-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathconfig.txt
110 lines (88 loc) · 2.73 KB
/
config.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
# ----- GENERAL PARAMETERS ---------------------------
tag: test
comment: test
architecture: SourceFilterMixtureAutoencoder2
# wst-model = trained_model
# output
# -------- DATASET PARAMETERS -----------------------
dataset = BCBQ
one-song = True
one-example = False
one-batch = False
# parallel = True
# confidence-threshold = 0.4
samplerate = 16000
example-length = 64000
# crepe-hop-size = 256
# crepe-center = True
f0-cuesta = False
voices = satb # satb
# pre-emphasis-factor = 0.95
# split = train # if dataset = nus
# source2-offset: 0.8 # if dataset = nus_mix
# ---- CSD ----
# train-song = Nino Dios
# val-song = Locus Iste
# ---- Cantoria ----
# train-song = CEA # train_songs = ['LBM1', 'LBM2', 'LJT1', 'LJT2', 'LNG', 'RRC', 'SSS', 'HCB']
# val-song = EJB1 # val_songs = ['THM', 'VBP', 'YSM']
# -------- TRAINING PARAMETERS -----------------------
epochs = 8000
batch-size = 15
lr = 0.0001
patience = 200
lr-decay-patience = 80
lr-decay-gamma = 0.98
weight-decay = 0.00001
seed = 4
nb-workers = 10
quiet = False
no-cuda = False
supervised = False
reconstruction-loss-weight = 1
loss-nfft = [2048, 1024, 512, 256, 128, 64]
loss-mag-weight = 1
loss-logmag-weight = 1
loss-logmel-weight = 0
loss-delta-freq-weight = 0
loss-delta-time-weight = 0
loss-lsf-weight = 0 # LSF regularization loss
loss-saliences-weight = 10000
loss-voices-weight = 100000
loss-comittment-weight = 10000
loss-f0-weight = 1
loss-1voice-per-salience-weight = 0
ss-loss-weight = 0
harmonic-amp-loss-weight = 1
f0-hz-loss-weight = 0
harmonics-roll-off-loss-weight = 0
lsf-loss-weight = 10 # self supervision loss
noise-gain-loss-weight = 100
noise-mags-loss-weight = 1
# -------- MODEL PARAMETERS ------------------------
nfft = 512
nhop = 256
filter-order = 20 # for vocal tract filter in source filter model
noise-filter-mags = 40 # for noise filter in source filter model
# nb-filter-magnitudes = 65 # for noise filter in harmonics plus noise model
encoder = MixEncoderSimple
n-sources = 4 # this parameter also determines the number of sources drawn from 'voices' in CSD dataset
encoder-hidden-size = 256
embedding-size = 128
decoder-hidden-size = 512
decoder-output-size = 512
unidirectional = True
estimate-lsf = True
voiced-unvoiced-same-noise = True
harmonic-roll-off = 6 # estimated by model if set to -1 (time-varying) or to -2 (time-invariant)
estimate-noise-mags = True # estimate constant noise shape for voiced frames
# estimate-f0 = False
# supervised-f0 = False
# switch-off-noise = False
f-ref-source-spec = 200
# source-spectrum = flat
return-sources = True
F0-models = True
F0-models-trainable = True
MultiF0-estimator-trainable = False
method = reconstruction # other possibilities: sigmoid, amplitude, ste, reconstruction, softmax, threshold