TensorFlow.js notes

Contents:

• About TFjs
• Image classification
  • Image classification with float16 data (textures)
    Tiny "Cat detector" with MobileNet.
  • What did MobileNet learn?
    Interactive image classification by MobileNet. Copy/Paste (then Crop/Rescale) images from the Net or your PC for classification. 6 different CNN models. Uses simplified TFjs MobileNet sources.
  • Snapshot and Stream classification in mobile phone. Comments
    Interactive classification of images from your mobile camera.
  • CNN translation/scaling invariance
    Interactive test of translation/scaling invariance. Noise. Size does matter!
    "Making convolutional networks shift-invariant again"?
• WebGL procedural flowers (demo)
  • Daisy classification by MobileNet
    Why CNNs like daisy? Rotation invariance. Blind Zone effect.
  • "Daisy" and "Daisy" Bouquet Dataset test
    Numbers. "Daisy" Bouquet Dataset is much better.
• MobileNet Heat Maps CAM
  • "Daisy" Class Activation Mapping
    Global Average Pooling. CNNs surgery.
  • Interactive Heat map demo
    Explore images from the Net.
  • How to trick CNNs?
    Classification of randomly splitted images :)
• MobileNet Transfer Learning
  • Tulips with 5 vs 6 petals
    Generation of procedural datasets with embeddings. MobileNet counts much faster than me!
  • Classifying Flowers with CNNs and Transfer Learning
    Port of Roshan Adusumilli's Colab model. tfFlowers dataset.
  • More procedural flowers: Daisy, Tulip, Rose
  • Rose vs Tulip

• Object detection
  • Object Detection coco ssd demo   17 Jan
    Copy/Paste images from Internet.
  • Animated Fruits Detection (coco-ssd)   21 Jan
    Coco likes big fruits (me too :)
  • Triangle vs. Rectangle
    TFjs team demo (Simple Object Detection) + a few comments.
  • Simple Object Detection retrain on Roses? :)
  • "Training Object Detectors with No Real Data using Domain Randomization" by Oliver Gyldenberg Hjermitslev   29 Jan
    Hope, it is not difficult to make demos with this model (weights).

• ToDo: Small object (<30% of the full image) classification/detection failure

• Weight quantization
  • Data-Free Quantization
    Comments to "Data-Free Quantization Through Weight Equalization and Bias Correction" https://arxiv.org/abs/1906.04721
  • "Daisy" weight quantization test
    Inference degradation test for 1-byte quantization! 'relu6' tfjs_converter error.
  • Layers model and Graph model MobileNet_1 weight analysis
    FusedBatchNorm. Too much weights :)
  • Reduced Layers model and Graph model analysis
    BatchNorm layers are folded with Convolution ones. Keras + KITO + Colab.

• How to use TFjs converter
  Weight quantization demo on Windows 10. Quantization-Aware models. Unsupported Ops error!
• Drafts

About TFjs

But Layers models may be highly optimized by TFjs converter (www.tensorflow.org/model_optimization/guide). Therefore we need not only editor? (I usually need "python -m SimpleHTTPServer" or "python -m http.server 8000" too.)

Image classification applications are tested in desktop Chrome and Firefox Beta (32bit) or Dev (64bit) + Windows 10.