Practical Equivariances via Relational Conditional Neural Processes

This repository provides the implementation and code used in the experiments for the article Practical Equivariances via Relational Conditional Neural Processes (Huang et al., NeurIPS 2023) [1]. The full paper can be found at arXiv.

Note: Our implementations build on the neuralprocesses library by Wessel Bruinsma. You can find further details on its structure and the dependencies at the github website.

Installation

``` git clone https://github.com/acerbilab/relational-neural-processes.git

cd relational-neural-processes

conda create --name np python=3.10.10

conda activate np

pip install -r requirements.txt ```

To use our RCNP models, you can specify --model as one of [rcnp, rgnp, fullrcnp, fullrgnp], where the latter two represents the full version of our models. You can also choose comparison function for relational encoding by using --comparison-function, where difference stands for encoding translational equivariance, and distance stands for encoding equivariance for rigid transformations.

We also support partial encoding, where some dimensions of the data do not need relational encoding, you can specify the dimensions that do not require relational encoding through --non-equivariant-dim, e.g., --non-equivariant-dim=“0,1” means the first and the second dimension will not be encoded by relational encoder.

Synthetic Regression

To run the synthetic regression experiment, you can use the following template command lines: angular2html python train.py \ --model='rcnp' \ --data='eq' \ --dim-x=2 \ --dim-y=1 \ --seed=1 \ --comparison-function='difference' For the above template command, we use RCNP model with difference comparison function on x2_y1 eq data.

Bayesian Optimization

To run the Bayesian optimization experiment first run python bayesopt/bo_train.py \ --target=hartmann3d \ --exp="bo_fixed" \ --model="rcnp" to pretrain a model with the desired set of kernels. Here, --target is one of [hartmann3d, hartmann6d], --exp is one of [bo_fixed, bo_matern, bo_single, bo_sumprod] corresponding to scenario (i)-(iv) in the paper. --model is one of [cnp, gnp, acnp, agnp, rcnp, rgnp].

Once trained, apply it to the BO task as python bayesopt/bo_apply.py \ --exp="bo_fixed" \ --target_name="hartmann3d" \ --n_rep=10 \ --n_steps=50 where --exp and --target_name as above and --n_rep, --n_steps specify the number of replications and the number of query steps respectively.

Lotka-Volterra Model

For the Lotka-Volterra model experiments, we used the "predprey" experiment available in the original package. To rerun our CNP experiments, use python train.py \ --data="predprey" \ --seed=x \ --model=m \ --enc-same

with x in (1, 10) and m in (rcnp, cnp, acnp, convcnp).

To rerun our GNP experiments, use python train.py \ --data="predprey" \ --seed=x \ --model=m \ --enc-same \ --num-basis-functions=32 with x in (1, 10) and m in (rgnp, gnp, agnp, convgnp)

Reaction-Diffusion Model

For the Reaction-Diffusion example, all the parameters selected for the experiments are listed in experiments/data/cancer.py.

The dataset used from the Reaction-Diffusion simulator is stored in the experiments/data/dataset_cancer folder, along with the python file that produces simulations (RD-simulator.py). To run experiments, use, with x the seed chosen and m the method chosen (rcnp, rgnp, cnp, gnp, acnp, agnp): python train.py \ --data="cancer" \ --seed=x \ --model=m For the additional study, run: python train.py \ --data="cancer_latent" \ --seed=x \ --model=m \ --non-equivariant-dim="3" \ --comparison-function="difference" with x in (1, 10) and m in (rgnp, gnp, agnp, convgnp)

Customize your own comparison function

Most examples in our paper use the difference and distance comparison functions that are available as built-in options. However, you can also write your own comparison function and pass it to the model constructor.

See neuralprocesses/comparison_functions.py for example comparison functions and experiments/data/image.py for an example on how to pass a custom comparison function to the model constructor.

Citation

Please cite our paper if you find this work useful for your research @article{huang2023practical, title={Practical Equivariances via Relational Conditional Neural Processes}, author={Huang, Daolang and Haussmann, Manuel and Remes, Ulpu and John, ST and Clart{\'e}, Gr{\'e}goire and Luck, Kevin Sebastian and Kaski, Samuel and Acerbi, Luigi}, journal={Advances in Neural Information Processing Systems}, year={2023} } For the full paper, please refer to our arXiv page.

License

This code is under the MIT License.