Upload 2 files
Browse files
dereverb-echo_128_4_4_mel_band_roformer.yaml
ADDED
@@ -0,0 +1,78 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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: 128
|
13 |
+
depth: 4
|
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: 441
|
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 |
+
|
43 |
+
training:
|
44 |
+
batch_size: 3
|
45 |
+
gradient_accumulation_steps: 8
|
46 |
+
grad_clip: 0
|
47 |
+
instruments:
|
48 |
+
- dry
|
49 |
+
- other
|
50 |
+
lr: 1.0e-04
|
51 |
+
patience: 2
|
52 |
+
reduce_factor: 0.95
|
53 |
+
target_instrument: dry
|
54 |
+
num_epochs: 1000
|
55 |
+
num_steps: 1000
|
56 |
+
q: 0.95
|
57 |
+
coarse_loss_clip: true
|
58 |
+
ema_momentum: 0.999
|
59 |
+
optimizer: adam
|
60 |
+
other_fix: false # it's needed for checking on multisong dataset if other is actually instrumental
|
61 |
+
use_amp: true # enable or disable usage of mixed precision (float16) - usually it must be true
|
62 |
+
|
63 |
+
augmentations:
|
64 |
+
enable: true # enable or disable all augmentations (to fast disable if needed)
|
65 |
+
loudness: true # randomly change loudness of each stem on the range (loudness_min; loudness_max)
|
66 |
+
loudness_min: 0.5
|
67 |
+
loudness_max: 1.5
|
68 |
+
mixup: false # mix several stems of same type with some probability (only works for dataset types: 1, 2, 3)
|
69 |
+
mixup_probs: !!python/tuple # 2 additional stems of the same type (1st with prob 0.2, 2nd with prob 0.02)
|
70 |
+
- 0.2
|
71 |
+
- 0.02
|
72 |
+
mixup_loudness_min: 0.5
|
73 |
+
mixup_loudness_max: 1.5
|
74 |
+
|
75 |
+
inference:
|
76 |
+
batch_size: 2
|
77 |
+
dim_t: 801
|
78 |
+
num_overlap: 4
|
dereverb-echo_128_4_4_mel_band_roformer_sdr_dry_12.4235.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:aea1752884622db55f6aa28808971088bcd1b833a3c0ec25b3376c6291b97b6d
|
3 |
+
size 128415988
|