# TensorFlow Time Series： TensorFlow时间序列（TFTS）

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 whatever the architecture. 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 (see examples/multivariate.py).

TensorFlow时间序列（TFTS）是一系列即用型经典模型（状态空间，自动回归）和灵活的基础架构，用于构建高性能时间序列模型，无论架构如何。它包括用于分块和批量系列的工具，并且用于在块之间保存模型状态，即使在长序列上训练序列模型（使用截断反向传播）也可以使用并行计算。

TFTS​​包括许多其他的建模工具，包括非线性自回归（参见estimators.py中的ARRegressor的hidden_​​layer_sizes参数）以及线性状态空间建模（级别，趋势，周期，向量自回归，移动平均值）的组件集合;请参阅StructuralEnsembleRegressor在estimateators.py中）。两个模型类都支持启发式，用于忽略训练数据中未标记的异常。受过训练的模型可以以SavedModel格式导出推理/服务（参见examples / multivariate.py）。

https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/timeseries