cemsubakan commited on
Commit
2ef3a73
1 Parent(s): 719fc92

hparams cleanup

Browse files
Files changed (1) hide show
  1. hyperparams.yaml +2 -110
hyperparams.yaml CHANGED
@@ -1,105 +1,24 @@
1
- # Generated 2021-07-13 from:
2
- # /home/mila/s/subakany/speechbrain_new/recipes/LibriMix/separation/hparams/sepformer-libri2mix.yaml
3
- # yamllint disable
4
  # ################################
5
- # Model: SepFormer for source separation
6
- # https://arxiv.org/abs/2010.13154
7
  # Dataset : Libri2mix
8
  # ################################
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
- # e.g. '/yourpath/Libri2Mix/train-clean-360/'
19
- # the data folder is needed even if dynamic mixing is applied
20
- data_folder: /miniscratch/subakany/LibriMixData_new/Libri2Mix/
21
-
22
- # this is the base folder for dynamic mixing
23
- base_folder_dm: /miniscratch/subakany/LibriMixData_new/LibriSpeech/train-clean-360_processed/
24
-
25
- experiment_name: sepformer-libri2mix
26
- output_folder: results/sepformer-libri2mix/1234
27
- train_log: results/sepformer-libri2mix/1234/train_log.txt
28
- save_folder: results/sepformer-libri2mix/1234/save
29
- train_data: results/sepformer-libri2mix/1234/save/libri2mix_train-360.csv
30
- valid_data: results/sepformer-libri2mix/1234/save/libri2mix_dev.csv
31
- test_data: results/sepformer-libri2mix/1234/save/libri2mix_test.csv
32
- skip_prep: false
33
-
34
- ckpt_interval_minutes: 60
35
 
36
  # Experiment params
37
- auto_mix_prec: true # Set it to True for mixed precision
38
- test_only: true
39
  num_spks: 2
40
- progressbar: true
41
- save_audio: false # Save estimated sources on disk
42
  sample_rate: 8000
43
 
44
- # Training parameters
45
- N_epochs: 200
46
- batch_size: 1
47
- lr: 0.00015
48
- clip_grad_norm: 5
49
- loss_upper_lim: 999999 # this is the upper limit for an acceptable loss
50
- # if True, the training sequences are cut to a specified length
51
- limit_training_signal_len: false
52
- # this is the length of sequences if we choose to limit
53
- # the signal length of training sequences
54
- training_signal_len: 32000000
55
-
56
- # Set it to True to dynamically create mixtures at training time
57
- dynamic_mixing: true
58
- use_wham_noise: false
59
-
60
- # Parameters for data augmentation
61
- use_wavedrop: false
62
- use_speedperturb: true
63
- use_speedperturb_sameforeachsource: false
64
- use_rand_shift: false
65
- min_shift: -8000
66
- max_shift: 8000
67
-
68
- speedperturb: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
69
- perturb_prob: 1.0
70
- drop_freq_prob: 0.0
71
- drop_chunk_prob: 0.0
72
- sample_rate: 8000
73
- speeds: [95, 100, 105]
74
-
75
- wavedrop: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
76
- perturb_prob: 0.0
77
- drop_freq_prob: 1.0
78
- drop_chunk_prob: 1.0
79
- sample_rate: 8000
80
-
81
- # loss thresholding -- this thresholds the training loss
82
- threshold_byloss: true
83
- threshold: -30
84
-
85
  # Encoder parameters
86
  N_encoder_out: 256
87
  out_channels: 256
88
  kernel_size: 16
89
  kernel_stride: 8
90
 
91
- # Dataloader options
92
- dataloader_opts:
93
- batch_size: 1
94
- num_workers: 0
95
-
96
-
97
  # Specifying the network
98
  Encoder: &id003 !new:speechbrain.lobes.models.dual_path.Encoder
99
  kernel_size: 16
100
  out_channels: 256
101
 
102
-
103
  SBtfintra: &id001 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
104
  num_layers: 8
105
  d_model: 256
@@ -119,7 +38,6 @@ SBtfinter: &id002 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
119
  norm_before: true
120
 
121
  MaskNet: &id005 !new:speechbrain.lobes.models.dual_path.Dual_Path_Model
122
-
123
  num_spks: 2
124
  in_channels: 256
125
  out_channels: 256
@@ -138,36 +56,10 @@ Decoder: &id004 !new:speechbrain.lobes.models.dual_path.Decoder
138
  stride: 8
139
  bias: false
140
 
141
- optimizer: !name:torch.optim.Adam
142
- lr: 0.00015
143
- weight_decay: 0
144
-
145
- loss: !name:speechbrain.nnet.losses.get_si_snr_with_pitwrapper
146
-
147
- lr_scheduler: !new:speechbrain.nnet.schedulers.ReduceLROnPlateau
148
- factor: 0.5
149
- patience: 2
150
- dont_halve_until_epoch: 5
151
-
152
- epoch_counter: &id006 !new:speechbrain.utils.epoch_loop.EpochCounter
153
- # lr_scheduler: !ref <lr_scheduler>
154
-
155
- limit: 200
156
-
157
  modules:
158
  encoder: *id003
159
  decoder: *id004
160
  masknet: *id005
161
- checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
162
- checkpoints_dir: results/sepformer-libri2mix/1234/save
163
- recoverables:
164
- encoder: *id003
165
- decoder: *id004
166
- masknet: *id005
167
- counter: *id006
168
- train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
169
- save_file: results/sepformer-libri2mix/1234/train_log.txt
170
-
171
 
172
  pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
173
  loadables:
 
1
+
 
 
2
  # ################################
3
+ # Model: Pretrained SepFormer for source separation
 
4
  # Dataset : Libri2mix
5
  # ################################
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6
 
7
  # Experiment params
 
 
8
  num_spks: 2
 
 
9
  sample_rate: 8000
10
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11
  # Encoder parameters
12
  N_encoder_out: 256
13
  out_channels: 256
14
  kernel_size: 16
15
  kernel_stride: 8
16
 
 
 
 
 
 
 
17
  # Specifying the network
18
  Encoder: &id003 !new:speechbrain.lobes.models.dual_path.Encoder
19
  kernel_size: 16
20
  out_channels: 256
21
 
 
22
  SBtfintra: &id001 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
23
  num_layers: 8
24
  d_model: 256
 
38
  norm_before: true
39
 
40
  MaskNet: &id005 !new:speechbrain.lobes.models.dual_path.Dual_Path_Model
 
41
  num_spks: 2
42
  in_channels: 256
43
  out_channels: 256
 
56
  stride: 8
57
  bias: false
58
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
59
  modules:
60
  encoder: *id003
61
  decoder: *id004
62
  masknet: *id005
 
 
 
 
 
 
 
 
 
 
63
 
64
  pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
65
  loadables: