listra92 commited on
Commit
6970dd2
·
verified ·
1 Parent(s): 44300c5

Upload 5 files

Browse files
_mdx_c_configs/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: aspiration
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: 1101
76
+ num_overlap: 2
_mdx_c_configs/config_mel_band_roformer_karaoke.yaml ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: 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
+ - karaoke
47
+ - other
48
+ lr: 1.0e-05
49
+ patience: 2
50
+ reduce_factor: 0.95
51
+ target_instrument: karaoke
52
+ num_epochs: 1000
53
+ num_steps: 2000
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: false # it's needed for checking on multisong dataset if other is actually instrumental
67
+
68
+ inference:
69
+ batch_size: 1
70
+ dim_t: 1101
71
+ num_overlap: 2
_mdx_c_configs/config_melbandroformer_big.yaml ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 352800
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: 0.0
10
+ model:
11
+ dim: 384
12
+ depth: 6
13
+ stereo: true
14
+ num_stems: 1
15
+ time_transformer_depth: 1
16
+ freq_transformer_depth: 1
17
+ num_bands: 60
18
+ dim_head: 64
19
+ heads: 8
20
+ attn_dropout: 0
21
+ ff_dropout: 0
22
+ flash_attn: true
23
+ dim_freqs_in: 1025
24
+ sample_rate: 44100
25
+ stft_n_fft: 2048
26
+ stft_hop_length: 441
27
+ stft_win_length: 2048
28
+ stft_normalized: false
29
+ mask_estimator_depth: 3
30
+ multi_stft_resolution_loss_weight: 1.0
31
+ multi_stft_resolutions_window_sizes: !!python/tuple
32
+ - 4096
33
+ - 2048
34
+ - 1024
35
+ - 512
36
+ - 256
37
+ multi_stft_hop_size: 147
38
+ multi_stft_normalized: false
39
+ training:
40
+ instruments:
41
+ - vocals
42
+ - other
43
+ target_instrument: vocals
44
+ use_amp: true
45
+ inference:
46
+ batch_size: 1
47
+ dim_t: 1101
48
+ num_overlap: 2
_mdx_c_configs/config_vocals_mel_band_roformer_kj.yaml ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: 0.0
10
+ model:
11
+ dim: 384
12
+ depth: 6
13
+ stereo: true
14
+ num_stems: 1
15
+ time_transformer_depth: 1
16
+ freq_transformer_depth: 1
17
+ num_bands: 60
18
+ dim_head: 64
19
+ heads: 8
20
+ attn_dropout: 0
21
+ ff_dropout: 0
22
+ flash_attn: true
23
+ dim_freqs_in: 1025
24
+ sample_rate: 44100
25
+ stft_n_fft: 2048
26
+ stft_hop_length: 441
27
+ stft_win_length: 2048
28
+ stft_normalized: false
29
+ mask_estimator_depth: 2
30
+ multi_stft_resolution_loss_weight: 1.0
31
+ multi_stft_resolutions_window_sizes: !!python/tuple
32
+ - 4096
33
+ - 2048
34
+ - 1024
35
+ - 512
36
+ - 256
37
+ multi_stft_hop_size: 147
38
+ multi_stft_normalized: false
39
+ training:
40
+ batch_size: 4
41
+ gradient_accumulation_steps: 1
42
+ grad_clip: 0
43
+ instruments:
44
+ - vocals
45
+ - other
46
+ lr: 1.0e-05
47
+ patience: 2
48
+ reduce_factor: 0.95
49
+ target_instrument: vocals
50
+ num_epochs: 1000
51
+ num_steps: 1000
52
+ augmentation: false
53
+ augmentation_type: null
54
+ use_mp3_compress: false
55
+ augmentation_mix: false
56
+ augmentation_loudness: false
57
+ augmentation_loudness_type: 1
58
+ augmentation_loudness_min: 0
59
+ augmentation_loudness_max: 0
60
+ q: 0.95
61
+ coarse_loss_clip: false
62
+ ema_momentum: 0.999
63
+ optimizer: adam
64
+ other_fix: true
65
+ use_amp: true
66
+ inference:
67
+ batch_size: 4
68
+ dim_t: 1101
69
+ num_overlap: 2
_mdx_c_configs/model_mel_band_roformer_denoise.yaml ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: 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: 2
43
+ gradient_accumulation_steps: 1
44
+ grad_clip: 0
45
+ instruments:
46
+ - dry
47
+ - other
48
+ lr: 1.0e-05
49
+ patience: 8
50
+ reduce_factor: 0.95
51
+ target_instrument: dry
52
+ num_epochs: 1000
53
+ num_steps: 4032
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: false # it's needed for checking on multisong dataset if other is actually instrumental
67
+
68
+ inference:
69
+ batch_size: 2
70
+ dim_t: 1101
71
+ num_overlap: 4