How to train on librispeech

[zhaojingxin123]

Hello, using LibriTTS960 training about how many epochs, the model can be used. Can you share some config or training information?

[Zhikang Niu]

Hello, using LibriTTS960 training about how many epochs, the model can be used. Can you share some config or training information?

maybe 20 epoch? you can check the hf model.

[zhaojingxin123]

Thank you for your answer. 感谢，我训练的非常慢，batchsize=2，log_interval: 100，tensor_cut: 240000，每100batch输出一次log，num_workers: 10，用时1min30秒左右，
然后batchsize=6的时候，用时是4min30s左右，好奇batch改大为什么时间没有变短，怎么能让训练变快呀？再次感谢
Why my training is so slow，

My common: save_interval: 2 test_interval: 5 log_interval: 100 max_epoch: 30 seed: 3401 amp: False datasets: # 训练测试数据位置 train_csv_path: '/home/zjx/home/zjx/encodec-pytorch-main/datasets/librispeech_train100h_train.csv' test_csv_path: '/home/zjx/home/zjx/encodec-pytorch-main/datasets/librispeech_train100h_test.csv' batch_size: 2 # 320000 tensor_cut: 240000 num_workers: 10 fixed_length: 0 pin_memory: True # 是否断点继续训练 checkpoint: resume: True checkpoint_path: '/home/zjx/home/zjx/encodec-pytorch-main/checkpoints/bs2_cut240000_length0_epoch14_lr0.0003.pt' disc_checkpoint_path: '/home/zjx/home/zjx/encodec-pytorch-main/checkpoints/bs2_cut240000_length0_epoch14_disc_lr0.0003.pt' # 保存位置 save_folder: '/home/zjx/home/zjx/encodec-pytorch-main/checkpoints' save_location: '${checkpoint.save_folder}/bs${datasets.batch_size}_cut${datasets.tensor_cut}_length${datasets.fixed_length}_'

optimization:
lr: 3e-4
disc_lr: 3e-4

lr_scheduler:
warmup_epoch: 5

model:
target_bandwidths: [1.5, 3., 6., 12., 24.]
sample_rate: 24_000
channels: 1
train_discriminator: True
audio_normalize: True
filters: 32
ratios: [8, 5, 4, 2]
disc_win_lengths: [1024, 2048, 512]
disc_hop_lengths: [256, 512, 128]
disc_n_ffts: [1024, 2048, 512]
distributed:
data_parallel: False
world_size: 4
find_unused_parameters: False
torch_distributed_debug: False
init_method: tmp

[zhaojingxin123]

[Zhikang Niu]

Thank you for your answer. 感谢，我训练的非常慢，batchsize=2，log_interval: 100，tensor_cut: 240000，每100batch输出一次log，num_workers: 10，用时1min30秒左右， 然后batchsize=6的时候，用时是4min30s左右，好奇batch改大为什么时间没有变短，怎么能让训练变快呀？再次感谢 Why my training is so slow，

My common: save_interval: 2 test_interval: 5 log_interval: 100 max_epoch: 30 seed: 3401 amp: False datasets: # 训练测试数据位置 train_csv_path: '/home/zjx/home/zjx/encodec-pytorch-main/datasets/librispeech_train100h_train.csv' test_csv_path: '/home/zjx/home/zjx/encodec-pytorch-main/datasets/librispeech_train100h_test.csv' batch_size: 2 # 320000 tensor_cut: 240000 num_workers: 10 fixed_length: 0 pin_memory: True # 是否断点继续训练 checkpoint: resume: True checkpoint_path: '/home/zjx/home/zjx/encodec-pytorch-main/checkpoints/bs2_cut240000_length0_epoch14_lr0.0003.pt' disc_checkpoint_path: '/home/zjx/home/zjx/encodec-pytorch-main/checkpoints/bs2_cut240000_length0_epoch14_disc_lr0.0003.pt' # 保存位置 save_folder: '/home/zjx/home/zjx/encodec-pytorch-main/checkpoints' save_location: '
𝑐
ℎ
𝑒
𝑐
𝑘
𝑝
𝑜
𝑖
𝑛
𝑡
.
𝑠
𝑎
𝑣
𝑒
𝑓
𝑜
𝑙
𝑑
𝑒
𝑟
/
𝑏
𝑠
{datasets.batch_size}cut
𝑑
𝑎
𝑡
𝑎
𝑠
𝑒
𝑡
𝑠
.
𝑡
𝑒
𝑛
𝑠
𝑜
𝑟
𝑐
𝑢
𝑡
𝑙
𝑒
𝑛
𝑔
𝑡
ℎ
{datasets.fixed_length}'
optimization: lr: 3e-4 disc_lr: 3e-4

lr_scheduler: warmup_epoch: 5

model: target_bandwidths: [1.5, 3., 6., 12., 24.] sample_rate: 24_000 channels: 1 train_discriminator: True audio_normalize: True filters: 32 ratios: [8, 5, 4, 2] disc_win_lengths: [1024, 2048, 512] disc_hop_lengths: [256, 512, 128] disc_n_ffts: [1024, 2048, 512] distributed: data_parallel: False world_size: 4 find_unused_parameters: False torch_distributed_debug: False init_method: tmp

你可以减少tensor_cut到1s ≈ 24000采样点。然后增大batch size训练，另外 numworkers不要设置那么大吧

[zhaojingxin123]

感谢大佬指点，这就是试一下

[zhaojingxin123]

 感谢，感谢！！！！！

[Zhikang Niu]

感谢，感谢！！！！！

但是这样对于音频质量是否有影响我暂时未探索过