Practical Conditional Neural Process Via Tractable Dependent Predictions

Authors: Stratis Markou, James Requeima, Wessel Bruinsma, Anna Vaughan, Richard E Turner

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

Reproducibility Variable Result LLM Response
Research Type Experimental We apply the proposed models to synthetic and real data. Our experiments with synthetic data comprise four Gaussian tasks and a non-Gaussian task. In our experiments with real data, we evaluate our models on electroencephalogram data as well as climate data.
Researcher Affiliation Academia Stratis Markou University of Cambridge em626@cam.ac.uk James Requeima University of Cambridge Invenia Labs jrr41@cam.ac.uk Wessel P. Bruinsma University of Cambridge Invenia Labs wpb23@cam.ac.uk Anna Vaughan University of Cambridge av555@cam.ac.uk Richard E. Turner University of Cambridge ret26@cam.ac.uk
Pseudocode No No pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes We publish the complete repository containing all our code, including the models, training scripts, pretrained models and Jupyter notebooks which will produce all plots in this paper (link3). https://github.com/censored/for-anonymity
Open Datasets Yes The EEG data is available via the UCI dataset website (link1). ... The climate modelling data is also available to the public, through the Copernicus European Climate Data Store (link2).
Dataset Splits Yes We sample 86 of the subjects for training, 10 for validation and 10 for testing. ... We train each of our meta-models for 1000 epochs, each consisting of 256 iterations, at each of which the model is presented with a batch 16 different tasks.
Hardware Specification Yes Table 3 shows the runtime and memory footprint of the GNP, AGNP, Conv GNP and Full Conv GNP models applied to data with one-dimensional inputs, D = 1. In particular, we measure the runtime and memory required to perform a single forward pass through the neural architecture of each model that was used for the one-dimensional Gaussian tasks, on an NVIDIA Ge Force RTX 2080 Ti GPU.
Software Dependencies No The paper mentions software like Adam optimizer, scipy's L-BFGS-B, and a specific Git Hub implementation (link5) but does not provide specific version numbers for these or other key software components used in the experiments.
Experiment Setup Yes We optimise all models with Adam (Kingma & Ba, 2014), using a learning rate of 5 10 4. ... For the Conv NP we use 10 latent samples to evaluate the loss during training, and 512 samples during testing. ... Each model is trained for 500 epochs.