Differentiable DAG Sampling

Authors: Bertrand Charpentier, Simon Kibler, Stephan Günnemann

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our extensive experiments, we compare VI-DP-DAG to other differentiable DAG learning baselines on synthetic and real datasets.
Researcher Affiliation Academia Bertrand Charpentier, Simon Kibler, Stephan Günnemann Department of Informatics & Munich Data Science Institute Technical University Munich {charpent, kibler, guennemann}@in.tum.de
Pseudocode Yes Figure 2: Differentiable DAG sampling in Python
Open Source Code Yes We provide all datasets and the model code at the project page 1. https://www.daml.in.tum.de/differentiable-dag-sampling
Open Datasets Yes We use the Sachs dataset which measures the expression level of different proteins and phospholipids in human cells (Sachs et al., 2005). We also use the pseudo-real Syn TRe N dataset sampled from a generator that was designed to create synthetic transcriptional regulatory networks and produces simulated gene expression data that approximates experimental data (Van den Bulcke et al., 2006). ... We provide all datasets and the model code at the project page 1.
Dataset Splits Yes We split all datasets in training/validation/test sets with 80%/10%/10%.
Hardware Specification Yes We evaluate the training time of all models on a single GPU (NVIDIA GTX 1080 Ti, 11 GB memory).
Software Dependencies No The paper mentions using optimizers like Adam (Kingma & Ba, 2015) and Rms Prop (Tieleman & Hinton, 2012), but it does not specify software versions for any programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes In our experiments, VI-DP-DAG parametrizes the permutation probability Pψ(Π) with Gumbel Sinkhorn or Gumbel-Top-k trick, the edge probability Pφ(U) with Gumbel-Softmax distribution and the causal mechanisms fi,θ with a 3 layers Multi-Layer Perceptron (MLP). We use early stopping and perform a grid search over the permutation probability parametrization (i.e. Gumbel-Sinkhorn or Gumbel-Top-k), the fixed prior probability Pprior(Uij) [1e 2, 1e 1] and the regularization factor λ [0, 1e 1]. Finally, all temperature parameters are fixed to τ = 1 in all experiments.