Adaptive Accompaniment with ReaLchords

Authors: Yusong Wu, Tim Cooijmans, Kyle Kastner, Adam Roberts, Ian Simon, Alexander Scarlatos, Chris Donahue, Cassie Tarakajian, Shayegan Omidshafiei, Aaron Courville, Pablo Samuel Castro, Natasha Jaques, Cheng-Zhi Anna Huang

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through quantitative experiments and listening tests, we demonstrate that the resulting model adapts well to unfamiliar input and produce fitting accompaniment.
Researcher Affiliation Collaboration 1Google Deep Mind 2Mila Quebec AI Institute, Universit e de Montr eal 3Google 4University of Massachusetts Amherst 5Carnegie Mellon University 6Work done while at Google 7Field AI 8Canada CIFAR AI Chair 9University of Washington.
Pseudocode No The paper does not contain any explicit pseudocode or algorithm blocks.
Open Source Code No The paper provides a link to audio examples ('Listen to audio examples here: https://storage.googleapis.com/realchords/index.html') but does not provide a link to the open-source code for the methodology described in the paper.
Open Datasets Yes We train our models on an updated version of the Hooktheory dataset (Donahue et al., 2022), which comprises crowdsourced analyses of monophonic melodies and chords from recordings and now contains 38K melody-chord pairs.
Dataset Splits Yes 20% of the data is held out and divided equally into validation and test sets. We develop on the validation set and report the test set results in the paper.
Hardware Specification No The paper does not explicitly describe the specific hardware used to run its experiments (e.g., specific GPU or CPU models).
Software Dependencies No The paper mentions software components like 'Adafactor optimizer', 'T5X framework', and 'Adam optimizer' but does not specify their version numbers or other crucial software dependencies required for reproducibility.
Experiment Setup Yes The online model is trained using Adafactor optimizer (Shazeer & Stern, 2018) and learning rate of 10 3 with a batch size of 256. The online model is trained for 50, 000 steps with 1000 steps of warmup. We apply a dropout with rate 0.1 to the online model during training. The coefficient β between reward maximization and KL loss in Equation 2 is fixed as 0.5 for all the experiments. We apply a coefficient of 50 to the reward produced by reward models. We apply a coefficient of 20 to the ending early penalty in all experiments used this penalty.