https://github.com/statmixedml/convts-mixer
Time Series Extension of "Patches Are All You Need"
https://github.com/statmixedml/convts-mixer
Last synced: 6 months ago
JSON representation
Time Series Extension of "Patches Are All You Need"
- Host: GitHub
- URL: https://github.com/statmixedml/convts-mixer
- Owner: StatMixedML
- License: apache-2.0
- Created: 2023-04-05T12:24:56.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2023-10-29T09:27:32.000Z (over 1 year ago)
- Last Synced: 2024-05-15T09:36:26.560Z (about 1 year ago)
- Language: Jupyter Notebook
- Size: 18.7 MB
- Stars: 6
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# ConvTSMixer, MlpTSMixer and TsT
Time Series Extension of "Patches Are All You Need"
Implementations of some prototypical time series mixers based on Conv, MLP, and ViT archs. modified for the probabilistic multivariate forecasting use case, where the emission head is currently an "independent same-family" distribution, e.g., diagonal Student-T.
In everything that follows, the inputs are typically 4-Tensors of shape `[Batch, Variate-dim, Context-length, 1+Features]`, and during training, the subsequent prediction window values are given `[B, Variate-dim, Pred-length]`. The inputs are embedded via 2d-conv to obtain patch embeddings:

## ConvTSMixer

## MlpTSMixer

## TsT (ViT style)

## Output head
