Skip to content

Commit a88b64d

Browse files
committed
add hifi gan ft config
1 parent 316fe8a commit a88b64d

File tree

1 file changed

+127
-0
lines changed

1 file changed

+127
-0
lines changed

configs/ft_hifigan.yaml

Lines changed: 127 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,127 @@
1+
# preprocessing
2+
binarizer_cls: preprocessing.BaseBinarizer
3+
raw_data_dir: []
4+
binary_data_dir: null
5+
binarization_args:
6+
num_workers: 8
7+
shuffle: true
8+
9+
DataIndexPath: data
10+
valid_set_name: valid
11+
train_set_name: train
12+
13+
14+
volume_aug: true
15+
volume_aug_prob: 0.5
16+
17+
18+
mel_vmin: -6. #-6.
19+
mel_vmax: 1.5
20+
21+
22+
audio_sample_rate: 44100
23+
audio_num_mel_bins: 128
24+
hop_size: 512 # Hop size.
25+
fft_size: 2048 # FFT size.
26+
win_size: 2048 # FFT size.
27+
fmin: 40
28+
fmax: 16000
29+
fmax_for_loss: null
30+
crop_mel_frames: 5
31+
test_prefixes: []
32+
33+
pe: rmvpe
34+
pe_ckpt: pretrained/rmvpe/model.pt
35+
36+
# global constants
37+
38+
39+
# neural networks
40+
41+
42+
#model_cls: training.nsf_HiFigan_task.nsf_HiFigan
43+
model_args:
44+
upsample_rates: [ 8, 8, 2, 2, 2 ]
45+
upsample_kernel_sizes: [ 16,16, 4, 4, 4 ]
46+
upsample_initial_channel: 512
47+
resblock_kernel_sizes: [ 3,7,11 ]
48+
resblock_dilation_sizes: [ [ 1,3,5 ], [ 1,3,5 ], [ 1,3,5 ] ]
49+
discriminator_periods: [ 3, 5, 7, 11, 17, 23, 37 ]
50+
resblock: "1"
51+
52+
# training
53+
use_midi_loss: true
54+
use_bound_loss: true
55+
task_cls: training.nsf_HiFigan_task.nsf_HiFigan
56+
57+
58+
#sort_by_len: true
59+
#optimizer_args:
60+
# optimizer_cls: torch.optim.AdamW
61+
# lr: 0.0001
62+
# beta1: 0.9
63+
# beta2: 0.98
64+
# weight_decay: 0
65+
#lab_aux_loss: 0.5
66+
discriminate_optimizer_args:
67+
optimizer_cls: torch.optim.AdamW
68+
lr: 0.00001
69+
beta1: 0.9
70+
beta2: 0.98
71+
weight_decay: 0
72+
73+
generater_optimizer_args:
74+
optimizer_cls: torch.optim.AdamW
75+
lr: 0.00001
76+
beta1: 0.9
77+
beta2: 0.98
78+
weight_decay: 0
79+
80+
lr_scheduler_args:
81+
scheduler_cls: lr_scheduler.scheduler.WarmupLR
82+
warmup_steps: 5000
83+
min_lr: 0.00001
84+
85+
clip_grad_norm: 1
86+
accumulate_grad_batches: 1
87+
sampler_frame_count_grid: 6
88+
ds_workers: 4
89+
dataloader_prefetch_factor: 2
90+
91+
batch_size: 2
92+
93+
94+
95+
num_valid_plots: 100
96+
log_interval: 100
97+
num_sanity_val_steps: 5 # steps of validation at the beginning
98+
val_check_interval: 8000
99+
num_ckpt_keep: 5
100+
max_updates: 100000
101+
permanent_ckpt_start: 200000
102+
permanent_ckpt_interval: 40000
103+
104+
###########
105+
# pytorch lightning
106+
# Read https://lightning.ai/docs/pytorch/stable/common/trainer.html#trainer-class-api for possible values
107+
###########
108+
pl_trainer_accelerator: 'auto'
109+
pl_trainer_devices: 'auto'
110+
pl_trainer_precision: '32-true'
111+
#pl_trainer_precision: 'bf16' #please do not use bf 16
112+
pl_trainer_num_nodes: 1
113+
pl_trainer_strategy: 'auto'
114+
ddp_backend: 'nccl' # choose from 'gloo', 'nccl', 'nccl_no_p2p'
115+
seed: 114514
116+
117+
###########
118+
# finetune
119+
###########
120+
121+
finetune_enabled: true
122+
finetune_ckpt_path: hifi.ckpt
123+
finetune_ignored_params: []
124+
finetune_strict_shapes: true
125+
126+
freezing_enabled: false
127+
frozen_params: []

0 commit comments

Comments
 (0)