EasyTSAD is a suite to facilitate the quick implementation and iteration of your time series anomaly detection algorithms. You can also easily develop a new set of evaluation metrics based on this suite and assess them against baseline methods.
We offer several training schemas to provide more practical perspectives for TSAD. The schemas are illustrated as follows:
We welcome you to send the algorithm code implemented based on this suite to our email. We will integrate your method into the algorithm library, making it convenient for researchers and practitioners to utilize your approach or evaluation protocol.
- Flexible interface for algorithm implementation, training and testing new algorithms on naive, all-in-one and zero-shot training schemas.
- Full pipeline for load dataset, run experiments, do evaluations and analysis(e.g. plots and compares anomaly scores, or generate CSVs for intuitional comparison) the performance of methods.
- Diversity evaluation protocols for preformance evaluations.
- Flexible interface for evaluation protocol implementation based on anomaly scores and ground truth labels.
- Easily perform evaluations on existing methods according to your protocol.
- Evaluation based on offline scores of methods, which generated by merely once training and test phase.
- Unified and clear Datasets format, easy for introduction of private datasets.
- Easy performance comparison of baselines on your dataset. E.g. Overall performance in CSV format based on protocols suitable for your applications; Plots of all methods on specify curves.
- Record runtime statistics (e.g. model parameter size, inference time) for performance, cost, and efficiency trade-off.
- An Evaluation protocol designed for real-time AD scenarios (EasyTSAD.Evaluations.Protocols.EventF1PA, for details please refer to our paper).
- We provide a continuous integrated leaderboard (https://adeval.cstcloud.cn/content/home) based on this suite and make it vivid to show state-of-the-art algorithms rankings based on various training schemas and evaluation protocols.
- Welcome to provide us your algorithms or evaluation criterion based on this suite by e-mails. We will add it into the leaderboard after checking, running, and obtaining your permission.
Class Name | Requirements | Description | Ref. |
---|---|---|---|
AR | pytorch | AutoRegression implemented by a torch linear (using first order difference) | Robust regression and outlier detection |
LSTMADalpha | pytorch | LSTMAD in a seq2seq manner | Long Short Term Memory Networks for Anomaly Detection in Time Series |
LSTMADbeta | pytorch | LSTMAD in a multi-step prediction manner | Long Short Term Memory Networks for Anomaly Detection in Time Series |
AE | pytorch | AutoEncoder | Sparse autoencoder |
EncDecAD | pytorch | Combine LSTM and AE | LSTM-based encoder-decoder for multi- sensor anomaly detection |
SRCNN | pytorch | Time-series anomaly detection service at microsoft | |
Amomaly Transformer | pytorch | Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy | |
TFAD | pytorch-lightning | TFAD: A decomposition time series anomaly detection architecture with time-frequency analysis | |
Donut | pytorch | Unsupervised anomaly 1032 detection via variational auto-encoder for seasonal kpis in web applications | |
FCVAE | pytorch-lightning | Revisiting VAE for Unsupervised Time Series Anomaly Detection: A Frequency Perspective | |
DCDetector | pytorch | DCdetector: Dual Attention Contrastive Representation Learning for Time Series Anomaly Detection | |
MatrixProfile | matrixprofile(pypi) | Matrix Profile XI: SCRIMP++: Time Series Motif Discovery at Interactive Speeds | |
SubLOF | sklearn | LOF in sequence manner | LOF: identifying density-based local outliers |
SAND | tslearn==0.4.1 | SAND: streaming subsequence anomaly detection | |
SubOCSVM | sklearn | OCSVM in sequence manner | Support Vector Method for Novelty Detection |
TimesNet | pytorch | TIMESNET: TEMPORAL 2D-VARIATION MODELING FOR GENERAL TIME SERIES ANALYSIS | |
OFA | pytorch | One Fits all, freezing some GPT2 params | One Fits All: Power General Time Series Analysis by Pretrained LM |
FITS | pytorch | FITS: Modeling Time Series with 10K Parameters |
For EVENT-BASED methods, there are two parameters when initializing the class:
mode (str): Defines the scale at which the anomaly segment is processed. \n
One of:\n
- 'squeeze': View an anomaly event lasting t timestamps as one timepoint.
- 'log': View an anomaly event lasting t timestamps as log(t) timepoint.
- 'sqrt': View an anomaly event lasting t timestamps as sqrt(t) timepoint.
- 'raw': View an anomaly event lasting t timestamps as t timepoint.
If using 'log', you can specify the param "base" to return the logarithm of x to the given base,
calculated as log(x) / log(base).
base (int): Default is 3.
For more details, please refer to the document
Class Name | Type | Description | Ref. |
---|---|---|---|
PointF1 | point-based | traditional F1 | |
PointPrc | point-based | traditional AUPRC | |
PointRoc | point-based | traditional AUROC | |
PointF1PA | point-based | F1 using point-adjustment | Robust Anomaly Detection for Multivariate Time Series through Stochastic Recurrent Neural Network |
PointKthF1PA | point-based | add k-delay constraint on F1PA | |
PointAuprcPA | point-based | ||
PointAurocPA | point-based | ||
VUS | range-based | Volume Under the Surface: A New Accuracy Evaluation Measure for Time-Series Anomaly Detection | |
EventDetect | event-based | Only suitable for dataset UCR | |
EventF1PA | event-based | view an anomaly segment as an event | |
EventPrcPA | event-based | event-based AUPRC | |
EventRocPA | event-based | event-based AUROC | |
EventKthPrcPA | event-based | event-base AUPRC under k-delay | |
EventKthRocPA | event-based | event-base AUROC under k-delay |
- python >= 3.9, < 3.13
pip install EasyTSAD
The documentation of EasyTSAD is hosted at https://dawnvince.github.io/EasyTSAD/.
NOTE: Some built-in algorithms are based on Pytorch 2.0 or Pytorch-lightning 2.0. You may need to install related packages (including but not limited to pytorch, pytorch-lightning, torchinfo, torch_optimizer, scikit-learn, tslearn) if you want to run the baselines.
There may be some version conflicts when using tslearn==0.4.1 (SAND). We recommend you to build and activate an independent conda environment using python 3.7 for these methods.
Original datasets can be downloaded from https://github.com/CSTCloudOps/datasets. The directory structure of the dataset is shown as follows:
datasets
└── UTS
├── dataset_1
│ ├── time_series_1
│ │ ├── train.npy (training set, 1-D ndarray)
│ │ ├── test.npy (test set, 1-D ndarray)
│ │ ├── train_label.npy (labels of training set, 1-D ndarray)
│ │ ├── test_label.npy (labels of test set, 1-D ndarray)
│ │ ├── train_timestamp.npy (timestamps of training set, 1-D ndarray)
│ │ ├── test_timestamp.npy (timestamps of test set, 1-D ndarray)
│ │ └── info.json (some additonal information, json)
│ │
│ ├── time_series_2
│ └── ...
│
├── dataset_2
└── ...
The file info.json
contains the information like:
{
"intervals": 300, # Minimum time interval
"training set anomaly ratio": 0.00148,
"testset anomaly ratio": 0.00808,
"total anomaly ratio": 0.00478
}
Preprocess your dataset to satisfy the above structure and format. Files labeled "necessary" must be offered. Then put it under the datasets/UTS/
path.
Examples of how to use the suite can be find here, including:
- run baselines with/without customized config files;
- implement your new algorithm with/without config files;
- implement your new evaluation protocol and evaluate the baselines;
- generate CSV including the overall performance of all trained methods;
- aggregate all methods' anomaly scores into one plot.
Also, you can refer to the documentation hosted at https://dawnvince.github.io/EasyTSAD/.
Prepare a global config toml file. If not provided, the default configuration will be applied:
# One example of GlobalCfg.toml.
# For more details please refer to the default configuration.
# The new items will overwrite the default ones.
[DatasetSetting]
train_proportion = 1 # Using the last x% of the training set as the new training set. 1 means use the full training set.
valid_proportion = 0.2 # The proportion of the validation set to the new training set.
from typing import Dict
import numpy as np
from EasyTSAD.Controller import TSADController
# if cfg_path is None, using default configuration
gctrl = TSADController(cfg_path="/path/to/GlobalCfg.toml")
# Specify certain curves in one dataset,
# e.g. AIOPS 0efb375b-b902-3661-ab23-9a0bb799f4e3 and ab216663-dcc2-3a24-b1ee-2c3e550e06c9
gctrl.set_dataset(
dataset_type="UTS",
dirname="/path/to/datasets", # The path to the parent directory of "UTS"
datasets="AIOPS",
specify_curves=True,
curve_names=[
"0efb375b-b902-3661-ab23-9a0bb799f4e3",
"ab216663-dcc2-3a24-b1ee-2c3e550e06c9"
]
)
# Use all curves in datasets:
datasets = ["AIOPS", "Yahoo"]
gctrl.set_dataset(
dataset_type="UTS",
dirname="/path/to/datasets", # The path to the parent directory of "UTS"
datasets=datasets,
)
The following class YourAlgo
just provides a skeleton, where you should implement several functions.
- The Spot instance will help you understand how to implement a statistic model;
- The ARLinear instance will help you understand how to implement a learning-based model (Implemented using PyTorch);
from EasyTSAD.Methods import BaseMethod
from EasyTSAD.DataFactory import TSData
class YourAlgo(BaseMethod):
def __init__(self, hparams) -> None:
super().__init__()
self.__anomaly_score = None
self.param_1 = hparams["param_1"]
def train_valid_phase(self, tsTrain: TSData):
...
def test_phase(self, tsData: TSData):
result = ...
self.__anomaly_score = result
def train_valid_phase_all_in_one(self, tsTrains: Dict[str, TSData]):
# used for all-in-one and zero-shot mode
...
def anomaly_score(self) -> np.ndarray:
return self.__anomaly_score
def param_statistic(self, save_file):
pass
We offer two options for algorithm setting configuration:
- use config file;
- specify the parameters in functions.
Note: Parameters defined within a function take higher priority than those specified in the configuration file.
- Prepare a toml file, which is a subset of Example.toml, for example:
# YourAlgo.toml
[Data_Params]
preprocess = "z-score"
[Model_Params.Default]
param_1 = false
- Load YourAlgo and the config file:
training_schema = "naive"
method = "YourAlgo" # string of your algo class
# run models
gctrl.run_exps(
method=method,
training_schema=training_schema,
cfg_path="path/to/YourAlgo.toml"
)
gctrl.run_exps(
method=method,
training_schema=training_schema,
hparams={
"param_1": False,
},
preprocess="z-score",
)
The Score Results can be founded in path workspace/Results/Scores
, and the runtime information can be founded in path workspace/Results/RunTime
from EasyTSAD.Evaluations.Protocols import EventF1PA, PointF1PA
# Specifying evaluation protocols
gctrl.set_evals(
[
PointF1PA(),
EventF1PA(),
EventF1PA(mode="squeeze")
]
)
gctrl.do_evals(
method=method,
training_schema=training_schema
)
The Evaluation Results can be founded in path workspace/Results/Evals
gctrl.plots(
method=method,
training_schema=training_schema
)
The Plot Results can be founded in path workspace/Results/Plots/score_only
@misc{si2024timeseriesbench,
title={TimeSeriesBench: An Industrial-Grade Benchmark for Time Series Anomaly Detection Models},
author={Haotian Si and Changhua Pei and Hang Cui and Jingwen Yang and Yongqian Sun and Shenglin Zhang and Jingjing Li and Haiming Zhang and Jing Han and Dan Pei and Jianhui Li and Gaogang Xie},
year={2024},
eprint={2402.10802},
archivePrefix={arXiv},
primaryClass={cs.LG}
}