Controllable Generative Modeling via Causal Reasoning
Authors: Joey Bose, Ricardo Pio Monti, Aditya Grover
TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through a series of large-scale synthetic and human evaluations, we demonstrate that generating counterfactual samples which respect the underlying causal relationships inferred via CAGE leads to subjectively more realistic images. |
| Researcher Affiliation | Collaboration | Avishek Joey Bose EMAIL Mc Gill University and Mila Ricardo Pio Monti EMAIL Meta Aditya Grover EMAIL UCLA |
| Pseudocode | Yes | Algorithm 1 Obtaining CAGE scores τ |
| Open Source Code | No | The paper does not provide any explicit statement about releasing the source code for the methodology described, nor does it provide a link to a code repository. It mentions using existing libraries or models but not providing their own implementation code. |
| Open Datasets | Yes | We study the consistency of the inferred causal directions for two high-dimensional image datasets, Morpho MNIST (Pawlowski et al., 2020) and Celeb AHQ (Karras et al., 2017), with known or biologically guessed prior cause-effect relationships. For our experiments, whenever possible, we used the default settings found in the original papers of all chosen models. In particular, we used the default settings for both Mintnet (Song et al., 2019) and Style GAN2 (Karras et al., 2020) which were pretrained on MNIST and Flickr Faces HQ respectively. |
| Dataset Splits | No | The paper mentions using a 'test set samples' for Morpho MNIST and finetuning on '10% of Celeb AHQ' but does not specify the exact percentages, sample counts, or methodology for creating comprehensive training, validation, and test splits for the datasets used in their experiments. |
| Hardware Specification | No | Finally, the authors would also like to thank Meta AI for the computing resources that made this work possible. The paper mentions computing resources provided by Meta AI but does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running the experiments. |
| Software Dependencies | No | For latent linear classifiers we use an SVM based classifier as found in the widely used Sci-kit learn library (Pedregosa et al., 2011). The paper mentions the use of 'Sci-kit learn library' and several models (Mintnet, Style GAN2, Masked Autoregressive Flow) but does not provide specific version numbers for these or any other software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | For Mintnet we finetuned on a Morpho Mnist dataset for 250 epochs using the Adam optimizer with default settings. Similarly, we also finetuned Style GAN2 on Celeb AHQ for 2000 iterations using 10% of Celeb AHQ. Our synthetic experiments on the other required us to train a Masked Auto Regressive Flow (Papamakarios et al., 2017) that consisted of 10 layers with 4 blocks per layer. The Masked Auto Regressive Flow was trained for 5000 iterations using 5000 data samples. |