chjn commited on
Commit
6c3a948
1 Parent(s): 1bd9c9d

Upload 3 files

Browse files
WonderAcute/WonderAcute.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d034eaf45b800dd200210a8469fb73840503368b7bcb1a539fbe783442f91654
3
+ size 627915739
WonderAcute/config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "log_interval": 200,
4
+ "eval_interval": 800,
5
+ "seed": 1234,
6
+ "epochs": 10000,
7
+ "learning_rate": 0.0001,
8
+ "betas": [
9
+ 0.8,
10
+ 0.99
11
+ ],
12
+ "eps": 1e-09,
13
+ "batch_size": 24,
14
+ "fp16_run": true,
15
+ "half_type": "fp16",
16
+ "lr_decay": 0.999875,
17
+ "segment_size": 10240,
18
+ "init_lr_ratio": 1,
19
+ "warmup_epochs": 0,
20
+ "c_mel": 45,
21
+ "c_kl": 1.0,
22
+ "use_sr": true,
23
+ "max_speclen": 512,
24
+ "port": "8001",
25
+ "keep_ckpts": 3,
26
+ "all_in_mem": true,
27
+ "vol_aug": true
28
+ },
29
+ "data": {
30
+ "training_files": "filelists/train.txt",
31
+ "validation_files": "filelists/val.txt",
32
+ "max_wav_value": 32768.0,
33
+ "sampling_rate": 44100,
34
+ "filter_length": 2048,
35
+ "hop_length": 512,
36
+ "win_length": 2048,
37
+ "n_mel_channels": 80,
38
+ "mel_fmin": 0.0,
39
+ "mel_fmax": 22050,
40
+ "unit_interpolate_mode": "nearest"
41
+ },
42
+ "model": {
43
+ "inter_channels": 192,
44
+ "hidden_channels": 192,
45
+ "filter_channels": 768,
46
+ "n_heads": 2,
47
+ "n_layers": 6,
48
+ "kernel_size": 3,
49
+ "p_dropout": 0.1,
50
+ "resblock": "1",
51
+ "resblock_kernel_sizes": [
52
+ 3,
53
+ 7,
54
+ 11
55
+ ],
56
+ "resblock_dilation_sizes": [
57
+ [
58
+ 1,
59
+ 3,
60
+ 5
61
+ ],
62
+ [
63
+ 1,
64
+ 3,
65
+ 5
66
+ ],
67
+ [
68
+ 1,
69
+ 3,
70
+ 5
71
+ ]
72
+ ],
73
+ "upsample_rates": [
74
+ 8,
75
+ 8,
76
+ 2,
77
+ 2,
78
+ 2
79
+ ],
80
+ "upsample_initial_channel": 512,
81
+ "upsample_kernel_sizes": [
82
+ 16,
83
+ 16,
84
+ 4,
85
+ 4,
86
+ 4
87
+ ],
88
+ "n_layers_q": 3,
89
+ "n_layers_trans_flow": 3,
90
+ "n_flow_layer": 4,
91
+ "use_spectral_norm": false,
92
+ "gin_channels": 768,
93
+ "ssl_dim": 768,
94
+ "n_speakers": 1,
95
+ "vocoder_name": "nsf-hifigan",
96
+ "speech_encoder": "vec768l12",
97
+ "speaker_embedding": false,
98
+ "vol_embedding": true,
99
+ "use_depthwise_conv": false,
100
+ "flow_share_parameter": false,
101
+ "use_automatic_f0_prediction": true,
102
+ "use_transformer_flow": false
103
+ },
104
+ "spk": {
105
+ "WonderAcute": 0
106
+ }
107
+ }
WonderAcute/diffusion.yaml ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ data:
2
+ block_size: 512
3
+ cnhubertsoft_gate: 10
4
+ duration: 2
5
+ encoder: vec768l12
6
+ encoder_hop_size: 320
7
+ encoder_out_channels: 768
8
+ encoder_sample_rate: 16000
9
+ extensions:
10
+ - wav
11
+ sampling_rate: 44100
12
+ training_files: filelists/train.txt
13
+ unit_interpolate_mode: nearest
14
+ validation_files: filelists/val.txt
15
+ device: cuda
16
+ env:
17
+ expdir: logs/44k/diffusion
18
+ gpu_id: 0
19
+ infer:
20
+ method: dpm-solver++
21
+ speedup: 10
22
+ model:
23
+ k_step_max: 0
24
+ n_chans: 512
25
+ n_hidden: 256
26
+ n_layers: 20
27
+ n_spk: 1
28
+ timesteps: 1000
29
+ type: Diffusion
30
+ use_pitch_aug: true
31
+ spk:
32
+ WonderAcute: 0
33
+ train:
34
+ amp_dtype: fp16
35
+ batch_size: 96
36
+ cache_all_data: true
37
+ cache_device: cuda
38
+ cache_fp16: true
39
+ decay_step: 6000
40
+ epochs: 100000
41
+ gamma: 0.9
42
+ interval_force_save: 5000
43
+ interval_log: 10
44
+ interval_val: 2000
45
+ lr: 0.0002
46
+ num_workers: 0
47
+ save_opt: true
48
+ weight_decay: 0.0001
49
+ vocoder:
50
+ ckpt: pretrain/nsf_hifigan/model
51
+ type: nsf-hifigan