Upload 2 files
Browse files- common.yaml +22 -0
- config.yaml +93 -0
common.yaml
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
path: ./logs/${hydra.job.config_name}/${now:%Y-%m-%d}/${now:%H-%M-%S}
|
| 2 |
+
|
| 3 |
+
log_level: INFO
|
| 4 |
+
seed: 1
|
| 5 |
+
tb_log_dir: tensorboard
|
| 6 |
+
tqdm: true
|
| 7 |
+
|
| 8 |
+
hydra:
|
| 9 |
+
run:
|
| 10 |
+
dir: ${path}
|
| 11 |
+
job_logging:
|
| 12 |
+
formatters:
|
| 13 |
+
colorlog:
|
| 14 |
+
format: '[%(cyan)s%(asctime)s%(reset)s][%(blue)s%(name)s:%(lineno)s:%(funcName)s()%(reset)s][%(log_color)s%(levelname)s%(reset)s]
|
| 15 |
+
- %(message)s'
|
| 16 |
+
handlers:
|
| 17 |
+
file:
|
| 18 |
+
filename: ${hydra.run.dir}/${hydra.job.name}_${now:%Y-%m-%d}_${now:%H-%M-%S}.log
|
| 19 |
+
|
| 20 |
+
defaults:
|
| 21 |
+
- override hydra/job_logging: colorlog
|
| 22 |
+
- override hydra/hydra_logging: colorlog
|
config.yaml
ADDED
|
@@ -0,0 +1,93 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
defaults:
|
| 2 |
+
- common
|
| 3 |
+
|
| 4 |
+
train:
|
| 5 |
+
batch_size: 128
|
| 6 |
+
betas: [0.8, 0.99]
|
| 7 |
+
c_kl: 1.0
|
| 8 |
+
c_mel: 45
|
| 9 |
+
distributed: false # BUG: multi-gpu is not working
|
| 10 |
+
use_multiprocessing: false # BUG: multi-gpu is not working
|
| 11 |
+
epochs: 20
|
| 12 |
+
eps: 1e-9
|
| 13 |
+
fp16_run: false
|
| 14 |
+
init_lr_ratio: 1
|
| 15 |
+
raise_error: false
|
| 16 |
+
learning_rate: 2e-4
|
| 17 |
+
log_interval: 10
|
| 18 |
+
log_level: ${log_level}
|
| 19 |
+
lr_decay: 0.98
|
| 20 |
+
max_speclen: 128
|
| 21 |
+
port: 8005
|
| 22 |
+
resume_training: false # set to false to finetune from a model
|
| 23 |
+
seed: 1234
|
| 24 |
+
segment_size: 8960
|
| 25 |
+
use_sr: false
|
| 26 |
+
valid_epoch_interval: 1
|
| 27 |
+
valid_steps_interval: 1000
|
| 28 |
+
save_epoch_interval: 10
|
| 29 |
+
save_steps_interval: 1000
|
| 30 |
+
warmup_epochs: 0
|
| 31 |
+
# weighted_batch_speaker_sampling : false
|
| 32 |
+
# weighted_batch_lang_sampling : false
|
| 33 |
+
weighted_batch_speaker_sampling : 0.5
|
| 34 |
+
weighted_batch_lang_sampling : 0.5
|
| 35 |
+
|
| 36 |
+
data:
|
| 37 |
+
dataset_dir: /raid/lucasgris/free-svc/data
|
| 38 |
+
filter_length: 1280
|
| 39 |
+
hop_length: 320
|
| 40 |
+
max_wav_value: 32768.0
|
| 41 |
+
mel_fmax: null
|
| 42 |
+
mel_fmin: 0.0
|
| 43 |
+
n_mel_channels: 80
|
| 44 |
+
num_workers: 64
|
| 45 |
+
# For pitch extraction, set the pitch_predictor (will compute in dataloader) or pitch_features_dir (will load from disk)
|
| 46 |
+
pitch_predictor: rmvpe # pm | crepe | harvest | dio | rmvpe | fcpe
|
| 47 |
+
pitch_features_dir: ${data.dataset_dir}/pitch_features/
|
| 48 |
+
sampling_rate: 24000
|
| 49 |
+
spectrogram_dir: null #${data.dataset_dir}/spectrograms # it is recommended NOT to use if you have small disk space
|
| 50 |
+
# For speaker embedding extraction, set the use_spk_emb to True and spk_embeddings_dir (will load from disk) or configure the model to compute it on forward
|
| 51 |
+
use_spk_emb: true
|
| 52 |
+
spk_embeddings_dir: ${data.dataset_dir}/spk_embeddings
|
| 53 |
+
# SR augmentation is deprecated, set use_sr to False
|
| 54 |
+
sr_min_max: [68, 92]
|
| 55 |
+
# For content feature extraction, set the content_feature_dir (will load from disk) or configure the model to compute it on forward
|
| 56 |
+
content_feature_dir: null
|
| 57 |
+
training_files: data/train.csv
|
| 58 |
+
validation_files: data/valid.csv
|
| 59 |
+
win_length: 1280
|
| 60 |
+
|
| 61 |
+
model:
|
| 62 |
+
save_dir: null
|
| 63 |
+
filter_channels: 768
|
| 64 |
+
finetune_from_model:
|
| 65 |
+
discriminator: /raid/lucasgris/free-svc/D-freevc-24.pth
|
| 66 |
+
generator: /raid/lucasgris/free-svc/freevc-24.pth
|
| 67 |
+
hidden_channels: 192
|
| 68 |
+
inter_channels: 192
|
| 69 |
+
kernel_size: 3
|
| 70 |
+
n_heads: 2
|
| 71 |
+
n_layers_q: 3
|
| 72 |
+
n_layers: 6
|
| 73 |
+
p_dropout: 0.1
|
| 74 |
+
resblock_dilation_sizes: [[1,3,5], [1,3,5], [1,3,5]]
|
| 75 |
+
resblock_kernel_sizes: [3,7,11]
|
| 76 |
+
resblock: 1
|
| 77 |
+
c_dim: 768
|
| 78 |
+
upsample_initial_channel: 512
|
| 79 |
+
upsample_kernel_sizes: [16,16,4,4]
|
| 80 |
+
upsample_rates: [10,8,2,2]
|
| 81 |
+
use_spectral_norm: false
|
| 82 |
+
freeze_external_spk: true
|
| 83 |
+
device: cuda
|
| 84 |
+
# For online speaker embedding extraction, set the use_spk_emb to True and spk_encoder_type
|
| 85 |
+
use_spk_emb: false
|
| 86 |
+
gin_channels: null # gin_channels = spk_encoder.embedding_dim
|
| 87 |
+
spk_encoder_type: null # ECAPA2SpeakerEncoder16k |
|
| 88 |
+
# For content feature extraction, set the content_encoder_type and content_encoder_ckpt
|
| 89 |
+
content_encoder_type: null # load from disk (data) - hubert | wavlm
|
| 90 |
+
content_encoder_ckpt: null # load from disk (data) - [path] | models/wavlm/WavLM-Large.pt | lengyue233/content-vec-best
|
| 91 |
+
post_content_encoder_type: vits-encoder-with-uv-emb # or freevc-bottleneck
|
| 92 |
+
coarse_f0: true
|
| 93 |
+
cond_f0_on_flow: false
|