Hugging Face's logo

Eddycrack864
/
audio-separator-models

ONNX
Model card Files
xet
Community
Eddycrack864 commited on
Commit
87699a3
·
verified ·
1 Parent(s): beeb78b

Upload 28 files

Browse files
Files changed (28) hide show
  1. roformers/config_aspiration_mel_band_roformer.yaml +76 -0
  2. roformers/config_bs_roformer_instrumental_resurrection_unwa.yaml +135 -0
  3. roformers/config_bs_roformer_vocals_gabox.yaml +133 -0
  4. roformers/config_bs_roformer_vocals_resurrection_unwa.yaml +135 -0
  5. roformers/config_bs_roformer_vocals_revive_unwa.yaml +134 -0
  6. roformers/config_chorus_male_female_bs_roformer.yaml +125 -0
  7. roformers/config_dereverb-echo_mel_band_roformer.yaml +76 -0
  8. roformers/config_dereverb-echo_mel_band_roformer_sdr_13.4843_v2.yaml +64 -0
  9. roformers/config_dereverb_echo_mel_band_roformer_v2.yaml +64 -0
  10. roformers/config_mel_band_roformer_bleed_suppressor_v1.yaml +51 -0
  11. roformers/config_mel_band_roformer_instrumental_becruily.yaml +72 -0
  12. roformers/config_mel_band_roformer_karaoke_becruily.yaml +72 -0
  13. roformers/config_mel_band_roformer_karaoke_gabox.yaml +72 -0
  14. roformers/config_mel_band_roformer_kim_ft_unwa.yaml +72 -0
  15. roformers/config_mel_band_roformer_vocal_fullness_aname.yaml +54 -0
  16. roformers/config_mel_band_roformer_vocals_becruily.yaml +72 -0
  17. roformers/config_mel_band_roformer_vocals_gabox.yaml +51 -0
  18. roformers/config_melband_roformer_big_beta5e.yaml +51 -0
  19. roformers/config_melbandroformer_big_beta4.yaml +51 -0
  20. roformers/config_melbandroformer_big_beta6.yaml +72 -0
  21. roformers/config_melbandroformer_big_beta6x.yaml +72 -0
  22. roformers/config_melbandroformer_inst.yaml +51 -0
  23. roformers/config_melbandroformer_inst_v2.yaml +51 -0
  24. roformers/config_melbandroformer_instvoc_duality.yaml +51 -0
  25. roformers/config_vocals_mel_band_roformer_big_v1_ft.yaml +51 -0
  26. roformers/config_vocals_mel_band_roformer_ft.yaml +72 -0
  27. roformers/dereverb_mel_band_roformer_anvuew.yaml +76 -0
  28. roformers/vocals_mel_band_roformer.yaml +50 -0
roformers/config_aspiration_mel_band_roformer.yaml ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 801 # don't work (use in model)
5
+ hop_length: 441 # don't work (use in model)
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 256
13
+ depth: 8
14
+ stereo: true
15
+ num_stems: 2
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0.1
22
+ ff_dropout: 0.1
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 8
44
+ grad_clip: 0
45
+ instruments:
46
+ - aspiration
47
+ - other
48
+ lr: 4.0e-05
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: null
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ q: 0.95
55
+ coarse_loss_clip: true
56
+ ema_momentum: 0.999
57
+ optimizer: adam
58
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
59
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
60
+
61
+ augmentations:
62
+ enable: true # enable or disable all augmentations (to fast disable if needed)
63
+ loudness: true # randomly change loudness of each stem on the range (loudness_min; loudness_max)
64
+ loudness_min: 0.5
65
+ loudness_max: 1.5
66
+ mixup: false # mix several stems of same type with some probability (only works for dataset types: 1, 2, 3)
67
+ mixup_probs: !!python/tuple # 2 additional stems of the same type (1st with prob 0.2, 2nd with prob 0.02)
68
+ - 0.2
69
+ - 0.02
70
+ mixup_loudness_min: 0.5
71
+ mixup_loudness_max: 1.5
72
+
73
+ inference:
74
+ batch_size: 4
75
+ dim_t: 801
76
+ num_overlap: 2
roformers/config_bs_roformer_instrumental_resurrection_unwa.yaml ADDED
@@ -0,0 +1,135 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 749259
3
+ dim_f: 1024
4
+ dim_t: 1700 # don't work (use in model)
5
+ hop_length: 441 # don't work (use in model)
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 256
13
+ depth: 12
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ linear_transformer_depth: 0
19
+ freqs_per_bands: !!python/tuple
20
+ - 2
21
+ - 2
22
+ - 2
23
+ - 2
24
+ - 2
25
+ - 2
26
+ - 2
27
+ - 2
28
+ - 2
29
+ - 2
30
+ - 2
31
+ - 2
32
+ - 2
33
+ - 2
34
+ - 2
35
+ - 2
36
+ - 2
37
+ - 2
38
+ - 2
39
+ - 2
40
+ - 2
41
+ - 2
42
+ - 2
43
+ - 2
44
+ - 4
45
+ - 4
46
+ - 4
47
+ - 4
48
+ - 4
49
+ - 4
50
+ - 4
51
+ - 4
52
+ - 4
53
+ - 4
54
+ - 4
55
+ - 4
56
+ - 12
57
+ - 12
58
+ - 12
59
+ - 12
60
+ - 12
61
+ - 12
62
+ - 12
63
+ - 12
64
+ - 24
65
+ - 24
66
+ - 24
67
+ - 24
68
+ - 24
69
+ - 24
70
+ - 24
71
+ - 24
72
+ - 48
73
+ - 48
74
+ - 48
75
+ - 48
76
+ - 48
77
+ - 48
78
+ - 48
79
+ - 48
80
+ - 128
81
+ - 129
82
+ dim_head: 64
83
+ heads: 8
84
+ attn_dropout: 0.
85
+ ff_dropout: 0.
86
+ flash_attn: true
87
+ dim_freqs_in: 1025
88
+ stft_n_fft: 2048
89
+ stft_hop_length: 441
90
+ stft_win_length: 2048
91
+ stft_normalized: false
92
+ mask_estimator_depth: 2
93
+ multi_stft_resolution_loss_weight: 1.0
94
+ multi_stft_resolutions_window_sizes: !!python/tuple
95
+ - 4096
96
+ - 2048
97
+ - 1024
98
+ - 512
99
+ - 256
100
+ multi_stft_hop_size: 147
101
+ multi_stft_normalized: False
102
+
103
+ training:
104
+ batch_size: 2
105
+ gradient_accumulation_steps: 1
106
+ grad_clip: 0
107
+ instruments: ['vocals', 'other']
108
+ patience: 3
109
+ reduce_factor: 0.95
110
+ target_instrument: other
111
+ num_epochs: 1000
112
+ num_steps: 1000
113
+ augmentation: false # enable augmentations by audiomentations and pedalboard
114
+ augmentation_type: simple1
115
+ use_mp3_compress: false # Deprecated
116
+ augmentation_mix: true # Mix several stems of the same type with some probability
117
+ augmentation_loudness: true # randomly change loudness of each stem
118
+ augmentation_loudness_type: 1 # Type 1 or 2
119
+ augmentation_loudness_min: 0.5
120
+ augmentation_loudness_max: 1.5
121
+ q: 0.95
122
+ coarse_loss_clip: true
123
+ ema_momentum: 0.999
124
+ # optimizer: prodigy
125
+ optimizer: adam
126
+ # lr: 1.0
127
+ lr: 1.0e-5
128
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
129
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
130
+
131
+ inference:
132
+ batch_size: 2
133
+ dim_t: 1700
134
+ num_overlap: 2
135
+ normalize: false
roformers/config_bs_roformer_vocals_gabox.yaml ADDED
@@ -0,0 +1,133 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 801 # don't work (use in model)
5
+ hop_length: 441 # don't work (use in model)
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.001
10
+
11
+ model:
12
+ dim: 512
13
+ depth: 12
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ freqs_per_bands: !!python/tuple
19
+ - 2
20
+ - 2
21
+ - 2
22
+ - 2
23
+ - 2
24
+ - 2
25
+ - 2
26
+ - 2
27
+ - 2
28
+ - 2
29
+ - 2
30
+ - 2
31
+ - 2
32
+ - 2
33
+ - 2
34
+ - 2
35
+ - 2
36
+ - 2
37
+ - 2
38
+ - 2
39
+ - 2
40
+ - 2
41
+ - 2
42
+ - 2
43
+ - 4
44
+ - 4
45
+ - 4
46
+ - 4
47
+ - 4
48
+ - 4
49
+ - 4
50
+ - 4
51
+ - 4
52
+ - 4
53
+ - 4
54
+ - 4
55
+ - 12
56
+ - 12
57
+ - 12
58
+ - 12
59
+ - 12
60
+ - 12
61
+ - 12
62
+ - 12
63
+ - 24
64
+ - 24
65
+ - 24
66
+ - 24
67
+ - 24
68
+ - 24
69
+ - 24
70
+ - 24
71
+ - 48
72
+ - 48
73
+ - 48
74
+ - 48
75
+ - 48
76
+ - 48
77
+ - 48
78
+ - 48
79
+ - 128
80
+ - 129
81
+ dim_head: 64
82
+ heads: 8
83
+ attn_dropout: 0.1
84
+ ff_dropout: 0.1
85
+ flash_attn: true
86
+ dim_freqs_in: 1025
87
+ stft_n_fft: 2048
88
+ stft_hop_length: 441
89
+ stft_win_length: 2048
90
+ stft_normalized: false
91
+ mask_estimator_depth: 2
92
+ multi_stft_resolution_loss_weight: 1.0
93
+ multi_stft_resolutions_window_sizes: !!python/tuple
94
+ - 4096
95
+ - 2048
96
+ - 1024
97
+ - 512
98
+ - 256
99
+ multi_stft_hop_size: 147
100
+ multi_stft_normalized: False
101
+
102
+ training:
103
+ batch_size: 16
104
+ gradient_accumulation_steps: 1
105
+ grad_clip: 0
106
+ instruments:
107
+ - Vocals
108
+ - Instrumental
109
+ lr: 5.0e-05
110
+ patience: 2
111
+ reduce_factor: 0.95
112
+ target_instrument: Vocals
113
+ num_epochs: 1000
114
+ num_steps: 1000
115
+ augmentation: false # enable augmentations by audiomentations and pedalboard
116
+ augmentation_type: simple1
117
+ use_mp3_compress: false # Deprecated
118
+ augmentation_mix: true # Mix several stems of the same type with some probability
119
+ augmentation_loudness: true # randomly change loudness of each stem
120
+ augmentation_loudness_type: 1 # Type 1 or 2
121
+ augmentation_loudness_min: 0.5
122
+ augmentation_loudness_max: 1.5
123
+ q: 0.95
124
+ coarse_loss_clip: true
125
+ ema_momentum: 0.999
126
+ optimizer: adam
127
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
128
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
129
+
130
+ inference:
131
+ batch_size: 1
132
+ dim_t: 801
133
+ num_overlap: 4
roformers/config_bs_roformer_vocals_resurrection_unwa.yaml ADDED
@@ -0,0 +1,135 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 785920
3
+ dim_f: 1024
4
+ dim_t: 1536 # don't work (use in model)
5
+ hop_length: 441 # don't work (use in model)
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 256
13
+ depth: 12
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ linear_transformer_depth: 0
19
+ freqs_per_bands: !!python/tuple
20
+ - 2
21
+ - 2
22
+ - 2
23
+ - 2
24
+ - 2
25
+ - 2
26
+ - 2
27
+ - 2
28
+ - 2
29
+ - 2
30
+ - 2
31
+ - 2
32
+ - 2
33
+ - 2
34
+ - 2
35
+ - 2
36
+ - 2
37
+ - 2
38
+ - 2
39
+ - 2
40
+ - 2
41
+ - 2
42
+ - 2
43
+ - 2
44
+ - 4
45
+ - 4
46
+ - 4
47
+ - 4
48
+ - 4
49
+ - 4
50
+ - 4
51
+ - 4
52
+ - 4
53
+ - 4
54
+ - 4
55
+ - 4
56
+ - 12
57
+ - 12
58
+ - 12
59
+ - 12
60
+ - 12
61
+ - 12
62
+ - 12
63
+ - 12
64
+ - 24
65
+ - 24
66
+ - 24
67
+ - 24
68
+ - 24
69
+ - 24
70
+ - 24
71
+ - 24
72
+ - 48
73
+ - 48
74
+ - 48
75
+ - 48
76
+ - 48
77
+ - 48
78
+ - 48
79
+ - 48
80
+ - 128
81
+ - 129
82
+ dim_head: 64
83
+ heads: 8
84
+ attn_dropout: 0.
85
+ ff_dropout: 0.
86
+ flash_attn: true
87
+ dim_freqs_in: 1025
88
+ stft_n_fft: 2048
89
+ stft_hop_length: 441
90
+ stft_win_length: 2048
91
+ stft_normalized: false
92
+ mask_estimator_depth: 2
93
+ multi_stft_resolution_loss_weight: 1.0
94
+ multi_stft_resolutions_window_sizes: !!python/tuple
95
+ - 4096
96
+ - 2048
97
+ - 1024
98
+ - 512
99
+ - 256
100
+ multi_stft_hop_size: 147
101
+ multi_stft_normalized: False
102
+
103
+ training:
104
+ batch_size: 2
105
+ gradient_accumulation_steps: 1
106
+ grad_clip: 0
107
+ instruments: ['vocals', 'other']
108
+ patience: 3
109
+ reduce_factor: 0.95
110
+ target_instrument: vocals
111
+ num_epochs: 1000
112
+ num_steps: 1000
113
+ augmentation: false # enable augmentations by audiomentations and pedalboard
114
+ augmentation_type: simple1
115
+ use_mp3_compress: false # Deprecated
116
+ augmentation_mix: true # Mix several stems of the same type with some probability
117
+ augmentation_loudness: true # randomly change loudness of each stem
118
+ augmentation_loudness_type: 1 # Type 1 or 2
119
+ augmentation_loudness_min: 0.5
120
+ augmentation_loudness_max: 1.5
121
+ q: 0.95
122
+ coarse_loss_clip: true
123
+ ema_momentum: 0.999
124
+ # optimizer: prodigy
125
+ optimizer: adam
126
+ # lr: 1.0
127
+ lr: 1.0e-5
128
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
129
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
130
+
131
+ inference:
132
+ batch_size: 2
133
+ dim_t: 1536
134
+ num_overlap: 2
135
+ normalize: false
roformers/config_bs_roformer_vocals_revive_unwa.yaml ADDED
@@ -0,0 +1,134 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100 #352800 #485100
3
+ dim_f: 1024
4
+ dim_t: 1101
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.
10
+
11
+ model:
12
+ dim: 512
13
+ depth: 12
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ linear_transformer_depth: 0
19
+ freqs_per_bands: !!python/tuple
20
+ - 2
21
+ - 2
22
+ - 2
23
+ - 2
24
+ - 2
25
+ - 2
26
+ - 2
27
+ - 2
28
+ - 2
29
+ - 2
30
+ - 2
31
+ - 2
32
+ - 2
33
+ - 2
34
+ - 2
35
+ - 2
36
+ - 2
37
+ - 2
38
+ - 2
39
+ - 2
40
+ - 2
41
+ - 2
42
+ - 2
43
+ - 2
44
+ - 4
45
+ - 4
46
+ - 4
47
+ - 4
48
+ - 4
49
+ - 4
50
+ - 4
51
+ - 4
52
+ - 4
53
+ - 4
54
+ - 4
55
+ - 4
56
+ - 12
57
+ - 12
58
+ - 12
59
+ - 12
60
+ - 12
61
+ - 12
62
+ - 12
63
+ - 12
64
+ - 24
65
+ - 24
66
+ - 24
67
+ - 24
68
+ - 24
69
+ - 24
70
+ - 24
71
+ - 24
72
+ - 48
73
+ - 48
74
+ - 48
75
+ - 48
76
+ - 48
77
+ - 48
78
+ - 48
79
+ - 48
80
+ - 128
81
+ - 129
82
+ dim_head: 64
83
+ heads: 8
84
+ attn_dropout: 0.
85
+ ff_dropout: 0.
86
+ flash_attn: true
87
+ dim_freqs_in: 1025
88
+ stft_n_fft: 2048
89
+ stft_hop_length: 441
90
+ stft_win_length: 2048
91
+ stft_normalized: false
92
+ mask_estimator_depth: 2
93
+ multi_stft_resolution_loss_weight: 1.0
94
+ multi_stft_resolutions_window_sizes: !!python/tuple
95
+ - 4096
96
+ - 2048
97
+ - 1024
98
+ - 512
99
+ - 256
100
+ multi_stft_hop_size: 147
101
+ multi_stft_normalized: False
102
+
103
+ training:
104
+ batch_size: 1
105
+ gradient_accumulation_steps: 1
106
+ grad_clip: 0
107
+ instruments:
108
+ - vocals
109
+ - other
110
+ lr: 1.0e-05
111
+ patience: 2
112
+ reduce_factor: 0.95
113
+ target_instrument: vocals
114
+ num_epochs: 1000
115
+ num_steps: 1000
116
+ augmentation: false # enable augmentations by audiomentations and pedalboard
117
+ augmentation_type: null
118
+ use_mp3_compress: false # Deprecated
119
+ augmentation_mix: false # Mix several stems of the same type with some probability
120
+ augmentation_loudness: false # randomly change loudness of each stem
121
+ augmentation_loudness_type: 1 # Type 1 or 2
122
+ augmentation_loudness_min: 0
123
+ augmentation_loudness_max: 0
124
+ q: 0.95
125
+ coarse_loss_clip: false
126
+ ema_momentum: 0.999
127
+ optimizer: adam
128
+ other_fix: true # it's needed for checking on multisong dataset if other is actually instrumental
129
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
130
+
131
+ inference:
132
+ batch_size: 2
133
+ dim_t: 1101
134
+ num_overlap: 2
roformers/config_chorus_male_female_bs_roformer.yaml ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 801 # don't work (use in model)
5
+ hop_length: 441 # don't work (use in model)
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 8
14
+ stereo: true
15
+ num_stems: 2
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ freqs_per_bands: !!python/tuple
19
+ - 2
20
+ - 2
21
+ - 2
22
+ - 2
23
+ - 2
24
+ - 2
25
+ - 2
26
+ - 2
27
+ - 2
28
+ - 2
29
+ - 2
30
+ - 2
31
+ - 2
32
+ - 2
33
+ - 2
34
+ - 2
35
+ - 2
36
+ - 2
37
+ - 2
38
+ - 2
39
+ - 2
40
+ - 2
41
+ - 2
42
+ - 2
43
+ - 4
44
+ - 4
45
+ - 4
46
+ - 4
47
+ - 4
48
+ - 4
49
+ - 4
50
+ - 4
51
+ - 4
52
+ - 4
53
+ - 4
54
+ - 4
55
+ - 12
56
+ - 12
57
+ - 12
58
+ - 12
59
+ - 12
60
+ - 12
61
+ - 12
62
+ - 12
63
+ - 24
64
+ - 24
65
+ - 24
66
+ - 24
67
+ - 24
68
+ - 24
69
+ - 24
70
+ - 24
71
+ - 48
72
+ - 48
73
+ - 48
74
+ - 48
75
+ - 48
76
+ - 48
77
+ - 48
78
+ - 48
79
+ - 128
80
+ - 129
81
+ dim_head: 64
82
+ heads: 8
83
+ attn_dropout: 0.0
84
+ ff_dropout: 0.0
85
+ flash_attn: true
86
+ dim_freqs_in: 1025
87
+ stft_n_fft: 2048
88
+ stft_hop_length: 441
89
+ stft_win_length: 2048
90
+ stft_normalized: false
91
+ mask_estimator_depth: 2
92
+ multi_stft_resolution_loss_weight: 1.0
93
+ multi_stft_resolutions_window_sizes: !!python/tuple
94
+ - 4096
95
+ - 2048
96
+ - 1024
97
+ - 512
98
+ - 256
99
+ multi_stft_hop_size: 147
100
+ multi_stft_normalized: False
101
+
102
+ training:
103
+ batch_size: 1
104
+ gradient_accumulation_steps: 1
105
+ grad_clip: 0
106
+ instruments:
107
+ - male
108
+ - female
109
+ lr: 1.0e-05
110
+ patience: 2
111
+ reduce_factor: 0.95
112
+ target_instrument: null
113
+ num_epochs: 1000
114
+ num_steps: 1000
115
+ q: 0.95
116
+ coarse_loss_clip: true
117
+ ema_momentum: 0.999
118
+ optimizer: adam
119
+ other_fix: true # it's needed for checking on multisong dataset if other is actually instrumental
120
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
121
+
122
+ inference:
123
+ batch_size: 1
124
+ dim_t: 801
125
+ num_overlap: 2
roformers/config_dereverb-echo_mel_band_roformer.yaml ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 801 # don't work (use in model)
5
+ hop_length: 441 # don't work (use in model)
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 256
13
+ depth: 8
14
+ stereo: true
15
+ num_stems: 2
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0.1
22
+ ff_dropout: 0.1
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 8
44
+ grad_clip: 0
45
+ instruments:
46
+ - dry
47
+ - No dry
48
+ lr: 4.0e-05
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: null
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ q: 0.95
55
+ coarse_loss_clip: true
56
+ ema_momentum: 0.999
57
+ optimizer: adam
58
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
59
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
60
+
61
+ augmentations:
62
+ enable: true # enable or disable all augmentations (to fast disable if needed)
63
+ loudness: true # randomly change loudness of each stem on the range (loudness_min; loudness_max)
64
+ loudness_min: 0.5
65
+ loudness_max: 1.5
66
+ mixup: false # mix several stems of same type with some probability (only works for dataset types: 1, 2, 3)
67
+ mixup_probs: !!python/tuple # 2 additional stems of the same type (1st with prob 0.2, 2nd with prob 0.02)
68
+ - 0.2
69
+ - 0.02
70
+ mixup_loudness_min: 0.5
71
+ mixup_loudness_max: 1.5
72
+
73
+ inference:
74
+ batch_size: 4
75
+ dim_t: 801
76
+ num_overlap: 4
roformers/config_dereverb-echo_mel_band_roformer_sdr_13.4843_v2.yaml ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 801
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 256
13
+ depth: 8
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0.1
22
+ ff_dropout: 0.1
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 8
44
+ grad_clip: 0
45
+ instruments:
46
+ - dry
47
+ - No dry
48
+ lr: 1.0e-05
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: dry
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ q: 0.95
55
+ coarse_loss_clip: true
56
+ ema_momentum: 0.999
57
+ optimizer: adam
58
+ other_fix: false
59
+ use_amp: true
60
+
61
+ inference:
62
+ batch_size: 1
63
+ dim_t: 801
64
+ num_overlap: 4
roformers/config_dereverb_echo_mel_band_roformer_v2.yaml ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 801
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 256
13
+ depth: 8
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0.1
22
+ ff_dropout: 0.1
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 8
44
+ grad_clip: 0
45
+ instruments:
46
+ - dry
47
+ - other
48
+ lr: 1.0e-05
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: dry
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ q: 0.95
55
+ coarse_loss_clip: true
56
+ ema_momentum: 0.999
57
+ optimizer: adam
58
+ other_fix: false
59
+ use_amp: true
60
+
61
+ inference:
62
+ batch_size: 1
63
+ dim_t: 801
64
+ num_overlap: 4
roformers/config_mel_band_roformer_bleed_suppressor_v1.yaml ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100
3
+ dim_f: 1024
4
+ dim_t: 801
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ instruments:
43
+ - Instrumental
44
+ - Bleed
45
+ target_instrument: Instrumental
46
+ use_amp: True
47
+
48
+ inference:
49
+ batch_size: 1
50
+ dim_t: 801
51
+ num_overlap: 2
roformers/config_mel_band_roformer_instrumental_becruily.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 1
44
+ grad_clip: 0
45
+ instruments:
46
+ - Instrumental
47
+ - Vocals
48
+ lr: 0.0005
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: Instrumental
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ augmentation: false # enable augmentations by audiomentations and pedalboard
55
+ augmentation_type: null
56
+ use_mp3_compress: false # Deprecated
57
+ augmentation_mix: false # Mix several stems of the same type with some probability
58
+ augmentation_loudness: false # randomly change loudness of each stem
59
+ augmentation_loudness_type: 1 # Type 1 or 2
60
+ augmentation_loudness_min: 0
61
+ augmentation_loudness_max: 0
62
+ q: 0.95
63
+ coarse_loss_clip: false
64
+ ema_momentum: 0.999
65
+ optimizer: adamw
66
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
67
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
68
+
69
+ inference:
70
+ batch_size: 1
71
+ dim_t: 1101
72
+ num_overlap: 2
roformers/config_mel_band_roformer_karaoke_becruily.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 2
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: true
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: false
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: false
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 1
44
+ grad_clip: 0
45
+ instruments:
46
+ - Vocals
47
+ - Instrumental
48
+ lr: 0.0005
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: null
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ augmentation: false # enable augmentations by audiomentations and pedalboard
55
+ augmentation_type:
56
+ use_mp3_compress: false # Deprecated
57
+ augmentation_mix: false # Mix several stems of the same type with some probability
58
+ augmentation_loudness: false # randomly change loudness of each stem
59
+ augmentation_loudness_type: 1 # Type 1 or 2
60
+ augmentation_loudness_min: 0
61
+ augmentation_loudness_max: 0
62
+ q: 0.95
63
+ coarse_loss_clip: false
64
+ ema_momentum: 0.999
65
+ optimizer: adamw
66
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
67
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
68
+
69
+ inference:
70
+ batch_size: 1
71
+ dim_t: 1101
72
+ num_overlap: 8
roformers/config_mel_band_roformer_karaoke_gabox.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: true
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: false
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: true
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 1
44
+ grad_clip: 0
45
+ instruments:
46
+ - Vocals
47
+ - Instrumental
48
+ lr: 0.0005
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: Vocals
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ augmentation: false # enable augmentations by audiomentations and pedalboard
55
+ augmentation_type:
56
+ use_mp3_compress: false # Deprecated
57
+ augmentation_mix: false # Mix several stems of the same type with some probability
58
+ augmentation_loudness: false # randomly change loudness of each stem
59
+ augmentation_loudness_type: 1 # Type 1 or 2
60
+ augmentation_loudness_min: 0
61
+ augmentation_loudness_max: 0
62
+ q: 0.95
63
+ coarse_loss_clip: false
64
+ ema_momentum: 0.999
65
+ optimizer: adamw
66
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
67
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
68
+
69
+ inference:
70
+ batch_size: 1
71
+ dim_t: 1101
72
+ num_overlap: 8
roformers/config_mel_band_roformer_kim_ft_unwa.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100
3
+ dim_f: 1024
4
+ dim_t: 801
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 1
44
+ grad_clip: 0
45
+ instruments:
46
+ - vocals
47
+ - other
48
+ lr: 1.0e-05
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: vocals
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ augmentation: false # enable augmentations by audiomentations and pedalboard
55
+ augmentation_type: null
56
+ use_mp3_compress: false # Deprecated
57
+ augmentation_mix: false # Mix several stems of the same type with some probability
58
+ augmentation_loudness: false # randomly change loudness of each stem
59
+ augmentation_loudness_type: 1 # Type 1 or 2
60
+ augmentation_loudness_min: 0
61
+ augmentation_loudness_max: 0
62
+ q: 0.95
63
+ coarse_loss_clip: false
64
+ ema_momentum: 0.999
65
+ optimizer: adam
66
+ other_fix: true # it's needed for checking on multisong dataset if other is actually instrumental
67
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
68
+
69
+ inference:
70
+ batch_size: 1
71
+ dim_t: 801
72
+ num_overlap: 8
roformers/config_mel_band_roformer_vocal_fullness_aname.yaml ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 661500
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.001
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 1
44
+ grad_clip: 0
45
+ instruments:
46
+ - vocals
47
+ - other
48
+ target_instrument: vocals
49
+ use_amp: true
50
+
51
+ inference:
52
+ batch_size: 4
53
+ dim_t: 1101
54
+ num_overlap: 4
roformers/config_mel_band_roformer_vocals_becruily.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 1
44
+ grad_clip: 0
45
+ instruments:
46
+ - vocals
47
+ - other
48
+ lr: 0.0005
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: vocals
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ augmentation: false # enable augmentations by audiomentations and pedalboard
55
+ augmentation_type: null
56
+ use_mp3_compress: false # Deprecated
57
+ augmentation_mix: false # Mix several stems of the same type with some probability
58
+ augmentation_loudness: false # randomly change loudness of each stem
59
+ augmentation_loudness_type: 1 # Type 1 or 2
60
+ augmentation_loudness_min: 0
61
+ augmentation_loudness_max: 0
62
+ q: 0.95
63
+ coarse_loss_clip: false
64
+ ema_momentum: 0.999
65
+ optimizer: adamw
66
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
67
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
68
+
69
+ inference:
70
+ batch_size: 1
71
+ dim_t: 1101
72
+ num_overlap: 2
roformers/config_mel_band_roformer_vocals_gabox.yaml ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.001
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ instruments:
43
+ - Vocals
44
+ - Instrumental
45
+ target_instrument: Vocals
46
+
47
+ inference:
48
+ batch_size: 1
49
+ dim_t: 1101
50
+ num_overlap: 1
51
+ chunk_size: 352800
roformers/config_melband_roformer_big_beta5e.yaml ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100
3
+ dim_f: 1024
4
+ dim_t: 801
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 3
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ instruments:
43
+ - vocals
44
+ - other
45
+ target_instrument: vocals
46
+ use_amp: True
47
+
48
+ inference:
49
+ batch_size: 1
50
+ dim_t: 801
51
+ num_overlap: 2
roformers/config_melbandroformer_big_beta4.yaml ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100
3
+ dim_f: 1024
4
+ dim_t: 1101
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 12
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 3
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ instruments:
43
+ - vocals
44
+ - other
45
+ target_instrument: vocals
46
+ use_amp: True
47
+
48
+ inference:
49
+ batch_size: 1
50
+ dim_t: 1101
51
+ num_overlap: 2
roformers/config_melbandroformer_big_beta6.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 529200
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 512
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 1
44
+ grad_clip: 0
45
+ instruments:
46
+ - vocals
47
+ - other
48
+ lr: 1.0e-05
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: vocals
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ augmentation: false # enable augmentations by audiomentations and pedalboard
55
+ augmentation_type: null
56
+ use_mp3_compress: false # Deprecated
57
+ augmentation_mix: false # Mix several stems of the same type with some probability
58
+ augmentation_loudness: false # randomly change loudness of each stem
59
+ augmentation_loudness_type: 1 # Type 1 or 2
60
+ augmentation_loudness_min: 0
61
+ augmentation_loudness_max: 0
62
+ q: 0.95
63
+ coarse_loss_clip: false
64
+ ema_momentum: 0.999
65
+ optimizer: adam
66
+ other_fix: true # it's needed for checking on multisong dataset if other is actually instrumental
67
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
68
+
69
+ inference:
70
+ batch_size: 2
71
+ dim_t: 1201
72
+ num_overlap: 2
roformers/config_melbandroformer_big_beta6x.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 529200
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 512
13
+ depth: 12
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 1
44
+ grad_clip: 0
45
+ instruments:
46
+ - vocals
47
+ - other
48
+ lr: 1.0e-05
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: vocals
52
+ num_epochs: 1000
53
+ num_steps: 1000
54
+ augmentation: false # enable augmentations by audiomentations and pedalboard
55
+ augmentation_type: null
56
+ use_mp3_compress: false # Deprecated
57
+ augmentation_mix: false # Mix several stems of the same type with some probability
58
+ augmentation_loudness: false # randomly change loudness of each stem
59
+ augmentation_loudness_type: 1 # Type 1 or 2
60
+ augmentation_loudness_min: 0
61
+ augmentation_loudness_max: 0
62
+ q: 0.95
63
+ coarse_loss_clip: false
64
+ ema_momentum: 0.999
65
+ optimizer: adam
66
+ other_fix: true # it's needed for checking on multisong dataset if other is actually instrumental
67
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
68
+
69
+ inference:
70
+ batch_size: 2
71
+ dim_t: 1201
72
+ num_overlap: 2
roformers/config_melbandroformer_inst.yaml ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100
3
+ dim_f: 1024
4
+ dim_t: 1101
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ instruments:
43
+ - other
44
+ - vocals
45
+ target_instrument: other
46
+ use_amp: True
47
+
48
+ inference:
49
+ batch_size: 1
50
+ dim_t: 1101
51
+ num_overlap: 2
roformers/config_melbandroformer_inst_v2.yaml ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100
3
+ dim_f: 1024
4
+ dim_t: 1101
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 12
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 3
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ instruments:
43
+ - Instrumental
44
+ - Vocals
45
+ target_instrument: Instrumental
46
+ use_amp: True
47
+
48
+ inference:
49
+ batch_size: 1
50
+ dim_t: 1101
51
+ num_overlap: 2
roformers/config_melbandroformer_instvoc_duality.yaml ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 2
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ instruments:
43
+ - Vocals
44
+ - Instrumental
45
+ target_instrument: null
46
+ use_amp: True
47
+
48
+ inference:
49
+ batch_size: 1
50
+ dim_t: 1101
51
+ num_overlap: 2
roformers/config_vocals_mel_band_roformer_big_v1_ft.yaml ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 485100
3
+ dim_f: 1024
4
+ dim_t: 801
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 3
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ instruments:
43
+ - vocals
44
+ - other
45
+ target_instrument: vocals
46
+ use_amp: True
47
+
48
+ inference:
49
+ batch_size: 1
50
+ dim_t: 801
51
+ num_overlap: 2
roformers/config_vocals_mel_band_roformer_ft.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 1
43
+ gradient_accumulation_steps: 8
44
+ grad_clip: 0
45
+ instruments:
46
+ - vocals
47
+ - other
48
+ lr: 1.0e-04
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: vocals
52
+ num_epochs: 1000
53
+ num_steps: 100
54
+ augmentation: true # enable augmentations by audiomentations and pedalboard
55
+ augmentation_type: null
56
+ use_mp3_compress: false # Deprecated
57
+ augmentation_mix: true # Mix several stems of the same type with some probability
58
+ augmentation_loudness: true # randomly change loudness of each stem
59
+ augmentation_loudness_type: 1 # Type 1 or 2
60
+ augmentation_loudness_min: 0
61
+ augmentation_loudness_max: 0
62
+ q: 0.95
63
+ coarse_loss_clip: false
64
+ ema_momentum: 0.999
65
+ optimizer: adamw8bit
66
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
67
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
68
+
69
+ inference:
70
+ batch_size: 4
71
+ dim_t: 256
72
+ num_overlap: 2
roformers/dereverb_mel_band_roformer_anvuew.yaml ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 801
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ batch_size: 3
43
+ gradient_accumulation_steps: 1
44
+ grad_clip: 0
45
+ instruments:
46
+ - noreverb
47
+ - reverb
48
+ lr: 5.0e-05
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: noreverb
52
+ num_epochs: 1000
53
+ num_steps: 4000
54
+ q: 0.95
55
+ coarse_loss_clip: false
56
+ ema_momentum: 0.999
57
+ optimizer: adamw
58
+ other_fix: true # it's needed for checking on multisong dataset if other is actually instrumental
59
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
60
+
61
+ augmentations:
62
+ enable: true # enable or disable all augmentations (to fast disable if needed)
63
+ loudness: true # randomly change loudness of each stem on the range (loudness_min; loudness_max)
64
+ loudness_min: 0.1
65
+ loudness_max: 1.0
66
+ mixup: false # mix several stems of same type with some probability (only works for dataset types: 1, 2, 3)
67
+ mixup_probs: !!python/tuple # 2 additional stems of the same type (1st with prob 0.2, 2nd with prob 0.02)
68
+ - 0.2
69
+ - 0.02
70
+ mixup_loudness_min: 0.5
71
+ mixup_loudness_max: 1.5
72
+
73
+ inference:
74
+ batch_size: 1
75
+ dim_t: 801
76
+ num_overlap: 2
roformers/vocals_mel_band_roformer.yaml ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 441
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.001
10
+
11
+ model:
12
+ dim: 384
13
+ depth: 6
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ num_bands: 60
19
+ dim_head: 64
20
+ heads: 8
21
+ attn_dropout: 0
22
+ ff_dropout: 0
23
+ flash_attn: True
24
+ dim_freqs_in: 1025
25
+ sample_rate: 44100 # needed for mel filter bank from librosa
26
+ stft_n_fft: 2048
27
+ stft_hop_length: 441
28
+ stft_win_length: 2048
29
+ stft_normalized: False
30
+ mask_estimator_depth: 2
31
+ multi_stft_resolution_loss_weight: 1.0
32
+ multi_stft_resolutions_window_sizes: !!python/tuple
33
+ - 4096
34
+ - 2048
35
+ - 1024
36
+ - 512
37
+ - 256
38
+ multi_stft_hop_size: 147
39
+ multi_stft_normalized: False
40
+
41
+ training:
42
+ instruments:
43
+ - vocals
44
+ - other
45
+ target_instrument: vocals
46
+
47
+ inference:
48
+ dim_t: 1101
49
+ num_overlap: 1
50
+ chunk_size: 352800