Project Description

Offline black-box optimization aims to maximize a black-box function using an offline dataset of designs and their measured properties. Our project propose Robust Guided Diffusion (RGD) for Offline Black-box Optimization, melding the advantages of proxy and proxy-free diffusion for robust guided diffusion.

RGD is comprised of two key modules:

1. Proxy-enhanced Sampling: utilizes the explicit guidance from a trained proxy to bolster the proxy-free diffusion with a stronger control.
2. Diffusion-based Proxy Refinement: integrate insights from the proxy-free diffusion back into the proxy.

Installation

The project relies on several essential libraries, including PyTorch and Design-Bench. Install the required dependencies with:

pip install -r requirements.txt

Reproducing Performance

To replicate our method's performance for specific tasks, follow these steps:

For the AntMorphology-Exact-v0 task:

• Train a proxy:

  python main.py --task AntMorphology-Exact-v0 --mode classifier_training --epochs 200

• Train a proxy-free diffusion model:

  python main.py --task AntMorphology-Exact-v0 --mode diffusion_training_elbo --epochs 1000

• Fine-tune the classifier with diffusion-based proxy refinement:

  python main.py --task AntMorphology-Exact-v0 --mode classifier_finetuning

• Run proxy-enhanced sampling:

  python main.py --task AntMorphology-Exact-v0 --mode both

The same steps apply to the discrete task TFBind10-Exact-v0:

python main.py --task TFBind10-Exact-v0 --mode [MODE] --epochs [EPOCHS]

Replace [MODE] with the desired mode and [EPOCHS] with the number of training epochs.

Hyperparameter Sensitivity

Evaluate the impact of various hyperparameters such as the number of diffusion steps $T$, the condition label $y$, and the sample learning rate $\eta$ on the model's performance.

• For different $T$ values (500, 750, 1000, 1250, 1500):

  T_values=(500 750 1000 1250 1500)
  for T in ${T_values[*]};
  do
      python main.py --task AntMorphology-Exact-v0 --num_steps $T
  done

• For different $y$ values (0.5, 1.0, 1.5, 2.0, 2.5):

  y_values=(0.5 1.0 1.5 2.0 2.5)
  for y in ${y_values[*]};
  do
      python main.py --task AntMorphology-Exact-v0 --y_ratio $y
  done

• For different sample learning rate $\eta$ [2.5e-3, 5.0e-3, 1.0e-2, 2.0e-2, 4.0e-2]:

  sample_lrs=(2.5e-3 5.0e-3 1.0e-2 2.0e-2 4.0e-2)
  for sample_lr in ${sample_lrs[*]};
  do
      python main.py --task AntMorphology-Exact-v0 --sample_lr $sample_lr
  done

Acknowledgements

We extend our gratitude to the following resources for their invaluable contributions to our project:

1. Design-Bench Library
2. Score-based SDE Repository
3. DDOM
4. Parallel Mentoring