Hugging Face's logo

FashionFlora
/
PitchExtractor

TensorBoard
Model card Files
xet
 Metrics Community
FashionFlora commited on
Commit
9a964a6
·
verified ·
1 Parent(s): be4545a

Upload folder using huggingface_hub

Browse files
Files changed (24) hide show
  1. .gitattributes +1 -0
  2. Checkpoint/config.yml +17 -0
  3. Checkpoint/epoch_00020.pth +3 -0
  4. Checkpoint/epoch_00030.pth +3 -0
  5. Checkpoint/epoch_00050.pth +3 -0
  6. Checkpoint/epoch_00060.pth +3 -0
  7. Checkpoint/epoch_00070.pth +3 -0
  8. Checkpoint/tensorboard/events.out.tfevents.1748789103.34a32647fcb7.2057.0 +3 -0
  9. Checkpoint/train.log +560 -0
  10. Configs/config.yml +17 -0
  11. Data/OOD_texts.txt +3 -0
  12. Data/train_list.txt +0 -0
  13. Data/val_list.txt +0 -0
  14. LICENSE +21 -0
  15. README.md +36 -0
  16. __pycache__/meldataset.cpython-312.pyc +0 -0
  17. __pycache__/model.cpython-312.pyc +0 -0
  18. __pycache__/optimizers.cpython-312.pyc +0 -0
  19. __pycache__/trainer.cpython-312.pyc +0 -0
  20. meldataset.py +191 -0
  21. model.py +175 -0
  22. optimizers.py +86 -0
  23. train.py +144 -0
  24. trainer.py +186 -0
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ Data/OOD_texts.txt filter=lfs diff=lfs merge=lfs -text
Checkpoint/config.yml ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ log_dir: "Checkpoint"
2
+ save_freq: 10
3
+ device: "cuda"
4
+ epochs: 100
5
+ batch_size: 32
6
+ pretrained_model: ""
7
+ train_data: "Data/train_list.txt"
8
+ val_data: "Data/val_list.txt"
9
+ num_workers: 16
10
+
11
+
12
+ optimizer_params:
13
+ lr: 0.0003
14
+
15
+ loss_params:
16
+ lambda_f0: 0.1
17
+
Checkpoint/epoch_00020.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1cd8434a6e1c8c964be6316c1933405d6f0bb61bcd11366b713259bad3545064
3
+ size 63056961
Checkpoint/epoch_00030.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a04b0d41ccce996a18c5f2264ff86e11607f4c5261a5866e8806c31540caaa86
3
+ size 63056961
Checkpoint/epoch_00050.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:acbb1d7f347135999be385042560ece77a54f54b179d181569a9ea1dbfaada84
3
+ size 63056961
Checkpoint/epoch_00060.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6023df18d4bc3d923e8716e1b482fd260806960ad41b95f595c70882d1c39caa
3
+ size 63056961
Checkpoint/epoch_00070.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:23ebcb6ba4b45c686d88219f6113d49dab198733d1b8188374b02989364f5e5f
3
+ size 63056961
Checkpoint/tensorboard/events.out.tfevents.1748789103.34a32647fcb7.2057.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f05710112970e3d24c848eba1a632aa5ad0aa9bbd4bfc594f7c582017d3b32ce
3
+ size 23608
Checkpoint/train.log ADDED
@@ -0,0 +1,560 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ INFO:2025-06-01 16:40:51,971: --- epoch 1 ---
2
+ INFO:2025-06-01 16:40:51,971: train/loss : 3.0885
3
+ INFO:2025-06-01 16:40:51,972: train/f0 : 2.8800
4
+ INFO:2025-06-01 16:40:51,972: train/sil : 0.2085
5
+ INFO:2025-06-01 16:40:51,972: train/learning_rate: 0.0003
6
+ INFO:2025-06-01 16:40:51,972: eval/loss : 1.7819
7
+ INFO:2025-06-01 16:40:51,973: eval/f0 : 1.5807
8
+ INFO:2025-06-01 16:40:51,973: eval/sil : 0.2012
9
+ INFO:2025-06-01 16:44:27,487: --- epoch 2 ---
10
+ INFO:2025-06-01 16:44:27,487: train/loss : 1.6526
11
+ INFO:2025-06-01 16:44:27,488: train/f0 : 1.4625
12
+ INFO:2025-06-01 16:44:27,488: train/sil : 0.1901
13
+ INFO:2025-06-01 16:44:27,488: train/learning_rate: 0.0003
14
+ INFO:2025-06-01 16:44:27,488: eval/loss : 1.6477
15
+ INFO:2025-06-01 16:44:27,488: eval/f0 : 1.4575
16
+ INFO:2025-06-01 16:44:27,489: eval/sil : 0.1902
17
+ INFO:2025-06-01 16:48:02,829: --- epoch 3 ---
18
+ INFO:2025-06-01 16:48:02,829: train/loss : 1.5773
19
+ INFO:2025-06-01 16:48:02,829: train/f0 : 1.3918
20
+ INFO:2025-06-01 16:48:02,830: train/sil : 0.1855
21
+ INFO:2025-06-01 16:48:02,830: train/learning_rate: 0.0003
22
+ INFO:2025-06-01 16:48:02,830: eval/loss : 1.6337
23
+ INFO:2025-06-01 16:48:02,830: eval/f0 : 1.4448
24
+ INFO:2025-06-01 16:48:02,830: eval/sil : 0.1889
25
+ INFO:2025-06-01 16:51:38,771: --- epoch 4 ---
26
+ INFO:2025-06-01 16:51:38,771: train/loss : 1.5464
27
+ INFO:2025-06-01 16:51:38,772: train/f0 : 1.3635
28
+ INFO:2025-06-01 16:51:38,772: train/sil : 0.1829
29
+ INFO:2025-06-01 16:51:38,772: train/learning_rate: 0.0003
30
+ INFO:2025-06-01 16:51:38,773: eval/loss : 1.5930
31
+ INFO:2025-06-01 16:51:38,773: eval/f0 : 1.4066
32
+ INFO:2025-06-01 16:51:38,773: eval/sil : 0.1865
33
+ INFO:2025-06-01 16:55:15,816: --- epoch 5 ---
34
+ INFO:2025-06-01 16:55:15,816: train/loss : 1.5211
35
+ INFO:2025-06-01 16:55:15,817: train/f0 : 1.3406
36
+ INFO:2025-06-01 16:55:15,817: train/sil : 0.1806
37
+ INFO:2025-06-01 16:55:15,817: train/learning_rate: 0.0003
38
+ INFO:2025-06-01 16:55:15,817: eval/loss : 1.5643
39
+ INFO:2025-06-01 16:55:15,817: eval/f0 : 1.3794
40
+ INFO:2025-06-01 16:55:15,817: eval/sil : 0.1849
41
+ INFO:2025-06-01 16:58:48,728: --- epoch 6 ---
42
+ INFO:2025-06-01 16:58:48,728: train/loss : 1.5044
43
+ INFO:2025-06-01 16:58:48,728: train/f0 : 1.3255
44
+ INFO:2025-06-01 16:58:48,729: train/sil : 0.1789
45
+ INFO:2025-06-01 16:58:48,729: train/learning_rate: 0.0003
46
+ INFO:2025-06-01 16:58:48,729: eval/loss : 1.5230
47
+ INFO:2025-06-01 16:58:48,730: eval/f0 : 1.3438
48
+ INFO:2025-06-01 16:58:48,730: eval/sil : 0.1791
49
+ INFO:2025-06-01 17:02:21,431: --- epoch 7 ---
50
+ INFO:2025-06-01 17:02:21,432: train/loss : 1.4880
51
+ INFO:2025-06-01 17:02:21,432: train/f0 : 1.3106
52
+ INFO:2025-06-01 17:02:21,432: train/sil : 0.1774
53
+ INFO:2025-06-01 17:02:21,433: train/learning_rate: 0.0003
54
+ INFO:2025-06-01 17:02:21,433: eval/loss : 1.5487
55
+ INFO:2025-06-01 17:02:21,433: eval/f0 : 1.3655
56
+ INFO:2025-06-01 17:02:21,433: eval/sil : 0.1832
57
+ INFO:2025-06-01 17:05:53,728: --- epoch 8 ---
58
+ INFO:2025-06-01 17:05:53,728: train/loss : 1.4784
59
+ INFO:2025-06-01 17:05:53,729: train/f0 : 1.3018
60
+ INFO:2025-06-01 17:05:53,729: train/sil : 0.1766
61
+ INFO:2025-06-01 17:05:53,729: train/learning_rate: 0.0003
62
+ INFO:2025-06-01 17:05:53,729: eval/loss : 1.5616
63
+ INFO:2025-06-01 17:05:53,730: eval/f0 : 1.3765
64
+ INFO:2025-06-01 17:05:53,730: eval/sil : 0.1851
65
+ INFO:2025-06-01 17:09:25,194: --- epoch 9 ---
66
+ INFO:2025-06-01 17:09:25,194: train/loss : 1.4688
67
+ INFO:2025-06-01 17:09:25,194: train/f0 : 1.2929
68
+ INFO:2025-06-01 17:09:25,195: train/sil : 0.1759
69
+ INFO:2025-06-01 17:09:25,195: train/learning_rate: 0.0003
70
+ INFO:2025-06-01 17:09:25,196: eval/loss : 1.5449
71
+ INFO:2025-06-01 17:09:25,196: eval/f0 : 1.3628
72
+ INFO:2025-06-01 17:09:25,196: eval/sil : 0.1821
73
+ INFO:2025-06-01 17:12:56,663: --- epoch 10 ---
74
+ INFO:2025-06-01 17:12:56,663: train/loss : 1.4583
75
+ INFO:2025-06-01 17:12:56,664: train/f0 : 1.2836
76
+ INFO:2025-06-01 17:12:56,664: train/sil : 0.1747
77
+ INFO:2025-06-01 17:12:56,664: train/learning_rate: 0.0003
78
+ INFO:2025-06-01 17:12:56,664: eval/loss : 1.4957
79
+ INFO:2025-06-01 17:12:56,664: eval/f0 : 1.3204
80
+ INFO:2025-06-01 17:12:56,664: eval/sil : 0.1753
81
+ INFO:2025-06-01 17:16:28,703: --- epoch 11 ---
82
+ INFO:2025-06-01 17:16:28,703: train/loss : 1.4496
83
+ INFO:2025-06-01 17:16:28,703: train/f0 : 1.2758
84
+ INFO:2025-06-01 17:16:28,703: train/sil : 0.1738
85
+ INFO:2025-06-01 17:16:28,704: train/learning_rate: 0.0003
86
+ INFO:2025-06-01 17:16:28,704: eval/loss : 1.5090
87
+ INFO:2025-06-01 17:16:28,704: eval/f0 : 1.3294
88
+ INFO:2025-06-01 17:16:28,704: eval/sil : 0.1796
89
+ INFO:2025-06-01 17:19:59,525: --- epoch 12 ---
90
+ INFO:2025-06-01 17:19:59,525: train/loss : 1.4421
91
+ INFO:2025-06-01 17:19:59,526: train/f0 : 1.2688
92
+ INFO:2025-06-01 17:19:59,526: train/sil : 0.1733
93
+ INFO:2025-06-01 17:19:59,526: train/learning_rate: 0.0003
94
+ INFO:2025-06-01 17:19:59,527: eval/loss : 1.5067
95
+ INFO:2025-06-01 17:19:59,527: eval/f0 : 1.3310
96
+ INFO:2025-06-01 17:19:59,527: eval/sil : 0.1757
97
+ INFO:2025-06-01 17:23:31,629: --- epoch 13 ---
98
+ INFO:2025-06-01 17:23:31,629: train/loss : 1.4307
99
+ INFO:2025-06-01 17:23:31,629: train/f0 : 1.2586
100
+ INFO:2025-06-01 17:23:31,630: train/sil : 0.1721
101
+ INFO:2025-06-01 17:23:31,630: train/learning_rate: 0.0003
102
+ INFO:2025-06-01 17:23:31,630: eval/loss : 1.5103
103
+ INFO:2025-06-01 17:23:31,631: eval/f0 : 1.3347
104
+ INFO:2025-06-01 17:23:31,631: eval/sil : 0.1756
105
+ INFO:2025-06-01 17:27:03,433: --- epoch 14 ---
106
+ INFO:2025-06-01 17:27:03,433: train/loss : 1.4232
107
+ INFO:2025-06-01 17:27:03,433: train/f0 : 1.2516
108
+ INFO:2025-06-01 17:27:03,433: train/sil : 0.1717
109
+ INFO:2025-06-01 17:27:03,434: train/learning_rate: 0.0003
110
+ INFO:2025-06-01 17:27:03,434: eval/loss : 1.4607
111
+ INFO:2025-06-01 17:27:03,434: eval/f0 : 1.2861
112
+ INFO:2025-06-01 17:27:03,434: eval/sil : 0.1745
113
+ INFO:2025-06-01 17:30:35,679: --- epoch 15 ---
114
+ INFO:2025-06-01 17:30:35,679: train/loss : 1.4136
115
+ INFO:2025-06-01 17:30:35,679: train/f0 : 1.2427
116
+ INFO:2025-06-01 17:30:35,679: train/sil : 0.1709
117
+ INFO:2025-06-01 17:30:35,679: train/learning_rate: 0.0003
118
+ INFO:2025-06-01 17:30:35,680: eval/loss : 1.4952
119
+ INFO:2025-06-01 17:30:35,680: eval/f0 : 1.3199
120
+ INFO:2025-06-01 17:30:35,680: eval/sil : 0.1753
121
+ INFO:2025-06-01 17:34:08,277: --- epoch 16 ---
122
+ INFO:2025-06-01 17:34:08,277: train/loss : 1.4039
123
+ INFO:2025-06-01 17:34:08,277: train/f0 : 1.2339
124
+ INFO:2025-06-01 17:34:08,277: train/sil : 0.1700
125
+ INFO:2025-06-01 17:34:08,278: train/learning_rate: 0.0003
126
+ INFO:2025-06-01 17:34:08,278: eval/loss : 1.5022
127
+ INFO:2025-06-01 17:34:08,278: eval/f0 : 1.3243
128
+ INFO:2025-06-01 17:34:08,278: eval/sil : 0.1779
129
+ INFO:2025-06-01 17:37:40,058: --- epoch 17 ---
130
+ INFO:2025-06-01 17:37:40,058: train/loss : 1.3960
131
+ INFO:2025-06-01 17:37:40,058: train/f0 : 1.2269
132
+ INFO:2025-06-01 17:37:40,058: train/sil : 0.1691
133
+ INFO:2025-06-01 17:37:40,059: train/learning_rate: 0.0003
134
+ INFO:2025-06-01 17:37:40,059: eval/loss : 1.5321
135
+ INFO:2025-06-01 17:37:40,059: eval/f0 : 1.3505
136
+ INFO:2025-06-01 17:37:40,059: eval/sil : 0.1815
137
+ INFO:2025-06-01 17:41:12,982: --- epoch 18 ---
138
+ INFO:2025-06-01 17:41:12,982: train/loss : 1.3908
139
+ INFO:2025-06-01 17:41:12,982: train/f0 : 1.2222
140
+ INFO:2025-06-01 17:41:12,983: train/sil : 0.1687
141
+ INFO:2025-06-01 17:41:12,983: train/learning_rate: 0.0003
142
+ INFO:2025-06-01 17:41:12,983: eval/loss : 1.4792
143
+ INFO:2025-06-01 17:41:12,983: eval/f0 : 1.3021
144
+ INFO:2025-06-01 17:41:12,983: eval/sil : 0.1771
145
+ INFO:2025-06-01 17:44:48,797: --- epoch 19 ---
146
+ INFO:2025-06-01 17:44:48,797: train/loss : 1.3829
147
+ INFO:2025-06-01 17:44:48,798: train/f0 : 1.2149
148
+ INFO:2025-06-01 17:44:48,798: train/sil : 0.1680
149
+ INFO:2025-06-01 17:44:48,798: train/learning_rate: 0.0003
150
+ INFO:2025-06-01 17:44:48,798: eval/loss : 1.4834
151
+ INFO:2025-06-01 17:44:48,798: eval/f0 : 1.3093
152
+ INFO:2025-06-01 17:44:48,799: eval/sil : 0.1741
153
+ INFO:2025-06-01 17:48:20,820: --- epoch 20 ---
154
+ INFO:2025-06-01 17:48:20,820: train/loss : 1.3726
155
+ INFO:2025-06-01 17:48:20,820: train/f0 : 1.2057
156
+ INFO:2025-06-01 17:48:20,821: train/sil : 0.1669
157
+ INFO:2025-06-01 17:48:20,821: train/learning_rate: 0.0003
158
+ INFO:2025-06-01 17:48:20,821: eval/loss : 1.5440
159
+ INFO:2025-06-01 17:48:20,821: eval/f0 : 1.3609
160
+ INFO:2025-06-01 17:48:20,821: eval/sil : 0.1831
161
+ INFO:2025-06-01 17:51:53,163: --- epoch 21 ---
162
+ INFO:2025-06-01 17:51:53,163: train/loss : 1.3631
163
+ INFO:2025-06-01 17:51:53,164: train/f0 : 1.1967
164
+ INFO:2025-06-01 17:51:53,164: train/sil : 0.1664
165
+ INFO:2025-06-01 17:51:53,164: train/learning_rate: 0.0003
166
+ INFO:2025-06-01 17:51:53,164: eval/loss : 1.5264
167
+ INFO:2025-06-01 17:51:53,164: eval/f0 : 1.3471
168
+ INFO:2025-06-01 17:51:53,164: eval/sil : 0.1793
169
+ INFO:2025-06-01 17:55:25,512: --- epoch 22 ---
170
+ INFO:2025-06-01 17:55:25,513: train/loss : 1.3546
171
+ INFO:2025-06-01 17:55:25,513: train/f0 : 1.1893
172
+ INFO:2025-06-01 17:55:25,513: train/sil : 0.1653
173
+ INFO:2025-06-01 17:55:25,513: train/learning_rate: 0.0003
174
+ INFO:2025-06-01 17:55:25,513: eval/loss : 1.4943
175
+ INFO:2025-06-01 17:55:25,514: eval/f0 : 1.3182
176
+ INFO:2025-06-01 17:55:25,514: eval/sil : 0.1761
177
+ INFO:2025-06-01 17:58:57,803: --- epoch 23 ---
178
+ INFO:2025-06-01 17:58:57,803: train/loss : 1.3466
179
+ INFO:2025-06-01 17:58:57,803: train/f0 : 1.1820
180
+ INFO:2025-06-01 17:58:57,804: train/sil : 0.1646
181
+ INFO:2025-06-01 17:58:57,804: train/learning_rate: 0.0003
182
+ INFO:2025-06-01 17:58:57,804: eval/loss : 1.4973
183
+ INFO:2025-06-01 17:58:57,804: eval/f0 : 1.3192
184
+ INFO:2025-06-01 17:58:57,804: eval/sil : 0.1780
185
+ INFO:2025-06-01 18:02:30,350: --- epoch 24 ---
186
+ INFO:2025-06-01 18:02:30,350: train/loss : 1.3341
187
+ INFO:2025-06-01 18:02:30,350: train/f0 : 1.1706
188
+ INFO:2025-06-01 18:02:30,350: train/sil : 0.1635
189
+ INFO:2025-06-01 18:02:30,351: train/learning_rate: 0.0003
190
+ INFO:2025-06-01 18:02:30,351: eval/loss : 1.5060
191
+ INFO:2025-06-01 18:02:30,351: eval/f0 : 1.3270
192
+ INFO:2025-06-01 18:02:30,351: eval/sil : 0.1790
193
+ INFO:2025-06-01 18:06:02,040: --- epoch 25 ---
194
+ INFO:2025-06-01 18:06:02,041: train/loss : 1.3278
195
+ INFO:2025-06-01 18:06:02,041: train/f0 : 1.1650
196
+ INFO:2025-06-01 18:06:02,041: train/sil : 0.1628
197
+ INFO:2025-06-01 18:06:02,041: train/learning_rate: 0.0003
198
+ INFO:2025-06-01 18:06:02,041: eval/loss : 1.5066
199
+ INFO:2025-06-01 18:06:02,042: eval/f0 : 1.3271
200
+ INFO:2025-06-01 18:06:02,042: eval/sil : 0.1795
201
+ INFO:2025-06-01 18:09:33,910: --- epoch 26 ---
202
+ INFO:2025-06-01 18:09:33,910: train/loss : 1.3176
203
+ INFO:2025-06-01 18:09:33,910: train/f0 : 1.1559
204
+ INFO:2025-06-01 18:09:33,911: train/sil : 0.1617
205
+ INFO:2025-06-01 18:09:33,911: train/learning_rate: 0.0003
206
+ INFO:2025-06-01 18:09:33,911: eval/loss : 1.5134
207
+ INFO:2025-06-01 18:09:33,911: eval/f0 : 1.3332
208
+ INFO:2025-06-01 18:09:33,911: eval/sil : 0.1802
209
+ INFO:2025-06-01 18:13:05,634: --- epoch 27 ---
210
+ INFO:2025-06-01 18:13:05,634: train/loss : 1.3055
211
+ INFO:2025-06-01 18:13:05,634: train/f0 : 1.1451
212
+ INFO:2025-06-01 18:13:05,635: train/sil : 0.1605
213
+ INFO:2025-06-01 18:13:05,635: train/learning_rate: 0.0002
214
+ INFO:2025-06-01 18:13:05,635: eval/loss : 1.4904
215
+ INFO:2025-06-01 18:13:05,635: eval/f0 : 1.3130
216
+ INFO:2025-06-01 18:13:05,636: eval/sil : 0.1774
217
+ INFO:2025-06-01 18:16:36,743: --- epoch 28 ---
218
+ INFO:2025-06-01 18:16:36,744: train/loss : 1.2989
219
+ INFO:2025-06-01 18:16:36,744: train/f0 : 1.1392
220
+ INFO:2025-06-01 18:16:36,744: train/sil : 0.1598
221
+ INFO:2025-06-01 18:16:36,744: train/learning_rate: 0.0002
222
+ INFO:2025-06-01 18:16:36,745: eval/loss : 1.5058
223
+ INFO:2025-06-01 18:16:36,745: eval/f0 : 1.3258
224
+ INFO:2025-06-01 18:16:36,745: eval/sil : 0.1799
225
+ INFO:2025-06-01 18:20:09,380: --- epoch 29 ---
226
+ INFO:2025-06-01 18:20:09,380: train/loss : 1.2861
227
+ INFO:2025-06-01 18:20:09,381: train/f0 : 1.1278
228
+ INFO:2025-06-01 18:20:09,381: train/sil : 0.1583
229
+ INFO:2025-06-01 18:20:09,381: train/learning_rate: 0.0002
230
+ INFO:2025-06-01 18:20:09,381: eval/loss : 1.4993
231
+ INFO:2025-06-01 18:20:09,381: eval/f0 : 1.3186
232
+ INFO:2025-06-01 18:20:09,382: eval/sil : 0.1807
233
+ INFO:2025-06-01 18:23:41,136: --- epoch 30 ---
234
+ INFO:2025-06-01 18:23:41,136: train/loss : 1.2751
235
+ INFO:2025-06-01 18:23:41,136: train/f0 : 1.1178
236
+ INFO:2025-06-01 18:23:41,137: train/sil : 0.1573
237
+ INFO:2025-06-01 18:23:41,137: train/learning_rate: 0.0002
238
+ INFO:2025-06-01 18:23:41,137: eval/loss : 1.4968
239
+ INFO:2025-06-01 18:23:41,137: eval/f0 : 1.3155
240
+ INFO:2025-06-01 18:23:41,138: eval/sil : 0.1813
241
+ INFO:2025-06-01 18:27:13,509: --- epoch 31 ---
242
+ INFO:2025-06-01 18:27:13,509: train/loss : 1.2640
243
+ INFO:2025-06-01 18:27:13,509: train/f0 : 1.1079
244
+ INFO:2025-06-01 18:27:13,509: train/sil : 0.1561
245
+ INFO:2025-06-01 18:27:13,509: train/learning_rate: 0.0002
246
+ INFO:2025-06-01 18:27:13,510: eval/loss : 1.5054
247
+ INFO:2025-06-01 18:27:13,510: eval/f0 : 1.3248
248
+ INFO:2025-06-01 18:27:13,510: eval/sil : 0.1806
249
+ INFO:2025-06-01 18:30:45,279: --- epoch 32 ---
250
+ INFO:2025-06-01 18:30:45,279: train/loss : 1.2605
251
+ INFO:2025-06-01 18:30:45,279: train/f0 : 1.1052
252
+ INFO:2025-06-01 18:30:45,279: train/sil : 0.1553
253
+ INFO:2025-06-01 18:30:45,280: train/learning_rate: 0.0002
254
+ INFO:2025-06-01 18:30:45,280: eval/loss : 1.5038
255
+ INFO:2025-06-01 18:30:45,280: eval/f0 : 1.3237
256
+ INFO:2025-06-01 18:30:45,280: eval/sil : 0.1802
257
+ INFO:2025-06-01 18:34:16,950: --- epoch 33 ---
258
+ INFO:2025-06-01 18:34:16,950: train/loss : 1.2508
259
+ INFO:2025-06-01 18:34:16,950: train/f0 : 1.0965
260
+ INFO:2025-06-01 18:34:16,951: train/sil : 0.1543
261
+ INFO:2025-06-01 18:34:16,951: train/learning_rate: 0.0002
262
+ INFO:2025-06-01 18:34:16,951: eval/loss : 1.5009
263
+ INFO:2025-06-01 18:34:16,951: eval/f0 : 1.3176
264
+ INFO:2025-06-01 18:34:16,952: eval/sil : 0.1833
265
+ INFO:2025-06-01 18:37:49,655: --- epoch 34 ---
266
+ INFO:2025-06-01 18:37:49,655: train/loss : 1.2386
267
+ INFO:2025-06-01 18:37:49,655: train/f0 : 1.0856
268
+ INFO:2025-06-01 18:37:49,655: train/sil : 0.1529
269
+ INFO:2025-06-01 18:37:49,655: train/learning_rate: 0.0002
270
+ INFO:2025-06-01 18:37:49,656: eval/loss : 1.5012
271
+ INFO:2025-06-01 18:37:49,656: eval/f0 : 1.3192
272
+ INFO:2025-06-01 18:37:49,656: eval/sil : 0.1820
273
+ INFO:2025-06-01 18:41:21,765: --- epoch 35 ---
274
+ INFO:2025-06-01 18:41:21,766: train/loss : 1.2292
275
+ INFO:2025-06-01 18:41:21,766: train/f0 : 1.0773
276
+ INFO:2025-06-01 18:41:21,766: train/sil : 0.1518
277
+ INFO:2025-06-01 18:41:21,766: train/learning_rate: 0.0002
278
+ INFO:2025-06-01 18:41:21,767: eval/loss : 1.4754
279
+ INFO:2025-06-01 18:41:21,767: eval/f0 : 1.2952
280
+ INFO:2025-06-01 18:41:21,767: eval/sil : 0.1802
281
+ INFO:2025-06-01 18:44:53,596: --- epoch 36 ---
282
+ INFO:2025-06-01 18:44:53,597: train/loss : 1.2208
283
+ INFO:2025-06-01 18:44:53,597: train/f0 : 1.0702
284
+ INFO:2025-06-01 18:44:53,597: train/sil : 0.1505
285
+ INFO:2025-06-01 18:44:53,597: train/learning_rate: 0.0002
286
+ INFO:2025-06-01 18:44:53,597: eval/loss : 1.5341
287
+ INFO:2025-06-01 18:44:53,598: eval/f0 : 1.3461
288
+ INFO:2025-06-01 18:44:53,598: eval/sil : 0.1880
289
+ INFO:2025-06-01 18:48:25,343: --- epoch 37 ---
290
+ INFO:2025-06-01 18:48:25,344: train/loss : 1.2100
291
+ INFO:2025-06-01 18:48:25,344: train/f0 : 1.0607
292
+ INFO:2025-06-01 18:48:25,344: train/sil : 0.1494
293
+ INFO:2025-06-01 18:48:25,344: train/learning_rate: 0.0002
294
+ INFO:2025-06-01 18:48:25,345: eval/loss : 1.5276
295
+ INFO:2025-06-01 18:48:25,345: eval/f0 : 1.3412
296
+ INFO:2025-06-01 18:48:25,345: eval/sil : 0.1864
297
+ INFO:2025-06-01 18:51:57,791: --- epoch 38 ---
298
+ INFO:2025-06-01 18:51:57,791: train/loss : 1.2062
299
+ INFO:2025-06-01 18:51:57,791: train/f0 : 1.0574
300
+ INFO:2025-06-01 18:51:57,792: train/sil : 0.1487
301
+ INFO:2025-06-01 18:51:57,792: train/learning_rate: 0.0002
302
+ INFO:2025-06-01 18:51:57,792: eval/loss : 1.5113
303
+ INFO:2025-06-01 18:51:57,793: eval/f0 : 1.3246
304
+ INFO:2025-06-01 18:51:57,793: eval/sil : 0.1867
305
+ INFO:2025-06-01 18:55:29,730: --- epoch 39 ---
306
+ INFO:2025-06-01 18:55:29,730: train/loss : 1.1953
307
+ INFO:2025-06-01 18:55:29,730: train/f0 : 1.0477
308
+ INFO:2025-06-01 18:55:29,730: train/sil : 0.1475
309
+ INFO:2025-06-01 18:55:29,731: train/learning_rate: 0.0002
310
+ INFO:2025-06-01 18:55:29,731: eval/loss : 1.5343
311
+ INFO:2025-06-01 18:55:29,731: eval/f0 : 1.3457
312
+ INFO:2025-06-01 18:55:29,731: eval/sil : 0.1886
313
+ INFO:2025-06-01 18:59:01,120: --- epoch 40 ---
314
+ INFO:2025-06-01 18:59:01,121: train/loss : 1.1866
315
+ INFO:2025-06-01 18:59:01,121: train/f0 : 1.0401
316
+ INFO:2025-06-01 18:59:01,121: train/sil : 0.1466
317
+ INFO:2025-06-01 18:59:01,121: train/learning_rate: 0.0002
318
+ INFO:2025-06-01 18:59:01,121: eval/loss : 1.5472
319
+ INFO:2025-06-01 18:59:01,122: eval/f0 : 1.3564
320
+ INFO:2025-06-01 18:59:01,122: eval/sil : 0.1908
321
+ INFO:2025-06-01 19:02:32,997: --- epoch 41 ---
322
+ INFO:2025-06-01 19:02:32,997: train/loss : 1.1796
323
+ INFO:2025-06-01 19:02:32,997: train/f0 : 1.0338
324
+ INFO:2025-06-01 19:02:32,997: train/sil : 0.1457
325
+ INFO:2025-06-01 19:02:32,998: train/learning_rate: 0.0002
326
+ INFO:2025-06-01 19:02:32,998: eval/loss : 1.5317
327
+ INFO:2025-06-01 19:02:32,998: eval/f0 : 1.3403
328
+ INFO:2025-06-01 19:02:32,998: eval/sil : 0.1913
329
+ INFO:2025-06-01 19:06:07,793: --- epoch 42 ---
330
+ INFO:2025-06-01 19:06:07,793: train/loss : 1.1701
331
+ INFO:2025-06-01 19:06:07,794: train/f0 : 1.0259
332
+ INFO:2025-06-01 19:06:07,794: train/sil : 0.1441
333
+ INFO:2025-06-01 19:06:07,794: train/learning_rate: 0.0002
334
+ INFO:2025-06-01 19:06:07,795: eval/loss : 1.5308
335
+ INFO:2025-06-01 19:06:07,795: eval/f0 : 1.3399
336
+ INFO:2025-06-01 19:06:07,795: eval/sil : 0.1909
337
+ INFO:2025-06-01 19:09:44,178: --- epoch 43 ---
338
+ INFO:2025-06-01 19:09:44,179: train/loss : 1.1607
339
+ INFO:2025-06-01 19:09:44,179: train/f0 : 1.0177
340
+ INFO:2025-06-01 19:09:44,179: train/sil : 0.1430
341
+ INFO:2025-06-01 19:09:44,179: train/learning_rate: 0.0002
342
+ INFO:2025-06-01 19:09:44,179: eval/loss : 1.5487
343
+ INFO:2025-06-01 19:09:44,180: eval/f0 : 1.3513
344
+ INFO:2025-06-01 19:09:44,180: eval/sil : 0.1974
345
+ INFO:2025-06-01 19:13:15,424: --- epoch 44 ---
346
+ INFO:2025-06-01 19:13:15,425: train/loss : 1.1534
347
+ INFO:2025-06-01 19:13:15,425: train/f0 : 1.0113
348
+ INFO:2025-06-01 19:13:15,425: train/sil : 0.1421
349
+ INFO:2025-06-01 19:13:15,425: train/learning_rate: 0.0002
350
+ INFO:2025-06-01 19:13:15,425: eval/loss : 1.5332
351
+ INFO:2025-06-01 19:13:15,426: eval/f0 : 1.3421
352
+ INFO:2025-06-01 19:13:15,426: eval/sil : 0.1911
353
+ INFO:2025-06-01 19:16:47,751: --- epoch 45 ---
354
+ INFO:2025-06-01 19:16:47,751: train/loss : 1.1434
355
+ INFO:2025-06-01 19:16:47,752: train/f0 : 1.0026
356
+ INFO:2025-06-01 19:16:47,752: train/sil : 0.1409
357
+ INFO:2025-06-01 19:16:47,752: train/learning_rate: 0.0002
358
+ INFO:2025-06-01 19:16:47,752: eval/loss : 1.5308
359
+ INFO:2025-06-01 19:16:47,753: eval/f0 : 1.3354
360
+ INFO:2025-06-01 19:16:47,753: eval/sil : 0.1954
361
+ INFO:2025-06-01 19:20:19,443: --- epoch 46 ---
362
+ INFO:2025-06-01 19:20:19,443: train/loss : 1.1354
363
+ INFO:2025-06-01 19:20:19,444: train/f0 : 0.9956
364
+ INFO:2025-06-01 19:20:19,444: train/sil : 0.1398
365
+ INFO:2025-06-01 19:20:19,444: train/learning_rate: 0.0002
366
+ INFO:2025-06-01 19:20:19,444: eval/loss : 1.5555
367
+ INFO:2025-06-01 19:20:19,444: eval/f0 : 1.3573
368
+ INFO:2025-06-01 19:20:19,445: eval/sil : 0.1982
369
+ INFO:2025-06-01 19:23:52,515: --- epoch 47 ---
370
+ INFO:2025-06-01 19:23:52,515: train/loss : 1.1252
371
+ INFO:2025-06-01 19:23:52,515: train/f0 : 0.9868
372
+ INFO:2025-06-01 19:23:52,515: train/sil : 0.1384
373
+ INFO:2025-06-01 19:23:52,515: train/learning_rate: 0.0002
374
+ INFO:2025-06-01 19:23:52,516: eval/loss : 1.5228
375
+ INFO:2025-06-01 19:23:52,516: eval/f0 : 1.3254
376
+ INFO:2025-06-01 19:23:52,516: eval/sil : 0.1975
377
+ INFO:2025-06-01 19:27:25,164: --- epoch 48 ---
378
+ INFO:2025-06-01 19:27:25,164: train/loss : 1.1223
379
+ INFO:2025-06-01 19:27:25,164: train/f0 : 0.9844
380
+ INFO:2025-06-01 19:27:25,165: train/sil : 0.1380
381
+ INFO:2025-06-01 19:27:25,165: train/learning_rate: 0.0002
382
+ INFO:2025-06-01 19:27:25,165: eval/loss : 1.5512
383
+ INFO:2025-06-01 19:27:25,165: eval/f0 : 1.3500
384
+ INFO:2025-06-01 19:27:25,165: eval/sil : 0.2012
385
+ INFO:2025-06-01 19:30:56,640: --- epoch 49 ---
386
+ INFO:2025-06-01 19:30:56,640: train/loss : 1.1139
387
+ INFO:2025-06-01 19:30:56,640: train/f0 : 0.9771
388
+ INFO:2025-06-01 19:30:56,640: train/sil : 0.1368
389
+ INFO:2025-06-01 19:30:56,641: train/learning_rate: 0.0002
390
+ INFO:2025-06-01 19:30:56,641: eval/loss : 1.5470
391
+ INFO:2025-06-01 19:30:56,641: eval/f0 : 1.3484
392
+ INFO:2025-06-01 19:30:56,641: eval/sil : 0.1986
393
+ INFO:2025-06-01 19:34:28,781: --- epoch 50 ---
394
+ INFO:2025-06-01 19:34:28,782: train/loss : 1.1048
395
+ INFO:2025-06-01 19:34:28,782: train/f0 : 0.9692
396
+ INFO:2025-06-01 19:34:28,782: train/sil : 0.1356
397
+ INFO:2025-06-01 19:34:28,782: train/learning_rate: 0.0002
398
+ INFO:2025-06-01 19:34:28,783: eval/loss : 1.5572
399
+ INFO:2025-06-01 19:34:28,783: eval/f0 : 1.3543
400
+ INFO:2025-06-01 19:34:28,783: eval/sil : 0.2029
401
+ INFO:2025-06-01 19:38:00,874: --- epoch 51 ---
402
+ INFO:2025-06-01 19:38:00,874: train/loss : 1.0971
403
+ INFO:2025-06-01 19:38:00,874: train/f0 : 0.9625
404
+ INFO:2025-06-01 19:38:00,874: train/sil : 0.1346
405
+ INFO:2025-06-01 19:38:00,875: train/learning_rate: 0.0001
406
+ INFO:2025-06-01 19:38:00,875: eval/loss : 1.5085
407
+ INFO:2025-06-01 19:38:00,875: eval/f0 : 1.3116
408
+ INFO:2025-06-01 19:38:00,875: eval/sil : 0.1969
409
+ INFO:2025-06-01 19:41:32,149: --- epoch 52 ---
410
+ INFO:2025-06-01 19:41:32,150: train/loss : 1.0895
411
+ INFO:2025-06-01 19:41:32,150: train/f0 : 0.9559
412
+ INFO:2025-06-01 19:41:32,150: train/sil : 0.1336
413
+ INFO:2025-06-01 19:41:32,151: train/learning_rate: 0.0001
414
+ INFO:2025-06-01 19:41:32,151: eval/loss : 1.5259
415
+ INFO:2025-06-01 19:41:32,151: eval/f0 : 1.3250
416
+ INFO:2025-06-01 19:41:32,151: eval/sil : 0.2009
417
+ INFO:2025-06-01 19:45:04,016: --- epoch 53 ---
418
+ INFO:2025-06-01 19:45:04,017: train/loss : 1.0835
419
+ INFO:2025-06-01 19:45:04,017: train/f0 : 0.9510
420
+ INFO:2025-06-01 19:45:04,017: train/sil : 0.1326
421
+ INFO:2025-06-01 19:45:04,017: train/learning_rate: 0.0001
422
+ INFO:2025-06-01 19:45:04,018: eval/loss : 1.5333
423
+ INFO:2025-06-01 19:45:04,018: eval/f0 : 1.3327
424
+ INFO:2025-06-01 19:45:04,018: eval/sil : 0.2006
425
+ INFO:2025-06-01 19:48:35,506: --- epoch 54 ---
426
+ INFO:2025-06-01 19:48:35,506: train/loss : 1.0736
427
+ INFO:2025-06-01 19:48:35,507: train/f0 : 0.9423
428
+ INFO:2025-06-01 19:48:35,507: train/sil : 0.1313
429
+ INFO:2025-06-01 19:48:35,508: train/learning_rate: 0.0001
430
+ INFO:2025-06-01 19:48:35,508: eval/loss : 1.5590
431
+ INFO:2025-06-01 19:48:35,508: eval/f0 : 1.3546
432
+ INFO:2025-06-01 19:48:35,508: eval/sil : 0.2043
433
+ INFO:2025-06-01 19:52:07,120: --- epoch 55 ---
434
+ INFO:2025-06-01 19:52:07,120: train/loss : 1.0664
435
+ INFO:2025-06-01 19:52:07,120: train/f0 : 0.9360
436
+ INFO:2025-06-01 19:52:07,120: train/sil : 0.1303
437
+ INFO:2025-06-01 19:52:07,121: train/learning_rate: 0.0001
438
+ INFO:2025-06-01 19:52:07,121: eval/loss : 1.5577
439
+ INFO:2025-06-01 19:52:07,121: eval/f0 : 1.3521
440
+ INFO:2025-06-01 19:52:07,121: eval/sil : 0.2056
441
+ INFO:2025-06-01 19:55:38,786: --- epoch 56 ---
442
+ INFO:2025-06-01 19:55:38,786: train/loss : 1.0589
443
+ INFO:2025-06-01 19:55:38,786: train/f0 : 0.9296
444
+ INFO:2025-06-01 19:55:38,787: train/sil : 0.1294
445
+ INFO:2025-06-01 19:55:38,787: train/learning_rate: 0.0001
446
+ INFO:2025-06-01 19:55:38,787: eval/loss : 1.5557
447
+ INFO:2025-06-01 19:55:38,787: eval/f0 : 1.3509
448
+ INFO:2025-06-01 19:55:38,787: eval/sil : 0.2048
449
+ INFO:2025-06-01 19:59:10,771: --- epoch 57 ---
450
+ INFO:2025-06-01 19:59:10,771: train/loss : 1.0538
451
+ INFO:2025-06-01 19:59:10,771: train/f0 : 0.9252
452
+ INFO:2025-06-01 19:59:10,771: train/sil : 0.1286
453
+ INFO:2025-06-01 19:59:10,772: train/learning_rate: 0.0001
454
+ INFO:2025-06-01 19:59:10,772: eval/loss : 1.5371
455
+ INFO:2025-06-01 19:59:10,772: eval/f0 : 1.3306
456
+ INFO:2025-06-01 19:59:10,772: eval/sil : 0.2065
457
+ INFO:2025-06-01 20:02:43,395: --- epoch 58 ---
458
+ INFO:2025-06-01 20:02:43,395: train/loss : 1.0495
459
+ INFO:2025-06-01 20:02:43,395: train/f0 : 0.9216
460
+ INFO:2025-06-01 20:02:43,396: train/sil : 0.1279
461
+ INFO:2025-06-01 20:02:43,396: train/learning_rate: 0.0001
462
+ INFO:2025-06-01 20:02:43,396: eval/loss : 1.5604
463
+ INFO:2025-06-01 20:02:43,396: eval/f0 : 1.3508
464
+ INFO:2025-06-01 20:02:43,397: eval/sil : 0.2097
465
+ INFO:2025-06-01 20:06:15,038: --- epoch 59 ---
466
+ INFO:2025-06-01 20:06:15,038: train/loss : 1.0400
467
+ INFO:2025-06-01 20:06:15,039: train/f0 : 0.9134
468
+ INFO:2025-06-01 20:06:15,039: train/sil : 0.1266
469
+ INFO:2025-06-01 20:06:15,039: train/learning_rate: 0.0001
470
+ INFO:2025-06-01 20:06:15,039: eval/loss : 1.5236
471
+ INFO:2025-06-01 20:06:15,039: eval/f0 : 1.3177
472
+ INFO:2025-06-01 20:06:15,040: eval/sil : 0.2059
473
+ INFO:2025-06-01 20:09:47,579: --- epoch 60 ---
474
+ INFO:2025-06-01 20:09:47,579: train/loss : 1.0314
475
+ INFO:2025-06-01 20:09:47,579: train/f0 : 0.9061
476
+ INFO:2025-06-01 20:09:47,579: train/sil : 0.1253
477
+ INFO:2025-06-01 20:09:47,580: train/learning_rate: 0.0001
478
+ INFO:2025-06-01 20:09:47,580: eval/loss : 1.5278
479
+ INFO:2025-06-01 20:09:47,580: eval/f0 : 1.3203
480
+ INFO:2025-06-01 20:09:47,580: eval/sil : 0.2075
481
+ INFO:2025-06-01 20:13:19,349: --- epoch 61 ---
482
+ INFO:2025-06-01 20:13:19,349: train/loss : 1.0231
483
+ INFO:2025-06-01 20:13:19,349: train/f0 : 0.8988
484
+ INFO:2025-06-01 20:13:19,349: train/sil : 0.1243
485
+ INFO:2025-06-01 20:13:19,350: train/learning_rate: 0.0001
486
+ INFO:2025-06-01 20:13:19,350: eval/loss : 1.5889
487
+ INFO:2025-06-01 20:13:19,350: eval/f0 : 1.3730
488
+ INFO:2025-06-01 20:13:19,350: eval/sil : 0.2159
489
+ INFO:2025-06-01 20:16:50,725: --- epoch 62 ---
490
+ INFO:2025-06-01 20:16:50,726: train/loss : 1.0208
491
+ INFO:2025-06-01 20:16:50,726: train/f0 : 0.8970
492
+ INFO:2025-06-01 20:16:50,726: train/sil : 0.1238
493
+ INFO:2025-06-01 20:16:50,726: train/learning_rate: 0.0001
494
+ INFO:2025-06-01 20:16:50,726: eval/loss : 1.5667
495
+ INFO:2025-06-01 20:16:50,726: eval/f0 : 1.3523
496
+ INFO:2025-06-01 20:16:50,727: eval/sil : 0.2145
497
+ INFO:2025-06-01 20:20:23,749: --- epoch 63 ---
498
+ INFO:2025-06-01 20:20:23,750: train/loss : 1.0141
499
+ INFO:2025-06-01 20:20:23,750: train/f0 : 0.8910
500
+ INFO:2025-06-01 20:20:23,750: train/sil : 0.1231
501
+ INFO:2025-06-01 20:20:23,750: train/learning_rate: 0.0001
502
+ INFO:2025-06-01 20:20:23,750: eval/loss : 1.5848
503
+ INFO:2025-06-01 20:20:23,751: eval/f0 : 1.3679
504
+ INFO:2025-06-01 20:20:23,751: eval/sil : 0.2169
505
+ INFO:2025-06-01 20:23:55,185: --- epoch 64 ---
506
+ INFO:2025-06-01 20:23:55,185: train/loss : 1.0078
507
+ INFO:2025-06-01 20:23:55,185: train/f0 : 0.8856
508
+ INFO:2025-06-01 20:23:55,186: train/sil : 0.1221
509
+ INFO:2025-06-01 20:23:55,186: train/learning_rate: 0.0001
510
+ INFO:2025-06-01 20:23:55,186: eval/loss : 1.5571
511
+ INFO:2025-06-01 20:23:55,186: eval/f0 : 1.3411
512
+ INFO:2025-06-01 20:23:55,186: eval/sil : 0.2160
513
+ INFO:2025-06-01 20:27:26,913: --- epoch 65 ---
514
+ INFO:2025-06-01 20:27:26,914: train/loss : 1.0027
515
+ INFO:2025-06-01 20:27:26,914: train/f0 : 0.8815
516
+ INFO:2025-06-01 20:27:26,914: train/sil : 0.1212
517
+ INFO:2025-06-01 20:27:26,914: train/learning_rate: 0.0001
518
+ INFO:2025-06-01 20:27:26,914: eval/loss : 1.5837
519
+ INFO:2025-06-01 20:27:26,915: eval/f0 : 1.3687
520
+ INFO:2025-06-01 20:27:26,915: eval/sil : 0.2150
521
+ INFO:2025-06-01 20:30:58,156: --- epoch 66 ---
522
+ INFO:2025-06-01 20:30:58,156: train/loss : 0.9969
523
+ INFO:2025-06-01 20:30:58,156: train/f0 : 0.8763
524
+ INFO:2025-06-01 20:30:58,157: train/sil : 0.1207
525
+ INFO:2025-06-01 20:30:58,157: train/learning_rate: 0.0001
526
+ INFO:2025-06-01 20:30:58,157: eval/loss : 1.5586
527
+ INFO:2025-06-01 20:30:58,157: eval/f0 : 1.3453
528
+ INFO:2025-06-01 20:30:58,157: eval/sil : 0.2133
529
+ INFO:2025-06-01 20:34:30,658: --- epoch 67 ---
530
+ INFO:2025-06-01 20:34:30,658: train/loss : 0.9925
531
+ INFO:2025-06-01 20:34:30,659: train/f0 : 0.8726
532
+ INFO:2025-06-01 20:34:30,659: train/sil : 0.1199
533
+ INFO:2025-06-01 20:34:30,659: train/learning_rate: 0.0001
534
+ INFO:2025-06-01 20:34:30,660: eval/loss : 1.5443
535
+ INFO:2025-06-01 20:34:30,660: eval/f0 : 1.3298
536
+ INFO:2025-06-01 20:34:30,660: eval/sil : 0.2146
537
+ INFO:2025-06-01 20:38:03,364: --- epoch 68 ---
538
+ INFO:2025-06-01 20:38:03,364: train/loss : 0.9872
539
+ INFO:2025-06-01 20:38:03,365: train/f0 : 0.8680
540
+ INFO:2025-06-01 20:38:03,365: train/sil : 0.1192
541
+ INFO:2025-06-01 20:38:03,365: train/learning_rate: 0.0001
542
+ INFO:2025-06-01 20:38:03,365: eval/loss : 1.5761
543
+ INFO:2025-06-01 20:38:03,365: eval/f0 : 1.3552
544
+ INFO:2025-06-01 20:38:03,366: eval/sil : 0.2209
545
+ INFO:2025-06-01 20:41:38,174: --- epoch 69 ---
546
+ INFO:2025-06-01 20:41:38,174: train/loss : 0.9811
547
+ INFO:2025-06-01 20:41:38,174: train/f0 : 0.8629
548
+ INFO:2025-06-01 20:41:38,175: train/sil : 0.1183
549
+ INFO:2025-06-01 20:41:38,175: train/learning_rate: 0.0001
550
+ INFO:2025-06-01 20:41:38,175: eval/loss : 1.5786
551
+ INFO:2025-06-01 20:41:38,175: eval/f0 : 1.3560
552
+ INFO:2025-06-01 20:41:38,176: eval/sil : 0.2226
553
+ INFO:2025-06-01 20:45:09,164: --- epoch 70 ---
554
+ INFO:2025-06-01 20:45:09,165: train/loss : 0.9775
555
+ INFO:2025-06-01 20:45:09,165: train/f0 : 0.8599
556
+ INFO:2025-06-01 20:45:09,165: train/sil : 0.1176
557
+ INFO:2025-06-01 20:45:09,165: train/learning_rate: 0.0001
558
+ INFO:2025-06-01 20:45:09,165: eval/loss : 1.5624
559
+ INFO:2025-06-01 20:45:09,166: eval/f0 : 1.3397
560
+ INFO:2025-06-01 20:45:09,166: eval/sil : 0.2227
Configs/config.yml ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ log_dir: "Checkpoint"
2
+ save_freq: 10
3
+ device: "cuda"
4
+ epochs: 100
5
+ batch_size: 32
6
+ pretrained_model: ""
7
+ train_data: "Data/train_list.txt"
8
+ val_data: "Data/val_list.txt"
9
+ num_workers: 16
10
+
11
+
12
+ optimizer_params:
13
+ lr: 0.0003
14
+
15
+ loss_params:
16
+ lambda_f0: 0.1
17
+
Data/OOD_texts.txt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:97ee709b86948c4e708e99f3f64e645ad3950594caff4d6ee171eba598fb4930
3
+ size 15671597
Data/train_list.txt ADDED
The diff for this file is too large to render. See raw diff
 
Data/val_list.txt ADDED
The diff for this file is too large to render. See raw diff
 
LICENSE ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2022 Aaron (Yinghao) Li
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
README.md ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # JDC-PitchExtractor
2
+ This repo contains the training code for deep neural pitch extractor for Voice Conversion (VC) and TTS used in [StarGANv2-VC](https://github.com/yl4579/StarGANv2-VC) and [StyleTTS](https://github.com/yl4579/StyleTTS). This is the F0 network in StarGANv2-VC and pitch extractor in StyleTTS.
3
+
4
+ ## Pre-requisites
5
+ 1. Python >= 3.7
6
+ 2. Clone this repository:
7
+ ```bash
8
+ git clone https://github.com/yl4579/PitchExtractor.git
9
+ cd PitchExtractor
10
+ ```
11
+ 3. Install python requirements:
12
+ ```bash
13
+ pip install SoundFile torchaudio torch pyyaml click matplotlib librosa pyworld
14
+ ```
15
+ 4. Prepare your own dataset and put the `train_list.txt` and `val_list.txt` in the `Data` folder (see Training section for more details).
16
+
17
+ ## Training
18
+ ```bash
19
+ python train.py --config_path ./Configs/config.yml
20
+ ```
21
+ Please specify the training and validation data in `config.yml` file. The data list format needs to be `filename.wav|anything`, see [train_list.txt](https://github.com/yl4579/StarGANv2-VC/blob/main/Data/train_list.txt) as an example (a subset of VCTK). Note that you can put anything after the filename because the training labels are generated ad-hoc.
22
+
23
+ Checkpoints and Tensorboard logs will be saved at `log_dir`. To speed up training, you may want to make `batch_size` as large as your GPU RAM can take.
24
+
25
+ ### IMPORTANT: DATA FOLDER NEEDS WRITE PERMISSION
26
+ Since both `harvest` and `dio` are relatively slow, we do have to save the computed F0 ground truth for later use. In [meldataset.py](https://github.com/yl4579/PitchExtractor/blob/main/meldataset.py#L77-L89), it will write the computed F0 curve `_f0.npy` for each `.wav` file. This requires write permission in your data folder.
27
+
28
+ ### F0 Computation Details
29
+ In [meldataset.py](https://github.com/yl4579/PitchExtractor/blob/main/meldataset.py#L83-L87), the F0 curves are computated using [PyWorld](https://github.com/JeremyCCHsu/Python-Wrapper-for-World-Vocoder), one with `harvest` and another with `dio`. Both methods are acoustic-based and are unstable under certain conditions. `harvest` is faster but fails more than `dio`, so we first try `harvest`. When `harvest` fails (determined by number of frames with non-zero values), it will compute the ground truth F0 labels with `dio`. If `dio` fails, the computed F0 will have `NaN` and will be replaced with 0. This is supposed to occur only occasionally and should not affect training because these samples are treated as noises by the neural network and deep learning models are kwown to even benefit from slightly noisy datasets. However, if a lot of your samples have this problem (say > 5%), please remove them from the training set so that the model does not learn from the failed samples.
30
+
31
+ ### Data Augmentation
32
+ Data augmentation is not included in this code. For better voice conversion results, please add your own data augmentation in [meldataset.py](https://github.com/yl4579/PitchExtractor/blob/main/meldataset.py) with [audiomentations](https://github.com/iver56/audiomentations).
33
+
34
+ ## References
35
+ - [keums/melodyExtraction_JDC](https://github.com/keums/melodyExtraction_JDC)
36
+ - [kan-bayashi/ParallelWaveGAN](https://github.com/kan-bayashi/ParallelWaveGAN)
__pycache__/meldataset.cpython-312.pyc ADDED
Binary file (8.49 kB). View file
 
__pycache__/model.cpython-312.pyc ADDED
Binary file (8.49 kB). View file
 
__pycache__/optimizers.cpython-312.pyc ADDED
Binary file (5.25 kB). View file
 
__pycache__/trainer.cpython-312.pyc ADDED
Binary file (10.6 kB). View file
 
meldataset.py ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #coding: utf-8
2
+ """
3
+ TODO:
4
+ - make TestDataset
5
+ - separate transforms
6
+ """
7
+
8
+ import os
9
+ import os.path as osp
10
+ import time
11
+ import random
12
+ import numpy as np
13
+ import random
14
+ import soundfile as sf
15
+ import torch
16
+ from torch import nn
17
+ import torch.nn.functional as F
18
+ import torchaudio
19
+ from torch.utils.data import DataLoader
20
+
21
+ import pyworld as pw
22
+
23
+ import logging
24
+ logger = logging.getLogger(__name__)
25
+ logger.setLevel(logging.DEBUG)
26
+
27
+ np.random.seed(1)
28
+ random.seed(1)
29
+
30
+ SPECT_PARAMS = {
31
+ "n_fft": 2048,
32
+ "win_length": 1200,
33
+ "hop_length": 300
34
+ }
35
+ MEL_PARAMS = {
36
+ "n_mels": 80,
37
+ "n_fft": 2048,
38
+ "win_length": 1200,
39
+ "hop_length": 300
40
+ }
41
+
42
+ class MelDataset(torch.utils.data.Dataset):
43
+ def __init__(self,
44
+ data_list,
45
+ sr=24000,
46
+ data_augmentation=False,
47
+ validation=False,
48
+ verbose=True
49
+ ):
50
+
51
+ _data_list = [l[:-1].split('|') for l in data_list]
52
+ self.data_list = [d[0] for d in _data_list]
53
+
54
+ self.sr = sr
55
+ self.to_melspec = torchaudio.transforms.MelSpectrogram(**MEL_PARAMS)
56
+
57
+ self.mean, self.std = -4, 4
58
+ self.data_augmentation = data_augmentation and (not validation)
59
+ self.max_mel_length = 192
60
+ self.mean, self.std = -4, 4
61
+
62
+ self.verbose = verbose
63
+
64
+ # for silence detection
65
+ self.zero_value = -10 # what the zero value is
66
+ self.bad_F0 = 5 # if less than 5 frames are non-zero, it's a bad F0, try another algorithm
67
+
68
+ def __len__(self):
69
+ return len(self.data_list)
70
+
71
+ def path_to_mel_and_label(self, path):
72
+ wave_tensor = self._load_tensor(path)
73
+
74
+ # use pyworld to get F0
75
+ output_file = path + "_f0.npy"
76
+ # check if the file exists
77
+ if os.path.isfile(output_file): # if exists, load it directly
78
+ f0 = np.load(output_file)
79
+ else: # if not exist, create F0 file
80
+ if self.verbose:
81
+ print('Computing F0 for ' + path + '...')
82
+ x = wave_tensor.numpy().astype("double")
83
+ frame_period = MEL_PARAMS['hop_length'] * 1000 / self.sr
84
+ _f0, t = pw.harvest(x, self.sr, frame_period=frame_period)
85
+ if sum(_f0 != 0) < self.bad_F0: # this happens when the algorithm fails
86
+ _f0, t = pw.dio(x, self.sr, frame_period=frame_period) # if harvest fails, try dio
87
+ f0 = pw.stonemask(x, _f0, t, self.sr)
88
+ # save the f0 info for later use
89
+ np.save(output_file, f0)
90
+
91
+ f0 = torch.from_numpy(f0).float()
92
+
93
+ if self.data_augmentation:
94
+ random_scale = 0.5 + 0.5 * np.random.random()
95
+ wave_tensor = random_scale * wave_tensor
96
+
97
+ mel_tensor = self.to_melspec(wave_tensor)
98
+ mel_tensor = (torch.log(1e-5 + mel_tensor) - self.mean) / self.std
99
+ mel_length = mel_tensor.size(1)
100
+
101
+ f0_zero = (f0 == 0)
102
+
103
+ #######################################
104
+ # You may want your own silence labels here
105
+ # The more accurate the label, the better the resultss
106
+ is_silence = torch.zeros(f0.shape)
107
+ is_silence[f0_zero] = 1
108
+ #######################################
109
+
110
+ if mel_length > self.max_mel_length:
111
+ random_start = np.random.randint(0, mel_length - self.max_mel_length)
112
+ mel_tensor = mel_tensor[:, random_start:random_start + self.max_mel_length]
113
+ f0 = f0[random_start:random_start + self.max_mel_length]
114
+ is_silence = is_silence[random_start:random_start + self.max_mel_length]
115
+
116
+ if torch.any(torch.isnan(f0)): # failed
117
+ f0[torch.isnan(f0)] = self.zero_value # replace nan value with 0
118
+
119
+ return mel_tensor, f0, is_silence
120
+
121
+
122
+ def __getitem__(self, idx):
123
+ data = self.data_list[idx]
124
+ mel_tensor, f0, is_silence = self.path_to_mel_and_label(data)
125
+ return mel_tensor, f0, is_silence
126
+
127
+ def _load_tensor(self, data):
128
+ wave_path = data
129
+ wave, sr = sf.read(wave_path)
130
+ wave_tensor = torch.from_numpy(wave).float()
131
+ return wave_tensor
132
+
133
+ class Collater(object):
134
+ """
135
+ Args:
136
+ adaptive_batch_size (bool): if true, decrease batch size when long data comes.
137
+ """
138
+
139
+ def __init__(self, return_wave=False):
140
+ self.text_pad_index = 0
141
+ self.return_wave = return_wave
142
+ self.min_mel_length = 192
143
+ self.max_mel_length = 192
144
+ self.mel_length_step = 16
145
+ self.latent_dim = 16
146
+
147
+ def __call__(self, batch):
148
+ # batch[0] = wave, mel, text, f0, speakerid
149
+ batch_size = len(batch)
150
+ nmels = batch[0][0].size(0)
151
+ mels = torch.zeros((batch_size, nmels, self.max_mel_length)).float()
152
+ f0s = torch.zeros((batch_size, self.max_mel_length)).float()
153
+ is_silences = torch.zeros((batch_size, self.max_mel_length)).float()
154
+
155
+ for bid, (mel, f0, is_silence) in enumerate(batch):
156
+ mel_size = mel.size(1)
157
+ mels[bid, :, :mel_size] = mel
158
+ f0s[bid, :mel_size] = f0
159
+ is_silences[bid, :mel_size] = is_silence
160
+
161
+ if self.max_mel_length > self.min_mel_length:
162
+ random_slice = np.random.randint(
163
+ self.min_mel_length//self.mel_length_step,
164
+ 1+self.max_mel_length//self.mel_length_step) * self.mel_length_step + self.min_mel_length
165
+ mels = mels[:, :, :random_slice]
166
+ f0 = f0[:, :random_slice]
167
+
168
+ mels = mels.unsqueeze(1)
169
+ return mels, f0s, is_silences
170
+
171
+
172
+ def build_dataloader(path_list,
173
+ validation=False,
174
+ batch_size=4,
175
+ num_workers=1,
176
+ device='cpu',
177
+ collate_config={},
178
+ dataset_config={}):
179
+
180
+ dataset = MelDataset(path_list, validation=validation, **dataset_config)
181
+ collate_fn = Collater(**collate_config)
182
+
183
+ data_loader = DataLoader(dataset,
184
+ batch_size=batch_size,
185
+ shuffle=(not validation),
186
+ num_workers=num_workers,
187
+ drop_last=(not validation),
188
+ collate_fn=collate_fn,
189
+ pin_memory=(device != 'cpu'))
190
+
191
+ return data_loader
model.py ADDED
@@ -0,0 +1,175 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Implementation of model from:
3
+ Kum et al. - "Joint Detection and Classification of Singing Voice Melody Using
4
+ Convolutional Recurrent Neural Networks" (2019)
5
+ Link: https://www.semanticscholar.org/paper/Joint-Detection-and-Classification-of-Singing-Voice-Kum-Nam/60a2ad4c7db43bace75805054603747fcd062c0d
6
+ """
7
+ import torch
8
+ from torch import nn
9
+
10
+
11
+ class JDCNet(nn.Module):
12
+ """
13
+ Joint Detection and Classification Network model for singing voice melody.
14
+ """
15
+ def __init__(self, num_class=722, leaky_relu_slope=0.01):
16
+ super().__init__()
17
+ self.num_class = num_class
18
+
19
+ # input = (b, 1, 31, 513), b = batch size
20
+ self.conv_block = nn.Sequential(
21
+ nn.Conv2d(in_channels=1, out_channels=64, kernel_size=3, padding=1, bias=False), # out: (b, 64, 31, 513)
22
+ nn.BatchNorm2d(num_features=64),
23
+ nn.LeakyReLU(leaky_relu_slope, inplace=True),
24
+ nn.Conv2d(64, 64, 3, padding=1, bias=False), # (b, 64, 31, 513)
25
+ )
26
+
27
+ # res blocks
28
+ self.res_block1 = ResBlock(in_channels=64, out_channels=128) # (b, 128, 31, 128)
29
+ self.res_block2 = ResBlock(in_channels=128, out_channels=192) # (b, 192, 31, 32)
30
+ self.res_block3 = ResBlock(in_channels=192, out_channels=256) # (b, 256, 31, 8)
31
+
32
+ # pool block
33
+ self.pool_block = nn.Sequential(
34
+ nn.BatchNorm2d(num_features=256),
35
+ nn.LeakyReLU(leaky_relu_slope, inplace=True),
36
+ nn.MaxPool2d(kernel_size=(1, 4)), # (b, 256, 31, 2)
37
+ nn.Dropout(p=0.5),
38
+ )
39
+
40
+ # maxpool layers (for auxiliary network inputs)
41
+ # in = (b, 128, 31, 513) from conv_block, out = (b, 128, 31, 2)
42
+ self.maxpool1 = nn.MaxPool2d(kernel_size=(1, 40))
43
+ # in = (b, 128, 31, 128) from res_block1, out = (b, 128, 31, 2)
44
+ self.maxpool2 = nn.MaxPool2d(kernel_size=(1, 20))
45
+ # in = (b, 128, 31, 32) from res_block2, out = (b, 128, 31, 2)
46
+ self.maxpool3 = nn.MaxPool2d(kernel_size=(1, 10))
47
+
48
+ # in = (b, 640, 31, 2), out = (b, 256, 31, 2)
49
+ self.detector_conv = nn.Sequential(
50
+ nn.Conv2d(640, 256, 1, bias=False),
51
+ nn.BatchNorm2d(256),
52
+ nn.LeakyReLU(leaky_relu_slope, inplace=True),
53
+ nn.Dropout(p=0.5),
54
+ )
55
+
56
+ # input: (b, 31, 512) - resized from (b, 256, 31, 2)
57
+ self.bilstm_classifier = nn.LSTM(
58
+ input_size=512, hidden_size=256,
59
+ batch_first=True, dropout=0.3, bidirectional=True) # (b, 31, 512)
60
+
61
+ # input: (b, 31, 512) - resized from (b, 256, 31, 2)
62
+ self.bilstm_detector = nn.LSTM(
63
+ input_size=512, hidden_size=256,
64
+ batch_first=True, dropout=0.3, bidirectional=True) # (b, 31, 512)
65
+
66
+ # input: (b * 31, 512)
67
+ self.classifier = nn.Linear(in_features=512, out_features=self.num_class) # (b * 31, num_class)
68
+
69
+ # input: (b * 31, 512)
70
+ self.detector = nn.Linear(in_features=512, out_features=2) # (b * 31, 2) - binary classifier
71
+
72
+ # initialize weights
73
+ self.apply(self.init_weights)
74
+
75
+ def forward(self, x):
76
+ """
77
+ Returns:
78
+ classification_prediction, detection_prediction
79
+ sizes: (b, 31, 722), (b, 31, 2)
80
+ """
81
+ seq_len = x.shape[-2]
82
+ ###############################
83
+ # forward pass for classifier #
84
+ ###############################
85
+ convblock_out = self.conv_block(x)
86
+
87
+ resblock1_out = self.res_block1(convblock_out)
88
+ resblock2_out = self.res_block2(resblock1_out)
89
+ resblock3_out = self.res_block3(resblock2_out)
90
+ poolblock_out = self.pool_block(resblock3_out)
91
+
92
+ # (b, 256, 31, 2) => (b, 31, 256, 2) => (b, 31, 512)
93
+ classifier_out = poolblock_out.permute(0, 2, 1, 3).contiguous().view((-1, seq_len, 512))
94
+ classifier_out, _ = self.bilstm_classifier(classifier_out) # ignore the hidden states
95
+
96
+ classifier_out = classifier_out.contiguous().view((-1, 512)) # (b * 31, 512)
97
+ classifier_out = self.classifier(classifier_out)
98
+ classifier_out = classifier_out.view((-1, seq_len, self.num_class)) # (b, 31, num_class)
99
+
100
+ #############################
101
+ # forward pass for detector #
102
+ #############################
103
+ mp1_out = self.maxpool1(convblock_out)
104
+ mp2_out = self.maxpool2(resblock1_out)
105
+ mp3_out = self.maxpool3(resblock2_out)
106
+
107
+ # out = (b, 640, 31, 2)
108
+ concat_out = torch.cat((mp1_out, mp2_out, mp3_out, poolblock_out), dim=1)
109
+ detector_out = self.detector_conv(concat_out)
110
+
111
+ # (b, 256, 31, 2) => (b, 31, 256, 2) => (b, 31, 512)
112
+ detector_out = detector_out.permute(0, 2, 1, 3).contiguous().view((-1, seq_len, 512))
113
+ detector_out, _ = self.bilstm_detector(detector_out) # (b, 31, 512)
114
+
115
+ detector_out = detector_out.contiguous().view((-1, 512))
116
+ detector_out = self.detector(detector_out)
117
+ detector_out = detector_out.view((-1, seq_len, 2)).sum(axis=-1) # binary classifier - (b, 31, 2)
118
+
119
+ # sizes: (b, 31, 722), (b, 31, 2)
120
+ # classifier output consists of predicted pitch classes per frame
121
+ # detector output consists of: (isvoice, notvoice) estimates per frame
122
+ return classifier_out, detector_out
123
+
124
+ @staticmethod
125
+ def init_weights(m):
126
+ if isinstance(m, nn.Linear):
127
+ nn.init.kaiming_uniform_(m.weight)
128
+ if m.bias is not None:
129
+ nn.init.constant_(m.bias, 0)
130
+ elif isinstance(m, nn.Conv2d):
131
+ nn.init.xavier_normal_(m.weight)
132
+ elif isinstance(m, nn.LSTM) or isinstance(m, nn.LSTMCell):
133
+ for p in m.parameters():
134
+ if p.data is None:
135
+ continue
136
+
137
+ if len(p.shape) >= 2:
138
+ nn.init.orthogonal_(p.data)
139
+ else:
140
+ nn.init.normal_(p.data)
141
+
142
+
143
+ class ResBlock(nn.Module):
144
+ def __init__(self, in_channels: int, out_channels: int, leaky_relu_slope=0.01):
145
+ super().__init__()
146
+ self.downsample = in_channels != out_channels
147
+
148
+ # BN / LReLU / MaxPool layer before the conv layer - see Figure 1b in the paper
149
+ self.pre_conv = nn.Sequential(
150
+ nn.BatchNorm2d(num_features=in_channels),
151
+ nn.LeakyReLU(leaky_relu_slope, inplace=True),
152
+ nn.MaxPool2d(kernel_size=(1, 2)), # apply downsampling on the y axis only
153
+ )
154
+
155
+ # conv layers
156
+ self.conv = nn.Sequential(
157
+ nn.Conv2d(in_channels=in_channels, out_channels=out_channels,
158
+ kernel_size=3, padding=1, bias=False),
159
+ nn.BatchNorm2d(out_channels),
160
+ nn.LeakyReLU(leaky_relu_slope, inplace=True),
161
+ nn.Conv2d(out_channels, out_channels, 3, padding=1, bias=False),
162
+ )
163
+
164
+ # 1 x 1 convolution layer to match the feature dimensions
165
+ self.conv1by1 = None
166
+ if self.downsample:
167
+ self.conv1by1 = nn.Conv2d(in_channels, out_channels, 1, bias=False)
168
+
169
+ def forward(self, x):
170
+ x = self.pre_conv(x)
171
+ if self.downsample:
172
+ x = self.conv(x) + self.conv1by1(x)
173
+ else:
174
+ x = self.conv(x) + x
175
+ return x
optimizers.py ADDED
@@ -0,0 +1,86 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #coding:utf-8
2
+ import os, sys
3
+ import os.path as osp
4
+ import numpy as np
5
+ import torch
6
+ from torch import nn
7
+ from torch.optim import Optimizer
8
+ from functools import reduce
9
+ from torch.optim import AdamW
10
+
11
+ class MultiOptimizer:
12
+ def __init__(self, optimizers={}, schedulers={}):
13
+ self.optimizers = optimizers
14
+ self.schedulers = schedulers
15
+ self.keys = list(optimizers.keys())
16
+ self.param_groups = reduce(lambda x,y: x+y, [v.param_groups for v in self.optimizers.values()])
17
+
18
+ def state_dict(self):
19
+ state_dicts = [(key, self.optimizers[key].state_dict())\
20
+ for key in self.keys]
21
+ return state_dicts
22
+
23
+ def load_state_dict(self, state_dict):
24
+ for key, val in state_dict:
25
+ try:
26
+ self.optimizers[key].load_state_dict(val)
27
+ except:
28
+ print("Unloaded %s" % key)
29
+
30
+
31
+ def step(self, key=None):
32
+ if key is not None:
33
+ self.optimizers[key].step()
34
+ else:
35
+ _ = [self.optimizers[key].step() for key in self.keys]
36
+
37
+ def zero_grad(self, key=None):
38
+ if key is not None:
39
+ self.optimizers[key].zero_grad()
40
+ else:
41
+ _ = [self.optimizers[key].zero_grad() for key in self.keys]
42
+
43
+ def scheduler(self, *args, key=None):
44
+ if key is not None:
45
+ self.schedulers[key].step(*args)
46
+ else:
47
+ _ = [self.schedulers[key].step(*args) for key in self.keys]
48
+
49
+
50
+ def build_optimizer(parameters):
51
+ optimizer, scheduler = _define_optimizer(parameters)
52
+ return optimizer, scheduler
53
+
54
+ def _define_optimizer(params):
55
+ optimizer_params = params['optimizer_params']
56
+ sch_params = params['scheduler_params']
57
+ optimizer = AdamW(
58
+ params['params'],
59
+ lr=optimizer_params.get('lr', 1e-4),
60
+ weight_decay=optimizer_params.get('weight_decay', 5e-4),
61
+ betas=(0.9, 0.98),
62
+ eps=1e-9)
63
+ scheduler = _define_scheduler(optimizer, sch_params)
64
+ return optimizer, scheduler
65
+
66
+ def _define_scheduler(optimizer, params):
67
+ print(params)
68
+ scheduler = torch.optim.lr_scheduler.OneCycleLR(
69
+ optimizer,
70
+ max_lr=params.get('max_lr', 5e-4),
71
+ epochs=params.get('epochs', 200),
72
+ steps_per_epoch=params.get('steps_per_epoch', 1000),
73
+ pct_start=params.get('pct_start', 0.0),
74
+ final_div_factor=5)
75
+
76
+ return scheduler
77
+
78
+ def build_multi_optimizer(parameters_dict, scheduler_params):
79
+ optim = dict([(key, AdamW(params, lr=1e-4, weight_decay=1e-6, betas=(0.9, 0.98), eps=1e-9))
80
+ for key, params in parameters_dict.items()])
81
+
82
+ schedulers = dict([(key, _define_scheduler(opt, scheduler_params)) \
83
+ for key, opt in optim.items()])
84
+
85
+ multi_optim = MultiOptimizer(optim, schedulers)
86
+ return multi_optim
train.py ADDED
@@ -0,0 +1,144 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from model import JDCNet
2
+ from meldataset import build_dataloader
3
+ from optimizers import build_optimizer
4
+ from trainer import Trainer
5
+
6
+ import time
7
+ import os
8
+ import os.path as osp
9
+ import re
10
+ import sys
11
+ import yaml
12
+ import shutil
13
+ import numpy as np
14
+ import torch
15
+ import torch.nn as nn
16
+ from torch.utils.tensorboard import SummaryWriter
17
+ import click
18
+ from tqdm import tqdm
19
+
20
+ import logging
21
+ from logging import StreamHandler
22
+ logger = logging.getLogger(__name__)
23
+ logger.setLevel(logging.DEBUG)
24
+ handler = StreamHandler()
25
+ handler.setLevel(logging.DEBUG)
26
+ logger.addHandler(handler)
27
+
28
+ torch.backends.cudnn.benchmark = True
29
+
30
+ def get_data_path_list(train_path=None, val_path=None):
31
+ if train_path is None:
32
+ train_path = "Data/train_list.txt"
33
+ if val_path is None:
34
+ val_path = "Data/val_list.txt"
35
+
36
+ with open(train_path, 'r') as f:
37
+ train_list = f.readlines()
38
+ with open(val_path, 'r') as f:
39
+ val_list = f.readlines()
40
+
41
+ # train_list = train_list[-500:]
42
+ # val_list = train_list[:500]
43
+ return train_list, val_list
44
+
45
+ @click.command()
46
+ @click.option('-p', '--config_path', default='./Configs/config.yml', type=str)
47
+ def main(config_path):
48
+ config = yaml.safe_load(open(config_path))
49
+ log_dir = config['log_dir']
50
+ if not osp.exists(log_dir): os.mkdir(log_dir)
51
+ shutil.copy(config_path, osp.join(log_dir, osp.basename(config_path)))
52
+
53
+ writer = SummaryWriter(log_dir + "/tensorboard")
54
+
55
+ # write logs
56
+ file_handler = logging.FileHandler(osp.join(log_dir, 'train.log'))
57
+ file_handler.setLevel(logging.DEBUG)
58
+ file_handler.setFormatter(logging.Formatter('%(levelname)s:%(asctime)s: %(message)s'))
59
+ logger.addHandler(file_handler)
60
+
61
+ batch_size = config.get('batch_size', 32)
62
+ device = config.get('device', 'cpu')
63
+ epochs = config.get('epochs', 100)
64
+ save_freq = config.get('save_freq', 10)
65
+ train_path = config.get('train_data', None)
66
+ val_path = config.get('val_data', None)
67
+ num_workers = config.get('num_workers', 8)
68
+
69
+ train_list, val_list = get_data_path_list(train_path, val_path)
70
+
71
+ train_dataloader = build_dataloader(train_list,
72
+ batch_size=batch_size,
73
+ num_workers=num_workers,
74
+ dataset_config=config.get('dataset_params', {}),
75
+ device=device)
76
+
77
+ val_dataloader = build_dataloader(val_list,
78
+ batch_size=batch_size,
79
+ validation=True,
80
+ num_workers=num_workers // 2,
81
+ device=device,
82
+ dataset_config=config.get('dataset_params', {}))
83
+
84
+ # define model
85
+ model = JDCNet(num_class=1) # num_class = 1 means regression
86
+
87
+ scheduler_params = {
88
+ "max_lr": float(config['optimizer_params'].get('lr', 5e-4)),
89
+ "pct_start": float(config['optimizer_params'].get('pct_start', 0.0)),
90
+ "epochs": epochs,
91
+ "steps_per_epoch": len(train_dataloader),
92
+ }
93
+
94
+ model.to(device)
95
+ optimizer, scheduler = build_optimizer(
96
+ {"params": model.parameters(), "optimizer_params":{}, "scheduler_params": scheduler_params})
97
+
98
+ criterion = {'l1': nn.SmoothL1Loss(), # F0 loss (regression)
99
+ 'ce': nn.BCEWithLogitsLoss() # silence loss (binary classification)
100
+ }
101
+
102
+ loss_config = config['loss_params']
103
+
104
+ trainer = Trainer(model=model,
105
+ criterion=criterion,
106
+ optimizer=optimizer,
107
+ scheduler=scheduler,
108
+ device=device,
109
+ train_dataloader=train_dataloader,
110
+ val_dataloader=val_dataloader,
111
+ loss_config=loss_config,
112
+ logger=logger)
113
+
114
+ if config.get('pretrained_model', '') != '':
115
+ trainer.load_checkpoint(config['pretrained_model'],
116
+ load_only_params=config.get('load_only_params', True))
117
+
118
+ # compute all F0 for training and validation data
119
+ print('Checking if all F0 data is computed...')
120
+ for _ in enumerate(train_dataloader):
121
+ continue
122
+ for _ in enumerate(val_dataloader):
123
+ continue
124
+ print('All F0 data is computed.')
125
+
126
+ for epoch in range(1, epochs+1):
127
+ train_results = trainer._train_epoch()
128
+ eval_results = trainer._eval_epoch()
129
+ results = train_results.copy()
130
+ results.update(eval_results)
131
+ logger.info('--- epoch %d ---' % epoch)
132
+ for key, value in results.items():
133
+ if isinstance(value, float):
134
+ logger.info('%-15s: %.4f' % (key, value))
135
+ writer.add_scalar(key, value, epoch)
136
+ else:
137
+ writer.add_figure(key, (v), epoch)
138
+ if (epoch % save_freq) == 0:
139
+ trainer.save_checkpoint(osp.join(log_dir, 'epoch_%05d.pth' % epoch))
140
+
141
+ return 0
142
+
143
+ if __name__=="__main__":
144
+ main()
trainer.py ADDED
@@ -0,0 +1,186 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+
3
+ import os
4
+ import os.path as osp
5
+ import sys
6
+ import time
7
+ from collections import defaultdict
8
+
9
+ import numpy as np
10
+ import torch
11
+ from torch import nn
12
+ from PIL import Image
13
+ from tqdm import tqdm
14
+
15
+ import matplotlib.pyplot as plt
16
+
17
+ import logging
18
+ logger = logging.getLogger(__name__)
19
+ logger.setLevel(logging.DEBUG)
20
+
21
+ class Trainer(object):
22
+ def __init__(self,
23
+ model=None,
24
+ criterion=None,
25
+ optimizer=None,
26
+ scheduler=None,
27
+ config={},
28
+ loss_config={},
29
+ device=torch.device("cpu"),
30
+ logger=logger,
31
+ train_dataloader=None,
32
+ val_dataloader=None,
33
+ initial_steps=0,
34
+ initial_epochs=0):
35
+
36
+ self.steps = initial_steps
37
+ self.epochs = initial_epochs
38
+ self.model = model
39
+ self.criterion = criterion
40
+ self.optimizer = optimizer
41
+ self.scheduler = scheduler
42
+ self.train_dataloader = train_dataloader
43
+ self.val_dataloader = val_dataloader
44
+ self.config = config
45
+ self.loss_config = loss_config
46
+ self.device = device
47
+ self.finish_train = False
48
+ self.logger = logger
49
+ self.fp16_run = False
50
+
51
+ def save_checkpoint(self, checkpoint_path):
52
+ """Save checkpoint.
53
+ Args:
54
+ checkpoint_path (str): Checkpoint path to be saved.
55
+ """
56
+ state_dict = {
57
+ "optimizer": self.optimizer.state_dict(),
58
+ "scheduler": self.scheduler.state_dict(),
59
+ "steps": self.steps,
60
+ "epochs": self.epochs,
61
+ }
62
+ state_dict["model"] = self.model.state_dict()
63
+
64
+ if not os.path.exists(os.path.dirname(checkpoint_path)):
65
+ os.makedirs(os.path.dirname(checkpoint_path))
66
+ torch.save(state_dict, checkpoint_path)
67
+
68
+ def load_checkpoint(self, checkpoint_path, load_only_params=False):
69
+ """Load checkpoint.
70
+ Args:
71
+ checkpoint_path (str): Checkpoint path to be loaded.
72
+ load_only_params (bool): Whether to load only model parameters.
73
+ """
74
+ state_dict = torch.load(checkpoint_path, map_location="cpu")
75
+ self._load(state_dict["model"], self.model)
76
+
77
+ if not load_only_params:
78
+ self.steps = state_dict["steps"]
79
+ self.epochs = state_dict["epochs"]
80
+ self.optimizer.load_state_dict(state_dict["optimizer"])
81
+
82
+ # overwrite schedular argument parameters
83
+ state_dict["scheduler"].update(**self.config.get("scheduler_params", {}))
84
+ self.scheduler.load_state_dict(state_dict["scheduler"])
85
+
86
+ def _load(self, states, model, force_load=True):
87
+ model_states = model.state_dict()
88
+ for key, val in states.items():
89
+ try:
90
+ if key not in model_states:
91
+ continue
92
+ if isinstance(val, nn.Parameter):
93
+ val = val.data
94
+
95
+ if val.shape != model_states[key].shape:
96
+ self.logger.info("%s does not have same shape" % key)
97
+ print(val.shape, model_states[key].shape)
98
+ if not force_load:
99
+ continue
100
+
101
+ min_shape = np.minimum(np.array(val.shape), np.array(model_states[key].shape))
102
+ slices = [slice(0, min_index) for min_index in min_shape]
103
+ model_states[key][slices].copy_(val[slices])
104
+ else:
105
+ model_states[key].copy_(val)
106
+ except:
107
+ self.logger.info("not exist :%s" % key)
108
+ print("not exist ", key)
109
+
110
+ @staticmethod
111
+ def get_gradient_norm(model):
112
+ total_norm = 0
113
+ for p in model.parameters():
114
+ param_norm = p.grad.data.norm(2)
115
+ total_norm += param_norm.item() ** 2
116
+
117
+ total_norm = np.sqrt(total_norm)
118
+ return total_norm
119
+
120
+ @staticmethod
121
+ def length_to_mask(lengths):
122
+ mask = torch.arange(lengths.max()).unsqueeze(0).expand(lengths.shape[0], -1).type_as(lengths)
123
+ mask = torch.gt(mask+1, lengths.unsqueeze(1))
124
+ return mask
125
+
126
+ def _get_lr(self):
127
+ for param_group in self.optimizer.param_groups:
128
+ lr = param_group['lr']
129
+ break
130
+ return lr
131
+
132
+ def run(self, batch):
133
+ self.optimizer.zero_grad()
134
+ batch = [b.to(self.device) for b in batch]
135
+
136
+ x, f0, sil = batch
137
+ f0_pred, sil_pred = self.model(x.transpose(-1, -2))
138
+
139
+ loss_f0 = self.loss_config['lambda_f0'] * self.criterion['l1'](f0_pred.squeeze(), f0)
140
+ loss_sil = self.criterion['ce'](sil_pred, sil)
141
+ loss = loss_f0 + loss_sil
142
+
143
+ loss.backward()
144
+ self.optimizer.step()
145
+ self.scheduler.step()
146
+
147
+ return {'loss': loss.item(),
148
+ 'f0': loss_f0.item(),
149
+ 'sil': loss_sil.item()}
150
+
151
+ def _train_epoch(self):
152
+ self.epochs += 1
153
+ train_losses = defaultdict(list)
154
+ self.model.train()
155
+ for train_steps_per_epoch, batch in enumerate(tqdm(self.train_dataloader, desc="[train]"), 1):
156
+ losses = self.run(batch)
157
+ for key, value in losses.items():
158
+ train_losses["train/%s" % key].append(value)
159
+
160
+ train_losses = {key: np.mean(value) for key, value in train_losses.items()}
161
+ train_losses['train/learning_rate'] = self._get_lr()
162
+ return train_losses
163
+
164
+ @torch.no_grad()
165
+ def _eval_epoch(self):
166
+ self.model.eval()
167
+ eval_losses = defaultdict(list)
168
+ eval_images = defaultdict(list)
169
+ for eval_steps_per_epoch, batch in enumerate(tqdm(self.val_dataloader, desc="[eval]"), 1):
170
+ batch = [b.to(self.device) for b in batch]
171
+ x, f0, sil = batch
172
+
173
+ f0_pred, sil_pred = self.model(x.transpose(-1, -2))
174
+
175
+ loss_f0 = self.loss_config['lambda_f0'] * self.criterion['l1'](f0_pred.squeeze(), f0)
176
+ loss_sil = self.criterion['ce'](sil_pred, sil)
177
+ loss = loss_f0 + loss_sil
178
+
179
+
180
+ eval_losses["eval/loss"].append(loss.item())
181
+ eval_losses["eval/f0"].append(loss_f0.item())
182
+ eval_losses["eval/sil"].append(loss_sil.item())
183
+
184
+ eval_losses = {key: np.mean(value) for key, value in eval_losses.items()}
185
+ eval_losses.update(eval_images)
186
+ return eval_losses