"Cat detector" with float16 textures

probability / className Model loading...

As I understand forceHalfFloat works only for TFjs ≤ 1.7.4 but what about TFjs Wasm backend? :)

You can Crop/Rescale the cat and Copy/Paste images from the Net for classification at What did MobileNet learn?

Comments

• This script with TFjs mobilenet model is enough for simple image classification.

    <script>
      tf.webgl.forceHalfFloat();
      const img = document.getElementById('image');
      const predictImage = async () => {
        const model = await mobilenet.load();
        const predictions = await model.classify(img);
        console.log(predictions);
      }
      predictImage();
    </script>
  

• This tiny script is not very useful, but if you just add   "tf.webgl.forceHalfFloat()"
  you will halve GPU memory footprint, get ~10% acceleration and save battery.
  See Mechanism to enforce use of gl.HALF_FLOAT with both WebGL backends.
• Surprisingly I get the same 16.98 MB for f16/f32 textures (see console for tf.memory() ). https://github.com/tensorflow/tfjs/issues/2065
• This application is tested in Chrome and Firefox Nightly.