Table 4 The experimental results of batch size optimization.
From: Channel equalization in ultraviolet communication based on LSTM-DNN hybrid model
Batch size | Training time (seconds/epoch) | The final accuracy of the test set | Training stability |
---|---|---|---|
32 | 6.2 | 92.9% | Stable training, but high computational cost |
64 | 4.1 | 93.1% | Stable training and moderate computational efficiency |
128 | 2.7 | 91.5% | Unstable training, significant loss oscillation |