Neural Dynamic Programming for Musical Self Similarity
Authors: Christian Walder, Dongwoo Kim
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate our model on real and synthetic data; in all cases it out-performs a strong stacked long short-term memory benchmark. Section 6 provides experiments |
| Researcher Affiliation | Collaboration | 1CSIRO Data61, Black Mountain, Australia 2The Australian National University 3Data to Decisions CRC, Kent Town, SA, Australia. |
| Pseudocode | Yes | Algorithm 1 Motif Net generalised distance. |
| Open Source Code | No | No explicit statement about releasing the source code or a link to a repository is provided. |
| Open Datasets | Yes | We used the same four sets of midi files as (Boulanger-Lewandowski et al., 2012), but rather than deriving simplified piano rolls, we derived simplified note onset sequences. The Bach chorale midis of (Boulanger Lewandowski et al., 2012) lack valid channel data, so we downloaded the analogous files from (Muse Data) for that dataset. The four datasets are JBM (J.S. Bach chorales from (Muse Data)), MUS (the Muse Data set of Boulanger Lewandowski et al. (2012)), NOT (Nottingham chord data of Shlien converted to midi by Boulanger-Lewandowski et al. (2012)) and PMD (piano midis provided by Krueger). |
| Dataset Splits | Yes | Methodology. We used a train/validation/test scheme based on log likelihoods. For each process and the five generation schemes below, we generate 300 sequences each for training, validation and testing. |
| Hardware Specification | No | GPUs yielded only modest speed ups, so we worked with a CPU cluster. This is too vague for a specific hardware specification. |
| Software Dependencies | No | Our implementation relied heavily on the dynamic graph of the Py Torch software; nonetheless we found the tree based implementation of Motif Net to be rather more involved than, say, the LSTM. We trained with stochastic gradient descent using Adam (Kingma & Ba, 2014). No specific version numbers are provided for the software mentioned. |
| Experiment Setup | Yes | Training and validation was performed for a range of hyper-parameters. We let the spaces E, C, D have the same dimension, which we varied as 2, 4, 8, . . . , 2048. Three algorithms are compared throughout: the LSTM (Hochreiter & Schmidhuber, 1997), our Motif Net, and the combination Motif Net+LSTM of subsection 4.3. We allowed the LSTM variants an advantage by letting the LSTM cell be stacked (Schmidhuber, 1992; El Hihi & Bengio, 1995) with number of layers ranging 1, 2, 3, 4, whereas for the Motif Net we fixed f A to be a GRU with one layer. For the toy problems of subsection 6.1 we used an exact Motif Net; for the music data of subsection 6.2 we used an approximate Motif Net with the crude setting of dmax = 4, and npriority ranging 2, 4, 8, . . . , 256. |