Open-source implementations behind my research, covering time-series foundation models, forecasting and generative models of functions. Each card links to the code and the paper it accompanies.
A foundation model for macroeconomic forecasting, built on Moirai and fine-tuned on multi-country economic data for probabilistic, zero-shot forecasts of economic indicators.
$ git clone https://github.com/bkoyuncu/mtsfm.git && cd mtsfm
$ pip install -e '.[notebook]'
# pull pre-trained weights from the Hub
$ huggingface-cli download bkoyuncu/MacroEconModels --local-dir ./pretrainedProbabilistic implicit neural representations for irregular multivariate time series, giving individualized imputation and forecasting in a single forward pass, with no per-series fine-tuning.
$ pip install -r requirements_pip.txt && pip install -e .
$ cd run
$ python main_wandb.py --cfg ./configs/traffic_tvinr_200.yamlLatent diffusion over implicit neural representations, using a Transformer hypernetwork decoder that can be trained from scratch or "hyper-transformed" from a pre-trained LDM.
$ conda env create -f environment.yaml && conda activate ldmi
$ python main.py --base configs/ldmi/<config_spec>.yaml -t --gpus 0,A lighter probabilistic transformer for multivariate forecasting, released with the first public re-implementation of the ProTran baseline, including a corrected ELBO.
$ pip install -r requirements.txt
$ python run.py --yaml_file example_cfgs/eprotran_electricity.yamlLearns distributions over functions by combining hypernetworks with a mixture of generators, so the same model handles images, 3D shapes and climate data, including point-dropout settings.
$ pip install -r requirements_pip.txt && pip install -e .
$ cd run
$ python main.py --cfg ./configs/models/shapes.yaml device "cuda:0" dataset.name shapes3d_10A compact notebook that builds a neural ODE from first principles with an explicit Euler solver. It was a stepping stone to my work on deep generative second-order ODEs (ODE2VAE).
# z_{t+1} = z_t + h * f_theta(z_t, t)
$ jupyter notebook Playground_neuralODE_euler.ipynb