Adel-Moumen commited on
Commit
8f98994
1 Parent(s): a22900d

Update hyperparams.yaml

Browse files
Files changed (1) hide show
  1. hyperparams.yaml +23 -9
hyperparams.yaml CHANGED
@@ -12,6 +12,8 @@
12
  sample_rate: 16000
13
  n_fft: 400
14
  n_mels: 80
 
 
15
 
16
  ####################### Model parameters ###########################
17
  # Transformer
@@ -83,8 +85,6 @@ decoder: !new:speechbrain.decoders.S2STransformerBeamSearch
83
  max_decode_ratio: !ref <max_decode_ratio>
84
  beam_size: !ref <beam_size>
85
  ctc_weight: !ref <ctc_weight_decode>
86
- lm_weight: !ref <lm_weight>
87
- lm_modules: !ref <lm_model>
88
  temperature: 1.15
89
  temperature_lm: 1.15
90
  using_eos_threshold: False
@@ -93,7 +93,7 @@ decoder: !new:speechbrain.decoders.S2STransformerBeamSearch
93
  log_softmax: !new:torch.nn.LogSoftmax
94
  dim: -1
95
 
96
- normalize: !new:speechbrain.processing.features.InputNormalization
97
  norm_type: global
98
  update_until_epoch: 4
99
 
@@ -105,16 +105,30 @@ compute_features: !new:speechbrain.lobes.features.Fbank
105
 
106
  tokenizer: !new:sentencepiece.SentencePieceProcessor
107
 
108
- modules:
109
- CNN: !ref <CNN>
110
- Transformer: !ref <Transformer>
111
- seq_lin: !ref <seq_lin>
112
- ctc_lin: !ref <ctc_lin>
113
- normalize: !ref <normalize>
 
 
 
114
 
115
  model: !new:torch.nn.ModuleList
116
  - [!ref <CNN>, !ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]
117
 
 
 
 
 
 
 
 
 
 
 
 
118
  # The pretrainer allows a mapping between pretrained files and instances that
119
  # are declared in the yaml.
120
  pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
 
12
  sample_rate: 16000
13
  n_fft: 400
14
  n_mels: 80
15
+ win_length: 25
16
+ n_time_mask: 7
17
 
18
  ####################### Model parameters ###########################
19
  # Transformer
 
85
  max_decode_ratio: !ref <max_decode_ratio>
86
  beam_size: !ref <beam_size>
87
  ctc_weight: !ref <ctc_weight_decode>
 
 
88
  temperature: 1.15
89
  temperature_lm: 1.15
90
  using_eos_threshold: False
 
93
  log_softmax: !new:torch.nn.LogSoftmax
94
  dim: -1
95
 
96
+ normalizer: !new:speechbrain.processing.features.InputNormalization
97
  norm_type: global
98
  update_until_epoch: 4
99
 
 
105
 
106
  tokenizer: !new:sentencepiece.SentencePieceProcessor
107
 
108
+ Tencoder: !new:speechbrain.lobes.models.transformer.TransformerASR.EncoderWrapper
109
+ transformer: !ref <Transformer>
110
+
111
+ encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
112
+ input_shape: [null, null, !ref <n_mels>]
113
+ compute_features: !ref <compute_features>
114
+ normalize: !ref <normalizer>
115
+ cnn: !ref <CNN>
116
+ transformer_encoder: !ref <Tencoder>
117
 
118
  model: !new:torch.nn.ModuleList
119
  - [!ref <CNN>, !ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]
120
 
121
+ modules:
122
+ pre_transformer: !ref <CNN>
123
+ transformer: !ref <Transformer>
124
+ seq_lin: !ref <seq_lin>
125
+ ctc_lin: !ref <ctc_lin>
126
+ normalizer: !ref <normalizer>
127
+ encoder: !ref <encoder>
128
+ compute_features: !ref <compute_features>
129
+ model: !ref <model>
130
+ decoder: !ref <decoder>
131
+
132
  # The pretrainer allows a mapping between pretrained files and instances that
133
  # are declared in the yaml.
134
  pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer