SimVP video prediction using ONNX's Wasm runtime. You can use the notebook linked below to train your own video prediction model that runs in real-time in the browser. SimVP is really impressive given how fast it runs - see the results that compare to other methods in the paper.
Note that these examples are probably under-trained. They're just proof-of-concepts to get the whole pipeline (from training to browser inference) working properly.
- Various models (color): https://josephrocca.github.io/SimVP-web/demo/simple.html
- Same as above, but infinite/recursive generation: https://josephrocca.github.io/SimVP-web/demo/simple-infinite.html
- Moving MNIST (greyscale): https://josephrocca.github.io/SimVP-web/demo/mmnist.html
- Moving MNIST: https://colab.research.google.com/gist/josephrocca/80c4a9814621d228dfb3d262a26a0806
- Train using your own custom video dataset: https://colab.research.google.com/gist/josephrocca/7be01186b14e07b5d54e80c245a35570
- https://colab.research.google.com/gist/josephrocca/eccbb3e5ff92b8326d7e2627827b6323/simvp2-diffusion.ipynb
- https://josephrocca.github.io/SimVP-web/demo/diffusion.html