subakany
commited on
Commit
•
e915e01
1
Parent(s):
df63b81
initial commit
Browse files- .gitattributes +7 -0
- CKPT.yaml +4 -0
- brain.ckpt +3 -0
- counter.ckpt +3 -0
- decoder.ckpt +3 -0
- encoder.ckpt +3 -0
- example_wham.wav +0 -0
- hyperparams.yaml +175 -0
- lr_scheduler.ckpt +3 -0
- masknet.ckpt +3 -0
- optimizer.ckpt +3 -0
.gitattributes
CHANGED
@@ -25,3 +25,10 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
25 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
26 |
*.zstandard filter=lfs diff=lfs merge=lfs -text
|
27 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
26 |
*.zstandard filter=lfs diff=lfs merge=lfs -text
|
27 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
28 |
+
masknet.ckpt filter=lfs diff=lfs merge=lfs -text
|
29 |
+
optimizer.ckpt filter=lfs diff=lfs merge=lfs -text
|
30 |
+
brain.ckpt filter=lfs diff=lfs merge=lfs -text
|
31 |
+
counter.ckpt filter=lfs diff=lfs merge=lfs -text
|
32 |
+
decoder.ckpt filter=lfs diff=lfs merge=lfs -text
|
33 |
+
encoder.ckpt filter=lfs diff=lfs merge=lfs -text
|
34 |
+
lr_scheduler.ckpt filter=lfs diff=lfs merge=lfs -text
|
CKPT.yaml
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# yamllint disable
|
2 |
+
end-of-epoch: true
|
3 |
+
si-snr: -15.359142046070115
|
4 |
+
unixtime: 1638387953.7461503
|
brain.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d9e24193f36931b7f57932532efbdcf64971f42732383ba6808825f77db258f6
|
3 |
+
size 28
|
counter.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:65a699905c02619370bcf9207f5a477c3d67130ca71ec6f750e07fe8d510b084
|
3 |
+
size 3
|
decoder.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4703b15d23ad5dd4c9b6b93b09539cf0048ba2e58a36c71a62fb860d5b0d343f
|
3 |
+
size 17272
|
encoder.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:6c6b3e53a4061b81b7b0abf6a7faac8ee9714e0509a49ac60d249488e347430c
|
3 |
+
size 17272
|
example_wham.wav
ADDED
Binary file (165 kB). View file
|
|
hyperparams.yaml
ADDED
@@ -0,0 +1,175 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Generated 2021-12-01 from:
|
2 |
+
# /home/mila/s/subakany/speechbrain-soundskrit/recipes/WHAMandWHAMR/enhancement/hparams/sepformer-wham.yaml
|
3 |
+
# yamllint disable
|
4 |
+
# ################################
|
5 |
+
# Model: SepFormer for source separation
|
6 |
+
# https://arxiv.org/abs/2010.13154
|
7 |
+
#
|
8 |
+
# Dataset : WHAM!
|
9 |
+
# ################################
|
10 |
+
# Basic parameters
|
11 |
+
# Seed needs to be set at top of yaml, before objects with parameters are made
|
12 |
+
#
|
13 |
+
seed: 1234
|
14 |
+
__set_seed: !apply:torch.manual_seed [1234]
|
15 |
+
|
16 |
+
# Data params
|
17 |
+
|
18 |
+
# the data folder for the wham dataset
|
19 |
+
# needs to end with wham_original for the wham dataset
|
20 |
+
# respecting this convention effects the code functionality
|
21 |
+
data_folder: /network/tmp1/subakany/wham_original
|
22 |
+
task: enhancement
|
23 |
+
dereverberate: false
|
24 |
+
|
25 |
+
# the path for wsj0/si_tr_s/ folder -- only needed if dynamic mixing is used
|
26 |
+
# e.g. /yourpath/wsj0-processed/si_tr_s/
|
27 |
+
## you need to convert the original wsj0 to 8k
|
28 |
+
# you can do this conversion with the script ../meta/preprocess_dynamic_mixing.py
|
29 |
+
base_folder_dm: /network/tmp1/subakany/wsj0-processed/si_tr_s/
|
30 |
+
|
31 |
+
experiment_name: sepformer-wham-enhancement
|
32 |
+
output_folder: results/sepformer-wham-enhancement/1234
|
33 |
+
train_log: results/sepformer-wham-enhancement/1234/train_log.txt
|
34 |
+
save_folder: results/sepformer-wham-enhancement/1234/save
|
35 |
+
|
36 |
+
# the file names should start with whamr instead of whamorg
|
37 |
+
train_data: results/sepformer-wham-enhancement/1234/save/whamorg_tr.csv
|
38 |
+
valid_data: results/sepformer-wham-enhancement/1234/save/whamorg_cv.csv
|
39 |
+
test_data: results/sepformer-wham-enhancement/1234/save/whamorg_tt.csv
|
40 |
+
skip_prep: false
|
41 |
+
|
42 |
+
|
43 |
+
# Experiment params
|
44 |
+
auto_mix_prec: true # Set it to True for mixed precision
|
45 |
+
test_only: true
|
46 |
+
num_spks: 1 # set to 3 for wsj0-3mix
|
47 |
+
progressbar: true
|
48 |
+
save_audio: false # Save estimated sources on disk
|
49 |
+
sample_rate: 8000
|
50 |
+
|
51 |
+
# Training parameters
|
52 |
+
N_epochs: 200
|
53 |
+
batch_size: 1
|
54 |
+
lr: 0.00015
|
55 |
+
clip_grad_norm: 5
|
56 |
+
loss_upper_lim: 999999 # this is the upper limit for an acceptable loss
|
57 |
+
# if True, the training sequences are cut to a specified length
|
58 |
+
limit_training_signal_len: false
|
59 |
+
# this is the length of sequences if we choose to limit
|
60 |
+
# the signal length of training sequences
|
61 |
+
training_signal_len: 32000000
|
62 |
+
|
63 |
+
# Set it to True to dynamically create mixtures at training time
|
64 |
+
dynamic_mixing: true
|
65 |
+
|
66 |
+
# Parameters for data augmentation
|
67 |
+
use_wavedrop: false
|
68 |
+
use_speedperturb: true
|
69 |
+
use_speedperturb_sameforeachsource: false
|
70 |
+
use_rand_shift: false
|
71 |
+
min_shift: -8000
|
72 |
+
max_shift: 8000
|
73 |
+
|
74 |
+
speedperturb: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
|
75 |
+
perturb_prob: 1.0
|
76 |
+
drop_freq_prob: 0.0
|
77 |
+
drop_chunk_prob: 0.0
|
78 |
+
sample_rate: 8000
|
79 |
+
speeds: [95, 100, 105]
|
80 |
+
|
81 |
+
wavedrop: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
|
82 |
+
perturb_prob: 0.0
|
83 |
+
drop_freq_prob: 1.0
|
84 |
+
drop_chunk_prob: 1.0
|
85 |
+
sample_rate: 8000
|
86 |
+
|
87 |
+
# loss thresholding -- this thresholds the training loss
|
88 |
+
threshold_byloss: true
|
89 |
+
threshold: -30
|
90 |
+
|
91 |
+
# Encoder parameters
|
92 |
+
N_encoder_out: 256
|
93 |
+
out_channels: 256
|
94 |
+
kernel_size: 16
|
95 |
+
kernel_stride: 8
|
96 |
+
|
97 |
+
# Dataloader options
|
98 |
+
dataloader_opts:
|
99 |
+
batch_size: 1
|
100 |
+
num_workers: 3
|
101 |
+
|
102 |
+
|
103 |
+
# Specifying the network
|
104 |
+
Encoder: &id003 !new:speechbrain.lobes.models.dual_path.Encoder
|
105 |
+
kernel_size: 16
|
106 |
+
out_channels: 256
|
107 |
+
|
108 |
+
|
109 |
+
SBtfintra: &id001 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
|
110 |
+
num_layers: 8
|
111 |
+
d_model: 256
|
112 |
+
nhead: 8
|
113 |
+
d_ffn: 1024
|
114 |
+
dropout: 0
|
115 |
+
use_positional_encoding: true
|
116 |
+
norm_before: true
|
117 |
+
|
118 |
+
SBtfinter: &id002 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
|
119 |
+
num_layers: 8
|
120 |
+
d_model: 256
|
121 |
+
nhead: 8
|
122 |
+
d_ffn: 1024
|
123 |
+
dropout: 0
|
124 |
+
use_positional_encoding: true
|
125 |
+
norm_before: true
|
126 |
+
|
127 |
+
MaskNet: &id005 !new:speechbrain.lobes.models.dual_path.Dual_Path_Model
|
128 |
+
|
129 |
+
num_spks: 1
|
130 |
+
in_channels: 256
|
131 |
+
out_channels: 256
|
132 |
+
num_layers: 2
|
133 |
+
K: 250
|
134 |
+
intra_model: *id001
|
135 |
+
inter_model: *id002
|
136 |
+
norm: ln
|
137 |
+
linear_layer_after_inter_intra: false
|
138 |
+
skip_around_intra: true
|
139 |
+
|
140 |
+
Decoder: &id004 !new:speechbrain.lobes.models.dual_path.Decoder
|
141 |
+
in_channels: 256
|
142 |
+
out_channels: 1
|
143 |
+
kernel_size: 16
|
144 |
+
stride: 8
|
145 |
+
bias: false
|
146 |
+
|
147 |
+
optimizer: !name:torch.optim.Adam
|
148 |
+
lr: 0.00015
|
149 |
+
weight_decay: 0
|
150 |
+
|
151 |
+
loss: !name:speechbrain.nnet.losses.get_si_snr_with_pitwrapper
|
152 |
+
|
153 |
+
lr_scheduler: &id007 !new:speechbrain.nnet.schedulers.ReduceLROnPlateau
|
154 |
+
|
155 |
+
factor: 0.5
|
156 |
+
patience: 2
|
157 |
+
dont_halve_until_epoch: 65
|
158 |
+
|
159 |
+
epoch_counter: &id006 !new:speechbrain.utils.epoch_loop.EpochCounter
|
160 |
+
limit: 200
|
161 |
+
|
162 |
+
modules:
|
163 |
+
encoder: *id003
|
164 |
+
decoder: *id004
|
165 |
+
masknet: *id005
|
166 |
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
|
167 |
+
checkpoints_dir: results/sepformer-wham-enhancement/1234/save
|
168 |
+
recoverables:
|
169 |
+
encoder: *id003
|
170 |
+
decoder: *id004
|
171 |
+
masknet: *id005
|
172 |
+
counter: *id006
|
173 |
+
lr_scheduler: *id007
|
174 |
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
|
175 |
+
save_file: results/sepformer-wham-enhancement/1234/train_log.txt
|
lr_scheduler.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:38d5d2d95d97c05f614716f700a35ac53b1a117d96c1a89ea8e2b7b4ab8ebb1b
|
3 |
+
size 1583
|
masknet.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:111312d682aee6b72610b83edc0dcf253d7a62f745136cd2828113fb75fbb6e4
|
3 |
+
size 112849478
|
optimizer.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c17d266b227ee5e2a9b09d1c87d40ebc93c6883e32c1541e7d6605c9890d2373
|
3 |
+
size 205168377
|