Deep Gaussian Markov Random Fields

Authors: Per Sidén, Fredrik Lindsten

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the behaviour of our model for an inpainting problem on the two toy datasets in Figure 2... We compare our method against some popular methods for large data sets in spatial statistics, by considering the satellite data of daytime land surface temperatures, used in the competition by Heaton et al. (2018). Table 1 compares different instances of our model with the methods in the competition...
Researcher Affiliation Academia 1Division of Statistics and Machine Learning, Department of Computer and Information Science, Link oping University, Link oping, Sweden.
Pseudocode Yes Algorithm 1 Inference algorithm
Open Source Code Yes Code for our methods and experiments are available at https://bitbucket.org/psiden/deepgmrf.
Open Datasets Yes We compare our method against some popular methods for large data sets in spatial statistics, by considering the satellite data of daytime land surface temperatures, used in the competition by Heaton et al. (2018). The data and code for some of the methods can be found at https://github. com/finnlindgren/heatoncomparison.
Dataset Splits No The data are on a 500 x 300 grid, with 105,569 non-missing observations as training set. The test set consists of 42,740 observations and have been selected as the pixels that were missing due to cloud cover on a different date. No explicit validation set or split information is provided.
Hardware Specification Yes our method takes roughly 2.5h for the seq5x5,L=5 model using a Tesla K40 GPU
Software Dependencies No We have implemented DGMRF in Tensor Flow (Abadi et al., 2016), taking advantage of autodiff and GPU computations. No specific version number for TensorFlow or other software dependencies is provided.
Experiment Setup No The paper mentions using Adam for optimization, the reparameterization trick, and details regarding convolution padding ('same'). However, specific hyperparameter values such as learning rate, batch size, or number of epochs are not provided in the main text.