6.883VAEDemo

VAE Demo Page for MIT Spring 2018 6.883

Content

.
├── index.html
├── model
│   ├── decoder_generator.png
│   ├── encoder.py
│   ├── mnist-cnn-vae.py
│   ├── model.bin
│   ├── model.h5
│   └── model_pb2.py
└── README.md

• index.html: all html, javascript, css in one file
• minst-cnn-vae.py: define and train CNN-VAE model, run with python3 minst-cnn-vae.py, will produce decoder_generator.png and model.h5
• decoder_generator.png: a visualization of the generated images
• encoder.py: (which use model_pb2.py): weight-wise quantization to 8-bit float, convert model.h5 to model.bin, which is used by the javascript in index.html

Replacing the Model

• Train a Keras model and save as model.h5
• Convert the model by encoder.py, get model.bin, place in ./model/ folder
• Change the name of input layer and output layer line 43 and line 45 in index.html
• Check it works

Improve the UI

• If you have any ideas and suggestions on improving the UI, please create a issue!