北邮人智-音频感知技术团队用户500
012-Effect of nperspeaker/batchsize for BCE
输入“/”快速插入内容
012-Effect of nperspeaker/batchsize for BCE
012-Effect of nperspeaker/batchsize for BCE
用户500用户500
用户2707用户27072023年8月17日修改
控制 batch每个人的样本数 batchBCE vs globalBCE
配置
batch_BCE_lam0.5_sampler
global_BCE_lam0.9_sampler
batch_size = 128
loss.sum
Clean test
batch_size = 128 n_per_speaker | Clean test | | | ||
| Top1 acc | Top5 acc | Top10 acc | ||
Batch_BCE_1 | 88.024% | 96.218% | 97.697% | ||
Batch_BCE_2 | 85.394% | 94.594% | 96.533% | ||
Batch_BCE_4 | 68.679% | 85.915% | 90.909% | ||
global_BCE_1 | 86.400% | 96.012% | 97.842% | ||
global_BCE_2 | 74.364% | 91.382% | 94.970% | ||
global_BCE_4 | 40.303% | 62.570% | 73.006% | ||
Why increasing n_per_speaker, the performance reduced significantly?
1.
If bug? Check it (dataloader or loss modules)
2.
The convergence of loss?
The code of Batch_BCE_4 has some problems. Be solved and changed in loss.mean test.
Noisy test
batch_size = 128 n_per_speaker | Noisy test | | | ||
| Top1 acc | Top5 acc | Top10 acc | ||
Batch_BCE_1 | 34.978% | 48.346% | 53.727% | ||
Batch_BCE_2 | 32.723% | 46.152% | 51.400% | ||
Batch_BCE_4 | 22.494% | 37.402% | 44.164% | ||
| | | | ||
global_BCE_1 | 37.656% | 52.963% | 58.781% | ||
global_BCE_2 | 29.051% | 46.031% | 52.818% | ||
global_BCE_4 | 11.683% | 24.967% | 32.263% | ||
For the Clean test, it can be observed that Batch_BCE outperforms Global_BCE. While for the Noisy test, Global_BCE wins Batch_BCE.
WHY?
loss.mean
Clean test
batch_size = 128 n_per_speaker | Clean test | | | ||
| Top1 acc | Top5 acc | Top10 acc | ||
Batch_BCE_1 | 88.024% | 96.218% | 97.697% | ||
Batch_BCE_2 | 88.327% | 96.206% | 97.758% | ||
Batch_BCE_4 | 82.242% | 93.564% | 95.770% | ||
| | | | ||
global_BCE_1 | 88.679% | 96.412% | 97.673% | ||
global_BCE_2 | 65.636% | 86.655% | 92.206% | ||
global_BCE_2_loss*2 | 81.309% | 94.061% | 96.594% | ||