Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/timetoai/timediffusion
Unified Framework for Multiple Time Series Tasks
https://github.com/timetoai/timediffusion
data-science deep-learning deep-neural-networks framework machine-learning machine-learning-algorithms multi-task multi-task-architecture multiple-tasks open-source pypi-package python python3 pytorch pytorch-implementation time-series time-series-forecasting time-series-imputation time-series-prediction time-series-simulation
Last synced: 2 days ago
JSON representation
Unified Framework for Multiple Time Series Tasks
- Host: GitHub
- URL: https://github.com/timetoai/timediffusion
- Owner: timetoai
- License: mit
- Created: 2023-08-30T10:44:27.000Z (about 1 year ago)
- Default Branch: main
- Last Pushed: 2023-10-13T09:53:32.000Z (about 1 year ago)
- Last Synced: 2024-10-05T19:04:52.393Z (about 1 month ago)
- Topics: data-science, deep-learning, deep-neural-networks, framework, machine-learning, machine-learning-algorithms, multi-task, multi-task-architecture, multiple-tasks, open-source, pypi-package, python, python3, pytorch, pytorch-implementation, time-series, time-series-forecasting, time-series-imputation, time-series-prediction, time-series-simulation
- Language: Python
- Homepage:
- Size: 2.04 MB
- Stars: 21
- Watchers: 2
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# TimeDiffusion - Unified Framework for Multiple Time Series Tasks
Supports 2D (image) and 3D (video) data as input for research purposes.
## Contents
* [Install](#install)
* [Quick Start](#quick-start)
* [Examples](#examples)
* [Philosophy](#philosophy)
* [Base Models](#base-models)
* [Model architecture](#model-architecture)
## Install
```bash
pip install timediffusion
```
## Quick Start
**Forecasting time series**
```python
# train sequence in shape [channels, sequence_length]
model = TD(input_dims=train.shape).to(device=device)
training_losses = model.fit(train)
# horizon : int - how many future values to forecast
predictions = model.forecast(horizon)
```
**Creating synthetic time series**
```python
# sequence in shape [channels, sequence_length]
model = TD(input_dims=seq.shape).to(device=device)
training_losses = model.fit(seq)
# proximity - how close to original, samples - total synthetic time series
synthetic_data = model.synth(proximity=0.9, samples=3, batch_size=2, step_granulation=10)
```
**Time series Imputation**
```python
# sequence in shape [channels, sequence_length]
model = TD(input_dims=seq.shape).to(device=device)
# mask - binary array of same shape, as sequence, with 1 in positions, that are unknown
training_losses = model.fit(seq, mask=mask)
restored_seq = model.restore(example=seq, mask=mask)
```
## Examples
[Time series: multiple tasks example](./examples/example_1d_data.ipynb)
[Forecasting bitcoin price example](./examples/bitcoin_forecast.ipynb)
## Philosophy
Main synopsis behind TimeDiffusion model is that in reality, when working with time series we don’t have many samples, as it could be in other machine learning fields (e.g. cv, nlp). Thus, classical autoregressive approaches like ARIMA has the most suitable approach of fitting / training only on original sequence (maybe with some exogenous data).
TimeDiffusion takes inspiration from these established methods and only trains on the input sample. Model incorporates most powerful modern deep learning techniques such as diffusion process, exponential dilated convolutions, residual connections and attention mechanism (in one of the versions)
## Base Models
* `TimeDiffusionProjector` - exponential dilated convolutions + residual connections.
Currently main model in use
* `TimeDiffusionAttention` - attention mechanism on top of `TimeDiffusionProjector` (q, k, v) projectors.
Currently not viable
* `TimeDiffusionLiquid` - exponential dilated convolutions with shared middle convolutional layer weight.
Lightweight, fast, but less accurate than the main model.
## Model architecture
Presented below are diagrams depicting model components, each new scheme representing a higher level of abstraction.
* Temporal Block
* TimeDiffusionProjector
* TimeDiffusionAttention
