mirco commited on
Commit
98d299b
1 Parent(s): c5c31bd

cleaned inference hyparam file

Browse files
Files changed (2) hide show
  1. hyperparams.yaml +23 -121
  2. hyperparams_train.yaml +163 -0
hyperparams.yaml CHANGED
@@ -1,92 +1,20 @@
1
- # Generated 2021-02-26 from:
2
- # /scratch/csubakan/speechbrain_new/recipes/WSJ2Mix/separation/yamls/dptransformer78.yaml
3
- # yamllint disable
4
  # ################################
5
- # Model: SepFormer for source separation
6
  # https://arxiv.org/abs/2010.13154
7
- #
8
- # Dataset : WSJ0-mix
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
- data_folder: /localscratch/csubakan.62709298.0/wsj0-mix/2speakers # wsj2mix or wsj3mix
18
- experiment_name: 78-speedchange-dynamicmix-hardcodegaussian
19
- output_folder: results/78-speedchange-dynamicmix-hardcodegaussian/1234
20
- train_log: results/78-speedchange-dynamicmix-hardcodegaussian/1234/train_log.txt
21
- save_folder: results/78-speedchange-dynamicmix-hardcodegaussian/1234/save
22
- train_data: results/78-speedchange-dynamicmix-hardcodegaussian/1234/save/wsj_tr.csv
23
- valid_data: results/78-speedchange-dynamicmix-hardcodegaussian/1234/save/wsj_cv.csv
24
- test_data: results/78-speedchange-dynamicmix-hardcodegaussian/1234/save/wsj_tt.csv
25
- wsj0_tr: /localscratch/csubakan.62709298.0/wsj0-processed/si_tr_s/
26
 
27
- # Experiment params
28
- auto_mix_prec: true
29
- test_only: false
30
- num_spks: 2 # set to 3 for wsj0-3mix
31
- progressbar: true
32
- save_audio: false # Save estimated sources on disk
33
  sample_rate: 8000
 
34
 
35
- # Training parameters
36
- N_epochs: 200
37
- batch_size: 1
38
- lr: 0.00015
39
- clip_grad_norm: 5
40
- loss_upper_lim: 999999 # this is the upper limit for an acceptable loss
41
- # if True, the training sequences are cut to a specified length
42
- limit_training_signal_len: false
43
- # this is the length of sequences if we choose to limit
44
- # the signal length of training sequences
45
- training_signal_len: 128000
46
- dynamic_mixing: regular
47
-
48
- # Augment parameters
49
- use_wavedrop: false
50
- use_speedperturb: true
51
- use_speedperturb_sameforeachsource: false
52
- use_rand_shift: false
53
- min_shift: -8000
54
- max_shift: 8000
55
-
56
- # Neural parameters
57
- N_encoder_out: 256
58
- out_channels: 256
59
- kernel_size: 16
60
- kernel_stride: 8
61
-
62
- threshold_byloss: true
63
- threshold: -30
64
-
65
- # Dataloader options
66
- dataloader_opts:
67
- batch_size: 1
68
- num_workers: 3
69
-
70
- speedperturb: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
71
- perturb_prob: 1.0
72
- drop_freq_prob: 0.0
73
- drop_chunk_prob: 0.0
74
- sample_rate: 8000
75
- speeds: [95, 100, 105]
76
-
77
- wavedrop: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
78
- perturb_prob: 0.0
79
- drop_freq_prob: 1.0
80
- drop_chunk_prob: 1.0
81
- sample_rate: 8000
82
-
83
-
84
- Encoder: &id003 !new:speechbrain.lobes.models.dual_path.Encoder
85
  kernel_size: 16
86
  out_channels: 256
87
 
88
-
89
- SBtfintra: &id001 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
90
  num_layers: 8
91
  d_model: 256
92
  nhead: 8
@@ -95,7 +23,7 @@ SBtfintra: &id001 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
95
  use_positional_encoding: true
96
  norm_before: true
97
 
98
- SBtfinter: &id002 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
99
  num_layers: 8
100
  d_model: 256
101
  nhead: 8
@@ -104,60 +32,34 @@ SBtfinter: &id002 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
104
  use_positional_encoding: true
105
  norm_before: true
106
 
107
- MaskNet: &id005 !new:speechbrain.lobes.models.dual_path.Dual_Path_Model
108
-
109
- num_spks: 2
110
  in_channels: 256
111
  out_channels: 256
112
  num_layers: 2
113
  K: 250
114
- intra_model: *id001
115
- inter_model: *id002
116
  norm: ln
117
  linear_layer_after_inter_intra: false
118
  skip_around_intra: true
119
 
120
- Decoder: &id004 !new:speechbrain.lobes.models.dual_path.Decoder
121
  in_channels: 256
122
  out_channels: 1
123
  kernel_size: 16
124
  stride: 8
125
  bias: false
126
 
127
- optimizer: !name:torch.optim.Adam
128
- lr: 0.00015
129
- weight_decay: 0
130
-
131
- loss: !name:speechbrain.nnet.losses.get_si_snr_with_pitwrapper
132
-
133
- lr_scheduler: &id007 !new:speechbrain.nnet.schedulers.ReduceLROnPlateau
134
-
135
- factor: 0.5
136
- patience: 4
137
- dont_halve_until_epoch: 100
138
-
139
- epoch_counter: &id006 !new:speechbrain.utils.epoch_loop.EpochCounter
140
- limit: 200
141
-
142
  modules:
143
- encoder: *id003
144
- decoder: *id004
145
- masknet: *id005
146
- checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
147
- checkpoints_dir: results/78-speedchange-dynamicmix-hardcodegaussian/1234/save
148
- recoverables:
149
- encoder: *id003
150
- decoder: *id004
151
- masknet: *id005
152
- counter: *id006
153
- lr_scheduler: *id007
154
- train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
155
- save_file: results/78-speedchange-dynamicmix-hardcodegaussian/1234/train_log.txt
156
 
 
 
 
 
 
157
 
158
 
159
- pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
160
- loadables:
161
- masknet: !ref <MaskNet>
162
- encoder: !ref <Encoder>
163
- decoder: !ref <Decoder>
 
 
 
 
1
  # ################################
2
+ # Model: Inference for source separation with SepFormer
3
  # https://arxiv.org/abs/2010.13154
4
+ # Generated from speechbrain/recipes/WSJ0Mix/separation/train/hparams/sepformer-wsj02mix.yaml
5
+ # Dataset : wsj02mix
6
+ # ###############################
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7
 
8
+ # Parameters
 
 
 
 
 
9
  sample_rate: 8000
10
+ num_spks: 2
11
 
12
+ # Specifying the network
13
+ Encoder: !new:speechbrain.lobes.models.dual_path.Encoder
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  kernel_size: 16
15
  out_channels: 256
16
 
17
+ SBtfintra: !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
 
18
  num_layers: 8
19
  d_model: 256
20
  nhead: 8
 
23
  use_positional_encoding: true
24
  norm_before: true
25
 
26
+ SBtfinter: !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
27
  num_layers: 8
28
  d_model: 256
29
  nhead: 8
 
32
  use_positional_encoding: true
33
  norm_before: true
34
 
35
+ MaskNet: !new:speechbrain.lobes.models.dual_path.Dual_Path_Model
36
+ num_spks: !ref <num_spks>
 
37
  in_channels: 256
38
  out_channels: 256
39
  num_layers: 2
40
  K: 250
41
+ intra_model: !ref <SBtfintra>
42
+ inter_model: !ref <SBtfinter>
43
  norm: ln
44
  linear_layer_after_inter_intra: false
45
  skip_around_intra: true
46
 
47
+ Decoder: !new:speechbrain.lobes.models.dual_path.Decoder
48
  in_channels: 256
49
  out_channels: 1
50
  kernel_size: 16
51
  stride: 8
52
  bias: false
53
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54
  modules:
55
+ encoder: !ref <Encoder>
56
+ decoder: !ref <Decoder>
57
+ masknet: !ref <MaskNet>
 
 
 
 
 
 
 
 
 
 
58
 
59
+ pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
60
+ loadables:
61
+ masknet: !ref <MaskNet>
62
+ encoder: !ref <Encoder>
63
+ decoder: !ref <Decoder>
64
 
65
 
 
 
 
 
 
hyperparams_train.yaml ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Generated 2021-02-26 from:
2
+ # /scratch/csubakan/speechbrain_new/recipes/WSJ2Mix/separation/yamls/dptransformer78.yaml
3
+ # yamllint disable
4
+ # ################################
5
+ # Model: SepFormer for source separation
6
+ # https://arxiv.org/abs/2010.13154
7
+ #
8
+ # Dataset : WSJ0-mix
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
+ data_folder: /localscratch/csubakan.62709298.0/wsj0-mix/2speakers # wsj2mix or wsj3mix
18
+ experiment_name: 78-speedchange-dynamicmix-hardcodegaussian
19
+ output_folder: results/78-speedchange-dynamicmix-hardcodegaussian/1234
20
+ train_log: results/78-speedchange-dynamicmix-hardcodegaussian/1234/train_log.txt
21
+ save_folder: results/78-speedchange-dynamicmix-hardcodegaussian/1234/save
22
+ train_data: results/78-speedchange-dynamicmix-hardcodegaussian/1234/save/wsj_tr.csv
23
+ valid_data: results/78-speedchange-dynamicmix-hardcodegaussian/1234/save/wsj_cv.csv
24
+ test_data: results/78-speedchange-dynamicmix-hardcodegaussian/1234/save/wsj_tt.csv
25
+ wsj0_tr: /localscratch/csubakan.62709298.0/wsj0-processed/si_tr_s/
26
+
27
+ # Experiment params
28
+ auto_mix_prec: true
29
+ test_only: false
30
+ num_spks: 2 # set to 3 for wsj0-3mix
31
+ progressbar: true
32
+ save_audio: false # Save estimated sources on disk
33
+ sample_rate: 8000
34
+
35
+ # Training parameters
36
+ N_epochs: 200
37
+ batch_size: 1
38
+ lr: 0.00015
39
+ clip_grad_norm: 5
40
+ loss_upper_lim: 999999 # this is the upper limit for an acceptable loss
41
+ # if True, the training sequences are cut to a specified length
42
+ limit_training_signal_len: false
43
+ # this is the length of sequences if we choose to limit
44
+ # the signal length of training sequences
45
+ training_signal_len: 128000
46
+ dynamic_mixing: regular
47
+
48
+ # Augment parameters
49
+ use_wavedrop: false
50
+ use_speedperturb: true
51
+ use_speedperturb_sameforeachsource: false
52
+ use_rand_shift: false
53
+ min_shift: -8000
54
+ max_shift: 8000
55
+
56
+ # Neural parameters
57
+ N_encoder_out: 256
58
+ out_channels: 256
59
+ kernel_size: 16
60
+ kernel_stride: 8
61
+
62
+ threshold_byloss: true
63
+ threshold: -30
64
+
65
+ # Dataloader options
66
+ dataloader_opts:
67
+ batch_size: 1
68
+ num_workers: 3
69
+
70
+ speedperturb: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
71
+ perturb_prob: 1.0
72
+ drop_freq_prob: 0.0
73
+ drop_chunk_prob: 0.0
74
+ sample_rate: 8000
75
+ speeds: [95, 100, 105]
76
+
77
+ wavedrop: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
78
+ perturb_prob: 0.0
79
+ drop_freq_prob: 1.0
80
+ drop_chunk_prob: 1.0
81
+ sample_rate: 8000
82
+
83
+
84
+ Encoder: &id003 !new:speechbrain.lobes.models.dual_path.Encoder
85
+ kernel_size: 16
86
+ out_channels: 256
87
+
88
+
89
+ SBtfintra: &id001 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
90
+ num_layers: 8
91
+ d_model: 256
92
+ nhead: 8
93
+ d_ffn: 1024
94
+ dropout: 0
95
+ use_positional_encoding: true
96
+ norm_before: true
97
+
98
+ SBtfinter: &id002 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
99
+ num_layers: 8
100
+ d_model: 256
101
+ nhead: 8
102
+ d_ffn: 1024
103
+ dropout: 0
104
+ use_positional_encoding: true
105
+ norm_before: true
106
+
107
+ MaskNet: &id005 !new:speechbrain.lobes.models.dual_path.Dual_Path_Model
108
+
109
+ num_spks: 2
110
+ in_channels: 256
111
+ out_channels: 256
112
+ num_layers: 2
113
+ K: 250
114
+ intra_model: *id001
115
+ inter_model: *id002
116
+ norm: ln
117
+ linear_layer_after_inter_intra: false
118
+ skip_around_intra: true
119
+
120
+ Decoder: &id004 !new:speechbrain.lobes.models.dual_path.Decoder
121
+ in_channels: 256
122
+ out_channels: 1
123
+ kernel_size: 16
124
+ stride: 8
125
+ bias: false
126
+
127
+ optimizer: !name:torch.optim.Adam
128
+ lr: 0.00015
129
+ weight_decay: 0
130
+
131
+ loss: !name:speechbrain.nnet.losses.get_si_snr_with_pitwrapper
132
+
133
+ lr_scheduler: &id007 !new:speechbrain.nnet.schedulers.ReduceLROnPlateau
134
+
135
+ factor: 0.5
136
+ patience: 4
137
+ dont_halve_until_epoch: 100
138
+
139
+ epoch_counter: &id006 !new:speechbrain.utils.epoch_loop.EpochCounter
140
+ limit: 200
141
+
142
+ modules:
143
+ encoder: *id003
144
+ decoder: *id004
145
+ masknet: *id005
146
+ checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
147
+ checkpoints_dir: results/78-speedchange-dynamicmix-hardcodegaussian/1234/save
148
+ recoverables:
149
+ encoder: *id003
150
+ decoder: *id004
151
+ masknet: *id005
152
+ counter: *id006
153
+ lr_scheduler: *id007
154
+ train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
155
+ save_file: results/78-speedchange-dynamicmix-hardcodegaussian/1234/train_log.txt
156
+
157
+
158
+
159
+ pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
160
+ loadables:
161
+ masknet: !ref <MaskNet>
162
+ encoder: !ref <Encoder>
163
+ decoder: !ref <Decoder>