Lifting Architectural Constraints of Injective Flows
Authors: Peter Sorrenson, Felix Draxler, Armand Rousselot, Sander Hummerich, Lea Zimmermann, Ullrich Koethe
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments on toy, tabular and image data, demonstrating the competitive performance of the resulting model. |
| Researcher Affiliation | Academia | Computer Vision and Learning Lab Heidelberg University firstname.lastname@iwr.uni-heidelberg.de |
| Pseudocode | No | The paper describes methods and estimators but does not contain any formally labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | We provide code to implement our model and reproduce our results at https://github.com/vislearn/ FFF. |
| Open Datasets | Yes | We evaluate our method on four of the tabular datasets used by Papamakarios et al. (2017)... We compare FIF against previous injective flows on Celeb A images (Liu et al., 2015)... Recently, Chadebec et al. (2022) proposed the Pythae benchmark for comparing generative autoencoders on image generation. They evaluate different training methods using two different architectures on MNIST (Le Cun et al., 2010), CIFAR10 (Krizhevsky, 2009), and Celeb A (Liu et al., 2015). |
| Dataset Splits | Yes | For MNIST and CIFAR10 this means training for 100 epochs with the Adam optimizer at a starting LR of 10 4, reserving the last 10k images of the training sets as validation sets. |
| Hardware Specification | Yes | a single RTX 2070 card, a single NVIDIA A40, an internal cluster of A100s. |
| Software Dependencies | Yes | Py Torch (Paszke et al., 2019), Py Torch Lightning (Falcon & The Py Torch Lightning team, 2019), Tensorflow (Abadi et al., 2015) for FID score evaluation, Numpy (Harris et al., 2020), Matplotlib (Hunter, 2007) for plotting and Pandas (Mc Kinney, 2010; The pandas development team, 2020) |
| Experiment Setup | Yes | We use a batch size of 512, add isotropic Gaussian noise with standard deviation 0.01, use K = 1 Hutchinson samples and a reconstruction weight β = 10 for all experiments. We use the Adam optimizer with the onecycle LR scheduler with LR of 10 4 (except for HEPMASS which has LR of 3 10 4) and weight decay of 10 4. |