Merge pull request #1225 from Jackwaterveg/new_config

[ASR][Config]refactor the train and test config

examples/aishell/asr0/conf/deepspeech2.yaml

@@ -1,68 +1,64 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
-  min_input_len: 0.0
-  max_input_len: 27.0 # second
-  min_output_len: 0.0
-  max_output_len: .inf
-  min_output_input_ratio: 0.00
-  max_output_input_ratio: .inf
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
+min_input_len: 0.0
+max_input_len: 27.0 # second
+min_output_len: 0.0
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
-  batch_size: 64 # one gpu
-  mean_std_filepath: data/mean_std.json
-  unit_type: char
-  vocab_filepath: data/lang_char/vocab.txt 
-  augmentation_config: conf/augmentation.json
-  random_seed: 0
-  spm_model_prefix: 
-  spectrum_type: linear
-  feat_dim: 
-  delta_delta: False
-  stride_ms: 10.0
-  window_ms: 20.0
-  n_fft: None
-  max_freq: None
-  target_sample_rate: 16000
-  use_dB_normalization: True
-  target_dB: -20
-  dither: 1.0
-  keep_transcription_text: False
-  sortagrad: True
-  shuffle_method: batch_shuffle
-  num_workers: 2
+###########################################
+#              Dataloader                 #
+###########################################
+batch_size: 64 # one gpu
+mean_std_filepath: data/mean_std.json
+unit_type: char
+vocab_filepath: data/lang_char/vocab.txt 
+augmentation_config: conf/augmentation.json
+random_seed: 0
+spm_model_prefix: 
+spectrum_type: linear
+feat_dim: 
+delta_delta: False
+stride_ms: 10.0
+window_ms: 20.0
+n_fft: None
+max_freq: None
+target_sample_rate: 16000
+use_dB_normalization: True
+target_dB: -20
+dither: 1.0
+keep_transcription_text: False
+sortagrad: True
+shuffle_method: batch_shuffle
+num_workers: 2
 
-model:
-  num_conv_layers: 2
-  num_rnn_layers: 3
-  rnn_layer_size: 1024
-  use_gru: True 
-  share_rnn_weights: False
-  blank_id: 0
-  ctc_grad_norm_type: instance 
+############################################
+#           Network Architecture           #
+############################################
+num_conv_layers: 2
+num_rnn_layers: 3
+rnn_layer_size: 1024
+use_gru: True 
+share_rnn_weights: False
+blank_id: 0
+ctc_grad_norm_type: instance 
 
-training:
-  n_epoch: 80
-  accum_grad: 1
-  lr: 2e-3
-  lr_decay: 0.83
-  weight_decay: 1e-06
-  global_grad_clip: 3.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-  
-decoding:
-  batch_size: 128
-  error_rate_type: cer 
-  decoding_method: ctc_beam_search
-  lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
-  alpha: 1.9
-  beta: 5.0
-  beam_size: 300
-  cutoff_prob: 0.99
-  cutoff_top_n: 40
-  num_proc_bsearch: 10
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 80
+accum_grad: 1
+lr: 2e-3
+lr_decay: 0.83
+weight_decay: 1e-06
+global_grad_clip: 3.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/aishell/asr0/conf/deepspeech2_online.yaml

@@ -1,70 +1,68 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
-  min_input_len: 0.0
-  max_input_len: 27.0 # second
-  min_output_len: 0.0
-  max_output_len: .inf
-  min_output_input_ratio: 0.00
-  max_output_input_ratio: .inf
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
+min_input_len: 0.0
+max_input_len: 27.0 # second
+min_output_len: 0.0
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
-  batch_size: 64 # one gpu
-  mean_std_filepath: data/mean_std.json
-  unit_type: char
-  vocab_filepath: data/lang_char/vocab.txt 
-  augmentation_config: conf/augmentation.json
-  random_seed: 0
-  spm_model_prefix: 
-  spectrum_type: linear #linear, mfcc, fbank
-  feat_dim: 
-  delta_delta: False
-  stride_ms: 10.0
-  window_ms: 20.0
-  n_fft: None
-  max_freq: None
-  target_sample_rate: 16000
-  use_dB_normalization: True
-  target_dB: -20
-  dither: 1.0
-  keep_transcription_text: False
-  sortagrad: True
-  shuffle_method: batch_shuffle
-  num_workers: 0
+###########################################
+#              Dataloader                 #
+###########################################
+batch_size: 64 # one gpu
+mean_std_filepath: data/mean_std.json
+unit_type: char
+vocab_filepath: data/lang_char/vocab.txt 
+augmentation_config: conf/augmentation.json
+random_seed: 0
+spm_model_prefix: 
+spectrum_type: linear #linear, mfcc, fbank
+feat_dim: 
+delta_delta: False
+stride_ms: 10.0
+window_ms: 20.0
+n_fft: None
+max_freq: None
+target_sample_rate: 16000
+use_dB_normalization: True
+target_dB: -20
+dither: 1.0
+keep_transcription_text: False
+sortagrad: True
+shuffle_method: batch_shuffle
+num_workers: 0
 
-model:
-  num_conv_layers: 2
-  num_rnn_layers: 5
-  rnn_layer_size: 1024
-  rnn_direction: forward # [forward, bidirect]
-  num_fc_layers: 0
-  fc_layers_size_list: -1,
-  use_gru: False 
-  blank_id: 0
+############################################
+#           Network Architecture           #
+############################################
+num_conv_layers: 2
+num_rnn_layers: 5
+rnn_layer_size: 1024
+rnn_direction: forward # [forward, bidirect]
+num_fc_layers: 0
+fc_layers_size_list: -1,
+use_gru: False 
+blank_id: 0
   
   
-training:
-  n_epoch: 65
-  accum_grad: 1
-  lr: 5e-4
-  lr_decay: 0.93
-  weight_decay: 1e-06
-  global_grad_clip: 3.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 65
+accum_grad: 1
+lr: 5e-4
+lr_decay: 0.93
+weight_decay: 1e-06
+global_grad_clip: 3.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
+
   
-decoding:
-  batch_size: 32
-  error_rate_type: cer 
-  decoding_method: ctc_beam_search
-  lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
-  alpha: 2.2 #1.9
-  beta: 4.3
-  beam_size: 300
-  cutoff_prob: 0.99
-  cutoff_top_n: 40
-  num_proc_bsearch: 10

examples/aishell/asr0/conf/tuning/chunk_decode.yaml

@@ -0,0 +1,10 @@
+chunk_batch_size: 32
+error_rate_type: cer 
+decoding_method: ctc_beam_search
+lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
+alpha: 2.2 #1.9
+beta: 4.3
+beam_size: 300
+cutoff_prob: 0.99
+cutoff_top_n: 40
+num_proc_bsearch: 10

examples/aishell/asr0/conf/tuning/decode.yaml

@@ -0,0 +1,10 @@
+decode_batch_size: 128
+error_rate_type: cer 
+decoding_method: ctc_beam_search
+lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
+alpha: 1.9
+beta: 5.0
+beam_size: 300
+cutoff_prob: 0.99
+cutoff_top_n: 40
+num_proc_bsearch: 10

examples/aishell/asr0/local/test.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-model_type=$3
+decode_config_path=$2
+ckpt_prefix=$3
+model_type=$4
 
 # download language model
 bash local/download_lm_ch.sh
@@ -21,6 +22,7 @@ fi
 python3 -u ${BIN_DIR}/test.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
 --model_type ${model_type}

examples/aishell/asr0/local/test_export.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-jit_model_export_path=$2
-model_type=$3
+decode_config_path=$2
+jit_model_export_path=$3
+model_type=$4
 
 # download language model
 bash local/download_lm_ch.sh > /dev/null 2>&1
@@ -21,6 +22,7 @@ fi
 python3 -u ${BIN_DIR}/test_export.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${jit_model_export_path}.rsl \
 --export_path ${jit_model_export_path} \
 --model_type ${model_type}

examples/aishell/asr0/local/test_wav.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 4 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type audio_file"
+if [ $# != 5 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type audio_file"
     exit -1
 fi
 
@@ -9,9 +9,10 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-model_type=$3
-audio_file=$4
+decode_config_path=$2
+ckpt_prefix=$3
+model_type=$4
+audio_file=$5
 
 mkdir -p data
 wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/
@@ -33,6 +34,7 @@ fi
 python3 -u ${BIN_DIR}/test_wav.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
 --model_type ${model_type} \

examples/aishell/asr0/run.sh

@@ -6,6 +6,7 @@ gpus=0,1,2,3
 stage=0
 stop_stage=100
 conf_path=conf/deepspeech2.yaml    #conf/deepspeech2.yaml or conf/deepspeeech2_online.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=1
 model_type=offline    # offline or online
 audio_file=data/demo_01_03.wav
@@ -34,7 +35,7 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type}|| exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type}|| exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
@@ -44,11 +45,11 @@ fi
 
 if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
     # test export ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test_export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt}.jit ${model_type}|| exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test_export.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt}.jit ${model_type}|| exit -1
 fi
 
 # Optionally, you can add LM and test it with runtime.
 if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
     # test a single .wav file
-    CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} ${audio_file} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} ${audio_file} || exit -1
 fi

examples/aishell/asr1/conf/chunk_conformer.yaml

@@ -1,103 +1,95 @@
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: conformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: True
-        cnn_module_kernel: 15
-        use_cnn_module: True
-        activation_type: 'swish'
-        pos_enc_layer_type: 'rel_pos'
-        selfattention_layer_type: 'rel_selfattn'
-        causal: true
-        use_dynamic_chunk: true
-        cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
-        use_dynamic_left_chunk: false
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: True
+    cnn_module_kernel: 15
+    use_cnn_module: True
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
+    causal: true
+    use_dynamic_chunk: true
+    cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
+    use_dynamic_left_chunk: false
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+###########################################
+#                   Data                  #
+###########################################
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
 
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
+###########################################
+#              Dataloader                 #
+###########################################
 
+vocab_filepath: data/lang_char/vocab.txt 
+spm_model_prefix: ''
+unit_type: 'char'
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 64
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'char'
-  augmentation_config: conf/preprocess.yaml
-  feat_dim: 80
-  stride_ms: 10.0
-  window_ms: 25.0
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  batch_size: 64
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  minibatches: 0 # for debug
-  batch_count: auto
-  batch_bins: 0 
-  batch_frames_in: 0
-  batch_frames_out: 0
-  batch_frames_inout: 0
-  num_workers: 0
-  subsampling_factor: 1
-  num_encs: 1
-
-
-training:
-  n_epoch: 240 
-  accum_grad: 2
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.002
-    weight_decay: 1e-6
-  scheduler: warmuplr
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-
-decoding:
-  beam_size: 10
-  batch_size: 128
-  error_rate_type: cer 
-  decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 240 
+accum_grad: 2
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.002
+  weight_decay: 1.0e-6
+scheduler: warmuplr
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
 

examples/aishell/asr1/conf/conformer.yaml

@@ -1,97 +1,89 @@
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: conformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: True
-        cnn_module_kernel: 15
-        use_cnn_module: True
-        activation_type: 'swish'
-        pos_enc_layer_type: 'rel_pos'
-        selfattention_layer_type: 'rel_selfattn'
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: True
+    cnn_module_kernel: 15
+    use_cnn_module: True
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt 
+spm_model_prefix: ''
+unit_type: 'char'
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 64
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 2
+subsampling_factor: 1
+num_encs: 1
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'char'
-  augmentation_config: conf/preprocess.yaml
-  feat_dim: 80
-  stride_ms: 10.0
-  window_ms: 25.0
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  batch_size: 64
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  minibatches: 0 # for debug
-  batch_count: auto
-  batch_bins: 0 
-  batch_frames_in: 0
-  batch_frames_out: 0
-  batch_frames_inout: 0
-  num_workers: 8
-  subsampling_factor: 1
-  num_encs: 1
-
-
-training:
-  n_epoch: 240 
-  accum_grad: 2
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.002
-    weight_decay: 1e-6
-  scheduler: warmuplr
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-
-decoding:
-  beam_size: 10
-  batch_size: 128
-  error_rate_type: cer 
-  decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 240 
+accum_grad: 2
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.002
+  weight_decay: 1.0e-6
+scheduler: warmuplr
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/aishell/asr1/conf/transformer.yaml

@@ -1,95 +1,85 @@
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
-
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
-
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
+###########################################
+#                   Data                  #
+###########################################
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
-
-
-collator:
-  unit_type: 'char'
-  vocab_filepath: data/lang_char/vocab.txt 
-  feat_dim: 80
-  stride_ms: 10.0
-  window_ms: 25.0
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  batch_size: 64 
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  minibatches: 0 # for debug
-  batch_count: auto
-  batch_bins: 0 
-  batch_frames_in: 0
-  batch_frames_out: 0
-  batch_frames_inout: 0
-  augmentation_config: conf/preprocess.yaml 
-  num_workers: 0
-  subsampling_factor: 1
-  num_encs: 1
-
-
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
-training:
-  n_epoch: 240 
-  accum_grad: 2
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.002
-    weight_decay: 1e-6
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
 
+###########################################
+#              Dataloader                 #
+###########################################
+unit_type: 'char'
+vocab_filepath: data/lang_char/vocab.txt 
+spm_model_prefix: ''
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 64 
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+preprocess_config: conf/preprocess.yaml 
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
-decoding:
-  beam_size: 10
-  batch_size: 128
-  error_rate_type: cer 
-  decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 240 
+accum_grad: 2
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.002
+  weight_decay: 1.0e-6
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/aishell/asr1/conf/tuning/chunk_decode.yaml

@@ -0,0 +1,11 @@
+beam_size: 10
+decode_batch_size: 128
+error_rate_type: cer 
+decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.

examples/aishell/asr1/conf/tuning/decode.yaml

@@ -0,0 +1,11 @@
+beam_size: 10
+decode_batch_size: 128
+error_rate_type: cer 
+decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.

examples/aishell/asr1/local/align.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 batch_size=1
 output_dir=${ckpt_prefix}
@@ -20,9 +21,10 @@ mkdir -p ${output_dir}
 python3 -u ${BIN_DIR}/alignment.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
---opts decoding.batch_size ${batch_size}
+--opts decode.decode_batch_size ${batch_size}
 
 if [ $? -ne 0 ]; then
     echo "Failed in ctc alignment!"

examples/aishell/asr1/local/test.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 chunk_mode=false
 if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
@@ -36,10 +37,11 @@ for type in attention ctc_greedy_search; do
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"
@@ -55,10 +57,11 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"

examples/aishell/asr1/local/test_wav.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix audio_file"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix audio_file"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-audio_file=$3
+decode_config_path=$2
+ckpt_prefix=$3
+audio_file=$4
 
 mkdir -p data
 wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/
@@ -42,10 +43,11 @@ for type in  attention_rescoring; do
     python3 -u ${BIN_DIR}/test_wav.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size} \
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size} \
     --audio_file ${audio_file}
 
     if [ $? -ne 0 ]; then

examples/aishell/asr1/run.sh

@@ -6,6 +6,7 @@ gpus=0,1,2,3
 stage=0
 stop_stage=50
 conf_path=conf/conformer.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=20
 audio_file=data/demo_01_03.wav
 
@@ -32,18 +33,18 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
     # ctc alignment of test data
-    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 # Optionally, you can add LM and test it with runtime.
 if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
     # test a single .wav file
-    CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1
 fi
 
 # Not supported at now!!!

examples/callcenter/asr1/conf/chunk_conformer.yaml

@@ -1,120 +1,98 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
-  min_input_len: 0.5
-  max_input_len: 20.0 # second
-  min_output_len: 0.0
-  max_output_len: 400.0
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
-  
-
-collator:
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'char'
-  spm_model_prefix: ''
-  augmentation_config: conf/preprocess.yaml
-  batch_size: 32
-  raw_wav: True  # use raw_wav or kaldi feature
-  spectrum_type: fbank #linear, mfcc, fbank
-  feat_dim: 80
-  delta_delta: False
-  dither: 1.0
-  target_sample_rate: 8000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 25.0
-  use_dB_normalization: True 
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
-
-
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: conformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: True
-        use_cnn_module: True
-        cnn_module_kernel: 15
-        activation_type: 'swish'
-        pos_enc_layer_type: 'rel_pos'
-        selfattention_layer_type: 'rel_selfattn'
-        causal: true
-        use_dynamic_chunk: true
-        cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
-        use_dynamic_left_chunk: false
-
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
-
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
+  
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt 
+unit_type: 'char'
+spm_model_prefix: ''
+preprocess_config: conf/preprocess.yaml
+batch_size: 32
+raw_wav: True  # use raw_wav or kaldi feature
+spectrum_type: fbank #linear, mfcc, fbank
+feat_dim: 80
+delta_delta: False
+dither: 1.0
+target_sample_rate: 8000
+max_freq: None
+n_fft: None
+stride_ms: 10.0
+window_ms: 25.0
+use_dB_normalization: True 
+target_dB: -20
+random_seed: 0
+keep_transcription_text: False
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 2
 
-training:
-  n_epoch: 240
-  accum_grad: 4
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.001
-    weight_decay: 1e-6
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
 
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: True
+    use_cnn_module: True
+    cnn_module_kernel: 15
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
+    causal: true
+    use_dynamic_chunk: true
+    cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
+    use_dynamic_left_chunk: false
 
-decoding:
-  batch_size: 128
-  error_rate_type: cer 
-  decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: true  # simulate streaming inference. Defaults to False.
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 240
+accum_grad: 4
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.001
+  weight_decay: 1.0e-6
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/callcenter/asr1/conf/conformer.yaml

@@ -1,117 +1,92 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
-  min_input_len: 0.5
-  max_input_len: 20.0 # second
-  min_output_len: 0.0
-  max_output_len: 400.0
-  min_output_input_ratio: 0.0
-  max_output_input_ratio: .inf 
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'char'
-  spm_model_prefix: ''
-  augmentation_config: conf/preprocess.yaml
-  batch_size: 32
-  raw_wav: True  # use raw_wav or kaldi feature
-  spectrum_type: fbank #linear, mfcc, fbank
-  feat_dim: 80
-  delta_delta: False
-  dither: 1.0
-  target_sample_rate: 8000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 25.0
-  use_dB_normalization: True 
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt 
+unit_type: 'char'
+spm_model_prefix: ''
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 64
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: conformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: True
-        use_cnn_module: True
-        cnn_module_kernel: 15
-        activation_type: 'swish'
-        pos_enc_layer_type: 'rel_pos'
-        selfattention_layer_type: 'rel_selfattn'
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: True
+    use_cnn_module: True
+    cnn_module_kernel: 15
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
-
-
-training:
-  n_epoch: 100 # 50 will be lowest 
-  accum_grad: 4
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.002
-    weight_decay: 1e-6
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-
-
-
-decoding:
-  batch_size: 128
-  error_rate_type: cer 
-  decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
 
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 100 # 50 will be lowest 
+accum_grad: 4
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.002
+  weight_decay: 1.0e-6
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/callcenter/asr1/conf/preprocess.yaml

@@ -1,7 +1,7 @@
 process:
   # extract kaldi fbank from PCM
   - type: fbank_kaldi
-    fs: 16000
+    fs: 8000
     n_mels: 80
     n_shift: 160
     win_length: 400

examples/callcenter/asr1/conf/tuning/chunk_decode.yaml

@@ -0,0 +1,11 @@
+decode_batch_size: 128
+error_rate_type: cer 
+decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+beam_size: 10
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: true  # simulate streaming inference. Defaults to False.

examples/callcenter/asr1/conf/tuning/decode.yaml

@@ -0,0 +1,13 @@
+decode_batch_size: 128
+error_rate_type: cer 
+decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+beam_size: 10
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.
+
+

examples/callcenter/asr1/local/align.sh

@@ -1,7 +1,7 @@
 #! /usr/bin/env bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 ckpt_name=$(basename ${ckpt_prefxi})
 
@@ -25,9 +26,10 @@ mkdir -p ${output_dir}
 python3 -u ${BIN_DIR}/alignment.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
---opts decoding.batch_size ${batch_size}
+--opts decode.decode_batch_size ${batch_size}
 
 if [ $? -ne 0 ]; then
     echo "Failed in ctc alignment!"

examples/callcenter/asr1/local/test.sh

@@ -1,7 +1,7 @@
 #! /usr/bin/env bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,7 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
+
 
 ckpt_name=$(basename ${ckpt_prefxi})
 
@@ -30,10 +32,11 @@ for type in attention ctc_greedy_search; do
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"
@@ -49,10 +52,11 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"

examples/callcenter/asr1/run.sh

@@ -4,8 +4,9 @@ source path.sh
 
 gpus=0,1,2,3
 stage=0
-stop_stage=100
+stop_stage=50
 conf_path=conf/conformer.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=20
 
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
@@ -31,15 +32,15 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
     # ctc alignment of test data
-    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
-if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
+if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then
     # export ckpt avg_n
     CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
 fi

examples/librispeech/asr0/conf/deepspeech2.yaml

@@ -1,68 +1,65 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev-clean
-  test_manifest: data/manifest.test-clean
-  min_input_len: 0.0
-  max_input_len: 30.0 # second
-  min_output_len: 0.0
-  max_output_len: .inf
-  min_output_input_ratio: 0.00
-  max_output_input_ratio: .inf
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev-clean
+test_manifest: data/manifest.test-clean
+min_input_len: 0.0
+max_input_len: 30.0 # second
+min_output_len: 0.0
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
-  batch_size: 20
-  mean_std_filepath: data/mean_std.json
-  unit_type: char
-  vocab_filepath: data/lang_char/vocab.txt
-  augmentation_config: conf/augmentation.json
-  random_seed: 0
-  spm_model_prefix: 
-  spectrum_type: linear
-  target_sample_rate: 16000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 20.0
-  delta_delta: False
-  dither: 1.0
-  use_dB_normalization: True 
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
+###########################################
+#              Dataloader                 #
+###########################################
+batch_size: 20
+mean_std_filepath: data/mean_std.json
+unit_type: char
+vocab_filepath: data/lang_char/vocab.txt
+augmentation_config: conf/augmentation.json
+random_seed: 0
+spm_model_prefix: 
+spectrum_type: linear
+feat_dim: 
+target_sample_rate: 16000
+max_freq: None
+n_fft: None
+stride_ms: 10.0
+window_ms: 20.0
+delta_delta: False
+dither: 1.0
+use_dB_normalization: True 
+target_dB: -20
+random_seed: 0
+keep_transcription_text: False
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 2
 
-model:
-  num_conv_layers: 2
-  num_rnn_layers: 3
-  rnn_layer_size: 2048
-  use_gru: False 
-  share_rnn_weights: True
-  blank_id: 0
+############################################
+#           Network Architecture           #
+############################################
+num_conv_layers: 2
+num_rnn_layers: 3
+rnn_layer_size: 2048
+use_gru: False 
+share_rnn_weights: True
+blank_id: 0
   
 
-training:
-  n_epoch: 50
-  accum_grad: 1
-  lr: 1e-3
-  lr_decay: 0.83
-  weight_decay: 1e-06
-  global_grad_clip: 5.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-decoding:
-  batch_size: 128
-  error_rate_type: wer
-  decoding_method: ctc_beam_search
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 1.9
-  beta: 0.3
-  beam_size: 500
-  cutoff_prob: 1.0
-  cutoff_top_n: 40
-  num_proc_bsearch: 8
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 50
+accum_grad: 1
+lr: 1e-3
+lr_decay: 0.83
+weight_decay: 1e-06
+global_grad_clip: 5.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/librispeech/asr0/conf/deepspeech2_online.yaml

@@ -1,70 +1,67 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev-clean
-  test_manifest: data/manifest.test-clean
-  min_input_len: 0.0
-  max_input_len: 30.0 # second
-  min_output_len: 0.0
-  max_output_len: .inf
-  min_output_input_ratio: 0.00
-  max_output_input_ratio: .inf
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev-clean
+test_manifest: data/manifest.test-clean
+min_input_len: 0.0
+max_input_len: 30.0 # second
+min_output_len: 0.0
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
-  batch_size: 15
-  mean_std_filepath: data/mean_std.json
-  unit_type: char
-  vocab_filepath: data/lang_char/vocab.txt
-  augmentation_config: conf/augmentation.json
-  random_seed: 0
-  spm_model_prefix: 
-  spectrum_type: linear
-  target_sample_rate: 16000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 20.0
-  delta_delta: False
-  dither: 1.0
-  use_dB_normalization: True 
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 0
+###########################################
+#              Dataloader                 #
+###########################################
+batch_size: 15
+mean_std_filepath: data/mean_std.json
+unit_type: char
+vocab_filepath: data/lang_char/vocab.txt
+augmentation_config: conf/augmentation.json
+random_seed: 0
+spm_model_prefix: 
+spectrum_type: linear
+feat_dim: 
+target_sample_rate: 16000
+max_freq: None
+n_fft: None
+stride_ms: 10.0
+window_ms: 20.0
+delta_delta: False
+dither: 1.0
+use_dB_normalization: True 
+target_dB: -20
+random_seed: 0
+keep_transcription_text: False
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 0
 
-model:
-  num_conv_layers: 2
-  num_rnn_layers: 3
-  rnn_layer_size: 2048
-  rnn_direction: forward
-  num_fc_layers: 2
-  fc_layers_size_list: 512, 256
-  use_gru: False 
-  blank_id: 0
+############################################
+#           Network Architecture           #
+############################################
+num_conv_layers: 2
+num_rnn_layers: 3
+rnn_layer_size: 2048
+rnn_direction: forward
+num_fc_layers: 2
+fc_layers_size_list: 512, 256
+use_gru: False 
+blank_id: 0
   
 
-training:
-  n_epoch: 50
-  accum_grad: 4
-  lr: 1e-3
-  lr_decay: 0.83
-  weight_decay: 1e-06
-  global_grad_clip: 5.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-decoding:
-  batch_size: 128
-  error_rate_type: wer
-  decoding_method: ctc_beam_search
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 1.9
-  beta: 0.3
-  beam_size: 500
-  cutoff_prob: 1.0
-  cutoff_top_n: 40
-  num_proc_bsearch: 8
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 50
+accum_grad: 4
+lr: 1e-3
+lr_decay: 0.83
+weight_decay: 1e-06
+global_grad_clip: 5.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/librispeech/asr0/conf/tuning/chunk_decode.yaml

@@ -0,0 +1,10 @@
+decode_batch_size: 128
+error_rate_type: wer
+decoding_method: ctc_beam_search
+lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
+alpha: 1.9
+beta: 0.3
+beam_size: 500
+cutoff_prob: 1.0
+cutoff_top_n: 40
+num_proc_bsearch: 8

examples/librispeech/asr0/conf/tuning/decode.yaml

@@ -0,0 +1,10 @@
+decode_batch_size: 128
+error_rate_type: wer
+decoding_method: ctc_beam_search
+lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
+alpha: 1.9
+beta: 0.3
+beam_size: 500
+cutoff_prob: 1.0
+cutoff_top_n: 40
+num_proc_bsearch: 8

examples/librispeech/asr0/local/test.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-model_type=$3
+decode_config_path=$2
+ckpt_prefix=$3
+model_type=$4
 
 # download language model
 bash local/download_lm_en.sh
@@ -21,6 +22,7 @@ fi
 python3 -u ${BIN_DIR}/test.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
 --model_type ${model_type}

examples/librispeech/asr0/local/test_wav.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 4 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type audio_file"
+if [ $# != 5 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type audio_file"
     exit -1
 fi
 
@@ -9,9 +9,10 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-model_type=$3
-audio_file=$4
+decode_config_path=$2
+ckpt_prefix=$3
+model_type=$4
+audio_file=$5
 
 mkdir -p data
 wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/en/demo_002_en.wav -P data/
@@ -33,6 +34,7 @@ fi
 python3 -u ${BIN_DIR}/test_wav.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
 --model_type ${model_type} \

examples/librispeech/asr0/run.sh

@@ -6,6 +6,7 @@ gpus=0,1,2,3,4,5,6,7
 stage=0
 stop_stage=100
 conf_path=conf/deepspeech2.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=30
 model_type=offline
 audio_file=data/demo_002_en.wav
@@ -33,7 +34,7 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
@@ -43,5 +44,5 @@ fi
 
 if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
     # test a single .wav file
-    CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} ${audio_file} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} ${audio_file} || exit -1
 fi

examples/librispeech/asr1/conf/chunk_conformer.yaml

@@ -1,103 +1,99 @@
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: conformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: True
-        use_cnn_module: True
-        cnn_module_kernel: 15
-        activation_type: 'swish'
-        pos_enc_layer_type: 'rel_pos'
-        selfattention_layer_type: 'rel_selfattn'
-        causal: True
-        use_dynamic_chunk: true
-        cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
-        use_dynamic_left_chunk: false
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: True
+    use_cnn_module: True
+    cnn_module_kernel: 15
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
+    causal: True
+    use_dynamic_chunk: true
+    cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
+    use_dynamic_left_chunk: false
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
 
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'spm'
-  spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
-  mean_std_filepath: ""
-  augmentation_config: conf/preprocess.yaml
-  feat_dim: 80
-  stride_ms: 10.0
-  window_ms: 25.0
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  batch_size: 16
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  minibatches: 0 # for debug
-  batch_count: auto
-  batch_bins: 0 
-  batch_frames_in: 0
-  batch_frames_out: 0
-  batch_frames_inout: 0
-  augmentation_config: conf/preprocess.yaml 
-  num_workers: 0
-  subsampling_factor: 1
-  num_encs: 1
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt 
+unit_type: 'spm'
+spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
+mean_std_filepath: ""
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 16
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0 
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
+
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 120
+accum_grad: 8
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.001
+  weight_decay: 1.0e-06 
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
 
-training:
-  n_epoch: 120
-  accum_grad: 8
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.001
-    weight_decay: 1e-06 
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
 
-decoding:
-  batch_size: 128
-  error_rate_type: wer
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  beam_size: 10
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: true  # simulate streaming inference. Defaults to False.
 
 

examples/librispeech/asr1/conf/chunk_transformer.yaml

@@ -1,103 +1,89 @@
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
-        use_dynamic_chunk: true
-        use_dynamic_left_chunk: false
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
+    use_dynamic_chunk: true
+    use_dynamic_left_chunk: false
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'spm'
-  spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
-  mean_std_filepath: ""
-  augmentation_config: conf/preprocess.yaml
-  feat_dim: 80
-  stride_ms: 10.0
-  window_ms: 25.0
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  batch_size: 64
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  minibatches: 0 # for debug
-  batch_count: auto
-  batch_bins: 0 
-  batch_frames_in: 0
-  batch_frames_out: 0
-  batch_frames_inout: 0
-  augmentation_config: conf/preprocess.yaml 
-  num_workers: 0
-  subsampling_factor: 1
-  num_encs: 1
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt 
+unit_type: 'spm'
+spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
+mean_std_filepath: ""
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 64
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
 
-training:
-  n_epoch: 120
-  accum_grad: 1
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.001
-    weight_decay: 1e-06
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-
-decoding:
-  batch_size: 64
-  error_rate_type: wer
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: true  # simulate streaming inference. Defaults to False.
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 120
+accum_grad: 1
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.001
+  weight_decay: 1.0e-06
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/librispeech/asr1/conf/conformer.yaml

@@ -1,104 +1,96 @@
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: conformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: True
-        use_cnn_module: True
-        cnn_module_kernel: 15
-        activation_type: 'swish'
-        pos_enc_layer_type: 'rel_pos'
-        selfattention_layer_type: 'rel_selfattn'
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: True
+    use_cnn_module: True
+    cnn_module_kernel: 15
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        ctc_grad_norm_type: null 
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    ctc_grad_norm_type: null 
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
 
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test-clean
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test-clean
 
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'spm'
-  spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
-  mean_std_filepath: ""
-  augmentation_config: conf/preprocess.yaml
-  feat_dim: 80
-  stride_ms: 10.0
-  window_ms: 25.0
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  batch_size: 16
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  minibatches: 0 # for debug
-  batch_count: auto
-  batch_bins: 0 
-  batch_frames_in: 0
-  batch_frames_out: 0
-  batch_frames_inout: 0
-  augmentation_config: conf/preprocess.yaml 
-  num_workers: 0
-  subsampling_factor: 1
-  num_encs: 1
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt 
+unit_type: 'spm'
+spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
+mean_std_filepath: ""
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 16
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
   
 
-training:
-  n_epoch: 70
-  accum_grad: 8
-  global_grad_clip: 3.0
-  optim: adam
-  optim_conf:
-    lr: 0.004
-    weight_decay: 1e-06
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-
-decoding:
-  batch_size: 64
-  error_rate_type: wer
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  beam_size: 10
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 70
+accum_grad: 8
+global_grad_clip: 3.0
+optim: adam
+optim_conf:
+  lr: 0.004
+  weight_decay: 1.0e-06
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
 
 

examples/librispeech/asr1/conf/transformer.yaml

@@ -1,110 +1,88 @@
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
 
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test-clean
-  min_input_len: 0.5  # second
-  max_input_len: 30.0 # second
-  min_output_len: 0.0 # tokens
-  max_output_len: 400.0 # tokens
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 100.0
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test-clean
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt
-  unit_type: 'spm'
-  spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
-  mean_std_filepath: ""
-  augmentation_config: conf/preprocess.yaml
-  feat_dim: 80
-  stride_ms: 10.0
-  window_ms: 25.0
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  batch_size: 32 
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  minibatches: 0 # for debug
-  batch_count: auto
-  batch_bins: 0 
-  batch_frames_in: 0
-  batch_frames_out: 0
-  batch_frames_inout: 0
-  augmentation_config: conf/preprocess.yaml 
-  num_workers: 0
-  subsampling_factor: 1
-  num_encs: 1
-
-
-training:
-  n_epoch: 120 
-  accum_grad: 4
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.004
-    weight_decay: 1e-06
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-
-decoding:
-  batch_size: 64
-  error_rate_type: wer
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt
+unit_type: 'spm'
+spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
+mean_std_filepath: ""
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 32 
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
 
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 120 
+accum_grad: 4
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.004
+  weight_decay: 1.0e-06
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/librispeech/asr1/conf/tuning/chunk_decode.yaml

@@ -0,0 +1,11 @@
+decode_batch_size: 128
+error_rate_type: wer
+decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+beam_size: 10
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: true  # simulate streaming inference. Defaults to False.

examples/librispeech/asr1/conf/tuning/decode.yaml

@@ -0,0 +1,11 @@
+decode_batch_size: 64
+error_rate_type: wer
+decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+beam_size: 10
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.

examples/librispeech/asr1/local/align.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 batch_size=1
 output_dir=${ckpt_prefix}
@@ -20,9 +21,10 @@ mkdir -p ${output_dir}
 python3 -u ${BIN_DIR}/alignment.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
---opts decoding.batch_size ${batch_size}
+--opts decode.decode_batch_size ${batch_size}
 
 if [ $? -ne 0 ]; then
     echo "Failed in ctc alignment!"

examples/librispeech/asr1/local/test.sh

@@ -15,8 +15,8 @@ recog_set="test-clean"
 stage=0
 stop_stage=100
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -24,7 +24,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 chunk_mode=false
 if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
@@ -52,10 +53,11 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
         python3 -u ${BIN_DIR}/test.py \
             --ngpu ${ngpu} \
             --config ${config_path} \
+            --decode_cfg ${decode_config_path} \
             --result_file ${ckpt_prefix}.${type}.rsl \
             --checkpoint_path ${ckpt_prefix} \
-            --opts decoding.decoding_method ${type} \
-            --opts decoding.batch_size ${batch_size}
+            --opts decode.decoding_method ${type} \
+            --opts decode.decode_batch_size ${batch_size}
 
         if [ $? -ne 0 ]; then
             echo "Failed in evaluation!"
@@ -76,10 +78,11 @@ for type in ctc_greedy_search; do
     python3 -u ${BIN_DIR}/test.py \
         --ngpu ${ngpu} \
         --config ${config_path} \
+        --decode_cfg ${decode_config_path} \
         --result_file ${ckpt_prefix}.${type}.rsl \
         --checkpoint_path ${ckpt_prefix} \
-        --opts decoding.decoding_method ${type} \
-        --opts decoding.batch_size ${batch_size}
+        --opts decode.decoding_method ${type} \
+        --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"
@@ -96,10 +99,11 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     python3 -u ${BIN_DIR}/test.py \
         --ngpu ${ngpu} \
         --config ${config_path} \
+        --decode_cfg ${decode_config_path} \
         --result_file ${ckpt_prefix}.${type}.rsl \
         --checkpoint_path ${ckpt_prefix} \
-        --opts decoding.decoding_method ${type} \
-        --opts decoding.batch_size ${batch_size}
+        --opts decode.decoding_method ${type} \
+        --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"

examples/librispeech/asr1/local/test_wav.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix audio_file"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix audio_file"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-audio_file=$3
+decode_config_path=$2
+ckpt_prefix=$3
+audio_file=$4
 
 mkdir -p data
 wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/en/demo_002_en.wav -P data/
@@ -49,10 +50,11 @@ for type in attention_rescoring; do
     python3 -u ${BIN_DIR}/test_wav.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size} \
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size} \
     --audio_file ${audio_file}
 
     #score_sclite.sh --bpe ${nbpe} --bpemodel ${bpemodel}.model --wer true ${expdir}/${decode_dir} ${dict}

examples/librispeech/asr1/run.sh

@@ -8,6 +8,7 @@ gpus=0,1,2,3
 stage=0
 stop_stage=50
 conf_path=conf/transformer.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=30
 audio_file=data/demo_002_en.wav
 
@@ -34,17 +35,17 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
     # ctc alignment of test data
-    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
     # test a single .wav file
-    CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1
 fi
 
 if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then

examples/librispeech/asr2/conf/decode/decode_base.yaml

@@ -0,0 +1,11 @@
+decode_batch_size: 1
+error_rate_type: wer
+decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+beam_size: 10
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.

examples/librispeech/asr2/conf/transformer.yaml

@@ -1,73 +1,80 @@
 # https://yaml.org/type/float.html
-# network architecture
-model:
-    cmvn_file:  
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file:  
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test-clean
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test-clean
 
-collator:
-  vocab_filepath: data/lang_char/train_960_unigram5000_units.txt
-  unit_type: spm
-  spm_model_prefix: data/lang_char/train_960_unigram5000
-  feat_dim: 83
-  stride_ms: 10.0
-  window_ms: 25.0
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  batch_size: 30 
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  minibatches: 0 # for debug
-  batch_count: auto
-  batch_bins: 0 
-  batch_frames_in: 0
-  batch_frames_out: 0
-  batch_frames_inout: 0
-  augmentation_config: conf/preprocess.yaml 
-  num_workers: 0
-  subsampling_factor: 1
-  num_encs: 1
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/train_960_unigram5000_units.txt
+unit_type: spm
+spm_model_prefix: data/lang_char/train_960_unigram5000
+feat_dim: 83
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 30 
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+preprocess_config: conf/preprocess.yaml 
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
 
-training:
-  n_epoch: 120
-  accum_grad: 2
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 120
+accum_grad: 2
+log_interval: 1
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
 
 optim: adam
 optim_conf:
@@ -79,23 +86,5 @@ scheduler_conf:
   warmup_steps: 25000
   lr_decay: 1.0
 
-decoding:
-  batch_size: 1
-  error_rate_type: wer
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
 
 

examples/librispeech/asr2/local/align.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path dict_path ckpt_path_prefix"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path dict_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-dict_path=$2
-ckpt_prefix=$3
+decode_config_path=$2
+dict_path=$3
+ckpt_prefix=$4
 
 batch_size=1
 output_dir=${ckpt_prefix}
@@ -24,9 +25,10 @@ python3 -u ${BIN_DIR}/test.py \
 --dict-path ${dict_path} \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result-file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
---opts decoding.batch_size ${batch_size}
+--opts decode.decode_batch_size ${batch_size}
 
 if [ $? -ne 0 ]; then
     echo "Failed in ctc alignment!"

examples/librispeech/asr2/local/test.sh

@@ -19,6 +19,7 @@ bpeprefix=data/lang_char/${train_set}_${bpemode}${nbpe}
 bpemodel=${bpeprefix}.model
 
 config_path=conf/transformer.yaml
+decode_config_path=conf/decode/decode_base.yaml
 dict=data/lang_char/${train_set}_${bpemode}${nbpe}_units.txt
 ckpt_prefix=
 
@@ -79,11 +80,12 @@ for dmethd in attention ctc_greedy_search ctc_prefix_beam_search attention_resco
             --ngpu ${ngpu} \
             --dict-path ${dict} \
             --config ${config_path} \
+            --decode_cfg ${decode_config_path} \
             --checkpoint_path ${ckpt_prefix} \
             --result-file ${decode_dir}/data.JOB.json \
-            --opts decoding.decoding_method ${dmethd} \
-            --opts decoding.batch_size ${batch_size} \
-            --opts data.test_manifest ${feat_recog_dir}/split${nj}/JOB/manifest.${rtask}
+            --opts decode.decoding_method ${dmethd} \
+            --opts decode.decode_batch_size ${batch_size} \
+            --opts test_manifest ${feat_recog_dir}/split${nj}/JOB/manifest.${rtask}
 
         score_sclite.sh --bpe ${nbpe} --bpemodel ${bpemodel} --wer false ${decode_dir} ${dict}
 

examples/librispeech/asr2/run.sh

@@ -9,7 +9,8 @@ gpus=0,1,2,3,4,5,6,7
 stage=0
 stop_stage=50
 conf_path=conf/transformer.yaml
-dict_path=lang_char/train_960_unigram5000_units.txt
+decode_conf_path=conf/decode/decode_base.yaml
+dict_path=data/lang_char/train_960_unigram5000_units.txt
 avg_num=10
 
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
@@ -35,7 +36,7 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # attetion resocre decoder
-    ./local/test.sh ${conf_path} ${dict_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    ./local/test.sh ${conf_path} ${decode_conf_path} ${dict_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
@@ -45,7 +46,7 @@ fi
 
 if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
     # ctc alignment of test data
-    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${dict_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${decode_conf_path} ${dict_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then

examples/other/1xt2x/aishell/conf/deepspeech2.yaml

@@ -1,67 +1,65 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
-  min_input_len: 0.0
-  max_input_len: 27.0 # second
-  min_output_len: 0.0
-  max_output_len: .inf
-  min_output_input_ratio: 0.00
-  max_output_input_ratio: .inf
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
+min_input_len: 0.0
+max_input_len: 27.0 # second
+min_output_len: 0.0
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
-  batch_size: 64 # one gpu
-  mean_std_filepath: data/mean_std.npz
-  unit_type: char
-  vocab_filepath: data/vocab.txt 
-  augmentation_config: conf/augmentation.json
-  random_seed: 0
-  spm_model_prefix: 
-  spectrum_type: linear
-  feat_dim: 
-  delta_delta: False
-  stride_ms: 10.0
-  window_ms: 20.0
-  n_fft: None
-  max_freq: None
-  target_sample_rate: 16000
-  use_dB_normalization: True
-  target_dB: -20
-  dither: 1.0
-  keep_transcription_text: False
-  sortagrad: True
-  shuffle_method: batch_shuffle
-  num_workers: 2
+###########################################
+#              Dataloader                 #
+###########################################
+batch_size: 64 # one gpu
+mean_std_filepath: data/mean_std.npz
+unit_type: char
+vocab_filepath: data/vocab.txt 
+augmentation_config: conf/augmentation.json
+random_seed: 0
+spm_model_prefix: 
+spectrum_type: linear
+feat_dim: 
+delta_delta: False
+stride_ms: 10.0
+window_ms: 20.0
+n_fft: None
+max_freq: None
+target_sample_rate: 16000
+use_dB_normalization: True
+target_dB: -20
+dither: 1.0
+keep_transcription_text: False
+sortagrad: True
+shuffle_method: batch_shuffle
+num_workers: 2
 
-model:
-  num_conv_layers: 2
-  num_rnn_layers: 3
-  rnn_layer_size: 1024
-  use_gru: True 
-  share_rnn_weights: False
-  blank_id: 4333
+############################################
+#           Network Architecture           #
+############################################
+num_conv_layers: 2
+num_rnn_layers: 3
+rnn_layer_size: 1024
+use_gru: True 
+share_rnn_weights: False
+blank_id: 4333
 
-training:
-  n_epoch: 80
-  accum_grad: 1
-  lr: 2e-3
-  lr_decay: 0.83
-  weight_decay: 1e-06
-  global_grad_clip: 3.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 80
+accum_grad: 1
+lr: 2e-3
+lr_decay: 0.83
+weight_decay: 1e-06
+global_grad_clip: 3.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
+  
   
-decoding:
-  batch_size: 32
-  error_rate_type: cer 
-  decoding_method: ctc_beam_search
-  lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
-  alpha: 2.6
-  beta: 5.0
-  beam_size: 300
-  cutoff_prob: 0.99
-  cutoff_top_n: 40
-  num_proc_bsearch: 8

examples/other/1xt2x/aishell/conf/tuning/decode.yaml

@@ -0,0 +1,10 @@
+decode_batch_size: 32
+error_rate_type: cer 
+decoding_method: ctc_beam_search
+lang_model_path: data/lm/zh_giga.no_cna_cmn.prune01244.klm
+alpha: 2.6
+beta: 5.0
+beam_size: 300
+cutoff_prob: 0.99
+cutoff_top_n: 40
+num_proc_bsearch: 8

examples/other/1xt2x/aishell/local/test.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-model_type=$3
+decode_config_path=$2
+ckpt_prefix=$3
+model_type=$4
 
 # download language model
 bash local/download_lm_ch.sh
@@ -21,6 +22,7 @@ fi
 python3 -u ${BIN_DIR}/test.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
 --model_type ${model_type}

examples/other/1xt2x/aishell/run.sh

@@ -5,6 +5,7 @@ source path.sh
 stage=0
 stop_stage=100
 conf_path=conf/deepspeech2.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=1
 model_type=offline
 gpus=2
@@ -23,6 +24,6 @@ fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
 fi
 

examples/other/1xt2x/baidu_en8k/conf/deepspeech2.yaml

@@ -1,67 +1,64 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test-clean
-  min_input_len: 0.0
-  max_input_len: .inf # second
-  min_output_len: 0.0
-  max_output_len: .inf
-  min_output_input_ratio: 0.00
-  max_output_input_ratio: .inf
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test-clean
+min_input_len: 0.0
+max_input_len: .inf # second
+min_output_len: 0.0
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
-  batch_size: 64 # one gpu
-  mean_std_filepath: data/mean_std.npz
-  unit_type: char
-  vocab_filepath: data/vocab.txt 
-  augmentation_config: conf/augmentation.json
-  random_seed: 0
-  spm_model_prefix: 
-  spectrum_type: linear
-  feat_dim: 
-  delta_delta: False
-  stride_ms: 10.0
-  window_ms: 20.0
-  n_fft: None
-  max_freq: None
-  target_sample_rate: 16000
-  use_dB_normalization: True
-  target_dB: -20
-  dither: 1.0
-  keep_transcription_text: False
-  sortagrad: True
-  shuffle_method: batch_shuffle
-  num_workers: 2
+###########################################
+#              Dataloader                 #
+###########################################
+batch_size: 64 # one gpu
+mean_std_filepath: data/mean_std.npz
+unit_type: char
+vocab_filepath: data/vocab.txt 
+augmentation_config: conf/augmentation.json
+random_seed: 0
+spm_model_prefix: 
+spectrum_type: linear
+feat_dim: 
+delta_delta: False
+stride_ms: 10.0
+window_ms: 20.0
+n_fft: None
+max_freq: None
+target_sample_rate: 16000
+use_dB_normalization: True
+target_dB: -20
+dither: 1.0
+keep_transcription_text: False
+sortagrad: True
+shuffle_method: batch_shuffle
+num_workers: 2
 
-model:
-  num_conv_layers: 2
-  num_rnn_layers: 3
-  rnn_layer_size: 1024
-  use_gru: True
-  share_rnn_weights: False
-  blank_id: 28
+############################################
+#           Network Architecture           #
+############################################
+num_conv_layers: 2
+num_rnn_layers: 3
+rnn_layer_size: 1024
+use_gru: True
+share_rnn_weights: False
+blank_id: 28
+
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 80
+accum_grad: 1
+lr: 2e-3
+lr_decay: 0.83
+weight_decay: 1e-06
+global_grad_clip: 3.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
 
-training:
-  n_epoch: 80
-  accum_grad: 1
-  lr: 2e-3
-  lr_decay: 0.83
-  weight_decay: 1e-06
-  global_grad_clip: 3.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-  
-decoding:
-  batch_size: 32
-  error_rate_type: wer 
-  decoding_method: ctc_beam_search
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 1.4
-  beta: 0.35
-  beam_size: 500
-  cutoff_prob: 1.0
-  cutoff_top_n: 40
-  num_proc_bsearch: 8

examples/other/1xt2x/baidu_en8k/conf/tuning/decode.yaml

@@ -0,0 +1,10 @@
+decode_batch_size: 32
+error_rate_type: wer 
+decoding_method: ctc_beam_search
+lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
+alpha: 1.4
+beta: 0.35
+beam_size: 500
+cutoff_prob: 1.0
+cutoff_top_n: 40
+num_proc_bsearch: 8

examples/other/1xt2x/baidu_en8k/local/test.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-model_type=$3
+decode_config_path=$2
+ckpt_prefix=$3
+model_type=$4
 
 # download language model
 bash local/download_lm_en.sh
@@ -21,6 +22,7 @@ fi
 python3 -u ${BIN_DIR}/test.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
 --model_type ${model_type}

examples/other/1xt2x/baidu_en8k/run.sh

@@ -5,6 +5,7 @@ source path.sh
 stage=0
 stop_stage=100
 conf_path=conf/deepspeech2.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=1
 model_type=offline
 gpus=0
@@ -23,6 +24,6 @@ fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
 fi
 

examples/other/1xt2x/librispeech/conf/deepspeech2.yaml

@@ -1,67 +1,64 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test-clean
-  min_input_len: 0.0
-  max_input_len: 1000.0 # second
-  min_output_len: 0.0
-  max_output_len: .inf
-  min_output_input_ratio: 0.00
-  max_output_input_ratio: .inf
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test-clean
+min_input_len: 0.0
+max_input_len: 1000.0 # second
+min_output_len: 0.0
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
-  batch_size: 64 # one gpu
-  mean_std_filepath: data/mean_std.npz
-  unit_type: char
-  vocab_filepath: data/vocab.txt 
-  augmentation_config: conf/augmentation.json
-  random_seed: 0
-  spm_model_prefix: 
-  spectrum_type: linear
-  feat_dim: 
-  delta_delta: False
-  stride_ms: 10.0
-  window_ms: 20.0
-  n_fft: None
-  max_freq: None
-  target_sample_rate: 16000
-  use_dB_normalization: True
-  target_dB: -20
-  dither: 1.0
-  keep_transcription_text: False
-  sortagrad: True
-  shuffle_method: batch_shuffle
-  num_workers: 2
+###########################################
+#              Dataloader                 #
+###########################################
+batch_size: 64 # one gpu
+mean_std_filepath: data/mean_std.npz
+unit_type: char
+vocab_filepath: data/vocab.txt 
+augmentation_config: conf/augmentation.json
+random_seed: 0
+spm_model_prefix: 
+spectrum_type: linear
+feat_dim: 
+delta_delta: False
+stride_ms: 10.0
+window_ms: 20.0
+n_fft: None
+max_freq: None
+target_sample_rate: 16000
+use_dB_normalization: True
+target_dB: -20
+dither: 1.0
+keep_transcription_text: False
+sortagrad: True
+shuffle_method: batch_shuffle
+num_workers: 2
 
-model:
-  num_conv_layers: 2
-  num_rnn_layers: 3
-  rnn_layer_size: 2048
-  use_gru: False
-  share_rnn_weights: True
-  blank_id: 28
+############################################
+#           Network Architecture           #
+############################################
+num_conv_layers: 2
+num_rnn_layers: 3
+rnn_layer_size: 2048
+use_gru: False
+share_rnn_weights: True
+blank_id: 28
+
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 80
+accum_grad: 1
+lr: 2e-3
+lr_decay: 0.83
+weight_decay: 1e-06
+global_grad_clip: 3.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
 
-training:
-  n_epoch: 80
-  accum_grad: 1
-  lr: 2e-3
-  lr_decay: 0.83
-  weight_decay: 1e-06
-  global_grad_clip: 3.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-  
-decoding:
-  batch_size: 32
-  error_rate_type: wer 
-  decoding_method: ctc_beam_search
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 500
-  cutoff_prob: 1.0
-  cutoff_top_n: 40
-  num_proc_bsearch: 8

examples/other/1xt2x/librispeech/conf/tuning/decode.yaml

@@ -0,0 +1,10 @@
+decode_batch_size: 32
+error_rate_type: wer 
+decoding_method: ctc_beam_search
+lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
+alpha: 2.5
+beta: 0.3
+beam_size: 500
+cutoff_prob: 1.0
+cutoff_top_n: 40
+num_proc_bsearch: 8

examples/other/1xt2x/librispeech/local/test.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-model_type=$3
+decode_config_path=$2
+ckpt_prefix=$3
+model_type=$4
 
 # download language model
 bash local/download_lm_en.sh
@@ -21,6 +22,7 @@ fi
 python3 -u ${BIN_DIR}/test.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
 --model_type ${model_type}

examples/other/1xt2x/librispeech/run.sh

@@ -5,6 +5,7 @@ source path.sh
 stage=0
 stop_stage=100
 conf_path=conf/deepspeech2.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=1
 model_type=offline
 gpus=1
@@ -23,5 +24,5 @@ fi
 
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${v18_ckpt} ${model_type}|| exit -1
 fi

examples/other/1xt2x/src_deepspeech2x/bin/test.py

@@ -13,8 +13,8 @@
 # limitations under the License.
 """Evaluation for DeepSpeech2 model."""
 from src_deepspeech2x.test_model import DeepSpeech2Tester as Tester
+from yacs.config import CfgNode
 
-from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults
 from paddlespeech.s2t.training.cli import default_argument_parser
 from paddlespeech.s2t.utils.utility import print_arguments
 
@@ -41,9 +41,13 @@ if __name__ == "__main__":
     print("model_type:{}".format(args.model_type))
 
     # https://yaml.org/type/float.html
-    config = get_cfg_defaults(args.model_type)
+    config = CfgNode(new_allowed=True)
     if args.config:
         config.merge_from_file(args.config)
+    if args.decode_cfg:
+        decode_confs = CfgNode(new_allowed=True)
+        decode_confs.merge_from_file(args.decode_cfg)
+        config.decode = decode_confs
     if args.opts:
         config.merge_from_list(args.opts)
     config.freeze()

examples/other/1xt2x/src_deepspeech2x/models/ds2/deepspeech2.py

@@ -120,20 +120,6 @@ class DeepSpeech2Model(nn.Layer):
     :rtype: tuple of LayerOutput
     """
 
-    @classmethod
-    def params(cls, config: Optional[CfgNode]=None) -> CfgNode:
-        default = CfgNode(
-            dict(
-                num_conv_layers=2,  #Number of stacking convolution layers.
-                num_rnn_layers=3,  #Number of stacking RNN layers.
-                rnn_layer_size=1024,  #RNN layer size (number of RNN cells).
-                use_gru=True,  #Use gru if set True. Use simple rnn if set False.
-                share_rnn_weights=True  #Whether to share input-hidden weights between forward and backward directional RNNs.Notice that for GRU, weight sharing is not supported.
-            ))
-        if config is not None:
-            config.merge_from_other_cfg(default)
-        return default
-
     def __init__(self,
                  feat_size,
                  dict_size,
@@ -233,11 +219,11 @@ class DeepSpeech2Model(nn.Layer):
         """
         model = cls(feat_size=dataloader.collate_fn.feature_size,
                     dict_size=len(dataloader.collate_fn.vocab_list),
-                    num_conv_layers=config.model.num_conv_layers,
-                    num_rnn_layers=config.model.num_rnn_layers,
-                    rnn_size=config.model.rnn_layer_size,
-                    use_gru=config.model.use_gru,
-                    share_rnn_weights=config.model.share_rnn_weights)
+                    num_conv_layers=config.num_conv_layers,
+                    num_rnn_layers=config.num_rnn_layers,
+                    rnn_size=config.rnn_layer_size,
+                    use_gru=config.use_gru,
+                    share_rnn_weights=config.share_rnn_weights)
         infos = Checkpoint().load_parameters(
             model, checkpoint_path=checkpoint_path)
         logger.info(f"checkpoint info: {infos}")
@@ -250,7 +236,7 @@ class DeepSpeech2Model(nn.Layer):
         Parameters
 
         config: yacs.config.CfgNode
-            config.model
+            config
         Returns
         -------
         DeepSpeech2Model

examples/other/1xt2x/src_deepspeech2x/test_model.py

@@ -44,27 +44,11 @@ logger = Log(__name__).getlog()
 
 
 class DeepSpeech2Trainer(Trainer):
-    @classmethod
-    def params(cls, config: Optional[CfgNode]=None) -> CfgNode:
-        # training config
-        default = CfgNode(
-            dict(
-                lr=5e-4,  # learning rate
-                lr_decay=1.0,  # learning rate decay
-                weight_decay=1e-6,  # the coeff of weight decay
-                global_grad_clip=5.0,  # the global norm clip
-                n_epoch=50,  # train epochs
-            ))
-
-        if config is not None:
-            config.merge_from_other_cfg(default)
-        return default
-
     def __init__(self, config, args):
         super().__init__(config, args)
 
     def train_batch(self, batch_index, batch_data, msg):
-        train_conf = self.config.training
+        train_conf = self.config
         start = time.time()
 
         # forward
@@ -98,7 +82,7 @@ class DeepSpeech2Trainer(Trainer):
         iteration_time = time.time() - start
 
         msg += "train time: {:>.3f}s, ".format(iteration_time)
-        msg += "batch size: {}, ".format(self.config.collator.batch_size)
+        msg += "batch size: {}, ".format(self.config.batch_size)
         msg += "accum: {}, ".format(train_conf.accum_grad)
         msg += ', '.join('{}: {:>.6f}'.format(k, v)
                          for k, v in losses_np.items())
@@ -126,7 +110,7 @@ class DeepSpeech2Trainer(Trainer):
                 total_loss += float(loss) * num_utts
                 valid_losses['val_loss'].append(float(loss))
 
-            if (i + 1) % self.config.training.log_interval == 0:
+            if (i + 1) % self.config.log_interval == 0:
                 valid_dump = {k: np.mean(v) for k, v in valid_losses.items()}
                 valid_dump['val_history_loss'] = total_loss / num_seen_utts
 
@@ -146,15 +130,15 @@ class DeepSpeech2Trainer(Trainer):
     def setup_model(self):
         config = self.config.clone()
         config.defrost()
-        config.model.feat_size = self.train_loader.collate_fn.feature_size
-        #config.model.dict_size = self.train_loader.collate_fn.vocab_size
-        config.model.dict_size = len(self.train_loader.collate_fn.vocab_list)
+        config.feat_size = self.train_loader.collate_fn.feature_size
+        #config.dict_size = self.train_loader.collate_fn.vocab_size
+        config.dict_size = len(self.train_loader.collate_fn.vocab_list)
         config.freeze()
 
         if self.args.model_type == 'offline':
-            model = DeepSpeech2Model.from_config(config.model)
+            model = DeepSpeech2Model.from_config(config)
         elif self.args.model_type == 'online':
-            model = DeepSpeech2ModelOnline.from_config(config.model)
+            model = DeepSpeech2ModelOnline.from_config(config)
         else:
             raise Exception("wrong model type")
         if self.parallel:
@@ -163,17 +147,13 @@ class DeepSpeech2Trainer(Trainer):
         logger.info(f"{model}")
         layer_tools.print_params(model, logger.info)
 
-        grad_clip = ClipGradByGlobalNormWithLog(
-            config.training.global_grad_clip)
+        grad_clip = ClipGradByGlobalNormWithLog(config.global_grad_clip)
         lr_scheduler = paddle.optimizer.lr.ExponentialDecay(
-            learning_rate=config.training.lr,
-            gamma=config.training.lr_decay,
-            verbose=True)
+            learning_rate=config.lr, gamma=config.lr_decay, verbose=True)
         optimizer = paddle.optimizer.Adam(
             learning_rate=lr_scheduler,
             parameters=model.parameters(),
-            weight_decay=paddle.regularizer.L2Decay(
-                config.training.weight_decay),
+            weight_decay=paddle.regularizer.L2Decay(config.weight_decay),
             grad_clip=grad_clip)
 
         self.model = model
@@ -184,59 +164,59 @@ class DeepSpeech2Trainer(Trainer):
     def setup_dataloader(self):
         config = self.config.clone()
         config.defrost()
-        config.collator.keep_transcription_text = False
+        config.keep_transcription_text = False
 
-        config.data.manifest = config.data.train_manifest
+        config.manifest = config.train_manifest
         train_dataset = ManifestDataset.from_config(config)
 
-        config.data.manifest = config.data.dev_manifest
+        config.manifest = config.dev_manifest
         dev_dataset = ManifestDataset.from_config(config)
 
-        config.data.manifest = config.data.test_manifest
+        config.manifest = config.test_manifest
         test_dataset = ManifestDataset.from_config(config)
 
         if self.parallel:
             batch_sampler = SortagradDistributedBatchSampler(
                 train_dataset,
-                batch_size=config.collator.batch_size,
+                batch_size=config.batch_size,
                 num_replicas=None,
                 rank=None,
                 shuffle=True,
                 drop_last=True,
-                sortagrad=config.collator.sortagrad,
-                shuffle_method=config.collator.shuffle_method)
+                sortagrad=config.sortagrad,
+                shuffle_method=config.shuffle_method)
         else:
             batch_sampler = SortagradBatchSampler(
                 train_dataset,
                 shuffle=True,
-                batch_size=config.collator.batch_size,
+                batch_size=config.batch_size,
                 drop_last=True,
-                sortagrad=config.collator.sortagrad,
-                shuffle_method=config.collator.shuffle_method)
+                sortagrad=config.sortagrad,
+                shuffle_method=config.shuffle_method)
 
         collate_fn_train = SpeechCollator.from_config(config)
 
-        config.collator.augmentation_config = ""
+        config.augmentation_config = ""
         collate_fn_dev = SpeechCollator.from_config(config)
 
-        config.collator.keep_transcription_text = True
-        config.collator.augmentation_config = ""
+        config.keep_transcription_text = True
+        config.augmentation_config = ""
         collate_fn_test = SpeechCollator.from_config(config)
 
         self.train_loader = DataLoader(
             train_dataset,
             batch_sampler=batch_sampler,
             collate_fn=collate_fn_train,
-            num_workers=config.collator.num_workers)
+            num_workers=config.num_workers)
         self.valid_loader = DataLoader(
             dev_dataset,
-            batch_size=config.collator.batch_size,
+            batch_size=config.batch_size,
             shuffle=False,
             drop_last=False,
             collate_fn=collate_fn_dev)
         self.test_loader = DataLoader(
             test_dataset,
-            batch_size=config.decoding.batch_size,
+            batch_size=config.decode.decode_batch_size,
             shuffle=False,
             drop_last=False,
             collate_fn=collate_fn_test)
@@ -250,31 +230,10 @@ class DeepSpeech2Trainer(Trainer):
 
 
 class DeepSpeech2Tester(DeepSpeech2Trainer):
-    @classmethod
-    def params(cls, config: Optional[CfgNode]=None) -> CfgNode:
-        # testing config
-        default = CfgNode(
-            dict(
-                alpha=2.5,  # Coef of LM for beam search.
-                beta=0.3,  # Coef of WC for beam search.
-                cutoff_prob=1.0,  # Cutoff probability for pruning.
-                cutoff_top_n=40,  # Cutoff number for pruning.
-                lang_model_path='models/lm/common_crawl_00.prune01111.trie.klm',  # Filepath for language model.
-                decoding_method='ctc_beam_search',  # Decoding method. Options: ctc_beam_search, ctc_greedy
-                error_rate_type='wer',  # Error rate type for evaluation. Options `wer`, 'cer'
-                num_proc_bsearch=8,  # # of CPUs for beam search.
-                beam_size=500,  # Beam search width.
-                batch_size=128,  # decoding batch size
-            ))
-
-        if config is not None:
-            config.merge_from_other_cfg(default)
-        return default
-
     def __init__(self, config, args):
 
         self._text_featurizer = TextFeaturizer(
-            unit_type=config.collator.unit_type, vocab_filepath=None)
+            unit_type=config.unit_type, vocab=None)
         super().__init__(config, args)
 
     def ordid2token(self, texts, texts_len):
@@ -293,7 +252,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
                         texts,
                         texts_len,
                         fout=None):
-        cfg = self.config.decoding
+        cfg = self.config.decode
         errors_sum, len_refs, num_ins = 0.0, 0, 0
         errors_func = error_rate.char_errors if cfg.error_rate_type == 'cer' else error_rate.word_errors
         error_rate_func = error_rate.cer if cfg.error_rate_type == 'cer' else error_rate.wer
@@ -399,31 +358,3 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
             self.export()
         except KeyboardInterrupt:
             exit(-1)
-
-    def setup(self):
-        """Setup the experiment.
-        """
-        paddle.set_device('gpu' if self.args.ngpu > 0 else 'cpu')
-
-        self.setup_output_dir()
-        self.setup_checkpointer()
-
-        self.setup_dataloader()
-        self.setup_model()
-
-        self.iteration = 0
-        self.epoch = 0
-
-    def setup_output_dir(self):
-        """Create a directory used for output.
-        """
-        # output dir
-        if self.args.output:
-            output_dir = Path(self.args.output).expanduser()
-            output_dir.mkdir(parents=True, exist_ok=True)
-        else:
-            output_dir = Path(
-                self.args.checkpoint_path).expanduser().parent.parent
-            output_dir.mkdir(parents=True, exist_ok=True)
-
-        self.output_dir = output_dir

examples/ted_en_zh/st0/conf/preprocess.yaml

@@ -0,0 +1,25 @@
+process:
+  # extract kaldi fbank from PCM
+  - type: fbank_kaldi
+    fs: 16000
+    n_mels: 80
+    n_shift: 160
+    win_length: 400
+    dither: 0.1
+  - type: cmvn_json
+    cmvn_path: data/mean_std.json
+  # these three processes are a.k.a. SpecAugument
+  - type: time_warp
+    max_time_warp: 5
+    inplace: true
+    mode: PIL
+  - type: freq_mask
+    F: 30
+    n_mask: 2
+    inplace: true
+    replace_with_zero: false
+  - type: time_mask
+    T: 40
+    n_mask: 2
+    inplace: true
+    replace_with_zero: false

examples/ted_en_zh/st0/conf/transformer.yaml

@@ -1,114 +1,98 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
-  min_input_len: 0.05  # second
-  max_input_len: 30.0 # second
-  min_output_len: 0.0 # tokens
-  max_output_len: 400.0 # tokens
-  min_output_input_ratio: 0.01
-  max_output_input_ratio: 20.0
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
+min_input_len: 0.05  # second
+max_input_len: 30.0 # second
+min_output_len: 0.0 # tokens
+max_output_len: 400.0 # tokens
+min_output_input_ratio: 0.01
+max_output_input_ratio: 20.0
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt
-  unit_type: 'spm'
-  spm_model_prefix: data/lang_char/bpe_unigram_8000
-  mean_std_filepath: ""
-  augmentation_config: conf/preprocess.yaml
-  batch_size: 16
-  maxlen_in: 5  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  raw_wav: True  # use raw_wav or kaldi feature
-  spectrum_type: fbank #linear, mfcc, fbank
-  feat_dim: 80
-  delta_delta: False
-  dither: 1.0
-  target_sample_rate: 16000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 25.0
-  use_dB_normalization: True
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt
+unit_type: 'spm'
+spm_model_prefix: data/lang_char/bpe_unigram_8000
+mean_std_filepath: ""
+preprocess_config: conf/preprocess.yaml
+batch_size: 16
+maxlen_in: 5  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+raw_wav: True  # use raw_wav or kaldi feature
+spectrum_type: fbank #linear, mfcc, fbank
+feat_dim: 80
+delta_delta: False
+dither: 1.0
+target_sample_rate: 16000
+max_freq: None
+n_fft: None
+stride_ms: 10.0
+window_ms: 25.0
+use_dB_normalization: True
+target_dB: -20
+random_seed: 0
+keep_transcription_text: False
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 2
 
 
-# network architecture
-model:
-    cmvn_file: "data/mean_std.json"
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: "data/mean_std.json"
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
-
-    # hybrid CTC/attention
-    model_conf:
-        asr_weight: 0.0
-        ctc_weight: 0.0
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
-
-
-training:
-  n_epoch: 120
-  accum_grad: 2
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 2.5
-    weight_decay: 1e-06
-  scheduler: noam    
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 50
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-
-decoding:
-  batch_size: 5
-  error_rate_type: char-bleu
-  decoding_method: fullsentence  # 'fullsentence', 'simultaneous'
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  word_reward: 0.7
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
+# hybrid CTC/attention
+model_conf:
+    asr_weight: 0.0
+    ctc_weight: 0.0
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 120
+accum_grad: 2
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 2.5
+  weight_decay: 1.0e-06
+scheduler: noam    
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 50
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/ted_en_zh/st0/conf/transformer_mtl_noam.yaml

@@ -1,114 +1,102 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
-  min_input_len: 0.05  # second
-  max_input_len: 30.0 # second
-  min_output_len: 0.0 # tokens
-  max_output_len: 400.0 # tokens
-  min_output_input_ratio: 0.01
-  max_output_input_ratio: 20.0
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
+min_input_len: 0.05  # second
+max_input_len: 30.0 # second
+min_output_len: 0.0 # tokens
+max_output_len: 400.0 # tokens
+min_output_input_ratio: 0.01
+max_output_input_ratio: 20.0
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt
-  unit_type: 'spm'
-  spm_model_prefix: data/lang_char/bpe_unigram_8000
-  mean_std_filepath: ""
-  augmentation_config: conf/preprocess.yaml
-  batch_size: 16
-  maxlen_in: 5  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  raw_wav: True  # use raw_wav or kaldi feature
-  spectrum_type: fbank #linear, mfcc, fbank
-  feat_dim: 80
-  delta_delta: False
-  dither: 1.0
-  target_sample_rate: 16000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 25.0
-  use_dB_normalization: True
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt
+unit_type: 'spm'
+spm_model_prefix: data/lang_char/bpe_unigram_8000
+mean_std_filepath: ""
+preprocess_config: conf/preprocess.yaml
+batch_size: 16
+maxlen_in: 5  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+raw_wav: True  # use raw_wav or kaldi feature
+spectrum_type: fbank #linear, mfcc, fbank
+feat_dim: 80
+delta_delta: False
+dither: 1.0
+target_sample_rate: 16000
+max_freq: None
+n_fft: None
+stride_ms: 10.0
+window_ms: 25.0
+use_dB_normalization: True
+target_dB: -20
+random_seed: 0
+keep_transcription_text: False
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 2
 
 
-# network architecture
-model:
-    cmvn_file: "data/mean_std.json"
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: "data/mean_std.json"
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        asr_weight: 0.5
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    asr_weight: 0.5
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
 
-training:
-  n_epoch: 120
-  accum_grad: 2
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 2.5
-    weight_decay: 1e-06
-  scheduler: noam    
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 50
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 120
+accum_grad: 2
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 2.5
+  weight_decay: 1.0e-06
+scheduler: noam    
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 50
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
 
 
-decoding:
-  batch_size: 5
-  error_rate_type: char-bleu
-  decoding_method: fullsentence  # 'fullsentence', 'simultaneous'
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  word_reward: 0.7
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
-
 

examples/ted_en_zh/st0/conf/tuning/decode.yaml

@@ -0,0 +1,11 @@
+batch_size: 5
+error_rate_type: char-bleu
+decoding_method: fullsentence  # 'fullsentence', 'simultaneous'
+beam_size: 10
+word_reward: 0.7
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.

examples/ted_en_zh/st0/local/test.sh

@@ -1,7 +1,7 @@
 #! /usr/bin/env bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,16 +9,18 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 for type in fullsentence; do
     echo "decoding ${type}"
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
+    --opts decode.decoding_method ${type} \
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"

examples/ted_en_zh/st0/run.sh

@@ -6,6 +6,7 @@ gpus=0,1,2,3
 stage=0
 stop_stage=50
 conf_path=conf/transformer_mtl_noam.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=5
 data_path=./TED_EnZh # path to unzipped data
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
@@ -32,7 +33,7 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then

examples/ted_en_zh/st1/conf/preprocess.yaml

@@ -0,0 +1,16 @@
+process:
+  # these three processes are a.k.a. SpecAugument
+  - type: time_warp
+    max_time_warp: 5
+    inplace: true
+    mode: PIL
+  - type: freq_mask
+    F: 30
+    n_mask: 2
+    inplace: true
+    replace_with_zero: false
+  - type: time_mask
+    T: 40
+    n_mask: 2
+    inplace: true
+    replace_with_zero: false

examples/ted_en_zh/st1/conf/transformer.yaml

@@ -1,104 +1,90 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
-collator:
-  vocab_filepath: data/lang_char/ted_en_zh_bpe8000.txt
-  unit_type: 'spm'
-  spm_model_prefix: data/lang_char/ted_en_zh_bpe8000
-  mean_std_filepath: ""
-  # augmentation_config: conf/augmentation.json
-  batch_size: 20
-  feat_dim: 83
-  stride_ms: 10.0
-  window_ms: 25.0
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  minibatches: 0 # for debug
-  batch_count: auto
-  batch_bins: 0 
-  batch_frames_in: 0
-  batch_frames_out: 0
-  batch_frames_inout: 0
-  augmentation_config:
-  num_workers: 0
-  subsampling_factor: 1
-  num_encs: 1
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/ted_en_zh_bpe8000.txt
+unit_type: 'spm'
+spm_model_prefix: data/lang_char/ted_en_zh_bpe8000
+mean_std_filepath: ""
+# preprocess_config: conf/augmentation.json
+batch_size: 20
+feat_dim: 83
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+preprocess_config:
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
 
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: None
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-# network architecture
-model:
-    cmvn_file: None
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# hybrid CTC/attention
+model_conf:
+    asr_weight: 0.0
+    ctc_weight: 0.0
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
-    # hybrid CTC/attention
-    model_conf:
-        asr_weight: 0.0
-        ctc_weight: 0.0
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
 
-
-training:
-  n_epoch: 40
-  accum_grad: 2
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 2.5
-    weight_decay: 0.
-  scheduler: noam    
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 50
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-
-decoding:
-  batch_size: 5
-  error_rate_type: char-bleu
-  decoding_method: fullsentence  # 'fullsentence', 'simultaneous'
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  word_reward: 0.7
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 40
+accum_grad: 2
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 2.5
+  weight_decay: 0.
+scheduler: noam    
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 50
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/ted_en_zh/st1/conf/transformer_mtl_noam.yaml

@@ -1,104 +1,90 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
-collator:
-  vocab_filepath: data/lang_char/ted_en_zh_bpe8000.txt
-  unit_type: 'spm'
-  spm_model_prefix: data/lang_char/ted_en_zh_bpe8000
-  mean_std_filepath: ""
-  # augmentation_config: conf/augmentation.json
-  batch_size: 20
-  feat_dim: 83
-  stride_ms: 10.0
-  window_ms: 25.0
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  minibatches: 0 # for debug
-  batch_count: auto
-  batch_bins: 0 
-  batch_frames_in: 0
-  batch_frames_out: 0
-  batch_frames_inout: 0
-  augmentation_config:
-  num_workers: 0
-  subsampling_factor: 1
-  num_encs: 1
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/ted_en_zh_bpe8000.txt
+unit_type: 'spm'
+spm_model_prefix: data/lang_char/ted_en_zh_bpe8000
+mean_std_filepath: ""
+# preprocess_config: conf/augmentation.json
+batch_size: 20
+feat_dim: 83
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+preprocess_config:
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
 
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: None
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-# network architecture
-model:
-    cmvn_file: None
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# hybrid CTC/attention
+model_conf:
+    asr_weight: 0.5
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
-    # hybrid CTC/attention
-    model_conf:
-        asr_weight: 0.5
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
 
-
-training:
-  n_epoch: 40
-  accum_grad: 2
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 2.5
-    weight_decay: 0.
-  scheduler: noam    
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 50
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-
-decoding:
-  batch_size: 5
-  error_rate_type: char-bleu
-  decoding_method: fullsentence  # 'fullsentence', 'simultaneous'
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  word_reward: 0.7
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 40
+accum_grad: 2
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 2.5
+  weight_decay: 0.
+scheduler: noam    
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 50
+checkpoint:
+  kbest_n: 50
+  latest_n: 5

examples/ted_en_zh/st1/conf/tuning/decode.yaml

@@ -0,0 +1,12 @@
+
+batch_size: 5
+error_rate_type: char-bleu
+decoding_method: fullsentence  # 'fullsentence', 'simultaneous'
+beam_size: 10
+word_reward: 0.7
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.

examples/ted_en_zh/st1/local/test.sh

@@ -1,7 +1,7 @@
 #! /usr/bin/env bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,16 +9,18 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 for type in fullsentence; do
     echo "decoding ${type}"
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
+    --opts decode.decoding_method ${type} \
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"

examples/ted_en_zh/st1/run.sh

@@ -7,6 +7,7 @@ gpus=0,1,2,3
 stage=1
 stop_stage=4
 conf_path=conf/transformer_mtl_noam.yaml
+decode_conf_path=conf/tuning/decode.yaml
 ckpt_path= # paddle.98 # (finetune from FAT-ST pretrained model)
 avg_num=5
 data_path=./TED_EnZh # path to unzipped data
@@ -27,7 +28,7 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     if [ -n "${ckpt_path}" ]; then
         echo "Finetune from Pretrained Model" ${ckpt_path}
         ./local/download_pretrain.sh || exit -1
-    fi 
+    fi
     CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${ckpt} "${ckpt_path}"
 fi
 
@@ -38,5 +39,5 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
-fi
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+fi

examples/timit/asr1/conf/transformer.yaml

@@ -1,110 +1,89 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
-  min_input_len: 0.0  # second
-  max_input_len: 10.0 # second
-  min_output_len: 0.0 # tokens
-  max_output_len: 150.0 # tokens
-  min_output_input_ratio: 0.005
-  max_output_input_ratio: 1000.0
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt
-  unit_type: "word"
-  mean_std_filepath: ""
-  augmentation_config: conf/preprocess.yaml
-  batch_size: 64
-  raw_wav: True  # use raw_wav or kaldi feature
-  spectrum_type: fbank #linear, mfcc, fbank
-  feat_dim: 80
-  delta_delta: False
-  dither: 1.0
-  target_sample_rate: 16000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 25.0
-  use_dB_normalization: True
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt
+spm_model_prefix: ''
+unit_type: "word"
+mean_std_filepath: ""
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 64
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
 
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 128    # dimension of attention
-        attention_heads: 4
-        linear_units: 1024  # the number of units of position-wise feed forward
-        num_blocks: 6      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 128    # dimension of attention
+    attention_heads: 4
+    linear_units: 1024  # the number of units of position-wise feed forward
+    num_blocks: 6      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 1024
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 1024
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.5
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.5
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
 
-training:
-  n_epoch: 50
-  accum_grad: 1
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.004
-    weight_decay: 1e-06
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 1200
-    lr_decay: 1.0
-  log_interval: 10
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-
-
-decoding:
-  batch_size: 64
-  error_rate_type: wer
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
-
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 50
+accum_grad: 1
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.004
+  weight_decay: 1.0e-6
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 1200
+  lr_decay: 1.0
+log_interval: 10
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
 

examples/timit/asr1/conf/tuning/decode.yaml

@@ -0,0 +1,11 @@
+decode_batch_size: 64
+error_rate_type: wer
+decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+beam_size: 10
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.

examples/timit/asr1/local/align.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 batch_size=1
 output_dir=${ckpt_prefix}
@@ -20,9 +21,10 @@ mkdir -p ${output_dir}
 python3 -u ${BIN_DIR}/alignment.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
---opts decoding.batch_size ${batch_size}
+--opts decode.decode_batch_size ${batch_size}
 
 if [ $? -ne 0 ]; then
     echo "Failed in ctc alignment!"

examples/timit/asr1/local/test.sh

@@ -7,8 +7,8 @@ stop_stage=50
 
 . ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -17,7 +17,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 chunk_mode=false
 if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
@@ -43,10 +44,11 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
         python3 -u ${BIN_DIR}/test.py \
         --ngpu ${ngpu} \
         --config ${config_path} \
+        --decode_cfg ${decode_config_path} \
         --result_file ${ckpt_prefix}.${type}.rsl \
         --checkpoint_path ${ckpt_prefix} \
-        --opts decoding.decoding_method ${type} \
-        --opts decoding.batch_size ${batch_size}
+        --opts decode.decoding_method ${type} \
+        --opts decode.decode_batch_size ${batch_size}
     
         if [ $? -ne 0 ]; then
             echo "Failed in evaluation!"
@@ -63,10 +65,11 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
         python3 -u ${BIN_DIR}/test.py \
         --ngpu ${ngpu}  \
         --config ${config_path} \
+        --decode_cfg ${decode_config_path} \
         --result_file ${ckpt_prefix}.${type}.rsl \
         --checkpoint_path ${ckpt_prefix} \
-        --opts decoding.decoding_method ${type} \
-        --opts decoding.batch_size ${batch_size}
+        --opts decode.decoding_method ${type} \
+        --opts decode.decode_batch_size ${batch_size}
     
         if [ $? -ne 0 ]; then
             echo "Failed in evaluation!"
@@ -82,10 +85,11 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
         python3 -u ${BIN_DIR}/test.py \
         --ngpu ${ngpu}  \
         --config ${config_path} \
+        --decode_cfg ${decode_config_path} \
         --result_file ${ckpt_prefix}.${type}.rsl \
         --checkpoint_path ${ckpt_prefix} \
-        --opts decoding.decoding_method ${type} \
-        --opts decoding.batch_size ${batch_size}
+        --opts decode.decoding_method ${type} \
+        --opts decode.decode_batch_size ${batch_size}
     
         if [ $? -ne 0 ]; then
             echo "Failed in evaluation!"

examples/timit/asr1/run.sh

@@ -7,6 +7,7 @@ gpus=0,1,2,3
 stage=0
 stop_stage=50
 conf_path=conf/transformer.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=10
 TIMIT_path=/path/to/TIMIT
 
@@ -34,15 +35,15 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
     # ctc alignment of test data
-    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
-# if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
-#     # export ckpt avg_n
-#     CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
-# fi
+if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then
+     # export ckpt avg_n
+     CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
+fi

examples/tiny/asr0/conf/deepspeech2.yaml

@@ -1,70 +1,67 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.tiny
-  dev_manifest: data/manifest.tiny
-  test_manifest: data/manifest.tiny 
-  min_input_len: 0.0
-  max_input_len: 30.0
-  min_output_len: 0.0
-  max_output_len: 400.0
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.tiny
+dev_manifest: data/manifest.tiny
+test_manifest: data/manifest.tiny 
+min_input_len: 0.0
+max_input_len: 30.0
+min_output_len: 0.0
+max_output_len: 400.0
+min_output_input_ratio: 0.05
+max_output_input_ratio: 10.0
 
 
-collator:
-  mean_std_filepath: data/mean_std.json
-  unit_type: char
-  vocab_filepath: data/lang_char/vocab.txt
-  augmentation_config: conf/augmentation.json
-  random_seed: 0
-  spm_model_prefix: 
-  spectrum_type: linear
-  feat_dim: 
-  delta_delta: False
-  stride_ms: 10.0
-  window_ms: 20.0
-  n_fft: None
-  max_freq: None
-  target_sample_rate: 16000
-  use_dB_normalization: True
-  target_dB: -20
-  dither: 1.0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
-  batch_size: 4
+###########################################
+#              Dataloader                 #
+###########################################
+mean_std_filepath: data/mean_std.json
+unit_type: char
+vocab_filepath: data/lang_char/vocab.txt
+augmentation_config: conf/augmentation.json
+random_seed: 0
+spm_model_prefix: 
+spectrum_type: linear
+feat_dim: 
+delta_delta: False
+stride_ms: 10.0
+window_ms: 20.0
+n_fft: None
+max_freq: None
+target_sample_rate: 16000
+use_dB_normalization: True
+target_dB: -20
+dither: 1.0
+keep_transcription_text: False
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 2
+batch_size: 4
   
-model:
-  num_conv_layers: 2
-  num_rnn_layers: 3
-  rnn_layer_size: 2048
-  use_gru: False 
-  share_rnn_weights: True 
-  blank_id: 0
+############################################
+#           Network Architecture           #
+############################################
+num_conv_layers: 2
+num_rnn_layers: 3
+rnn_layer_size: 2048
+use_gru: False 
+share_rnn_weights: True 
+blank_id: 0
   
 
-training:
-  n_epoch: 5
-  accum_grad: 1
-  lr: 1e-5 
-  lr_decay: 0.8 
-  weight_decay: 1e-06
-  global_grad_clip: 5.0
-  log_interval: 1
-  checkpoint:
-    kbest_n: 3
-    latest_n: 2
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 5
+accum_grad: 1
+lr: 1e-5 
+lr_decay: 0.8 
+weight_decay: 1e-06
+global_grad_clip: 5.0
+log_interval: 1
+checkpoint:
+  kbest_n: 3
+  latest_n: 2
 
 
-decoding:
-  batch_size: 128
-  error_rate_type: wer
-  decoding_method: ctc_beam_search
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 500
-  cutoff_prob: 1.0
-  cutoff_top_n: 40
-  num_proc_bsearch: 8

examples/tiny/asr0/conf/deepspeech2_online.yaml

@@ -1,72 +1,68 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.tiny
-  dev_manifest: data/manifest.tiny
-  test_manifest: data/manifest.tiny 
-  min_input_len: 0.0
-  max_input_len: 30.0
-  min_output_len: 0.0
-  max_output_len: 400.0
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.tiny
+dev_manifest: data/manifest.tiny
+test_manifest: data/manifest.tiny 
+min_input_len: 0.0
+max_input_len: 30.0
+min_output_len: 0.0
+max_output_len: 400.0
+min_output_input_ratio: 0.05
+max_output_input_ratio: 10.0
 
 
-collator:
-  mean_std_filepath: data/mean_std.json
-  unit_type: char
-  vocab_filepath: data/lang_char/vocab.txt
-  augmentation_config: conf/augmentation.json
-  random_seed: 0
-  spm_model_prefix: 
-  spectrum_type: linear
-  feat_dim: 
-  delta_delta: False
-  stride_ms: 10.0
-  window_ms: 20.0
-  n_fft: None
-  max_freq: None
-  target_sample_rate: 16000
-  use_dB_normalization: True
-  target_dB: -20
-  dither: 1.0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 0
-  batch_size: 4
+###########################################
+#              Dataloader                 #
+###########################################
+mean_std_filepath: data/mean_std.json
+unit_type: char
+vocab_filepath: data/lang_char/vocab.txt
+augmentation_config: conf/augmentation.json
+random_seed: 0
+spm_model_prefix: 
+spectrum_type: linear
+feat_dim: 
+delta_delta: False
+stride_ms: 10.0
+window_ms: 20.0
+n_fft: None
+max_freq: None
+target_sample_rate: 16000
+use_dB_normalization: True
+target_dB: -20
+dither: 1.0
+keep_transcription_text: False
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 0
+batch_size: 4
   
-model:
-  num_conv_layers: 2
-  num_rnn_layers: 4
-  rnn_layer_size: 2048
-  rnn_direction: forward
-  num_fc_layers: 2
-  fc_layers_size_list: 512, 256
-  use_gru: True 
-  blank_id: 0
+############################################
+#           Network Architecture           #
+############################################
+num_conv_layers: 2
+num_rnn_layers: 4
+rnn_layer_size: 2048
+rnn_direction: forward
+num_fc_layers: 2
+fc_layers_size_list: 512, 256
+use_gru: True 
+blank_id: 0
   
 
-training:
-  n_epoch: 5
-  accum_grad: 1
-  lr: 1e-5 
-  lr_decay: 1.0 
-  weight_decay: 1e-06
-  global_grad_clip: 5.0
-  log_interval: 1
-  checkpoint:
-    kbest_n: 3
-    latest_n: 2
+###########################################
+#                Training                 #
+###########################################
+n_epoch: 5
+accum_grad: 1
+lr: 1e-5 
+lr_decay: 1.0 
+weight_decay: 1e-06
+global_grad_clip: 5.0
+log_interval: 1
+checkpoint:
+  kbest_n: 3
+  latest_n: 2
 
-
-decoding:
-  batch_size: 128
-  error_rate_type: wer
-  decoding_method: ctc_beam_search
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 500
-  cutoff_prob: 1.0
-  cutoff_top_n: 40
-  num_proc_bsearch: 8

examples/tiny/asr0/conf/tuning/chunk_decode.yaml

@@ -0,0 +1,10 @@
+decode_batch_size: 128
+error_rate_type: wer
+decoding_method: ctc_beam_search
+lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
+alpha: 2.5
+beta: 0.3
+beam_size: 500
+cutoff_prob: 1.0
+cutoff_top_n: 40
+num_proc_bsearch: 8

examples/tiny/asr0/conf/tuning/decode.yaml

@@ -0,0 +1,10 @@
+decode_batch_size: 128
+error_rate_type: wer
+decoding_method: ctc_beam_search
+lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
+alpha: 2.5
+beta: 0.3
+beam_size: 500
+cutoff_prob: 1.0
+cutoff_top_n: 40
+num_proc_bsearch: 8

examples/tiny/asr0/local/test.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix model_type"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-model_type=$3
+decode_config_path=$2
+ckpt_prefix=$3
+model_type=$4
 
 # download language model
 bash local/download_lm_en.sh
@@ -21,6 +22,7 @@ fi
 python3 -u ${BIN_DIR}/test.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${ckpt_prefix}.rsl \
 --checkpoint_path ${ckpt_prefix} \
 --model_type ${model_type}

examples/tiny/asr0/run.sh

@@ -6,6 +6,7 @@ gpus=0
 stage=0
 stop_stage=100
 conf_path=conf/deepspeech2.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=1
 model_type=offline
 
@@ -32,7 +33,7 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} || exit -1
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then

examples/tiny/asr1/conf/chunk_confermer.yaml

@@ -1,120 +1,97 @@
-# https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.tiny
-  dev_manifest: data/manifest.tiny
-  test_manifest: data/manifest.tiny
-  min_input_len: 0.5  # second
-  max_input_len: 30.0 # second
-  min_output_len: 0.0 # tokens
-  max_output_len: 400.0 # tokens
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
-  
-collator:
-  mean_std_filepath: ""
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'spm'
-  spm_model_prefix: 'data/lang_char/bpe_unigram_200'
-  augmentation_config: conf/preprocess.yaml
-  batch_size: 4
-  raw_wav: True  # use raw_wav or kaldi feature
-  spectrum_type: fbank #linear, mfcc, fbank
-  feat_dim: 80
-  delta_delta: False
-  dither: 1.0
-  target_sample_rate: 16000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 25.0
-  use_dB_normalization: True
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
-
-
-# network architecture
-model:
-    cmvn_file: "data/mean_std.json"
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: conformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: True
-        use_cnn_module: True
-        cnn_module_kernel: 15
-        activation_type: 'swish'
-        pos_enc_layer_type: 'rel_pos'
-        selfattention_layer_type: 'rel_selfattn'
-        causal: True
-        use_dynamic_chunk: True
-        cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
-        use_dynamic_left_chunk: false
-
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: "data/mean_std.json"
+cmvn_file_type: "json"
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: True
+    use_cnn_module: True
+    cnn_module_kernel: 15
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
+    causal: True
+    use_dynamic_chunk: True
+    cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
+    use_dynamic_left_chunk: false
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
-training:
-  n_epoch: 5
-  accum_grad: 1
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.001
-    weight_decay: 1e-06
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 1
-  checkpoint:
-    kbest_n: 10
-    latest_n: 1
 
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.tiny
+dev_manifest: data/manifest.tiny
+test_manifest: data/manifest.tiny
 
-decoding:
-  batch_size: 64
-  error_rate_type: wer
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
-
+  
+###########################################
+#              Dataloader                 #
+###########################################
+mean_std_filepath: ""
+vocab_filepath: data/lang_char/vocab.txt 
+unit_type: 'spm'
+spm_model_prefix: 'data/lang_char/bpe_unigram_200'
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 4
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+preprocess_config: conf/preprocess.yaml 
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
+  
 
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 5
+accum_grad: 1
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.001
+  weight_decay: 1.0e-06
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 1
+checkpoint:
+  kbest_n: 10
+  latest_n: 1

examples/tiny/asr1/conf/chunk_transformer.yaml

@@ -1,113 +1,91 @@
-# https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.tiny
-  dev_manifest: data/manifest.tiny
-  test_manifest: data/manifest.tiny
-  min_input_len: 0.5  # second
-  max_input_len: 20.0 # second
-  min_output_len: 0.0 # tokens
-  max_output_len: 400.0 # tokens
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
-  
-collator:
-  mean_std_filepath: ""
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'spm'
-  spm_model_prefix: 'data/lang_char/bpe_unigram_200'
-  augmentation_config: conf/preprocess.yaml
-  batch_size: 4
-  raw_wav: True  # use raw_wav or kaldi feature
-  spectrum_type: fbank #linear, mfcc, fbank
-  feat_dim: 80
-  delta_delta: False
-  dither: 1.0
-  target_sample_rate: 16000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 25.0
-  use_dB_normalization: True
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
-
-
-# network architecture
-model:
-    cmvn_file: "data/mean_std.json"
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
-        use_dynamic_chunk: true
-        use_dynamic_left_chunk: false
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: "data/mean_std.json"
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
+    use_dynamic_chunk: true
+    use_dynamic_left_chunk: false
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
 
-training:
-  n_epoch: 5
-  accum_grad: 1
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.002
-    weight_decay: 1e-06
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 1
-  checkpoint:
-    kbest_n: 10
-    latest_n: 1
-
+# https://yaml.org/type/float.html
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.tiny
+dev_manifest: data/manifest.tiny
+test_manifest: data/manifest.tiny
+  
+###########################################
+#              Dataloader                 #
+###########################################
+mean_std_filepath: ""
+vocab_filepath: data/lang_char/vocab.txt 
+unit_type: 'spm'
+spm_model_prefix: 'data/lang_char/bpe_unigram_200'
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 4
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
-decoding:
-  batch_size: 64
-  error_rate_type: wer
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
 
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 5
+accum_grad: 1
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.002
+  weight_decay: 1.0e-06
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 1
+checkpoint:
+  kbest_n: 10
+  latest_n: 1
 

examples/tiny/asr1/conf/conformer.yaml

@@ -1,116 +1,97 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.tiny
-  dev_manifest: data/manifest.tiny
-  test_manifest: data/manifest.tiny
-  min_input_len: 0.5  # second
-  max_input_len: 20.0 # second
-  min_output_len: 0.0 # tokens
-  max_output_len: 400.0 # tokens
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
-  
-collator:
-  mean_std_filepath: ""
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'spm'
-  spm_model_prefix: 'data/lang_char/bpe_unigram_200'
-  augmentation_config: conf/preprocess.yaml
-  batch_size: 4
-  raw_wav: True  # use raw_wav or kaldi feature
-  spectrum_type: fbank #linear, mfcc, fbank
-  feat_dim: 80
-  delta_delta: False
-  dither: 1.0
-  target_sample_rate: 16000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 25.0
-  use_dB_normalization: True
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: "data/mean_std.json"
+cmvn_file_type: "json"
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
+    use_cnn_module: True
+    cnn_module_kernel: 15
+    activation_type: 'swish'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
 
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-# network architecture
-model:
-    cmvn_file: "data/mean_std.json"
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: conformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
-        use_cnn_module: True
-        cnn_module_kernel: 15
-        activation_type: 'swish'
-        pos_enc_layer_type: 'rel_pos'
-        selfattention_layer_type: 'rel_selfattn'
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.tiny
+dev_manifest: data/manifest.tiny
+test_manifest: data/manifest.tiny
+  
 
+###########################################
+#              Dataloader                 #
+###########################################
+mean_std_filepath: ""
+vocab_filepath: data/lang_char/vocab.txt 
+unit_type: 'spm'
+spm_model_prefix: 'data/lang_char/bpe_unigram_200'
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 4
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
-training:
-  n_epoch: 5
-  accum_grad: 4
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.002
-    weight_decay: 1e-06
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 1
-  checkpoint:
-    kbest_n: 10
-    latest_n: 1
 
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 5
+accum_grad: 4
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.002
+  weight_decay: 1.0e-06
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 1
+checkpoint:
+  kbest_n: 10
+  latest_n: 1
 
-decoding:
-  batch_size: 64
-  error_rate_type: wer
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
 
 

examples/tiny/asr1/conf/transformer.yaml

@@ -1,110 +1,90 @@
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.tiny
-  dev_manifest: data/manifest.tiny
-  test_manifest: data/manifest.tiny
-  min_input_len: 0.5  # second
-  max_input_len: 20.0 # second
-  min_output_len: 0.0 # tokens
-  max_output_len: 400.0 # tokens
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
-  
-collator:
-  mean_std_filepath: data/mean_std.json
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'spm'
-  spm_model_prefix: 'data/lang_char/bpe_unigram_200'
-  augmentation_config: conf/preprocess.yaml
-  batch_size: 4
-  raw_wav: True  # use raw_wav or kaldi feature
-  spectrum_type: fbank #linear, mfcc, fbank
-  feat_dim: 80
-  delta_delta: False
-  dither: 1.0
-  target_sample_rate: 16000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 25.0
-  use_dB_normalization: True
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
-
-# network architecture
-model:
-    cmvn_file: 
-    cmvn_file_type: "json"
-    # encoder related
-    encoder: transformer
-    encoder_conf:
-        output_size: 256    # dimension of attention
-        attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: true
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 4
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.tiny
+dev_manifest: data/manifest.tiny
+test_manifest: data/manifest.tiny
+  
+###########################################
+#              Dataloader                 #
+###########################################
+mean_std_filepath: data/mean_std.json
+vocab_filepath: data/lang_char/vocab.txt 
+unit_type: 'spm'
+spm_model_prefix: 'data/lang_char/bpe_unigram_200'
+preprocess_config: conf/preprocess.yaml
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 4
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
-training:
-  n_epoch: 5
-  accum_grad: 1
-  global_grad_clip: 5.0
-  optim: adam
-  optim_conf:
-    lr: 0.002
-    weight_decay: 1e-06
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 25000
-    lr_decay: 1.0
-  log_interval: 1
-  checkpoint:
-    kbest_n: 2
-    latest_n: 1
 
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 5
+accum_grad: 1
+global_grad_clip: 5.0
+optim: adam
+optim_conf:
+  lr: 0.002
+  weight_decay: 1.0e-06
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 25000
+  lr_decay: 1.0
+log_interval: 1
+checkpoint:
+  kbest_n: 2
+  latest_n: 1
 
-decoding:
-  batch_size: 8 #64
-  error_rate_type: wer
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
 
 

examples/tiny/asr1/conf/tuning/chunk_decode.yaml

@@ -0,0 +1,11 @@
+decode_batch_size: 8 #64
+error_rate_type: wer
+decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+beam_size: 10
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.

examples/tiny/asr1/conf/tuning/decode.yaml

@@ -0,0 +1,11 @@
+decode_batch_size: 8 #64
+error_rate_type: wer
+decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+beam_size: 10
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.

examples/tiny/asr1/local/align.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 batch_size=1
 output_dir=${ckpt_prefix}
@@ -20,9 +21,10 @@ mkdir -p ${output_dir}
 python3 -u ${BIN_DIR}/alignment.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
---opts decoding.batch_size ${batch_size}
+--opts decode.decode_batch_size ${batch_size}
 
 if [ $? -ne 0 ]; then
     echo "Failed in ctc alignment!"

examples/tiny/asr1/local/test.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 chunk_mode=false
 if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
@@ -33,10 +34,11 @@ for type in attention ctc_greedy_search; do
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"
@@ -50,10 +52,11 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"

examples/tiny/asr1/run.sh

@@ -6,6 +6,7 @@ gpus=0
 stage=0
 stop_stage=50
 conf_path=conf/transformer.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=1
 
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
@@ -31,12 +32,12 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
     # ctc alignment of test data
-    CUDA_VISIBLE_DEVICES=${gpus} ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/align.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then

examples/wenetspeech/asr1/conf/conformer.yaml

@@ -1,111 +1,92 @@
-# network architecture
-model:
-    # encoder related
-    encoder: conformer
-    encoder_conf:
-        output_size: 512    # dimension of attention
-        attention_heads: 8
-        linear_units: 2048  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        attention_dropout_rate: 0.0
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        normalize_before: True
-        use_cnn_module: True
-        cnn_module_kernel: 15
-        cnn_module_norm: layer_norm
-        activation_type: swish
-        pos_enc_layer_type: rel_pos
-        selfattention_layer_type: rel_selfattn
+############################################
+#           Network Architecture           #
+############################################
+cmvn_file: 
+cmvn_file_type: "json"
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 512    # dimension of attention
+    attention_heads: 8
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: True
+    use_cnn_module: True
+    cnn_module_kernel: 15
+    cnn_module_norm: layer_norm
+    activation_type: swish
+    pos_enc_layer_type: rel_pos
+    selfattention_layer_type: rel_selfattn
 
-    # decoder related
-    decoder: transformer
-    decoder_conf:
-        attention_heads: 8
-        linear_units: 2048
-        num_blocks: 6
-        dropout_rate: 0.1
-        positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 8
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
-    model_conf:
-        ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
 
 # https://yaml.org/type/float.html
-data:
-  train_manifest: data/manifest.train
-  dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
-  min_input_len: 0.1 # second
-  max_input_len: 12.0 # second
-  min_output_len: 1.0
-  max_output_len: 400.0
-  min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
+###########################################
+#                   Data                  #
+###########################################
+train_manifest: data/manifest.train
+dev_manifest: data/manifest.dev
+test_manifest: data/manifest.test
 
-collator:
-  vocab_filepath: data/lang_char/vocab.txt 
-  unit_type: 'char'
-  spm_model_prefix: ''
-  augmentation_config: conf/preprocess.yaml
-  batch_size: 64
-  raw_wav: True  # use raw_wav or kaldi feature
-  spectrum_type: fbank #linear, mfcc, fbank
-  feat_dim: 80
-  delta_delta: False
-  dither: 1.0
-  target_sample_rate: 16000
-  max_freq: None
-  n_fft: None
-  stride_ms: 10.0
-  window_ms: 25.0
-  use_dB_normalization: True 
-  target_dB: -20
-  random_seed: 0
-  keep_transcription_text: False
-  sortagrad: True 
-  shuffle_method: batch_shuffle
-  num_workers: 2
+###########################################
+#              Dataloader                 #
+###########################################
+vocab_filepath: data/lang_char/vocab.txt 
+unit_type: 'char'
+preprocess_config: conf/preprocess.yaml
+spm_model_prefix: ''
+feat_dim: 80
+stride_ms: 10.0
+window_ms: 25.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+batch_size: 64
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+minibatches: 0 # for debug
+batch_count: auto
+batch_bins: 0 
+batch_frames_in: 0
+batch_frames_out: 0
+batch_frames_inout: 0
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
 
-training:
-  n_epoch: 240 
-  accum_grad: 16
-  global_grad_clip: 5.0
-  log_interval: 100
-  checkpoint:
-    kbest_n: 50
-    latest_n: 5
-  optim: adam
-  optim_conf:
-    lr: 0.001
-    weight_decay: 1e-6
-  scheduler: warmuplr     
-  scheduler_conf:
-    warmup_steps: 5000
-    lr_decay: 1.0
-
-
-decoding:
-  batch_size: 128
-  error_rate_type: cer 
-  decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
-  alpha: 2.5
-  beta: 0.3
-  beam_size: 10
-  cutoff_prob: 1.0
-  cutoff_top_n: 0
-  num_proc_bsearch: 8
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
-      # <0: for decoding, use full chunk.
-      # >0: for decoding, use fixed chunk size as set.
-      # 0: used for training, it's prohibited here. 
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+###########################################
+#                 Training                #
+###########################################
+n_epoch: 240 
+accum_grad: 16
+global_grad_clip: 5.0
+log_interval: 100
+checkpoint:
+  kbest_n: 50
+  latest_n: 5
+optim: adam
+optim_conf:
+  lr: 0.001
+  weight_decay: 1.0e-6
+scheduler: warmuplr     
+scheduler_conf:
+  warmup_steps: 5000
+  lr_decay: 1.0

examples/wenetspeech/asr1/conf/tuning/decode.yaml

@@ -0,0 +1,11 @@
+decode_batch_size: 128
+error_rate_type: cer 
+decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+beam_size: 10
+ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+    # <0: for decoding, use full chunk.
+    # >0: for decoding, use fixed chunk size as set.
+    # 0: used for training, it's prohibited here. 
+num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+simulate_streaming: False  # simulate streaming inference. Defaults to False.

examples/wenetspeech/asr1/local/test.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 2 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+if [ $# != 3 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix"
     exit -1
 fi
 
@@ -9,7 +9,8 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
+decode_config_path=$2
+ckpt_prefix=$3
 
 chunk_mode=false
 if [[ ${config_path} =~ ^.*chunk_.*yaml$ ]];then
@@ -36,10 +37,11 @@ for type in attention ctc_greedy_search; do
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"
@@ -55,10 +57,11 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"

examples/wenetspeech/asr1/local/test_wav.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix audio_file"
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path decode_config_path ckpt_path_prefix audio_file"
     exit -1
 fi
 
@@ -9,8 +9,9 @@ ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
 echo "using $ngpu gpus..."
 
 config_path=$1
-ckpt_prefix=$2
-audio_file=$3
+decode_config_path=$2
+ckpt_prefix=$3
+audio_file=$4
 
 mkdir -p data
 wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/
@@ -43,10 +44,11 @@ for type in  attention_rescoring; do
     python3 -u ${BIN_DIR}/test_wav.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size} \
+    --opts decode.decoding_method ${type} \
+    --opts decode.decode_batch_size ${batch_size} \
     --audio_file ${audio_file}
 
     if [ $? -ne 0 ]; then

examples/wenetspeech/asr1/run.sh

@@ -7,7 +7,7 @@ gpus=0,1,2,3,4,5,6,7
 stage=0
 stop_stage=100
 conf_path=conf/conformer.yaml
-
+decode_conf_path=conf/tuning/decode.yaml
 average_checkpoint=true
 avg_num=10
 
@@ -36,12 +36,12 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
     # ctc alignment of test data
-    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
@@ -51,5 +51,5 @@ fi
 
 if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
     # test a single .wav file
-    CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1
 fi

paddlespeech/s2t/decoders/recog.py

@@ -85,7 +85,7 @@ def recog_v2(args):
         mode="asr",
         load_output=False,
         sort_in_input_length=False,
-        preprocess_conf=confs.collator.augmentation_config
+        preprocess_conf=confs.preprocess_config
         if args.preprocess_conf is None else args.preprocess_conf,
         preprocess_args={"train": False}, )