Commit
•
6ceca85
1
Parent(s):
ac23d45
Update hyperparams.yaml
Browse files- hyperparams.yaml +1 -119
hyperparams.yaml
CHANGED
@@ -1,92 +1,10 @@
|
|
1 |
-
# Generated 2022-06-30 from:
|
2 |
-
# /home/cem/Dropbox/speechbrain-1/recipes/WHAMandWHAMR/enhancement/yamls/sepformer-wham-16k.yaml
|
3 |
-
# yamllint disable
|
4 |
# ################################
|
5 |
-
# Model: SepFormer for
|
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: /data2/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 ../meta/preprocess_dynamic_mixing.py
|
29 |
-
base_folder_dm: /yourpath/wsj0-processed/si_tr_s/
|
30 |
-
|
31 |
-
experiment_name: sepformer-wham-16k
|
32 |
-
output_folder: results/sepformer-wham-16k/1234
|
33 |
-
train_log: results/sepformer-wham-16k/1234/train_log.txt
|
34 |
-
save_folder: results/sepformer-wham-16k/1234/save
|
35 |
-
|
36 |
-
# the file names should start with whamr instead of whamorg
|
37 |
-
train_data: results/sepformer-wham-16k/1234/save/whamorg_tr.csv
|
38 |
-
valid_data: results/sepformer-wham-16k/1234/save/whamorg_cv.csv
|
39 |
-
test_data: results/sepformer-wham-16k/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 |
-
noprogressbar: false
|
48 |
-
save_audio: true # Save estimated sources on disk
|
49 |
sample_rate: 16000
|
50 |
-
n_audio_to_save: 20
|
51 |
-
|
52 |
-
# Training parameters
|
53 |
-
N_epochs: 200
|
54 |
-
batch_size: 1
|
55 |
-
lr: 0.00015
|
56 |
-
clip_grad_norm: 5
|
57 |
-
loss_upper_lim: 999999 # this is the upper limit for an acceptable loss
|
58 |
-
# if True, the training sequences are cut to a specified length
|
59 |
-
limit_training_signal_len: true
|
60 |
-
# this is the length of sequences if we choose to limit
|
61 |
-
# the signal length of training sequences
|
62 |
-
training_signal_len: 64000
|
63 |
-
|
64 |
-
# Set it to True to dynamically create mixtures at training time
|
65 |
-
dynamic_mixing: false
|
66 |
-
|
67 |
-
# Parameters for data augmentation
|
68 |
-
use_wavedrop: false
|
69 |
-
use_speedperturb: true
|
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: 16000
|
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: 16000
|
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
|
@@ -94,21 +12,11 @@ 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 |
-
dataloader_opts_valid:
|
103 |
-
batch_size: 1
|
104 |
-
num_workers: 3
|
105 |
-
|
106 |
# Specifying the network
|
107 |
Encoder: &id003 !new:speechbrain.lobes.models.dual_path.Encoder
|
108 |
kernel_size: 16
|
109 |
out_channels: 256
|
110 |
|
111 |
-
|
112 |
SBtfintra: &id001 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
|
113 |
num_layers: 8
|
114 |
d_model: 256
|
@@ -128,7 +36,6 @@ SBtfinter: &id002 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
|
|
128 |
norm_before: true
|
129 |
|
130 |
MaskNet: &id005 !new:speechbrain.lobes.models.dual_path.Dual_Path_Model
|
131 |
-
|
132 |
num_spks: 1
|
133 |
in_channels: 256
|
134 |
out_channels: 256
|
@@ -147,35 +54,10 @@ Decoder: &id004 !new:speechbrain.lobes.models.dual_path.Decoder
|
|
147 |
stride: 8
|
148 |
bias: false
|
149 |
|
150 |
-
optimizer: !name:torch.optim.Adam
|
151 |
-
lr: 0.00015
|
152 |
-
weight_decay: 0
|
153 |
-
|
154 |
-
loss: !name:speechbrain.nnet.losses.get_si_snr_with_pitwrapper
|
155 |
-
|
156 |
-
lr_scheduler: &id007 !new:speechbrain.nnet.schedulers.ReduceLROnPlateau
|
157 |
-
|
158 |
-
factor: 0.5
|
159 |
-
patience: 2
|
160 |
-
dont_halve_until_epoch: 65
|
161 |
-
|
162 |
-
epoch_counter: &id006 !new:speechbrain.utils.epoch_loop.EpochCounter
|
163 |
-
limit: 200
|
164 |
-
|
165 |
modules:
|
166 |
encoder: *id003
|
167 |
decoder: *id004
|
168 |
masknet: *id005
|
169 |
-
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
|
170 |
-
checkpoints_dir: results/sepformer-wham-16k/1234/save
|
171 |
-
recoverables:
|
172 |
-
encoder: *id003
|
173 |
-
decoder: *id004
|
174 |
-
masknet: *id005
|
175 |
-
counter: *id006
|
176 |
-
lr_scheduler: *id007
|
177 |
-
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
|
178 |
-
save_file: results/sepformer-wham-16k/1234/train_log.txt
|
179 |
|
180 |
pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
|
181 |
loadables:
|
|
|
|
|
|
|
|
|
1 |
# ################################
|
2 |
+
# Model: Pretrained SepFormer for speech enhancement
|
|
|
|
|
3 |
# Dataset : WHAM!
|
4 |
# ################################
|
|
|
|
|
|
|
|
|
|
|
5 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
num_spks: 1 # set to 3 for wsj0-3mix
|
|
|
|
|
7 |
sample_rate: 16000
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
# Encoder parameters
|
10 |
N_encoder_out: 256
|
|
|
12 |
kernel_size: 16
|
13 |
kernel_stride: 8
|
14 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
# Specifying the network
|
16 |
Encoder: &id003 !new:speechbrain.lobes.models.dual_path.Encoder
|
17 |
kernel_size: 16
|
18 |
out_channels: 256
|
19 |
|
|
|
20 |
SBtfintra: &id001 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
|
21 |
num_layers: 8
|
22 |
d_model: 256
|
|
|
36 |
norm_before: true
|
37 |
|
38 |
MaskNet: &id005 !new:speechbrain.lobes.models.dual_path.Dual_Path_Model
|
|
|
39 |
num_spks: 1
|
40 |
in_channels: 256
|
41 |
out_channels: 256
|
|
|
54 |
stride: 8
|
55 |
bias: false
|
56 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
modules:
|
58 |
encoder: *id003
|
59 |
decoder: *id004
|
60 |
masknet: *id005
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
|
62 |
pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
|
63 |
loadables:
|