Intriguing Properties of Data Attribution on Diffusion Models

[Project Page] | [arXiv] | [Data Repository] 

---

TL, DR:

We report counter-intuitive observations that theoretically unjustified design choices 
for attributing diffusion models empirically outperform previous baselines 
by a large margin.

Proponents and opponents visualization on ArtBench-2 using TRAK and D-TRAK with various # of timesteps (10 or 100). For each sample of interest, 5 most positive influential training samples and 3 most negative influential training samples are given together with the influence scores (below each sample).

Counterfactual visualization on CIFAR-2	Counterfactual visualization on ArtBench-2

How to run

Quickstart

Check quickstart.ipynb to conduct data attribution on pre-trained diffusion models loaded from huggingface directly!

Replicating the paper's results

Setup

To get started, follow these steps:

1. Clone the GitHub Repository: Begin by cloning the repository using the command:

   git clone https://github.com/sail-sg/D-TRAK.git

2. Set Up Python Environment: Ensure you have a version 3.8. name:

   conda create -n dtrak python=3.8 -y
   conda activate dtrak

3. Install Dependencies: Install the necessary dependencies by running:

   pip install -r requirements.txt

Commands for LDS evaluation

We provide the commands to run experiments on CIFAR-2. It is easy to transfer to other datasets.

1. Data pre-processing:

   cd CIFAR2

   Run 00_EDA.ipynb to create dataset splits and subsets of the training set.

2. Train a diffusion model and generate images:

   bash scripts/run_train.sh 0 18888 5000-0.5
   bash scripts/run_gen.sh 0 0 5000-0.5

3. Construct the LDS benchmark:

   Train 64 models corresponding to 64 subsets of the training set

   bash scripts/run_lds_val_sub.sh 0 18888 5000-0.5 0 63

   Evaluate the model outputs on the validation set

   bash scripts/run_eval_lds_val_sub.sh 0 0 5000-0.5 idx_val.pkl 0 63
   bash scripts/run_eval_lds_val_sub.sh 0 1 5000-0.5 idx_val.pkl 0 63
   bash scripts/run_eval_lds_val_sub.sh 0 2 5000-0.5 idx_val.pkl 0 63

   Evaluate the model outputs on the generation set

   bash scripts/run_eval_lds_val_sub.sh 0 0 5000-0.5 idx_gen.pkl 0 63
   bash scripts/run_eval_lds_val_sub.sh 0 1 5000-0.5 idx_gen.pkl 0 63
   bash scripts/run_eval_lds_val_sub.sh 0 2 5000-0.5 idx_gen.pkl 0 63

4. Compute gradients:

   We shard the training set into 5 parts, each has 1000 examples.

   Use the following commands to compute the gradients to be used for TRAK.

   bash scripts/run_grad.sh 0 0 5000-0.5 idx-train.pkl 0 ddpm/checkpoint-8000 loss uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-train.pkl 1 ddpm/checkpoint-8000 loss uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-train.pkl 2 ddpm/checkpoint-8000 loss uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-train.pkl 3 ddpm/checkpoint-8000 loss uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-train.pkl 4 ddpm/checkpoint-8000 loss uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-val.pkl 0 ddpm/checkpoint-8000 loss uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-gen.pkl 0 ddpm/checkpoint-8000 loss uniform 10 32768

   Use the following commands to compute the gradients to be used for D-TRAK.

   bash scripts/run_grad.sh 0 0 5000-0.5 idx-train.pkl 0 ddpm/checkpoint-8000 mean-squared-l2-norm uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-train.pkl 1 ddpm/checkpoint-8000 mean-squared-l2-norm uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-train.pkl 2 ddpm/checkpoint-8000 mean-squared-l2-norm uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-train.pkl 3 ddpm/checkpoint-8000 mean-squared-l2-norm uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-train.pkl 4 ddpm/checkpoint-8000 mean-squared-l2-norm uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-val.pkl 0 ddpm/checkpoint-8000 mean-squared-l2-norm uniform 10 32768
   bash scripts/run_grad.sh 0 0 5000-0.5 idx-gen.pkl 0 ddpm/checkpoint-8000 mean-squared-l2-norm uniform 10 32768

5. Compute the TRAK/D-TRAK attributions and evaluate the LDS scores

   Run notebooks in methods/04_if.

   The implementations of other baselines can also be found in methods.

Commands for counterfactual evaluation

1. Data pre-processing

   Run this notebook first to get the indices of those training examples to be removed.

2. Retrain models after removing the top-influenctial training examples

   bash scripts/run_counter.sh 0 18888 5000-0.5 0 59

3. Generate images using the retrained models

   Run 02_counter.ipynb

4. Measure l2 distance

   Run 03_counter_eval_l2.ipynb

5. Measure CLIP cosine similarity

   Run 03_counter_eval_clip.ipynb

Bibtex

If you find this project useful in your research, please consider citing our paper:

@inproceedings{
zheng2023intriguing,
title={Intriguing Properties of Data Attribution on Diffusion Models},
  author={Zheng, Xiaosen and Pang, Tianyu and Du, Chao and Jiang, Jing and Lin, Min},
booktitle={International Conference on Learning Representations (ICLR)},
year={2024},
}