Skip to main content

Table 1 Performance of CRNNs. All results are given in macro average F1. Tamp amplitude threshold, lr learning rate. All results are measured in macro average F1

From: Towards cross-modal pre-training and learning tempo-spatial characteristics for audio recognition with convolutional and recurrent neural networks

CRNN
lr Batch size Tamp [dB] Devel Test
0.001 128 − 30 66.5
0.001 128 − 45 57.8
0.001 64 − 30 55.4
0.01 128 − 30 66.2
0.01 128 − 60 69.2
0.01 64 − 30 70.7 72.3
0.01 64 − 45 73.7 79.3
0.01 64 − 60 78.8 74.2
Fusion of best 3 CRNNs 81.4 82.2
DCASE 2018, task 5 baseline [8]
0.0001 256 84.5 83.1