- 2-layer RNN with gated recurrent units (GRU)
- 4 layer CNNs
- Assumption that temporal pattern can be aggregated better with RNNs than CNNs
- RNNs are used to aggregate temporal patterns
- Sub-sampling results in a feature map of N x 1 x 15
- Fed into 2-layer RNN

University of Michigan - Ann Arbor

Four different models were compared: One CRNN and 3 CNNs.  All convolutional and fully connected layers have identical optimization techniques and activation functions - batch normalization and ELU activation function. The CRNN architecture can be compared to the CNN architectures (k1c2, k2c1, k2c2) where 'k' specifies the kernel shape and 'c' specifies the convolutional dimension.

Models of CNN/CRNN in Music Categorization

Choi, K., Fazekas, G., Sandler, M., & Cho, K. (2017, March). Convolutional recurrent neural networks for music classification. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 2392-2396). IEEE.

CONVOLUTIONAL RECURRENT NEURAL NETWORKS FOR MUSIC CLASSIFICATION

CRNN Model

- For genre classification
- 4 convolutional layers, 2 fully-connected layers
- Alternate between 1D convolutional layers and max-pooling layers

k1c2 - CNN Model

- For music tagging and genre classification
- 5 convolutional layers, 2 fully-connected layers
- First layer learns 2D kernels that are applied to frequency band
- Then, alternate between 1D convolutional layers and max-pooling layers
- First layer reduces computation complexity

k2c1 - CNN Model

- For music tagging and genre classification
- 5 convolutional layers of 3x3 kernels and max-pooling layers
- Reduces size of feature maps to 1x1 at final layer, where each feature covers the whole input rather than each frequency band as in k1c1 and k2c1
- Allows time and frequency invariances in different scale by gradual 2D sub-samplings
- Fully-convolutional (fewer parameters)

Learn Before

Related