TimeFlow

Diffusion Experiments for Time Series

This directory contains the diffusion/flow-based time series experiments used in the project. The notes below only describe how to set up the environment and run training / sampling.

---

Environment setup

• Python 3.8+
• One CUDA‑capable GPU is recommended.

Create a new environment and install dependencies:

conda create -n timeflow python=3.8
conda activate timeflow
pip install -r requirements.txt

---

Data preparation

Place all datasets under:

Data/dataset/

The exact file names and splits are controlled by the YAML config files in Config/.

---

Training

All experiments are launched with main.py. The most important arguments are:

• --name: experiment name (used to create a subdirectory under OUTPUT/).
• --config_file: path to a YAML config in Config/.
• --train: enable training mode.

python main.py \
  --name {dataset} \
  --config_file Config/{dataset}.yaml \
  --train

During training, checkpoints and logs are saved to:

OUTPUT/{name}/

---

Sampling

After training, use main.py with --sample to generate sequences from a trained checkpoint.

regular sampling

python main.py \
  --name {dataset} \
  --config_file Config/dataset.yaml \
  --sample 0 \
  --milestone {checkpoint_number}

irregular sampling

python main.py \
  --name {dataset} \
  --config_file Config/energy_impute.yaml \
  --gpu 0 \
  --sample 1 \
  --mode infill \
  --missing_ratio {missing ratio} \
  --milestone {checkpoint_number}

Generated samples are saved as `.npy` files under `OUTPUT/{name}/` and can be directly used for evaluation and visualization.