Retrieval-based-Voice-Conversion-WebUI icon indicating copy to clipboard operation
Retrieval-based-Voice-Conversion-WebUI copied to clipboard

训练数据不收敛,可以调整哪些参数

Open glide-the opened this issue 2 years ago • 0 comments

我在2个3090上使用一个人的纯净语言,预估有7G的数据量,15小时的音频时间 我通过此仓库的webui加载了训练任务,参数如下,在尝试训练335后,发现loss并没有下降的情况,测试了下模型效果,部分语气下相似 对于优化训练的方法,哪里可以调整参数,已达到最优效果 训练参数 batch_size 20, enable ckpt save_every_epoch weights folder of 5 epoch pretrained_v2

image

训练日志

INFO:lulu-epoch:Train Epoch: 345 [83%]
INFO:lulu-epoch:[30000, 9.577890768671308e-05]
INFO:lulu-epoch:loss_disc=3.664, loss_gen=3.372, loss_fm=10.585,loss_mel=16.342, loss_kl=1.011
INFO:lulu-epoch:Saving model and optimizer state at epoch 345 to ./logs/lulu-epoch/G_2333333.pth
INFO:lulu-epoch:Saving model and optimizer state at epoch 345 to ./logs/lulu-epoch/D_2333333.pth
INFO:lulu-epoch:saving ckpt lulu-epoch_e345:Success.
INFO:lulu-epoch:====> Epoch: 345 [2023-06-24 01:39:45] | (0:01:17.178153)
INFO:lulu-epoch:====> Epoch: 346 [2023-06-24 01:40:59] | (0:01:14.771824)

glide-the avatar Jun 23 '23 17:06 glide-the