Sequential Neural Models with Stochastic Layers

Authors: Marco Fraccaro, Søren Kaae Sønderby, Ulrich Paquet, Ole Winther

NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 4 we test the performances of SRNN on speech and polyphonic music modeling tasks.
Researcher Affiliation Collaboration Technical University of Denmark University of Copenhagen * Google Deep Mind
Pseudocode Yes Algorithm 1 Inference of SRNN with Resq parameterization from (12).
Open Source Code Yes The code for SRNN is available at github.com/marcofraccaro/srnn.
Open Datasets Yes We test SRNN on the Blizzard [18] and TIMIT raw audio data sets (Table 1) used in [7]. Additionally, we test SRNN for modeling sequences of polyphonic music (Table 2), using the four data sets of MIDI songs introduced in [4].
Dataset Splits No The paper mentions evaluating on test sets and states that 'The preprocessing of the data sets and the testing performance measures are identical to those reported in [7],' implying standard splits, but does not explicitly provide specific percentages or counts for train/validation/test splits.
Hardware Specification Yes Training using a NVIDIA Titan X GPU took around 1.5 hours for TIMIT, 18 hours for Blizzard... and NVIDIA Corporation for the donation of TITAN X and Tesla K40 GPUs.
Software Dependencies No All models where implemented using Theano [2], Lasagne [9] and Parmesan1. No version numbers for the software dependencies are provided.
Experiment Setup No The paper mentions that 'Further implementation and experimental details can be found in the Supplementary Material' but does not include specific hyperparameter values or detailed training configurations in the main text.