WebNN (WNN) is a JavaScript framework for distribution and training of Neural Networks (NN) through the internet. This is a large part of my research, working alongside Ron Cotton into distributed training of neural networks at WSU Vancouver. WNN consists of a NodeJS HTTP server that facilitates the asynchronous distributed training of user-defined NN. It builds on TensorFlow.js, which leverages the clients' GPU using WebGL (or CPU if WebGL unavailable) to run and train a NN in any web browser that supports WebGL 1.0 or better--which is 96% of browsers in the US and 93% of browsers globally [Caniuse].
- Clone this repo onto the machine that will be your server.
- Navigate into the repo and run
npm installto install all necessary dependencies. - Run
node wnn train mnistto launch the server with the mnist model included with this repo - You'll see some output in your terminal as the server prepares the mnist model for training.
- Once you see the 'Server ready!' output: open your browser and navigate to the url
localhost:8888/watcherto open the watcher page. - In another page or tab, navigate to the url
localhost:8888/clientto begin training the model. You should soon see accuracy and loss metrics populate the charts of the watcher page showing you the training and validation progress.
WebNN comes with a simple JS command-line script to make certain workflows a little easier: new - Create blank model reset - Resets model back to initialized stated copy <src_name> <dest_name> - Copies model to a new clean version ready for training save <src_name> <dest_name> - Copies model including all logs and weights train - Runs training for model retrain - Shortcut for reset followed by train list - List all models
Use the Watcher to see how many clients are training, the current learning rate, and the running average (5 samples) of the model validation accuracy and loss. The first chart shows validation metrics, and the second shows actual training metrics. In both charts, accuracy is in orange with labels on the right y-axis, and loss is in blue with labels on the left y-axis.
WebNN prepares dumb clients for your convenience that simply request data from the server and train. But you can run and train models in real applications as well. Look inside /lib/server/clients/client.js to see how these clients function, and how you can use them in your own applications.
Models you define and create live in the ./models directory of the repo.
Use the node wnn new YOUR_MODEL_NAME command to initialize your model. This command will create a directory under ./models for your model, and will include the following template files:
- config.js - this file is where you configure how your model will train, set hyperparameters, and hook up training and validation data sources.
- model.js - this file is where you define your model architecture in a Keras-like format
Also included will be these empty CSV files:
- log.csv - where all server requests will be logged
- training.csv - where all client training metrics are logged
- validation.csv - where all validation metrics are logged
WebNN can be included into most front-ends painlessly. Simply import the ModelClient.js module into your client and instantiate a new ModelClient (singleton class). See the documentation for ModelClient to see the available functions and lifecycle.