-
Notifications
You must be signed in to change notification settings - Fork 1
/
run.sh
executable file
·211 lines (184 loc) · 7.5 KB
/
run.sh
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
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
#!/bin/bash
# Set bash to 'debug' mode, it will exit on :
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
set -e
set -u
set -o pipefail
function xrun () {
set -x
$@
set +x
}
script_dir=$(cd $(dirname ${BASH_SOURCE:-$0}); pwd)
NNSVS_ROOT=$script_dir/../../../
. $script_dir/utils/yaml_parser.sh || exit 1;
eval $(parse_yaml "./config.yaml" "config_")
train_set="train_no_dev"
dev_set="dev"
eval_set="eval"
datasets=($train_set $dev_set $eval_set)
testsets=($dev_set $eval_set)
dumpdir=dump
dump_org_dir=$dumpdir/$config_spk/org
dump_norm_dir=$dumpdir/$config_spk/norm
stage=0
stop_stage=0
. $NNSVS_ROOT/utils/parse_options.sh || exit 1;
# exp name
if [ -z ${config_tag:=} ]; then
expname=${config_spk}
else
expname=${config_spk}_${config_tag}
fi
expdir=exp/$expname
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
if [ ! -e $(eval echo $config_db_root) ]; then
cat<<EOF
stage -1: Downloading
This recipe does not download Natsume_Singing_DB.zip to
provide you the opportunity to read the original license.
Please visit https://amanokei.hatenablog.com/entry/2020/04/30/230003
and read the term of services, and then download the singing voice database
manually.
EOF
fi
fi
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
echo "stage 0: Data preparation"
sh utils/data_prep.sh
mkdir -p data/list
echo "train/dev/eval split"
find data/acoustic/ -type f -name "*.wav" -exec basename {} .wav \; \
| grep -v _low | sort > data/list/utt_list.txt
grep 50 data/list/utt_list.txt > data/list/$eval_set.list
grep 51 data/list/utt_list.txt > data/list/$dev_set.list
grep -v 50 data/list/utt_list.txt | grep -v 51 > data/list/$train_set.list
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
echo "stage 1: Feature generation"
for s in ${datasets[@]};
do
# Natsume_Singing_DB
# Max frequency: 587.3295358348153, Min frequency: 109.99999999999989
nnsvs-prepare-features utt_list=data/list/$s.list out_dir=$dump_org_dir/$s/ \
question_path=$config_question_path
done
# Compute normalization stats for each input/output
mkdir -p $dump_norm_dir
for inout in "in" "out"; do
if [ $inout = "in" ]; then
scaler_class="sklearn.preprocessing.MinMaxScaler"
else
scaler_class="sklearn.preprocessing.StandardScaler"
fi
for typ in timelag duration acoustic;
do
find $dump_org_dir/$train_set/${inout}_${typ} -name "*feats.npy" > train_list.txt
scaler_path=$dump_org_dir/${inout}_${typ}_scaler.joblib
nnsvs-fit-scaler list_path=train_list.txt scaler.class=$scaler_class \
out_path=$scaler_path
rm -f train_list.txt
cp -v $scaler_path $dump_norm_dir/${inout}_${typ}_scaler.joblib
done
done
# apply normalization
for s in ${datasets[@]}; do
for inout in "in" "out"; do
for typ in timelag duration acoustic;
do
nnsvs-preprocess-normalize in_dir=$dump_org_dir/$s/${inout}_${typ}/ \
scaler_path=$dump_org_dir/${inout}_${typ}_scaler.joblib \
out_dir=$dump_norm_dir/$s/${inout}_${typ}/
done
done
done
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
echo "stage 2: Training time-lag model"
if [ ! -z "${config_pretrained_expdir:=}" ]; then
resume_checkpoint=$config_pretrained_expdir/timelag/latest.pth
else
resume_checkpoint=
fi
xrun nnsvs-train data.train_no_dev.in_dir=$dump_norm_dir/$train_set/in_timelag/ \
data.train_no_dev.out_dir=$dump_norm_dir/$train_set/out_timelag/ \
data.dev.in_dir=$dump_norm_dir/$dev_set/in_timelag/ \
data.dev.out_dir=$dump_norm_dir/$dev_set/out_timelag/ \
model=timelag train.out_dir=$expdir/timelag \
data.batch_size=$config_batch_size \
resume.checkpoint=$resume_checkpoint
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
echo "stage 3: Training phoneme duration model"
if [ ! -z "${config_pretrained_expdir:=}" ]; then
resume_checkpoint=$config_pretrained_expdir/duration/latest.pth
else
resume_checkpoint=
fi
xrun nnsvs-train data.train_no_dev.in_dir=$dump_norm_dir/$train_set/in_duration/ \
data.train_no_dev.out_dir=$dump_norm_dir/$train_set/out_duration/ \
data.dev.in_dir=$dump_norm_dir/$dev_set/in_duration/ \
data.dev.out_dir=$dump_norm_dir/$dev_set/out_duration/ \
model=duration train.out_dir=$expdir/duration \
data.batch_size=$config_batch_size \
resume.checkpoint=$resume_checkpoint
fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
echo "stage 4: Training acoustic model"
if [ ! -z "${config_pretrained_expdir:=}" ]; then
resume_checkpoint=$config_pretrained_expdir/acoustic/latest.pth
else
resume_checkpoint=
fi
xrun nnsvs-train data.train_no_dev.in_dir=$dump_norm_dir/$train_set/in_acoustic/ \
data.train_no_dev.out_dir=$dump_norm_dir/$train_set/out_acoustic/ \
data.dev.in_dir=$dump_norm_dir/$dev_set/in_acoustic/ \
data.dev.out_dir=$dump_norm_dir/$dev_set/out_acoustic/ \
model=acoustic train.out_dir=$expdir/acoustic \
data.batch_size=$config_batch_size \
resume.checkpoint=$resume_checkpoint
fi
# NOTE: step 5 does not generate waveform. It just saves neural net's outputs.
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
echo "stage 5: Generation features from timelag/duration/acoustic models"
for s in ${testsets[@]}; do
for typ in timelag duration acoustic; do
checkpoint=$expdir/$typ/latest.pth
name=$(basename $checkpoint)
xrun nnsvs-generate model.checkpoint=$checkpoint \
model.model_yaml=$expdir/$typ/model.yaml \
out_scaler_path=$dump_norm_dir/out_${typ}_scaler.joblib \
in_dir=$dump_norm_dir/$s/in_${typ}/ \
out_dir=$expdir/$typ/predicted/$s/${name%.*}/
done
done
fi
if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
echo "stage 6: Synthesis waveforms"
for s in ${testsets[@]}; do
for input in label_phone_score label_phone_align; do
if [ $input = label_phone_score ]; then
ground_truth_duration=false
else
ground_truth_duration=true
fi
xrun nnsvs-synthesis question_path=$config_question_path \
timelag.checkpoint=$expdir/timelag/latest.pth \
timelag.in_scaler_path=$dump_norm_dir/in_timelag_scaler.joblib \
timelag.out_scaler_path=$dump_norm_dir/out_timelag_scaler.joblib \
timelag.model_yaml=$expdir/timelag/model.yaml \
duration.checkpoint=$expdir/duration/latest.pth \
duration.in_scaler_path=$dump_norm_dir/in_duration_scaler.joblib \
duration.out_scaler_path=$dump_norm_dir/out_duration_scaler.joblib \
duration.model_yaml=$expdir/duration/model.yaml \
acoustic.checkpoint=$expdir/acoustic/latest.pth \
acoustic.in_scaler_path=$dump_norm_dir/in_acoustic_scaler.joblib \
acoustic.out_scaler_path=$dump_norm_dir/out_acoustic_scaler.joblib \
acoustic.model_yaml=$expdir/acoustic/model.yaml \
utt_list=./data/list/$s.list \
in_dir=data/acoustic/$input/ \
out_dir=$expdir/synthesis/$s/latest/$input \
ground_truth_duration=$ground_truth_duration
done
done
fi