Calibdetector

🔍 Calibrating Foundation Models for Training-Free Time Series Anomaly Detection

Table of Contents

• 🏄 Overall Framework
• 📄 Main Results
• ⚙️ Setup
• 🗄️ Prepare Datasets
• 🤖 Prepare Foundation Models
• 🔁 Reproduce the Main Results
• 🔁 Reproduce the Baseline Results
• 🔁 Reproduce the Ablation Study Results

---

🏄 Overall Framework

📄 Main Results

---

⚙️ Setup

Installation

conda create -n CalibDetector python=3.11
conda activate CalibDetector
pip install -r requirements.txt

🗄️ Prepare Datasets

Download the datasets from Google Drive. Extract the downloaded files into the data directory in the project root and organize them as follows:

data/
├── UCR/
└── TSB_AD/

🤖 Prepare Foundation Models

Download the required foundation models from Hugging Face:

• MOMENT-1-large
• TiRex
• Chronos-2
• PatchTST-FM

After downloading the models, update the default values of --moment_checkpoint_path, --tirex_checkpoint_path, --chronos_checkpoint_path, and --patchtst_fm_checkpoint_path in meta_detection.py to match their local checkpoint directories.

---

🔁 Reproduce the Main Results

1. Generate the synthetic training data:

python ./data/generated_series_new.py generate --n_samples 6000

2. Run the main experiments:

sh runners/run_meta.sh

Follow the instructions in runners/run_meta.sh to change the backbone. The default backbone is MOMENT.

🔁 Reproduce the Baseline Results

Data Mining Baselines

sh runners/run_baseline_dm.sh

Other Deep Learning Baselines

sh runners/run_train_deep_detection.sh

Foundation Model Baselines

After downloading the models, update the corresponding checkpoint paths to match the local model directories before running the foundation model baselines.

sh runners/run_all_pretrain.sh

🔁 Reproduce the Ablation Study Results

Run each ablation study separately:

sh runners/run_ablation_patching.sh
sh runners/run_ablation_protofrac.sh
sh runners/run_ablation_sampling.sh
sh runners/run_ablation_composer.sh