Hugging Face's logo

Politrees
/
UVR_resources

ONNX
uvr
uvr5
ultimatevocalremover
demucs
vr-arch
mdx-net
mdx23c
roformer
scnet
bandit
Model card Files
xet
Community
3
Politrees commited on
Commit
f1f9c96
·
verified ·
1 Parent(s): caf9316

Adding new models (read description)

Browse files

- BandSplit Roformer | FNO by Unwa
- BandSplit Roformer | Karaoke Frazer by becruily
- BandSplit Roformer | SW by jarredou
- MelBand Roformer | Duality v1 by Aname

Files changed (8) hide show
  1. models/Roformer/BandSplit/config_BandSplit-Roformer_FNO_by-Unwa.yaml +136 -0
  2. models/Roformer/BandSplit/config_BandSplit-Roformer_Karaoke_Frazer_by-becruily.yaml +129 -0
  3. models/Roformer/BandSplit/config_BandSplit-Roformer_SW_by-jarredou.yaml +197 -0
  4. models/Roformer/BandSplit/model_BandSplit-Roformer_FNO_by-Unwa.ckpt +3 -0
  5. models/Roformer/BandSplit/model_BandSplit-Roformer_Karaoke_Frazer_by-becruily.ckpt +3 -0
  6. models/Roformer/BandSplit/model_BandSplit-Roformer_SW_by-jarredou.ckpt +3 -0
  7. models/Roformer/MelBand/config_MelBand-Roformer_Duality_v1_by-Aname.yaml +72 -0
  8. models/Roformer/MelBand/model_MelBand-Roformer_Duality_v1_by-Aname.ckpt +3 -0
models/Roformer/BandSplit/config_BandSplit-Roformer_FNO_by-Unwa.yaml ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ mlp_expansion_factor: 4
103
+
104
+ training:
105
+ batch_size: 2
106
+ gradient_accumulation_steps: 1
107
+ grad_clip: 0
108
+ instruments: ['vocals', 'other']
109
+ patience: 3
110
+ reduce_factor: 0.95
111
+ target_instrument: other
112
+ num_epochs: 1000
113
+ num_steps: 1000
114
+ augmentation: false # enable augmentations by audiomentations and pedalboard
115
+ augmentation_type: simple1
116
+ use_mp3_compress: false # Deprecated
117
+ augmentation_mix: true # Mix several stems of the same type with some probability
118
+ augmentation_loudness: true # randomly change loudness of each stem
119
+ augmentation_loudness_type: 1 # Type 1 or 2
120
+ augmentation_loudness_min: 0.5
121
+ augmentation_loudness_max: 1.5
122
+ q: 0.95
123
+ coarse_loss_clip: true
124
+ ema_momentum: 0.999
125
+ # optimizer: prodigy
126
+ optimizer: adam
127
+ # lr: 1.0
128
+ lr: 1.0e-5
129
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
130
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
131
+
132
+ inference:
133
+ batch_size: 2
134
+ dim_t: 1700
135
+ num_overlap: 2
136
+ normalize: false
models/Roformer/BandSplit/config_BandSplit-Roformer_Karaoke_Frazer_by-becruily.yaml ADDED
@@ -0,0 +1,129 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 882000
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: 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: 512
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
+ mlp_expansion_factor: 4
103
+
104
+ training:
105
+ batch_size: 1
106
+ gradient_accumulation_steps: 1
107
+ grad_clip: 0
108
+ instruments:
109
+ - Vocals
110
+ - Instrumental
111
+ patience: 2
112
+ reduce_factor: 0.95
113
+ target_instrument: Vocals
114
+ num_epochs: 1000
115
+ num_steps: 1000
116
+ q: 0.95
117
+ coarse_loss_clip: true
118
+ ema_momentum: 0.999
119
+ # optimizer: prodigy
120
+ optimizer: adam
121
+ lr: 1.0e-5
122
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
123
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
124
+
125
+ inference:
126
+ batch_size: 2
127
+ dim_t: 2001
128
+ num_overlap: 4
129
+ normalize: false
models/Roformer/BandSplit/config_BandSplit-Roformer_SW_by-jarredou.yaml ADDED
@@ -0,0 +1,197 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 588800 #882000
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: 12
14
+ stereo: true
15
+ num_stems: 6
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: 0.1
85
+ ff_dropout: 0.1
86
+ flash_attn: true
87
+ dim_freqs_in: 1025
88
+ stft_n_fft: 2048
89
+ stft_hop_length: 512
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
+ mlp_expansion_factor: 4
103
+ use_torch_checkpoint: False # it allows to greatly reduce GPU memory consumption during training (not fully tested)
104
+ skip_connection: False # Enable skip connection between transformer blocks - can solve problem with gradients and probably faster training
105
+
106
+ training:
107
+ batch_size: 2
108
+ gradient_accumulation_steps: 1
109
+ grad_clip: 0
110
+ instruments: ['bass', 'drums', 'other', 'vocals', 'guitar', 'piano']
111
+ patience: 3
112
+ reduce_factor: 0.95
113
+ target_instrument: null
114
+ num_epochs: 1000
115
+ num_steps: 1000
116
+ augmentation: false # enable augmentations by audiomentations and pedalboard
117
+ augmentation_type: simple1
118
+ use_mp3_compress: false # Deprecated
119
+ augmentation_mix: true # Mix several stems of the same type with some probability
120
+ augmentation_loudness: true # randomly change loudness of each stem
121
+ augmentation_loudness_type: 1 # Type 1 or 2
122
+ augmentation_loudness_min: 0.5
123
+ augmentation_loudness_max: 1.5
124
+ q: 0.95
125
+ coarse_loss_clip: true
126
+ ema_momentum: 0.999
127
+ # optimizer: prodigy
128
+ optimizer: adam
129
+ # lr: 1.0
130
+ lr: 1.0e-5
131
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
132
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
133
+
134
+ augmentations:
135
+ enable: true # enable or disable all augmentations (to fast disable if needed)
136
+ loudness: true # randomly change loudness of each stem on the range (loudness_min; loudness_max)
137
+ loudness_min: 0.5
138
+ loudness_max: 1.5
139
+ mixup: true # mix several stems of same type with some probability (only works for dataset types: 1, 2, 3)
140
+ mixup_probs: !!python/tuple # 2 additional stems of the same type (1st with prob 0.2, 2nd with prob 0.02)
141
+ - 0.2
142
+ - 0.02
143
+ mixup_loudness_min: 0.5
144
+ mixup_loudness_max: 1.5
145
+
146
+ all:
147
+ channel_shuffle: 0.5 # Set 0 or lower to disable
148
+ random_inverse: 0.1 # inverse track (better lower probability)
149
+ random_polarity: 0.5 # polarity change (multiply waveform to -1)
150
+
151
+ vocals:
152
+ pitch_shift: 0.1
153
+ pitch_shift_min_semitones: -5
154
+ pitch_shift_max_semitones: 5
155
+ seven_band_parametric_eq: 0.1
156
+ seven_band_parametric_eq_min_gain_db: -9
157
+ seven_band_parametric_eq_max_gain_db: 9
158
+ tanh_distortion: 0.1
159
+ tanh_distortion_min: 0.1
160
+ tanh_distortion_max: 0.7
161
+ bass:
162
+ pitch_shift: 0.1
163
+ pitch_shift_min_semitones: -2
164
+ pitch_shift_max_semitones: 2
165
+ seven_band_parametric_eq: 0.1
166
+ seven_band_parametric_eq_min_gain_db: -3
167
+ seven_band_parametric_eq_max_gain_db: 6
168
+ tanh_distortion: 0.1
169
+ tanh_distortion_min: 0.1
170
+ tanh_distortion_max: 0.5
171
+ drums:
172
+ pitch_shift: 0.1
173
+ pitch_shift_min_semitones: -5
174
+ pitch_shift_max_semitones: 5
175
+ seven_band_parametric_eq: 0.1
176
+ seven_band_parametric_eq_min_gain_db: -9
177
+ seven_band_parametric_eq_max_gain_db: 9
178
+ tanh_distortion: 0.1
179
+ tanh_distortion_min: 0.1
180
+ tanh_distortion_max: 0.6
181
+ other:
182
+ pitch_shift: 0.1
183
+ pitch_shift_min_semitones: -4
184
+ pitch_shift_max_semitones: 4
185
+ gaussian_noise: 0.1
186
+ gaussian_noise_min_amplitude: 0.001
187
+ gaussian_noise_max_amplitude: 0.015
188
+ time_stretch: 0.1
189
+ time_stretch_min_rate: 0.8
190
+ time_stretch_max_rate: 1.25
191
+
192
+
193
+ inference:
194
+ batch_size: 1
195
+ dim_t: 1101
196
+ num_overlap: 2
197
+ normalize: false
models/Roformer/BandSplit/model_BandSplit-Roformer_FNO_by-Unwa.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f35bf6d87b2863372388e85c2d9679e5b7651e5c2ddd23aab1480f7af10b90ca
3
+ size 332004435
models/Roformer/BandSplit/model_BandSplit-Roformer_Karaoke_Frazer_by-becruily.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb90ee24c1154d83fbcfd27e96182f19e061557cc6e4746953125e08c29389f9
3
+ size 204436907
models/Roformer/BandSplit/model_BandSplit-Roformer_SW_by-jarredou.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:24e7d35ee9c64415673d3fd33e06a67cac2c103c5df6267ba1576459c775916e
3
+ size 699412152
models/Roformer/MelBand/config_MelBand-Roformer_Duality_v1_by-Aname.yaml ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 661500
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
+ batch_size: 4
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: 4
71
+ dim_t: 1101
72
+ num_overlap: 4
models/Roformer/MelBand/model_MelBand-Roformer_Duality_v1_by-Aname.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:69c244e76d7142a948f0d41c67e293112ba84f12a28c89d24b22a3fecd9ef79e
3
+ size 913043907