Flowification: Everything is a normalizing flow
Authors: Bálint Máté, Samuel Klein, Tobias Golling, François Fleuret
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In 4 we flowify multi-layer perceptrons and convolutional networks and train them as normalizing flows using the likelihood. This demonstrates that models built from standard layers can be used for density estimation directly. To test the constructions described in the previous section we flowify multilayer perceptrons and convolutional architectures and train them to maximize the likelihood of different datasets. |
| Researcher Affiliation | Academia | Bálint Máté University of Geneva balint.mate@unige.ch Samuel Klein University of Geneva samuel.klein@unige.ch Tobias Golling University of Geneva tobias.golling@unige.ch François Fleuret University of Geneva francois.fleuret@unige.ch |
| Pseudocode | No | The paper does not contain any explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code for reproducing our experiments is available under MIT license at https://github.com/balintmate/flowification. |
| Open Datasets | Yes | Tabular data In this section we study a selection of UCI datasets [24] and the BSDS300 collection of natural images [25] using the preprocessed dataset used by masked autoregressive flows [17, 26]. Image Data In this section we use the MNIST [27] and CIFAR10 [28] datasets with the standard training and test splits. |
| Dataset Splits | No | The paper mentions 'standard training and test splits' for MNIST and CIFAR10, and using a 'preprocessed dataset' for UCI/BSDS300, but does not explicitly specify validation dataset splits or percentages. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions the use of 'rational quadratic splines [7]' and refers to 'PyTorch [37]' and 'nflows [39]' in its bibliography, but does not explicitly state specific version numbers for these or other software dependencies within the main text. |
| Experiment Setup | Yes | The data is uniformly dequantized as required to train on image data [29, 30]... The flowified layers sample from N(0, a) for dimension increasing operations, where a is a per-layer trainable parameter. We use rational quadratic splines [7] with 8 knots and a tail bound of 2 as activation functions, where the same function is applied per output node. |