Titouan commited on
Commit
9e5941e
1 Parent(s): 6fb11ea

EncoderDecoderASR update

Browse files

Files changed (1) hide show
  1. hyperparams.yaml +12 -5
hyperparams.yaml CHANGED
@@ -73,7 +73,7 @@ seq_lin: !new:speechbrain.nnet.linear.Linear
73
  input_size: !ref <d_model>
74
  n_neurons: !ref <output_neurons>
75
 
76
- beam_searcher: !new:speechbrain.decoders.S2STransformerBeamSearch
77
  modules: [!ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]
78
  bos_index: !ref <bos_index>
79
  eos_index: !ref <eos_index>
@@ -118,9 +118,15 @@ lm_model: !new:speechbrain.lobes.models.transformer.TransformerLM.TransformerLM
118
 
119
  tokenizer: !new:sentencepiece.SentencePieceProcessor
120
 
121
- asr_encoder: !new:speechbrain.utils.callchains.LengthsCapableChain
122
- - !ref <CNN>
123
- - !ref <Transformer.encode>
 
 
 
 
 
 
124
 
125
  # Models
126
  asr_model: !new:torch.nn.ModuleList
@@ -133,7 +139,8 @@ modules:
133
  asr_model: !ref <asr_model>
134
  normalize: !ref <normalize>
135
  lm_model: !ref <lm_model>
136
- beam_searcher: !ref <beam_searcher>
 
137
 
138
  # The pretrainer allows a mapping between pretrained files and instances that
139
  # are declared in the yaml.
73
  input_size: !ref <d_model>
74
  n_neurons: !ref <output_neurons>
75
 
76
+ decoder: !new:speechbrain.decoders.S2STransformerBeamSearch
77
  modules: [!ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]
78
  bos_index: !ref <bos_index>
79
  eos_index: !ref <eos_index>
118
 
119
  tokenizer: !new:sentencepiece.SentencePieceProcessor
120
 
121
+ Tencoder: !new:speechbrain.lobes.models.transformer.TransformerASR.EncoderWrapper
122
+ transformer: !ref <Transformer>
123
+
124
+ encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
125
+ input_shape: [null, null, !ref <n_mels>]
126
+ compute_features: !ref <compute_features>
127
+ normalize: !ref <normalize>
128
+ cnn: !ref <CNN>
129
+ transformer_encoder: !ref <Tencoder>
130
 
131
  # Models
132
  asr_model: !new:torch.nn.ModuleList
139
  asr_model: !ref <asr_model>
140
  normalize: !ref <normalize>
141
  lm_model: !ref <lm_model>
142
+ encoder: !ref <encoder>
143
+ decoder: !ref <decoder>
144
 
145
  # The pretrainer allows a mapping between pretrained files and instances that
146
  # are declared in the yaml.