Semi-Autoregressive Energy Flows: Exploring Likelihood-Free Training of Normalizing Flows

Authors: Phillip Si, Zeyi Chen, Subham Sekhar Sahoo, Yair Schiff, Volodymyr Kuleshov

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experiments We evaluate our framework on a range of UCI datasets (Dua & Graff, 2017), datasets of handwritten digits (Pedregosa et al., 2011; Deng, 2012), and real world images, such as CIFAR10 (Krizhevsky et al., 2009) and Celeb-A (Liu et al., 2015).
Researcher Affiliation Academia 1Computer and Information Science, Cornell University, Ithaca, NY, USA 2Machine Learning Department, Carnegie-Mellon University, Pittsburgh, PA, USA 3Tsinghua University, Beijing, China 4Department of Computer Science, Cornell Tech, NYC, NY, USA.
Pseudocode Yes F. Pseudocode for Semi-Autoregressive Flows In this section, we provide the algorithms for training (Algorithm 1) and sampling from (Algorithm 2) SAEFs.
Open Source Code Yes 1 Code is available at https://github.com/ps789/SAEF.
Open Datasets Yes We evaluate our framework on a range of UCI datasets (Dua & Graff, 2017), datasets of handwritten digits (Pedregosa et al., 2011; Deng, 2012), and real world images, such as CIFAR10 (Krizhevsky et al., 2009) and Celeb-A (Liu et al., 2015).
Dataset Splits Yes We split the original 10,000 test samples into 8,000 training data and 2,000 evaluation data for the logistic model.
Hardware Specification No Glow was trained with 3 levels and a depth of 1 on 8 GPUs for 250 epochs. No specific GPU model, CPU, or other hardware details are provided beyond the number of GPUs.
Software Dependencies No The paper mentions software like Keras and scikit-learn in the references and generally refers to optimizers like ADAM, but it does not specify version numbers for any software dependencies crucial for reproducibility.
Experiment Setup Yes All models were trained with a batch size of 200 and learning rate of 1e 3 using the ADAM optimizer (Kingma & Ba, 2014) for 200 epochs for the smaller datasets (Miniboone, Hepmass) and 20 epochs for the larger ones (Gas, Power, BSDS 300).