Skip to content

Latest commit

 

History

History
129 lines (102 loc) · 5.45 KB

README.md

File metadata and controls

129 lines (102 loc) · 5.45 KB

Variational Gaussian Process Diffusion Processes

This directory contains the experiments included in the publication:

  • P. Verma, V. Adam, and A. Solin, (2024). Variational Gaussian Process Diffusion Processes. In Proceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS) 2024. arXiv

Requirements

To install additional packages required for the experiments in this directory, run the following command:

pip install -r dp_requirements.txt

Hydra configurations

All the experiments are configured using the Hydra configuration system. The configuration files are available in the configs/ directory from where the hyperparameters for each experiments can be set. If you want to run the experiments with different hyperparameters, you can modify the configuration files or pass the hyperparameters as command line arguments.

Synthetic diffusion processes experiments

This section discuss the experiments to re-create Table 1-2, Fig. 1-4, and Fig. A8-A11 of the paper.

Generate the data

The data can be generated using the script generate_data.py. A sample command for generating data for a linear (Ornstein--Uhlenbeck) process is:

python generate_data.py -d=0.5 -q=1.0 -t0=0.0 -t1=100. -x0=1 -dt=0.01 -n=50 --o="data/linear" -s=15 -sde="ou" -k=5

More arguments can be seen by the command python generate_data.py --help

Note: The data used for the experiments in the paper can be found in docs/diffusion_processes/data.zip.

Inference experiments

We can run the experiments when only inference is performed using the following sample commands:

VDP (Archambeau et al., 2007)

python vi_markov_gp.py --config-name="vi_linear_process.yaml" data_path="data/linear/15/0.npz" prior_sde="ou" prior_sde.decay=1.2 dt=0.01 trainer.warmup_x0_itr=5 trainer.q_lr=0.1 trainer.x0_lr=0.1

The above command runs the VDP model for the linear DP (Ornstein-Uhlenbeck).

CVI-DP (Verma et al., 2024)

python cvi_dp.py --config-name="cvi_linear_process.yaml" data_path="data/linear/15/0.npz" prior_sde="ou" prior_sde.decay=1.2 dt=0.01

The above command runs the CVI-DP model for the linear DP (Ornstein-Uhlenbeck).

To run the CVI-DP model for the non-linear DP, use the following command:

python cvi_dp.py --config-name="cvi_non_linear_process.yaml" data_path="data/dw/3/0.npz" prior_sde="dw" dt=0.01 trainer.max_itr_sites_optim=20 trainer.girsanov_sites_lr=0.5 trainer.data_sites_lr=0.5 prior_sde.c=1.0

GPR model

python gpr_linear.py data_path="data/linear/15/0.npz" decay=1.2

The above command runs the GPR model for the linear DP (Ornstein-Uhlenbeck).

To run the GPR model for the non-linear DP, use the following command:

python gpr_non_linear.py data_path="data/dw/3/0.npz"

Learning experiments

We can run the experiments when both inference and learning are performed using the following sample commands:

VDP (Archambeau et al., 2007)

python vi_markov_gp.py --config-name="vi_linear_process.yaml" data_path="data/linear/15/0.npz" prior_sde="ou" prior_sde.decay=2.5 dt=0.01 trainer.learn_prior_sde=true trainer.warmup_x0_itr=0 trainer.optimize_prior_initial_state=true trainer.q_lr=0.1 trainer.x0_lr=0.1 trainer.prior_sde_lr=0.01

The above command runs the VDP model for the linear DP (Ornstein-Uhlenbeck).

CVI-DP (Verma et al., 2024)

python cvi_dp.py --config-name="cvi_linear_process.yaml" data_path="data/linear/15/0.npz" prior_sde="ou" prior_sde.decay=2.5 dt=0.01 trainer.learn_prior_sde=true trainer.learning_max_itr=1000 trainer.max_itr=10

The above command runs the CVI-DP model for the linear DP (Ornstein-Uhlenbeck).

GPR model

python gpr_linear.py data_path="data/linear/15/0.npz" decay=2.5 optimize=true

The above command runs the GPR model for the linear DP (Ornstein-Uhlenbeck).

Real world experiment

This section discusses the experiments to re-create Fig. 6,7, and A12 of the paper.

Finance data experiment

The Apple Inc. share price experiment can be performed using the following sample commands:

VDP (Archambeau et al., 2007)

python vi_markov_gp.py --config-name=vi_apple_stock_process.yaml data_path="data/apple_stock/0_data.npz" dt=0.001

CVI-DP (Verma et al., 2024)

python cvi_dp.py --config-name=cvi_apple_stock_process.yaml data_path="data/apple_stock/0_data.npz"

GPR

cd stock/
python gpr_stock.py data_path="../data/apple_stock/0_data.npz" constant_kernel_variance=0.25 linear_kernel_variance=0.25

SGPR

cd stock/
python sgpr_stock.py data_path="../data/apple_stock/0_data.npz"

Vehicle tracking experiment

The vehcile tracking experiment can be performed using the following sample comm+ands:

VDP (Archambeau et al., 2007)

python vi_markov_gp.py --config-name=vi_gps_process.yaml data_path="data/gps/0_data.npz"  dataloader.train_dim=0

CVI-DP (Verma et al., 2024)

python cvi_dp.py --config-name=cvi_gps_process.yaml data_path="data/gps/0_data.npz" dataloader.train_dim=0

Comparison with NeuralSDE method

The comparison experiment with NeuralSDE is implemented using the torchsde library. The code is available in the directory docs/difussion_process/neuralsde and a sample experiment can be run by the following command:

cd neuralsde
python main.py --data_dir="../data/dw/3/0.npz"