blob: 0e15d162dd2e082f3f0e3e0a99a742447b69fd41 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
|
# TensorFlow Time Series
TensorFlow Time Series (TFTS) is a collection of ready-to-use classic models
(state space, autoregressive), and flexible infrastructure for building
high-performance time series models with custom architectures. It includes tools
for chunking and batching a series, and for saving model state across chunks,
making use of parallel computation even when training sequential models on long
series (using truncated backpropagation).
To get started, take a look at the `examples/` directory, which includes:
- Making probabilistic forecasts (`examples/predict.py`)
- Using exogenous features to train on data with known anomalies/changepoints (`examples/known_anomaly.py`)
- Learning correlations between series (multivariate forecasting/anomaly
detection; `examples/multivariate.py`)
- More advanced custom model building (`examples/lstm.py`)
TFTS includes many other modeling tools, including non-linear autoregression
(see the `hidden_layer_sizes` argument to `ARRegressor` in `estimators.py`) and
a collection of components for linear state space modeling (level, trend,
period, vector autoregression, moving averages; see the
`StructuralEnsembleRegressor` in `estimators.py`). Both model classes support
heuristics for ignoring un-labeled anomalies in training data. Trained models
can be exported for inference/serving in
[SavedModel format](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md)
(see `examples/multivariate.py`).
|
