Dance Dance Convolution

Authors: Chris Donahue, Zachary C. Lipton, Julian McAuley

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental For both the Fraxtil and ITG datasets we apply 80%, 10%, 10% splits for training, validation, and test data, respectively. In Table 2, we list the results of our experiments for step placement. In Table 3 we present results for the step selection task.
Researcher Affiliation Academia 1 UCSD Department of Music, San Diego, CA 2 UCSD Department of Computer Science, San Diego, CA.
Pseudocode No The paper describes the architecture of the models and training procedures, but does not provide pseudocode or algorithm blocks.
Open Source Code Yes 1 https://github.com/chrisdonahue/ddc
Open Datasets Yes We contribute two large public datasets, which we consider to be of notably high quality and consistency.1 Each dataset is a collection of recordings and step charts. One contains charts by a single author and the other by multiple authors. 1https://github.com/chrisdonahue/ddc
Dataset Splits Yes For both the Fraxtil and ITG datasets we apply 80%, 10%, 10% splits for training, validation, and test data, respectively.
Hardware Specification Yes All models satisfy this criteria within 12 hours of training on a single machine with an NVIDIA Tesla K40m GPU.
Software Dependencies No The paper mentions using the 'ESSENTIA library (Bogdanov et al., 2013)', but does not provide specific version numbers for software dependencies.
Experiment Setup Yes We train with mini-batches of size 256, scaling down gradients when their l2 norm exceeds 5. We apply 50% dropout following each LSTM and fully connected layer. For recurrent neural networks, we calculate updates using backpropagation through time with 100 steps of unrolling. For all neural network models, we learn parameters by minimizing cross-entropy. We train with mini-batches of size 64, and scale gradients using the same scheme as for step placement. We use 50% dropout during training for both the MLP and RNN models.