ML Classifier is a machine learning engine for quickly training image classification models in your browser. Models can be saved with a single command, and the resulting models reused to make image classification predictions.

This package is intended as a companion for ml-classifier-ui, which provides a web frontend in React for uploading data and seeing results.

A walkthrough of the code can be found in the article Image Classification in the Browser with Javascript.

An interactive demo can be found here.

Screenshot of demo

ml-classifier can be installed via yarn or npm:

yarn add ml-classifier

or

npm install ml-classifier

Start by instantiating a new MLClassifier.

import MLClassifier from 'ml-classifier';

const mlClassifier = new MLClassifier();

Then, train the model:

await mlClassifier.train(imageData, {
  callbacks: {
    onTrainBegin: () => {
      console.log('training begins');
    },
    onBatchEnd: (batch: any,logs: any) => {
      console.log('Loss is: ' + logs.loss.toFixed(5));
    }
  },
});

And get predictions:

const prediction = await mlClassifier.predict(data);

When you have a trained model you're happy with, save it with:

mlClassifier.save();

When you hit save, Tensorflow.js will download a weights file and a model topology file.

You'll need to combine both into a single json file. Open up your model topology file and at the top level of the JSON file, make sure to add a weightsManifest key pointing to your weights, like:

{
  "weightsManifest": "ml-classifier-class1-class2.weights.bin",
  "modelTopology": {
    ...
  }
}

When using the model in your app, there's a few things to keep in mind:

1. You need to make sure you transform images into the correct dimensions, depending on the pretrained model it was trained with. (For MOBILENET, this would be 1x224x224x3).
2. You must create a pretrained model matching the dimensions used to train. An example is below for MOBILENET.
3. You must first run your images through the pretrained model to activate them.
4. After getting the final prediction, you must take the arg max.
5. You'll get back a number indicating your class.

Full example for MOBILENET:

    const loadImage = (src) => new Promise((resolve, reject) => {
      const image = new Image();
      image.src = src;
      image.crossOrigin = 'Anonymous';
      image.onload = () => resolve(image);
      image.onerror = (err) => reject(err);
    });

    const pretrainedModelURL = 'https://storage.googleapis.com/tfjs-models/tfjs/mobilenet_v1_0.25_224/model.json';

    tf.loadModel(pretrainedModelURL).then(model => {
      const layer = model.getLayer('conv_pw_13_relu');
      return tf.model({
        inputs: [model.inputs[0]],
        outputs: layer.output,
      });
    }).then(pretrainedModel => {
      return tf.loadModel('/model.json').then(model => {
        return loadImage('/trees/tree1.png').then(loadedImage => {
          const image = tf.reshape(tf.fromPixels(loadedImage), [1,224,224,3]);
          const pretrainedModelPrediction = pretrainedModel.predict(image);
          const modelPrediction = model.predict(pretrainedModelPrediction);
          const prediction = modelPrediction.as1D().argMax().dataSync()[0];
          console.log(prediction);
        });
      });
    }).catch(err => {
      console.error('Error', err);
    });

Start by instantiating a new instance of MLClassifier with:

const mlClassifier = new MLClassifier();

This will begin loading the pretrained model and provide you with an object onto which to add data and train.

MLClassifier accepts a number of callbacks for beginning and end of various methods.

You can provide a custom pretrained model as a pretrainedModel.

You can provide a custom training model as a trainingModel.

• pretrainedModel (string | tf.Model) Optional - A string denoting which pretrained model to load from an internal config. Valid strings can be found on the exported object PRETRAINED_MODELS. You can also specify a preloaded pretrained model directly.
• trainingModel (tf.Model | Function) Optional - A custom model to use during training. Can be provided as a tf.Model or as a function that accepts {xs: [...], ys: [...], number of classes, and params provided to train.
• onLoadStart (Function) Optional - A callback for when load (loading the pre-trained model) is first called.
• onLoadComplete (Function) Optional - A callback for when load (loading the pre-trained model) is complete.
• onAddDataStart (Function) Optional - A callback for when addData is first called.
• onAddDataComplete (Function) Optional - A callback for when addData is complete.
• onClearDataStart (Function) Optional - A callback for when clearData is first called.
• onClearDataComplete (Function) Optional - A callback for when clearData is complete.
• onTrainStart (Function) Optional - A callback for when train is first called.
• onTrainComplete (Function) Optional - A callback for when train is complete.
• onEvaluateStart (Function) Optional - A callback for when evaluate is first called.
• onEvaluateComplete (Function) Optional - A callback for when evaluate is complete.
• onPredictStart (Function) Optional - A callback for when predict is first called.
• onPredictComplete (Function) Optional - A callback for when predict is complete.
• onSaveStart (Function) Optional - A callback for when save is first called.
• onSaveComplete (Function) Optional - A callback for when save is complete.

import MLClassifier, {
  PRETRAINED_MODELS,
} from 'ml-classifier';

const mlClassifier = new MLClassifier({
  pretrainedModel: PRETRAINED_MODELS.MOBILENET,

  onLoadStart: () => console.log('onLoadStart'),
  onLoadComplete: () => console.log('onLoadComplete'),
  onAddDataStart: () => console.log('onAddDataStart'),
  onAddDataComplete: () => console.log('onAddDataComplete'),
  onClearDataStart: () => console.log('onClearDataStart'),
  onClearDataComplete: () => console.log('onClearDataComplete'),
  onTrainStart: () => console.log('onTrainStart'),
  onTrainComplete: () => console.log('onTrainComplete'),
  onEvaluateStart: () => console.log('onEvaluateStart'),
  onEvaluateComplete: () => console.log('onEvaluateComplete'),
  onPredictStart: () => console.log('onPredictStart'),
  onPredictComplete: () => console.log('onPredictComplete'),
  onSaveStart: () => console.log('onSaveStart'),
  onSaveComplete: () => console.log('onSaveComplete'),
});

Example of specifying a preloaded pretrained model:

import MLClassifier from 'ml-classifier';

const mlClassifier = tf.loadModel('... some pretrained model ...').then(model => {
  return new MLClassifier({
    pretrainedModel: model,
  });
});

This method takes an array of incoming images, an optional array of labels, and an optional dataType.

import MLClassifier from 'ml-classifier';
const mlClassifier = new MLClassifier();
mlClassifier.addData(images, labels, 'train');

• images (Array<tf.Tensor3D | ImageData | HTMLImageElement | string>) - an array of 3D tensors, ImageData (output from a canvas toPixels, a native browser Image, or a string representing the image src. Images can be any sizes, but will be cropped and sized down to match the pretrained model.
• labels (string[]) - an array of strings, matching the images passed above.
• dataType (string) Optional - an enum specifying which data type the images match. Data types can be train for data used in model.train(), and eval, for data used in model.evaluate(). If no argument is supplied, dataType will default to train.

Nothing.

train begins training on the given dataset.

import MLClassifier from 'ml-classifier';
const mlClassifier = new MLClassifier();
mlClassifier.addData(images, labels, DataType.TRAIN);
mlClassifier.train({
  callbacks: {
    onTrainBegin: () => {
      console.log('training begins');
    },
  },
});

• params (Object) Optional - a set of parameters that will be passed directly to model.fit. View the Tensorflow.JS docs for an up-to-date list of arguments.

train returns the resolved promise from fit, an object containing loss and accuracy.

evaluate is used to evaluate a model's performance.

import MLClassifier from 'ml-classifier';
const mlClassifier = new MLClassifier();
mlClassifier.addData(images, labels, DataType.TRAIN);
mlClassifier.train();
mlClassifier.addData(evaluationImages, labels, DataType.EVALUATE);
mlClassifier.evaluate();

• params (Object) Optional - a set of parameters that will be passed directly to model.evaluate. View the Tensorflow.JS docs for an up-to-date list of arguments.

evaluate returns a tf.Scalar representing the result of evaluate.

predict is used to make a specific prediction using a saved model.

import MLClassifier from 'ml-classifier';
const mlClassifier = new MLClassifier();
mlClassifier.addData(images, labels, DataType.TRAIN);
mlClassifier.train();
mlClassifier.predict(imageToPredict);

• image (tf.Tensor3D) - a single image encoded as a tf.Tensor3D. Image can be any size, but will be cropped and sized down to match the pretrained model.

predict will return a string matching the prediction.

save is a proxy to tf.model.save, and will initiate a download from the browser, or save to local storage.

import MLClassifier from 'ml-classifier';
const mlClassifier = new MLClassifier();
mlClassifier.addData(images, labels, DataType.TRAIN);
mlClassifier.train();
mlClassifier.save(('path-to-save');

• handlerOrUrl (io.IOHandler | string) Optional - an argument to be passed to model.save. If omitted, the model's unique labels will be concatenated together in the form of class1-class2-class3.
• params (Object) Optional - a set of parameters that will be passed directly to model.save. View the Tensorflow.JS docs for an up-to-date list of arguments.

getModel will return the trained Tensorflow.js model. Calling this method prior to calling mlClassifier.train will return null.

import MLClassifier from 'ml-classifier';
const mlClassifier = new MLClassifier();
mlClassifier.addData(images, labels, DataType.TRAIN);
mlClassifier.train();
mlClassifier.getModel();

None.

The saved Tensorflow.js model.

clearData will clear out saved data.

import MLClassifier from 'ml-classifier';
const mlClassifier = new MLClassifier();
mlClassifier.addData(images, labels, DataType.TRAIN);
mlClassifier.clearData(DataType.TRAIN);

• dataType (DataType) Optional - specifies which data to clear. If no argument is provided, all data will be cleared.

Nothing.

Contributions are welcome!

You can start up a local copy of ml-classifier with:

yarn watch

ml-classifier is written in Typescript.

Tests are a work in progress. Currently, the test suite only consists of unit tests. Pull requests for additional tests are welcome!

Run tests with:

yarn test

• Kevin Scott

This project is licensed under the MIT License - see the LICENSE file for details