Awesome

Introduction

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.

Features

For Algorithm Researches

• 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.

For Evaluation Researches

• 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.

For Practitioners of Community or Enterprise

• 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).

Leaderboard Representation

• 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.

Table of content

• Introduction
• Features
• Built-in Methods
• Built-in Evaluation Protocols
• Get Started
  • Installation
  • Prepare datasets
• Usage
• Citation

Built-in Methods

Built-in Evaluation Protocols

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

Get Started

Installation

Prerequisites (environment manager like conda, pipenv or poetry is recommended)

• python >= 3.9, < 3.13

Using pip to install the suite from Pypi

pip install EasyTSAD

The documentation of EasyTSAD is hosted at https://dawnvince.github.io/EasyTSAD/.

Additonal Dependencies

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.

Prepare datasets

Use default datasets

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
}

Add your datasets

Preprocess your dataset to satisfy the above structure and format. Files labeled "necessary" must be offered. Then put it under the datasets/UTS/ path.

Usage

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/.

An example that implements a new method.

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.

Define the Controller

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")

Load Dataset configurations

Option 1: Load certain time series in one dataset:

# 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"
    ]
)

Option 2: Load all time series in certain datasets:

# 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,
)

Implement your algorithm (inherit from class BaseMethod):

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

Do Experiments for your algorithm

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.

Option 1: Use config file for methods (Recommended)

• 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"
)

Option 2: Specify the parameters in functions

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

Perform evaluations (Based on the saved scores)

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

Plot the anomaly scores for each time series

gctrl.plots(
    method=method,
    training_schema=training_schema
)

The Plot Results can be founded in path workspace/Results/Plots/score_only

Citation

@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}
}