. The point is that the models are already trained, so a developer can drop them into a web page or Node project and call them without needing to know how to train a model from scratch. They are published on NPM and unpkg, so any project can install them out of the box. The README says the models can be used directly, or in a transfer-learning setup where someone fine-tunes them with their own data. Each model lives in its own subdirectory with its own README and its own API. The maintainers say they try to hide the underlying tensors from the API so people who are not machine learning experts can still use the code. A large table in the README lists what is available. For images, there is MobileNet for classifying images against the ImageNet label set, hand-pose detection, body pose detection, Coco SSD for object detection, DeepLab v3 for semantic segmentation, and face landmark detection for inferring the 3D geometry of a face in real time. For audio, there is a speech-commands model that classifies one-second clips. For text, there is the Universal Sentence Encoder, which turns text into a 512-dimensional embedding, and a toxicity model that scores comments. Other entries include portrait depth estimation, which estimates per-pixel distance from a single portrait image, and general utilities such as a K-Nearest-Neighbors classifier. Each row lists a live demo, a source link, a short description, and an npm install command. People interested in adding a new model are asked to file an issue on the main tfjs repo first.