In CNN structures, the size of the feature maps are most critical because the number of convolutional operations is proportional to the sizes. K2c1 is fastest, reducing the size of feature map in the first convolutional layer. K2c2 is faster than k1c2 because k1c2 only reduces the size in the temporal axis, while k2c2 reduces the sizes of the feature maps in both axes.


University of California, Santa Cruz

When compared to other architectures in terms of computation time, k2c2 and CRNN have similar results and are better than k1c1 and k2c1. Training speed, however, is ranked as such: k2c1 > k2c2 > k1c2 > CRNN. There are two factors that likely lead to this (child nodes).

Computation Time in CNN/CRNN Model Comparison

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

Effects of Feature Map Size in CNN/CRNN Model Comparison

The difference between CNN and CRNN structures is due to the negative correlation speed to depth of the networks. CRNN naturally has more depth due to the recursive structure, and thus induces heavier computation and more training time. 


Learn Before

Related