yyyang0/VETime-checkpoints

VETime

Vision-Enhanced Zero-Shot Time Series Anomaly Detection

This repository provides only the official pretrained weights for VETime. The complete code is available in the VETime code repository.

VETime is a vision-enhanced framework for zero-shot time-series anomaly detection. It combines temporal and visual representations to detect both short point anomalies and long-range contextual anomalies while retaining timestamp-level localization.

• Paper: arXiv:2602.16681
• Code: yyyang0/VETime
• Checkpoints: yyyang0/VETime-checkpoints

Model overview

VETime contains four main components:

1. Traceable Image Conversion (TIC) transforms a time series into a dense visual representation while preserving temporal correspondence.
2. Patch-Level Temporal Alignment (PTA) maps visual patches back to the temporal axis.
3. Anomaly Window Contrastive Learning (AWCL) improves discrimination of short and long anomaly patterns.
4. Task-Adaptive Multi-Modal Fusion (TMF) dynamically combines temporal, visual, and anomaly-enhanced features.

The model outputs an anomaly score for each timestamp and is evaluated without target-domain training or fine-tuning.

Repository contents

File	Approx. size	Purpose
VETime.pth	664.4 MB	Main VETime checkpoint used for evaluation
weight_v/mae_visualize_base.pth	447.7 MB	MAE-Base visual encoder checkpoint
weight_ts/full_mask_anomaly_head_pretrain_checkpoint_best.pth	444.8 MB	Pretrained time-series encoder/anomaly-head checkpoint, mainly used for initialization and training

VETime.pth and weight_v/mae_visualize_base.pth are required by the current evaluation path. The published Test_TSB.py accepts --ts_path, but the current implementation does not load that argument during evaluation.

Installation

The reference implementation was tested with Python 3.11, PyTorch 2.3.0, and CUDA 12.1.

git clone https://huggingface.co/yyyang0/VETime
cd VETime

conda create -n VETime python=3.11 -y
conda activate VETime

conda install pytorch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt
pip install -U huggingface_hub

Download checkpoints

Download the complete repository directly into the path expected by the code:

hf download yyyang0/VETime-checkpoints --local-dir ./checkpoints

Legacy CLI installations may use:

huggingface-cli download yyyang0/VETime-checkpoints --local-dir ./checkpoints

The resulting layout should be:

VETime/
└── checkpoints/
    ├── VETime.pth
    ├── weight_ts/
    │   └── full_mask_anomaly_head_pretrain_checkpoint_best.pth
    └── weight_v/
        └── mae_visualize_base.pth

Data preparation

VETime is evaluated on the TSB-AD benchmark. Download TSB-AD-U and/or TSB-AD-M and place them under your chosen dataset directory.

For TSB-AD CSV files, the current evaluation script expects:

• the final column to be named Label;
• all earlier columns to contain time-series values;
• the filename to follow the TSB-AD naming convention used by the benchmark code.

The path in config/run_vetime_test.sh may differ from your local layout. Set DATASET_DIR to the directory that actually contains the CSV files.

Evaluation

Option 1: use the provided script

Edit these values in config/run_vetime_test.sh before running it:

DATASET_DIR="./dataset/Datasets/TSB-AD-U"  # change to your actual path
VISION_NAME="mae_visualize_base.pth"
VETIME_PATH="./checkpoints/VETime.pth"

Then run:

bash config/run_vetime_test.sh

Option 2: call the evaluation entry point directly

python Test_TSB.py \
  --model_name VETime \
  --dataset_dir ./dataset/Datasets/TSB-AD-U \
  --save_dir ./output/metrics/uni/ \
  --device cuda:0 \
  --vision_name mae_visualize_base.pth \
  --vetime_path ./checkpoints/VETime.pth \
  --num_workers 10

The script writes per-series anomaly scores and aggregate evaluation CSV files under --save_dir.

Published results

The paper evaluates VETime on 26 public datasets from TSB-AD using Affiliation-F1, F1-T, Standard-F1, and VUS-PR. The macro-averages below are calculated from the per-dataset VUS-PR values in Table 1 of arXiv v3.

Setting	Number of datasets	Macro-average VUS-PR
Univariate	11	49.14%
Multivariate	15	32.31%

Please refer to the paper for per-dataset results, baseline comparisons, ablations, and efficiency analysis.

Intended use

VETime is intended for research on zero-shot univariate and multivariate time-series anomaly detection, including evaluation on TSB-AD-compatible data.

The current release is a PyTorch research checkpoint and is not yet packaged as a standalone Transformers model. Load it through the official VETime codebase rather than transformers.AutoModel.from_pretrained().

Limitations

• Seasonal-trend decomposition can be sensitive to irregular high-frequency noise.
• The channel-independent multivariate design may not fully capture cross-variable dependencies.
• Very high-dimensional inputs may lose fine-grained information when mapped to a fixed visual canvas.
• The vision backbone adds compute and memory overhead.
• The current public code focuses on benchmark evaluation and does not yet expose a minimal arbitrary-series inference API comparable to TimeRCDDetector.

Users should independently validate thresholds, calibration, and failure modes before applying the model to safety-critical or production systems.

License

The official code is released under the Apache License 2.0. Check the licenses and usage terms of TSB-AD and any other datasets separately.

Citation

@article{yang2026vetime,
  title   = {VETime: Vision Enhanced Zero-Shot Time Series Anomaly Detection},
  author  = {Yingyuan Yang and Tian Lan and Yifei Gao and Yimeng Lu and Xuming An and Wenjun He and Meng Wang and Chenghao Liu and Chen Zhang},
  journal = {arXiv preprint arXiv:2602.16681},
  year    = {2026}
}