ProDiffAD

This repository is the official implementation of ProDiffAD: Progressively Distilled Diffusion Models for Multivariate Time Series Anomaly Detection in JointCloud Environment

Requirements

To install requirements:

pip install -r requirements.txt

Training

To train the models in the paper, run this command:

python train_diffusion_val.py --training diffusion

Distillation

To distill model from teacher to student:

python train_diffusion_val.py --training distill

Inference

To perform inference (by onnx):

python train_diffusion_val.py --input <model> --use_onnx True (--onnx_name  <path_to_onnxfile>)

For example:

python train --dataset point_global --denoise_steps 64 --batch_size 8 --training distill --lr 1e-4 --epoch 5 --train_loss_begin 0.10 --window_size 64 --noise_steps 512 --input point_global_128_128_trial --output point_global_128_64_trial --use_onnx True --onnx_name 1.onnx --test_only True