Transfer Learning: Tulip with 5 vs 6 petals

The script generates 300 pairs of randomly rotated procedural tulips with 5 and 6 petals (corresponding predictions 0 or 1). MobileNet (~15 MB of weights) is used to get a 1024 features map (embeddings) from a 224x224 image. The script collects embeddings in a dataset and trains one perceptron (just 1024 weights) for 30 epochs. It all takes a few minutes (HP laptop with Ryzen 5 3500U).

This simple demo counts petals surprisingly well (much faster than me)!

It is based on TensorFlow.js Example: Transfer Learning to play Pacman via the Webcam. It uses a small neural network on the top of MobileNet_v1_1.0_224

  gl.drawElements(gl.TRIANGLES, di - di1, gl.UNSIGNED_INT, 4*di1)
  const embedding = await mobilenet.infer(cnv, true);
  const predictions = model.predict(embedding);

  async function train() {
    model = tf.sequential({
      layers: [
        tf.layers.dense({
          inputShape: [1024],  units: 1,  activation: 'sigmoid',
          kernelInitializer: 'varianceScaling',  useBias: true
        })
      ]
    });
    const optimizer = tf.train.adam(.01);
    model.compile({optimizer: optimizer, loss: 'binaryCrossentropy'});
    const batchSize = 32;
    model.fit(xs, ys, {
      batchSize,
      epochs: 40,  //    shuffle: true,
      callbacks: {
        onEpochEnd: async (epoch, logs) => {
          console.log(epoch + ' Loss: ' + logs.loss.toFixed(5));
        }
      }
    });
  }

async function data(x){
  const dx = await x.data();
  dataset.set(dx, dataPos);
  dataPos += 1024;
  x.dispose();
}

Dataset with WebGL procedural flowers