Conditional Generative Models are Sufficient to Sample from Any Causal Effect Estimand
Authors: Md Musfiqur Rahman, Matt Jordan, Murat Kocaoglu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments on a Colored MNIST dataset having both the treatment (X) and the target variables (Y ) as images and sample from P(y|do(x)). Our algorithm also enables us to conduct a causal analysis to evaluate spurious correlations among input features of generative models pre-trained on the Celeb A dataset. Finally, we generate high-dimensional interventional samples from the MIMIC-CXR dataset involving text and image variables. |
| Researcher Affiliation | Academia | Md Musfiqur Rahman Purdue University Matt Jordan University of Texas at Austin Murat Kocaoglu Purdue University |
| Pseudocode | Yes | Algorithm 1 ID-GEN (Y, X, G, D, ˆX, ˆG) |
| Open Source Code | Yes | Codes are available at github.com/musfiqshohan/idgen. |
| Open Datasets | Yes | We conduct experiments on a Colored MNIST dataset having both the treatment (X) and the target variables (Y ) as images and sample from P(y|do(x)). Our algorithm also enables us to conduct a causal analysis to evaluate spurious correlations among input features of generative models pre-trained on the Celeb A dataset. Finally, we generate high-dimensional interventional samples from the MIMIC-CXR dataset involving text and image variables." and "[29] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), December 2015." and "[24] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. |
| Dataset Splits | Yes | Finally, a random split of the 30K images is performed: keeping 20K to be used during training, and 10K to be used as validation images. |
| Hardware Specification | Yes | We performed some of our experiments on a machine with an RTX-3090 GPU. We also performed some training on 2 A100 GPU s which took roughly 9 hours for 1000 epochs. |
| Software Dependencies | Yes | For reproducibility purposes, we provide our anonimized source codes with instructions. |
| Experiment Setup | Yes | Batch sizes of 256 are used everywhere. Training is performed for 1000 epochs, which takes roughly 9 hours on 2 A100 GPU s. Sampling is performed using DDIM over 100 timesteps, with a conditioning weight of w = 1 (true conditional sampling) and noise σ = 0.3. |