Hugging Face's logo

am2460162
/
msst_failed_failed_test

Model card Files
xet
Community
am2460162 commited on
Commit
3fdb10e
·
verified ·
1 Parent(s): 5ace3a5

Upload 8 files

Browse files
Files changed (8) hide show
  1. config_4_bass_bs_roformer.yaml +134 -0
  2. config_4_drums_bs_roformer.yaml +134 -0
  3. config_drums_musdb18_moises_mel_band_roformer.yaml +71 -0
  4. config_orchestra_mdx23c.yaml +99 -0
  5. model_bass_bs_roformer_ep_10_sdr_5.7862.ckpt +3 -0
  6. model_drums_bs_roformer_ep_12_sdr_7.2279.ckpt +3 -0
  7. model_mdx23c_ep_120_sdr_4.4174.ckpt +3 -0
  8. model_mel_band_roformer_ep_11_sdr_7.6853.ckpt +3 -0
config_4_bass_bs_roformer.yaml ADDED
@@ -0,0 +1,134 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 131584
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 512
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 192
13
+ depth: 6
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: 0.1
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
+
103
+ training:
104
+ batch_size: 4
105
+ gradient_accumulation_steps: 1
106
+ grad_clip: 0
107
+ instruments:
108
+ - vocals
109
+ - bass
110
+ - drums
111
+ - other
112
+ lr: 5.0e-05
113
+ patience: 2
114
+ reduce_factor: 0.95
115
+ target_instrument: bass
116
+ num_epochs: 1000
117
+ num_steps: 1000
118
+ q: 0.95
119
+ coarse_loss_clip: true
120
+ ema_momentum: 0.999
121
+ optimizer: adam
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
+ augmentations:
126
+ enable: true # enable or disable all augmentations (to fast disable if needed)
127
+ loudness: true # randomly change loudness of each stem on the range (loudness_min; loudness_max)
128
+ loudness_min: 0.5
129
+ loudness_max: 1.5
130
+
131
+ inference:
132
+ batch_size: 1
133
+ dim_t: 256
134
+ num_overlap: 4
config_4_drums_bs_roformer.yaml ADDED
@@ -0,0 +1,134 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 131584
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 512
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.000
10
+
11
+ model:
12
+ dim: 192
13
+ depth: 6
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: 0.1
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
+
103
+ training:
104
+ batch_size: 4
105
+ gradient_accumulation_steps: 1
106
+ grad_clip: 0
107
+ instruments:
108
+ - vocals
109
+ - bass
110
+ - drums
111
+ - other
112
+ lr: 5.0e-05
113
+ patience: 2
114
+ reduce_factor: 0.95
115
+ target_instrument: drums
116
+ num_epochs: 1000
117
+ num_steps: 1000
118
+ q: 0.95
119
+ coarse_loss_clip: true
120
+ ema_momentum: 0.999
121
+ optimizer: adam
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
+ augmentations:
126
+ enable: true # enable or disable all augmentations (to fast disable if needed)
127
+ loudness: true # randomly change loudness of each stem on the range (loudness_min; loudness_max)
128
+ loudness_min: 0.5
129
+ loudness_max: 1.5
130
+
131
+ inference:
132
+ batch_size: 1
133
+ dim_t: 256
134
+ num_overlap: 4
config_drums_musdb18_moises_mel_band_roformer.yaml ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 131584
3
+ dim_f: 1024
4
+ dim_t: 256
5
+ hop_length: 512
6
+ n_fft: 2048
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.001
10
+
11
+ model:
12
+ dim: 192
13
+ depth: 8
14
+ stereo: true
15
+ num_stems: 1
16
+ time_transformer_depth: 1
17
+ freq_transformer_depth: 1
18
+ linear_transformer_depth: 0
19
+ num_bands: 60
20
+ dim_head: 64
21
+ heads: 8
22
+ attn_dropout: 0.1
23
+ ff_dropout: 0.1
24
+ flash_attn: True
25
+ dim_freqs_in: 1025
26
+ sample_rate: 44100 # needed for mel filter bank from librosa
27
+ stft_n_fft: 2048
28
+ stft_hop_length: 512
29
+ stft_win_length: 2048
30
+ stft_normalized: False
31
+ mask_estimator_depth: 2
32
+ multi_stft_resolution_loss_weight: 1.0
33
+ multi_stft_resolutions_window_sizes: !!python/tuple
34
+ - 4096
35
+ - 2048
36
+ - 1024
37
+ - 512
38
+ - 256
39
+ multi_stft_hop_size: 147
40
+ multi_stft_normalized: False
41
+
42
+ training:
43
+ batch_size: 2
44
+ gradient_accumulation_steps: 1
45
+ grad_clip: 0
46
+ instruments:
47
+ - percussions
48
+ - other
49
+ lr: 2.566710416397524e-05
50
+ patience: 2
51
+ reduce_factor: 0.95
52
+ target_instrument: percussions
53
+ num_epochs: 1000
54
+ num_steps: 20000
55
+ q: 0.95
56
+ coarse_loss_clip: true
57
+ ema_momentum: 0.999
58
+ optimizer: adam
59
+ other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
60
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
61
+
62
+ augmentations:
63
+ enable: true # enable or disable all augmentations (to fast disable if needed)
64
+ loudness: true # randomly change loudness of each stem on the range (loudness_min; loudness_max)
65
+ loudness_min: 0.5
66
+ loudness_max: 1.5
67
+
68
+ inference:
69
+ batch_size: 2
70
+ dim_t: 256
71
+ num_overlap: 4
config_orchestra_mdx23c.yaml ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ audio:
2
+ chunk_size: 261120
3
+ dim_f: 4096
4
+ dim_t: 256
5
+ hop_length: 1024
6
+ n_fft: 8192
7
+ num_channels: 2
8
+ sample_rate: 44100
9
+ min_mean_abs: 0.001
10
+
11
+ model:
12
+ act: gelu
13
+ bottleneck_factor: 4
14
+ growth: 128
15
+ norm: InstanceNorm
16
+ num_blocks_per_scale: 2
17
+ num_channels: 128
18
+ num_scales: 5
19
+ num_subbands: 4
20
+ scale:
21
+ - 2
22
+ - 2
23
+
24
+ training:
25
+ batch_size: 1
26
+ gradient_accumulation_steps: 3
27
+ grad_clip: 0
28
+ instruments:
29
+ - inst
30
+ - orch
31
+ lr: 9.0e-05
32
+ patience: 6
33
+ reduce_factor: 0.95
34
+ target_instrument: orch
35
+ num_epochs: 1000
36
+ num_steps: 750
37
+ q: 0.95
38
+ coarse_loss_clip: true
39
+ ema_momentum: 0.999
40
+ optimizer: adam
41
+ other_fix: true # it's needed for checking on multisong dataset if other is actually instrumental
42
+ use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
43
+
44
+ augmentations:
45
+ enable: true # enable or disable all augmentations (to fast disable if needed)
46
+ loudness: true # randomly change loudness of each stem on the range (loudness_min; loudness_max)
47
+ loudness_min: 0.5
48
+ loudness_max: 1.5
49
+ mixup: true # mix several stems of same type with some probability (only works for dataset types: 1, 2, 3)
50
+ mixup_probs: [0.4, 0.02]
51
+ mixup_loudness_min: 0.5
52
+ mixup_loudness_max: 1.5
53
+
54
+ # apply mp3 compression to mixture only (emulate downloading mp3 from internet)
55
+ mp3_compression_on_mixture: 0.01
56
+ mp3_compression_on_mixture_bitrate_min: 32
57
+ mp3_compression_on_mixture_bitrate_max: 320
58
+ mp3_compression_on_mixture_backend: "lameenc"
59
+
60
+ all:
61
+ channel_shuffle: 0.5 # Set 0 or lower to disable
62
+ random_inverse: 0.1 # inverse track (better lower probability)
63
+ random_polarity: 0.5 # polarity change (multiply waveform to -1)
64
+ mp3_compression: 0.01
65
+ mp3_compression_min_bitrate: 32
66
+ mp3_compression_max_bitrate: 320
67
+ mp3_compression_backend: "lameenc"
68
+
69
+ orch:
70
+ pitch_shift: 0.25
71
+ pitch_shift_min_semitones: -5
72
+ pitch_shift_max_semitones: 5
73
+ seven_band_parametric_eq: 1
74
+ seven_band_parametric_eq_min_gain_db: -6
75
+ seven_band_parametric_eq_max_gain_db: 6
76
+ tanh_distortion: 0.5
77
+ tanh_distortion_min: 0.05
78
+ tanh_distortion_max: 0.6
79
+ inst:
80
+ pitch_shift: 0.25
81
+ pitch_shift_min_semitones: -5
82
+ pitch_shift_max_semitones: 5
83
+ gaussian_noise: 0.1
84
+ gaussian_noise_min_amplitude: 0.001
85
+ gaussian_noise_max_amplitude: 0.015
86
+ time_stretch: 0.01
87
+ time_stretch_min_rate: 0.8
88
+ time_stretch_max_rate: 1.25
89
+ seven_band_parametric_eq: 1
90
+ seven_band_parametric_eq_min_gain_db: -6
91
+ seven_band_parametric_eq_max_gain_db: 6
92
+ tanh_distortion: 0.5
93
+ tanh_distortion_min: 0.05
94
+ tanh_distortion_max: 0.6
95
+
96
+ inference:
97
+ batch_size: 1
98
+ dim_t: 256
99
+ num_overlap: 4
model_bass_bs_roformer_ep_10_sdr_5.7862.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:57ab7828b32040b9c55ba03ac516e9d7a976c292247ffce2d8fb66fb7a68f993
3
+ size 98602680
model_drums_bs_roformer_ep_12_sdr_7.2279.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb534cb6b4b90e7dcbcbb741ba4111393ccf6083b372b554bba7b556121d104e
3
+ size 98603241
model_mdx23c_ep_120_sdr_4.4174.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:80ecfd131abe51a456c29df9f583f91ca55b5aaf72496f24c65f323b2064876b
3
+ size 448091321
model_mel_band_roformer_ep_11_sdr_7.6853.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8fede38c911b7fa87c05186111dcad788895439cbaa1eead656768163e13a49d
3
+ size 276715240