Skip to content

Commit 1027afb

Browse files
committed
New SCNet Tran with transformer blocks instead of LSTM was added.
1 parent 2b6f12d commit 1027afb

File tree

3 files changed

+825
-0
lines changed

3 files changed

+825
-0
lines changed
Lines changed: 154 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,154 @@
1+
audio:
2+
chunk_size: 485100 # 44100 * 11
3+
num_channels: 2
4+
sample_rate: 44100
5+
min_mean_abs: 0.000
6+
7+
model:
8+
sources: ['drums', 'bass', 'other', 'vocals']
9+
audio_channels: 2
10+
dims: [4, 32, 64, 128]
11+
nfft: 4096
12+
hop_size: 1024
13+
win_size: 4096
14+
normalized: True
15+
band_SR: [0.175, 0.392, 0.433]
16+
band_stride: [1, 4, 16]
17+
band_kernel: [3, 4, 16]
18+
conv_depths: [3, 2, 1]
19+
compress: 4
20+
conv_kernel: 3
21+
num_dplayer: 6
22+
expand: 1
23+
tran_rotary_embedding_dim: 64
24+
tran_depth: 1
25+
tran_heads: 8
26+
tran_dim_head: 64
27+
tran_attn_dropout: 0.0
28+
tran_ff_dropout: 0.0
29+
tran_flash_attn: False
30+
31+
training:
32+
batch_size: 5
33+
gradient_accumulation_steps: 1
34+
grad_clip: 0
35+
instruments: ['drums', 'bass', 'other', 'vocals']
36+
patience: 2
37+
reduce_factor: 0.95
38+
target_instrument: null
39+
num_epochs: 1000
40+
num_steps: 1000
41+
q: 0.95
42+
coarse_loss_clip: true
43+
ema_momentum: 0.999
44+
optimizer: adam
45+
lr: 5.0e-05
46+
# optimizer: prodigy
47+
# lr: 1.0
48+
normalize: false # perform normalization on input of model (use the same for inference!)
49+
other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
50+
use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
51+
52+
augmentations:
53+
enable: true # enable or disable all augmentations (to fast disable if needed)
54+
loudness: true # randomly change loudness of each stem on the range (loudness_min; loudness_max)
55+
loudness_min: 0.5
56+
loudness_max: 1.5
57+
mixup: true # mix several stems of same type with some probability (only works for dataset types: 1, 2, 3)
58+
mixup_probs: !!python/tuple # 2 additional stems of the same type (1st with prob 0.2, 2nd with prob 0.02)
59+
- 0.2
60+
- 0.02
61+
- 0.002
62+
mixup_loudness_min: 0.5
63+
mixup_loudness_max: 1.5
64+
65+
# apply mp3 compression to mixture only (emulate downloading mp3 from internet)
66+
mp3_compression_on_mixture: 0.01
67+
mp3_compression_on_mixture_bitrate_min: 32
68+
mp3_compression_on_mixture_bitrate_max: 320
69+
mp3_compression_on_mixture_backend: "lameenc"
70+
71+
all:
72+
channel_shuffle: 0.5 # Set 0 or lower to disable
73+
random_inverse: 0.01 # inverse track (better lower probability)
74+
random_polarity: 0.5 # polarity change (multiply waveform to -1)
75+
76+
vocals:
77+
pitch_shift: 0.1
78+
pitch_shift_min_semitones: -5
79+
pitch_shift_max_semitones: 5
80+
seven_band_parametric_eq: 0.1
81+
seven_band_parametric_eq_min_gain_db: -9
82+
seven_band_parametric_eq_max_gain_db: 9
83+
tanh_distortion: 0.1
84+
tanh_distortion_min: 0.1
85+
tanh_distortion_max: 0.7
86+
time_stretch: 0.01
87+
time_stretch_min_rate: 0.8
88+
time_stretch_max_rate: 1.25
89+
bass:
90+
pitch_shift: 0.01
91+
pitch_shift_min_semitones: -2
92+
pitch_shift_max_semitones: 2
93+
seven_band_parametric_eq: 0.01
94+
seven_band_parametric_eq_min_gain_db: -3
95+
seven_band_parametric_eq_max_gain_db: 6
96+
tanh_distortion: 0.01
97+
tanh_distortion_min: 0.1
98+
tanh_distortion_max: 0.5
99+
time_stretch: 0.1
100+
time_stretch_min_rate: 0.9
101+
time_stretch_max_rate: 1.1
102+
drums:
103+
pitch_shift: 0.1
104+
pitch_shift_min_semitones: -5
105+
pitch_shift_max_semitones: 5
106+
seven_band_parametric_eq: 0.1
107+
seven_band_parametric_eq_min_gain_db: -9
108+
seven_band_parametric_eq_max_gain_db: 9
109+
tanh_distortion: 0.1
110+
tanh_distortion_min: 0.1
111+
tanh_distortion_max: 0.6
112+
time_stretch: 0.01
113+
time_stretch_min_rate: 0.8
114+
time_stretch_max_rate: 1.25
115+
other:
116+
pitch_shift: 0.1
117+
pitch_shift_min_semitones: -4
118+
pitch_shift_max_semitones: 4
119+
gaussian_noise: 0.1
120+
gaussian_noise_min_amplitude: 0.001
121+
gaussian_noise_max_amplitude: 0.015
122+
time_stretch: 0.01
123+
time_stretch_min_rate: 0.8
124+
time_stretch_max_rate: 1.25
125+
126+
inference:
127+
batch_size: 2
128+
dim_t: 256
129+
num_overlap: 2
130+
normalize: false
131+
132+
loss_multistft:
133+
fft_sizes:
134+
- 1024
135+
- 2048
136+
- 4096
137+
hop_sizes:
138+
- 147
139+
- 256
140+
- 512
141+
win_lengths:
142+
- 1024
143+
- 2048
144+
- 4096
145+
window: "hann_window"
146+
scale: "mel"
147+
n_bins: 128
148+
sample_rate: 44100
149+
perceptual_weighting: true
150+
w_sc: 1.0
151+
w_log_mag: 1.0
152+
w_lin_mag: 0.0
153+
w_phs: 0.0
154+
mag_distance: "L1"

0 commit comments

Comments
 (0)