Diffusion based temporal convolutional model for synthetic time-series generation. Main focus is on coherence of autoregressive models' results on real and synthetic data.
Project also includes several time-series generation models implementation for performance comparison:
- QuantGAN
- TTS GAN
- RealNVP
- FourierFlow
Project structure
- results - folder with csv results files
- utils
- utils/dl.py - time-series deep learning models in pytorch with some decorators for training / inference
- utils/synth_eval.py - functions for models evaluation
- utils/timediffusion.py - TimeDiffusion model
- results - folder with csv results files
- <Model_Name>_train_synth.ipynb - jupyter notebooks with training and subsequent generation of synthetic data for specific model
- synth_model_evaluation.ipynb - jupyter notebook with example of generation model quality evaluation
- results_visualization.ipynb - jupyter notebook with visualization of generation models comparison