Bagel

Bagel is a robust and unsupervised KPI anomaly detection algorithm based on conditional variational autoencoder.

This is an implementation of Bagel in TensorFlow 2. The original PyTorch 0.4 implementation can be found at NetManAIOps/Bagel.

Install

pip will automatically install required PyPI dependencies when you install this package:

For development use:

git clone https://github.com/alumik/bagel-tensorflow.git
cd bagel-tensorflow
pip install -e .

For production use:

pip install git+https://github.com/alumik/bagel-tensorflow.git

An environment.yml is also provided if you prefer conda to manage dependencies:

conda env create -f environment.yml

Run

KPI Format

KPI data must be stored in csv files in the following format:

timestamp,   value,       label
1469376000,  0.847300274, 0
1469376300, -0.036137314, 0
1469376600,  0.074292384, 0
1469376900,  0.074292384, 0
1469377200, -0.036137314, 0
1469377500,  0.184722083, 0
1469377800, -0.036137314, 0
1469378100,  0.184722083, 0

timestamp: timestamps in seconds (10-digit).
label (optional): 0 for normal points, 1 for anomaly points.
Labels are used only for evaluation and are not required in model training and inference. However, if labels are provided, the model can still take labeled data to improve the performance.

Sample Script

A sample script can be found at sample/main.py:

Usage

To prepare the data:

import bagel

kpi = bagel.data.load_kpi('kpi.csv')
kpi.complete_timestamp()
train_kpi, valid_kpi, test_kpi = kpi.split((0.49, 0.21, 0.3))
train_kpi, mean, std = train_kpi.standardize()
valid_kpi, _, _ = valid_kpi.standardize(mean=mean, std=std)
test_kpi, _, _ = test_kpi.standardize(mean=mean, std=std)
dataset = bagel.data.KPIDataset(
    train_kpi.use_labels(0.),
    window_size=window_size,
    time_feature=time_feature,
    missing_injection_rate=missing_injection_rate,
)
valid_dataset = bagel.data.KPIDataset(valid_kpi, window_size=window_size, time_feature=time_feature)
test_dataset = bagel.data.KPIDataset(test_kpi.no_labels(), window_size=window_size, time_feature=time_feature)

To build and train a Bagel model:

model = bagel.Bagel(
    window_size=window_size,
    hidden_dims=hidden_dims,
    latent_dim=latent_dim,
    dropout_rate=dropout_rate,
)
lr_scheduler = tf.keras.optimizers.schedules.ExponentialDecay(
    initial_learning_rate=learning_rate,
    decay_steps=10 * len(dataset) // batch_size,
    decay_rate=0.75,
    staircase=True,
)
optimizer = tf.keras.optimizers.Adam(learning_rate=lr_scheduler, clipnorm=clipnorm)
model.compile(optimizer=optimizer, jit_compile=True)
model.fit(
    x=[dataset.values, dataset.time_code, dataset.normal],
    batch_size=batch_size,
    epochs=epochs,
    validation_data=([valid_dataset.values, valid_dataset.time_code, valid_dataset.normal], None),
    validation_batch_size=batch_size,
)

To use the trained model for prediction:

anomaly_scores = model.predict(
    x=[test_dataset.values, test_dataset.time_code, test_dataset.normal],
    batch_size=batch_size,
)

Use tf.keras.Model.save API to save the model.

Citation

@inproceedings{conf/ipccc/LiCP18,
    author    = {Zeyan Li and
                 Wenxiao Chen and
                 Dan Pei},
    title     = {Robust and Unsupervised {KPI} Anomaly Detection Based on Conditional
                 Variational Autoencoder},
    booktitle = {37th {IEEE} International Performance Computing and Communications
                 Conference, {IPCCC} 2018, Orlando, FL, USA, November 17-19, 2018},
    pages     = {1--9},
    publisher = {{IEEE}},
    year      = {2018},
    url       = {https://doi.org/10.1109/PCCC.2018.8710885},
    doi       = {10.1109/PCCC.2018.8710885}
}