change all recipes

examples/aishell/asr0/conf/deepspeech2.yaml

@@ -1,68 +1,64 @@
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.train
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train
-  min_input_len: 0.0
+dev_manifest: data/manifest.dev
-  max_input_len: 27.0 # second
+test_manifest: data/manifest.test
-  min_output_len: 0.0
+min_input_len: 0.0
-  max_output_len: .inf
+max_input_len: 27.0 # second
-  min_output_input_ratio: 0.00
+min_output_len: 0.0
-  max_output_input_ratio: .inf
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
+###########################################
-  batch_size: 64 # one gpu
+#              Dataloader                 #
-  mean_std_filepath: data/mean_std.json
+###########################################
-  unit_type: char
+batch_size: 64 # one gpu
-  vocab_filepath: data/lang_char/vocab.txt 
+mean_std_filepath: data/mean_std.json
-  augmentation_config: conf/augmentation.json
+unit_type: char
-  random_seed: 0
+vocab_filepath: data/lang_char/vocab.txt 
-  spm_model_prefix: 
+augmentation_config: conf/augmentation.json
-  spectrum_type: linear
+random_seed: 0
-  feat_dim: 
+spm_model_prefix: 
-  delta_delta: False
+spectrum_type: linear
-  stride_ms: 10.0
+feat_dim: 
-  window_ms: 20.0
+delta_delta: False
-  n_fft: None
+stride_ms: 10.0
-  max_freq: None
+window_ms: 20.0
-  target_sample_rate: 16000
+n_fft: None
-  use_dB_normalization: True
+max_freq: None
-  target_dB: -20
+target_sample_rate: 16000
-  dither: 1.0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True
+dither: 1.0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 2
+sortagrad: True
+shuffle_method: batch_shuffle
+num_workers: 2
 
-model:
+############################################
-  num_conv_layers: 2
+#           Network Architecture           #
-  num_rnn_layers: 3
+############################################
-  rnn_layer_size: 1024
+num_conv_layers: 2
-  use_gru: True 
+num_rnn_layers: 3
-  share_rnn_weights: False
+rnn_layer_size: 1024
-  blank_id: 0
+use_gru: True 
-  ctc_grad_norm_type: instance 
+share_rnn_weights: False
+blank_id: 0
+ctc_grad_norm_type: instance 
 
-training:
+###########################################
-  n_epoch: 80
+#                Training                 #
-  accum_grad: 1
+###########################################
-  lr: 2e-3
+n_epoch: 80
-  lr_decay: 0.83
+accum_grad: 1
-  weight_decay: 1e-06
+lr: 2e-3
-  global_grad_clip: 3.0
+lr_decay: 0.83
-  log_interval: 100
+weight_decay: 1e-06
-  checkpoint:
+global_grad_clip: 3.0
-    kbest_n: 50
+log_interval: 100
-    latest_n: 5
+checkpoint:
-  
+  kbest_n: 50
-decoding:
+  latest_n: 5
-  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/conf/deepspeech2_online.yaml

@@ -1,70 +1,68 @@
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.train
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train
-  min_input_len: 0.0
+dev_manifest: data/manifest.dev
-  max_input_len: 27.0 # second
+test_manifest: data/manifest.test
-  min_output_len: 0.0
+min_input_len: 0.0
-  max_output_len: .inf
+max_input_len: 27.0 # second
-  min_output_input_ratio: 0.00
+min_output_len: 0.0
-  max_output_input_ratio: .inf
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
+###########################################
-  batch_size: 64 # one gpu
+#              Dataloader                 #
-  mean_std_filepath: data/mean_std.json
+###########################################
-  unit_type: char
+batch_size: 64 # one gpu
-  vocab_filepath: data/lang_char/vocab.txt 
+mean_std_filepath: data/mean_std.json
-  augmentation_config: conf/augmentation.json
+unit_type: char
-  random_seed: 0
+vocab_filepath: data/lang_char/vocab.txt 
-  spm_model_prefix: 
+augmentation_config: conf/augmentation.json
-  spectrum_type: linear #linear, mfcc, fbank
+random_seed: 0
-  feat_dim: 
+spm_model_prefix: 
-  delta_delta: False
+spectrum_type: linear #linear, mfcc, fbank
-  stride_ms: 10.0
+feat_dim: 
-  window_ms: 20.0
+delta_delta: False
-  n_fft: None
+stride_ms: 10.0
-  max_freq: None
+window_ms: 20.0
-  target_sample_rate: 16000
+n_fft: None
-  use_dB_normalization: True
+max_freq: None
-  target_dB: -20
+target_sample_rate: 16000
-  dither: 1.0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True
+dither: 1.0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 0
+sortagrad: True
+shuffle_method: batch_shuffle
+num_workers: 0
 
-model:
+############################################
-  num_conv_layers: 2
+#           Network Architecture           #
-  num_rnn_layers: 5
+############################################
-  rnn_layer_size: 1024
+num_conv_layers: 2
-  rnn_direction: forward # [forward, bidirect]
+num_rnn_layers: 5
-  num_fc_layers: 0
+rnn_layer_size: 1024
-  fc_layers_size_list: -1,
+rnn_direction: forward # [forward, bidirect]
-  use_gru: False 
+num_fc_layers: 0
-  blank_id: 0
+fc_layers_size_list: -1,
+use_gru: False 
+blank_id: 0
   
   
-training:
+###########################################
-  n_epoch: 65
+#                Training                 #
-  accum_grad: 1
+###########################################
-  lr: 5e-4
+n_epoch: 65
-  lr_decay: 0.93
+accum_grad: 1
-  weight_decay: 1e-06
+lr: 5e-4
-  global_grad_clip: 3.0
+lr_decay: 0.93
-  log_interval: 100
+weight_decay: 1e-06
-  checkpoint:
+global_grad_clip: 3.0
-    kbest_n: 50
+log_interval: 100
-    latest_n: 5
+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
+if [ $# != 4 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+    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
+decode_config_path=$2
-model_type=$3
+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
+if [ $# != 4 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+    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
+decode_config_path=$2
-model_type=$3
+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_hub_ori

@@ -0,0 +1,47 @@
+#!/bin/bash
+
+if [ $# != 4 ];then
+    echo "usage: ${0} config_path ckpt_path_prefix model_type audio_file"
+    exit -1
+fi
+
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+echo "using $ngpu gpus..."
+
+config_path=$1
+ckpt_prefix=$2
+model_type=$3
+audio_file=$4
+
+mkdir -p data
+wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/zh/demo_01_03.wav -P data/
+if [ $? -ne 0 ]; then
+   exit 1
+fi
+
+if [ ! -f ${audio_file} ]; then
+    echo "Plase input the right audio_file path"
+    exit 1
+fi
+
+# download language model
+bash local/download_lm_ch.sh
+if [ $? -ne 0 ]; then
+   exit 1
+fi
+
+python3 -u ${BIN_DIR}/test_hub.py \
+--nproc ${ngpu} \
+--config ${config_path} \
+--result_file ${ckpt_prefix}.rsl \
+--checkpoint_path ${ckpt_prefix} \
+--model_type ${model_type} \
+--audio_file ${audio_file}
+
+if [ $? -ne 0 ]; then
+    echo "Failed in evaluation!"
+    exit 1
+fi
+
+
+exit 0

examples/aishell/asr0/local/test_wav.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-if [ $# != 4 ];then
+if [ $# != 5 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type audio_file"
+    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
+decode_config_path=$2
-model_type=$3
+ckpt_prefix=$3
-audio_file=$4
+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

@@ -54,8 +54,9 @@ test_manifest: data/manifest.test
 ###########################################
 
 vocab_filepath: data/lang_char/vocab.txt 
+spm_model_prefix: ''
 unit_type: 'char'
-augmentation_config: conf/preprocess.yaml
+preprocess_config: conf/preprocess.yaml
 feat_dim: 80
 stride_ms: 10.0
 window_ms: 25.0
@@ -74,7 +75,7 @@ subsampling_factor: 1
 num_encs: 1
 
 ###########################################
-#                 training                #
+#                 Training                #
 ###########################################
 n_epoch: 240 
 accum_grad: 2
@@ -82,7 +83,7 @@ global_grad_clip: 5.0
 optim: adam
 optim_conf:
   lr: 0.002
-  weight_decay: 1e-6
+  weight_decay: 1.0e-6
 scheduler: warmuplr
 scheduler_conf:
   warmup_steps: 25000

examples/aishell/asr1/conf/conformer.yaml

@@ -49,8 +49,9 @@ test_manifest: data/manifest.test
 #              Dataloader                 #
 ###########################################
 vocab_filepath: data/lang_char/vocab.txt 
+spm_model_prefix: ''
 unit_type: 'char'
-augmentation_config: conf/preprocess.yaml
+preprocess_config: conf/preprocess.yaml
 feat_dim: 80
 stride_ms: 10.0
 window_ms: 25.0
@@ -69,7 +70,7 @@ subsampling_factor: 1
 num_encs: 1
 
 ###########################################
-#                 training                #
+#                Training                 #
 ###########################################
 n_epoch: 240 
 accum_grad: 2

examples/aishell/asr1/conf/transformer.yaml

@@ -46,6 +46,7 @@ test_manifest: data/manifest.test
 ###########################################
 unit_type: 'char'
 vocab_filepath: data/lang_char/vocab.txt 
+spm_model_prefix: ''
 feat_dim: 80
 stride_ms: 10.0
 window_ms: 25.0
@@ -59,13 +60,13 @@ batch_bins: 0
 batch_frames_in: 0
 batch_frames_out: 0
 batch_frames_inout: 0
-augmentation_config: conf/preprocess.yaml 
+preprocess_config: conf/preprocess.yaml 
 num_workers: 0
 subsampling_factor: 1
 num_encs: 1
 
 ###########################################
-#                 training                #
+#                 Training                #
 ###########################################
 n_epoch: 240 
 accum_grad: 2
@@ -73,7 +74,7 @@ global_grad_clip: 5.0
 optim: adam
 optim_conf:
   lr: 0.002
-  weight_decay: 1e-6
+  weight_decay: 1.0e-6
 scheduler: warmuplr     
 scheduler_conf:
   warmup_steps: 25000

examples/aishell/asr1/local/align.sh

@@ -21,7 +21,7 @@ mkdir -p ${output_dir}
 python3 -u ${BIN_DIR}/alignment.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
---decode_config ${decode_config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
 --opts decode.decode_batch_size ${batch_size}

examples/aishell/asr1/local/test.sh

@@ -30,14 +30,14 @@ for type in attention ctc_greedy_search; do
         # stream decoding only support batchsize=1
         batch_size=1
     else
-        batch_size=1
+        batch_size=64
     fi
     output_dir=${ckpt_prefix}
     mkdir -p ${output_dir}
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
-    --decode_config ${decode_config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
     --opts decode.decoding_method ${type} \
@@ -57,7 +57,7 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     python3 -u ${BIN_DIR}/test.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
-    --decode_config ${decode_config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
     --opts decode.decoding_method ${type} \

examples/aishell/asr1/local/test_wav.sh

@@ -43,7 +43,7 @@ for type in  attention_rescoring; do
     python3 -u ${BIN_DIR}/test_wav.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
-    --decode_config ${decode_config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
     --opts decode.decoding_method ${type} \

examples/callcenter/asr1/conf/chunk_conformer.yaml

@@ -1,120 +1,98 @@
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.train
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train
-  min_input_len: 0.5
+dev_manifest: data/manifest.dev
-  max_input_len: 20.0 # second
+test_manifest: data/manifest.test
-  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
 
+  
+###########################################
+#              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:
+# decoder related
-  batch_size: 128
+decoder: transformer
-  error_rate_type: cer 
+decoder_conf:
-  decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+    attention_heads: 4
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
+    linear_units: 2048
-  alpha: 2.5
+    num_blocks: 6
-  beta: 0.3
+    dropout_rate: 0.1
-  beam_size: 10
+    positional_dropout_rate: 0.1
-  cutoff_prob: 1.0
+    self_attention_dropout_rate: 0.0
-  cutoff_top_n: 0
+    src_attention_dropout_rate: 0.0
-  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.
 
+# 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
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train
-  min_input_len: 0.5
+dev_manifest: data/manifest.dev
-  max_input_len: 20.0 # second
+test_manifest: data/manifest.test
-  min_output_len: 0.0
-  max_output_len: 400.0
-  min_output_input_ratio: 0.0
-  max_output_input_ratio: .inf 
 
 
-collator:
+###########################################
-  vocab_filepath: data/lang_char/vocab.txt 
+#              Dataloader                 #
-  unit_type: 'char'
+###########################################
-  spm_model_prefix: ''
+vocab_filepath: data/lang_char/vocab.txt 
-  augmentation_config: conf/preprocess.yaml
+unit_type: 'char'
-  batch_size: 32
+spm_model_prefix: ''
-  raw_wav: True  # use raw_wav or kaldi feature
+preprocess_config: conf/preprocess.yaml
-  spectrum_type: fbank #linear, mfcc, fbank
+feat_dim: 80
-  feat_dim: 80
+stride_ms: 10.0
-  delta_delta: False
+window_ms: 25.0
-  dither: 1.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  target_sample_rate: 8000
+batch_size: 64
-  max_freq: None
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  n_fft: None
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  stride_ms: 10.0
+minibatches: 0 # for debug
-  window_ms: 25.0
+batch_count: auto
-  use_dB_normalization: True 
+batch_bins: 0 
-  target_dB: -20
+batch_frames_in: 0
-  random_seed: 0
+batch_frames_out: 0
-  keep_transcription_text: False
+batch_frames_inout: 0
-  sortagrad: True 
+num_workers: 0
-  shuffle_method: batch_shuffle
+subsampling_factor: 1
-  num_workers: 2
+num_encs: 1
 
-# network architecture
+############################################
-model:
+#           Network Architecture           #
-    cmvn_file: 
+############################################
-    cmvn_file_type: "json"
+cmvn_file: 
-    # encoder related
+cmvn_file_type: "json"
-    encoder: conformer
+# encoder related
-    encoder_conf:
+encoder: conformer
-        output_size: 256    # dimension of attention
+encoder_conf:
-        attention_heads: 4
+    output_size: 256    # dimension of attention
-        linear_units: 2048  # the number of units of position-wise feed forward
+    attention_heads: 4
-        num_blocks: 12      # the number of encoder blocks
+    linear_units: 2048  # the number of units of position-wise feed forward
-        dropout_rate: 0.1
+    num_blocks: 12      # the number of encoder blocks
-        positional_dropout_rate: 0.1
+    dropout_rate: 0.1
-        attention_dropout_rate: 0.0
+    positional_dropout_rate: 0.1
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    attention_dropout_rate: 0.0
-        normalize_before: True
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        use_cnn_module: True
+    normalize_before: True
-        cnn_module_kernel: 15
+    use_cnn_module: True
-        activation_type: 'swish'
+    cnn_module_kernel: 15
-        pos_enc_layer_type: 'rel_pos'
+    activation_type: 'swish'
-        selfattention_layer_type: 'rel_selfattn'
+    pos_enc_layer_type: 'rel_pos'
+    selfattention_layer_type: 'rel_selfattn'
 
-    # decoder related
+# decoder related
-    decoder: transformer
+decoder: transformer
-    decoder_conf:
+decoder_conf:
-        attention_heads: 4
+    attention_heads: 4
-        linear_units: 2048
+    linear_units: 2048
-        num_blocks: 6
+    num_blocks: 6
-        dropout_rate: 0.1
+    dropout_rate: 0.1
-        positional_dropout_rate: 0.1
+    positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
+    self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
+# hybrid CTC/attention
-    model_conf:
+model_conf:
-        ctc_weight: 0.3
+    ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
+    lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+    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.
 
 
+###########################################
+#                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
+if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+    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
+if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+    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 decode.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --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 decode.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decode_batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"

examples/callcenter/asr1/run.sh

@@ -6,6 +6,7 @@ gpus=0,1,2,3
 stage=0
 stop_stage=100
 conf_path=conf/conformer.yaml
+decode_conf_path=conf/tuning/decode.yaml
 avg_num=20
 
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
@@ -31,12 +32,12 @@ 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

examples/librispeech/asr0/conf/deepspeech2.yaml

@@ -1,68 +1,65 @@
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.train
+#                   Data                  #
-  dev_manifest: data/manifest.dev-clean
+###########################################
-  test_manifest: data/manifest.test-clean
+train_manifest: data/manifest.train
-  min_input_len: 0.0
+dev_manifest: data/manifest.dev-clean
-  max_input_len: 30.0 # second
+test_manifest: data/manifest.test-clean
-  min_output_len: 0.0
+min_input_len: 0.0
-  max_output_len: .inf
+max_input_len: 30.0 # second
-  min_output_input_ratio: 0.00
+min_output_len: 0.0
-  max_output_input_ratio: .inf
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
+###########################################
-  batch_size: 20
+#              Dataloader                 #
-  mean_std_filepath: data/mean_std.json
+###########################################
-  unit_type: char
+batch_size: 20
-  vocab_filepath: data/lang_char/vocab.txt
+mean_std_filepath: data/mean_std.json
-  augmentation_config: conf/augmentation.json
+unit_type: char
-  random_seed: 0
+vocab_filepath: data/lang_char/vocab.txt
-  spm_model_prefix: 
+augmentation_config: conf/augmentation.json
-  spectrum_type: linear
+random_seed: 0
-  target_sample_rate: 16000
+spm_model_prefix: 
-  max_freq: None
+spectrum_type: linear
-  n_fft: None
+feat_dim: 
-  stride_ms: 10.0
+target_sample_rate: 16000
-  window_ms: 20.0
+max_freq: None
-  delta_delta: False
+n_fft: None
-  dither: 1.0
+stride_ms: 10.0
-  use_dB_normalization: True 
+window_ms: 20.0
-  target_dB: -20
+delta_delta: False
-  random_seed: 0
+dither: 1.0
-  keep_transcription_text: False
+use_dB_normalization: True 
-  sortagrad: True 
+target_dB: -20
-  shuffle_method: batch_shuffle
+random_seed: 0
-  num_workers: 2
+keep_transcription_text: False
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 2
 
-model:
+############################################
-  num_conv_layers: 2
+#           Network Architecture           #
-  num_rnn_layers: 3
+############################################
-  rnn_layer_size: 2048
+num_conv_layers: 2
-  use_gru: False 
+num_rnn_layers: 3
-  share_rnn_weights: True
+rnn_layer_size: 2048
-  blank_id: 0
+use_gru: False 
+share_rnn_weights: True
+blank_id: 0
   
 
-training:
+###########################################
-  n_epoch: 50
+#                Training                 #
-  accum_grad: 1
+###########################################
-  lr: 1e-3
+n_epoch: 50
-  lr_decay: 0.83
+accum_grad: 1
-  weight_decay: 1e-06
+lr: 1e-3
-  global_grad_clip: 5.0
+lr_decay: 0.83
-  log_interval: 100
+weight_decay: 1e-06
-  checkpoint:
+global_grad_clip: 5.0
-    kbest_n: 50
+log_interval: 100
-    latest_n: 5
+checkpoint:
-
+  kbest_n: 50
-decoding:
+  latest_n: 5
-  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/deepspeech2_online.yaml

@@ -1,70 +1,67 @@
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.train
+#                   Data                  #
-  dev_manifest: data/manifest.dev-clean
+###########################################
-  test_manifest: data/manifest.test-clean
+train_manifest: data/manifest.train
-  min_input_len: 0.0
+dev_manifest: data/manifest.dev-clean
-  max_input_len: 30.0 # second
+test_manifest: data/manifest.test-clean
-  min_output_len: 0.0
+min_input_len: 0.0
-  max_output_len: .inf
+max_input_len: 30.0 # second
-  min_output_input_ratio: 0.00
+min_output_len: 0.0
-  max_output_input_ratio: .inf
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
+###########################################
-  batch_size: 15
+#              Dataloader                 #
-  mean_std_filepath: data/mean_std.json
+###########################################
-  unit_type: char
+batch_size: 15
-  vocab_filepath: data/lang_char/vocab.txt
+mean_std_filepath: data/mean_std.json
-  augmentation_config: conf/augmentation.json
+unit_type: char
-  random_seed: 0
+vocab_filepath: data/lang_char/vocab.txt
-  spm_model_prefix: 
+augmentation_config: conf/augmentation.json
-  spectrum_type: linear
+random_seed: 0
-  target_sample_rate: 16000
+spm_model_prefix: 
-  max_freq: None
+spectrum_type: linear
-  n_fft: None
+feat_dim: 
-  stride_ms: 10.0
+target_sample_rate: 16000
-  window_ms: 20.0
+max_freq: None
-  delta_delta: False
+n_fft: None
-  dither: 1.0
+stride_ms: 10.0
-  use_dB_normalization: True 
+window_ms: 20.0
-  target_dB: -20
+delta_delta: False
-  random_seed: 0
+dither: 1.0
-  keep_transcription_text: False
+use_dB_normalization: True 
-  sortagrad: True 
+target_dB: -20
-  shuffle_method: batch_shuffle
+random_seed: 0
-  num_workers: 0
+keep_transcription_text: False
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 0
 
-model:
+############################################
-  num_conv_layers: 2
+#           Network Architecture           #
-  num_rnn_layers: 3
+############################################
-  rnn_layer_size: 2048
+num_conv_layers: 2
-  rnn_direction: forward
+num_rnn_layers: 3
-  num_fc_layers: 2
+rnn_layer_size: 2048
-  fc_layers_size_list: 512, 256
+rnn_direction: forward
-  use_gru: False 
+num_fc_layers: 2
-  blank_id: 0
+fc_layers_size_list: 512, 256
+use_gru: False 
+blank_id: 0
   
 
-training:
+###########################################
-  n_epoch: 50
+#                Training                 #
-  accum_grad: 4
+###########################################
-  lr: 1e-3
+n_epoch: 50
-  lr_decay: 0.83
+accum_grad: 4
-  weight_decay: 1e-06
+lr: 1e-3
-  global_grad_clip: 5.0
+lr_decay: 0.83
-  log_interval: 100
+weight_decay: 1e-06
-  checkpoint:
+global_grad_clip: 5.0
-    kbest_n: 50
+log_interval: 100
-    latest_n: 5
+checkpoint:
-
+  kbest_n: 50
-decoding:
+  latest_n: 5
-  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/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
+if [ $# != 4 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+    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
+decode_config_path=$2
-model_type=$3
+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
+if [ $# != 5 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type audio_file"
+    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
+decode_config_path=$2
-model_type=$3
+ckpt_prefix=$3
-audio_file=$4
+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

@@ -57,7 +57,7 @@ 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
+preprocess_config: conf/preprocess.yaml
 feat_dim: 80
 stride_ms: 10.0
 window_ms: 25.0
@@ -70,8 +70,7 @@ batch_count: auto
 batch_bins: 0 
 batch_frames_in: 0
 batch_frames_out: 0
-batch_frames_inout: 0
+batch_frames_inout: 0 
-augmentation_config: conf/preprocess.yaml 
 num_workers: 0
 subsampling_factor: 1
 num_encs: 1
@@ -85,10 +84,11 @@ global_grad_clip: 5.0
 optim: adam
 optim_conf:
   lr: 0.001
-  weight_decay: 1e-06 
+  weight_decay: 1.0e-06 
 scheduler: warmuplr     
 scheduler_conf:
   warmup_steps: 25000
+  lr_decay: 1.0
 log_interval: 100
 checkpoint:
   kbest_n: 50

examples/librispeech/asr1/conf/chunk_transformer.yaml

@@ -50,7 +50,7 @@ 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
+preprocess_config: conf/preprocess.yaml
 feat_dim: 80
 stride_ms: 10.0
 window_ms: 25.0
@@ -64,7 +64,6 @@ 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
@@ -79,7 +78,7 @@ global_grad_clip: 5.0
 optim: adam
 optim_conf:
   lr: 0.001
-  weight_decay: 1e-06
+  weight_decay: 1.0e-06
 scheduler: warmuplr     
 scheduler_conf:
   warmup_steps: 25000

examples/librispeech/asr1/conf/conformer.yaml

@@ -55,7 +55,7 @@ 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
+preprocess_config: conf/preprocess.yaml
 feat_dim: 80
 stride_ms: 10.0
 window_ms: 25.0
@@ -69,7 +69,6 @@ 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
@@ -84,7 +83,7 @@ global_grad_clip: 3.0
 optim: adam
 optim_conf:
   lr: 0.004
-  weight_decay: 1e-06
+  weight_decay: 1.0e-06
 scheduler: warmuplr     
 scheduler_conf:
   warmup_steps: 25000

examples/librispeech/asr1/conf/transformer.yaml

@@ -49,7 +49,7 @@ 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
+preprocess_config: conf/preprocess.yaml
 feat_dim: 80
 stride_ms: 10.0
 window_ms: 25.0
@@ -63,7 +63,6 @@ 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
@@ -78,7 +77,7 @@ global_grad_clip: 5.0
 optim: adam
 optim_conf:
   lr: 0.004
-  weight_decay: 1e-06
+  weight_decay: 1.0e-06
 scheduler: warmuplr     
 scheduler_conf:
   warmup_steps: 25000

examples/librispeech/asr1/local/align.sh

@@ -21,7 +21,7 @@ mkdir -p ${output_dir}
 python3 -u ${BIN_DIR}/alignment.py \
 --ngpu ${ngpu} \
 --config ${config_path} \
---decode_config ${decode_config_path} \
+--decode_cfg ${decode_config_path} \
 --result_file ${output_dir}/${type}.align \
 --checkpoint_path ${ckpt_prefix} \
 --opts decode.decode_batch_size ${batch_size}

examples/librispeech/asr1/local/test.sh

@@ -53,7 +53,7 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
         python3 -u ${BIN_DIR}/test.py \
             --ngpu ${ngpu} \
             --config ${config_path} \
-            --decode_config ${decode_config_path} \
+            --decode_cfg ${decode_config_path} \
             --result_file ${ckpt_prefix}.${type}.rsl \
             --checkpoint_path ${ckpt_prefix} \
             --opts decode.decoding_method ${type} \
@@ -78,7 +78,7 @@ for type in ctc_greedy_search; do
     python3 -u ${BIN_DIR}/test.py \
         --ngpu ${ngpu} \
         --config ${config_path} \
-        --decode_config ${decode_config_path} \
+        --decode_cfg ${decode_config_path} \
         --result_file ${ckpt_prefix}.${type}.rsl \
         --checkpoint_path ${ckpt_prefix} \
         --opts decode.decoding_method ${type} \
@@ -99,7 +99,7 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     python3 -u ${BIN_DIR}/test.py \
         --ngpu ${ngpu} \
         --config ${config_path} \
-        --decode_config ${decode_config_path} \
+        --decode_cfg ${decode_config_path} \
         --result_file ${ckpt_prefix}.${type}.rsl \
         --checkpoint_path ${ckpt_prefix} \
         --opts decode.decoding_method ${type} \

examples/librispeech/asr1/local/test_wav.sh

@@ -50,7 +50,7 @@ for type in attention_rescoring; do
     python3 -u ${BIN_DIR}/test_wav.py \
     --ngpu ${ngpu} \
     --config ${config_path} \
-    --decode_config ${decode_config_path} \
+    --decode_cfg ${decode_config_path} \
     --result_file ${output_dir}/${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
     --opts decode.decoding_method ${type} \

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:
+#           Network Architecture           #
-    cmvn_file:  
+############################################
-    cmvn_file_type: "json"
+cmvn_file:  
-    # encoder related
+cmvn_file_type: "json"
-    encoder: transformer
+# encoder related
-    encoder_conf:
+encoder: transformer
-        output_size: 256    # dimension of attention
+encoder_conf:
-        attention_heads: 4
+    output_size: 256    # dimension of attention
-        linear_units: 2048  # the number of units of position-wise feed forward
+    attention_heads: 4
-        num_blocks: 12      # the number of encoder blocks
+    linear_units: 2048  # the number of units of position-wise feed forward
-        dropout_rate: 0.1
+    num_blocks: 12      # the number of encoder blocks
-        positional_dropout_rate: 0.1
+    dropout_rate: 0.1
-        attention_dropout_rate: 0.0
+    positional_dropout_rate: 0.1
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    attention_dropout_rate: 0.0
-        normalize_before: true
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
+# decoder related
-    decoder: transformer
+decoder: transformer
-    decoder_conf:
+decoder_conf:
-        attention_heads: 4
+    attention_heads: 4
-        linear_units: 2048
+    linear_units: 2048
-        num_blocks: 6
+    num_blocks: 6
-        dropout_rate: 0.1
+    dropout_rate: 0.1
-        positional_dropout_rate: 0.1
+    positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
+    self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
+# hybrid CTC/attention
-    model_conf:
+model_conf:
-        ctc_weight: 0.3
+    ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
+    lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+    length_normalized_loss: false
 
-data:
+###########################################
-  train_manifest: data/manifest.train
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test-clean
+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
+#              Dataloader                 #
-  unit_type: spm
+###########################################
-  spm_model_prefix: data/lang_char/train_960_unigram5000
+vocab_filepath: data/lang_char/train_960_unigram5000_units.txt
-  feat_dim: 83
+unit_type: spm
-  stride_ms: 10.0
+spm_model_prefix: data/lang_char/train_960_unigram5000
-  window_ms: 25.0
+feat_dim: 83
-  sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
+stride_ms: 10.0
-  batch_size: 30 
+window_ms: 25.0
-  maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
+batch_size: 30 
-  minibatches: 0 # for debug
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  batch_count: auto
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  batch_bins: 0 
+minibatches: 0 # for debug
-  batch_frames_in: 0
+batch_count: auto
-  batch_frames_out: 0
+batch_bins: 0 
-  batch_frames_inout: 0
+batch_frames_in: 0
-  augmentation_config: conf/preprocess.yaml 
+batch_frames_out: 0
-  num_workers: 0
+batch_frames_inout: 0
-  subsampling_factor: 1
+preprocess_config: conf/preprocess.yaml 
-  num_encs: 1
+num_workers: 0
+subsampling_factor: 1
+num_encs: 1
 
 
-training:
+###########################################
-  n_epoch: 120
+#                Training                 #
-  accum_grad: 2
+###########################################
-  log_interval: 100
+n_epoch: 120
-  checkpoint:
+accum_grad: 2
-    kbest_n: 50
+log_interval: 1
-    latest_n: 5
+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
+if [ $# != 4 ];then
-    echo "usage: ${0} config_path dict_path ckpt_path_prefix"
+    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
+decode_config_path=$2
-ckpt_prefix=$3
+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

@@ -19,8 +19,9 @@ 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=
+ckpt_prefix=exp/transformer/checkpoints/init
 
 source ${MAIN_ROOT}/utils/parse_options.sh || exit 1;
 
@@ -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 decode.decoding_method ${dmethd} \
-            --opts decoding.batch_size ${batch_size} \
+            --opts decode.decode_batch_size ${batch_size} \
-            --opts data.test_manifest ${feat_recog_dir}/split${nj}/JOB/manifest.${rtask}
+            --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

@@ -9,12 +9,14 @@ 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;
 
 avg_ckpt=avg_${avg_num}
+avg_ckpt=init
 ckpt=$(basename ${conf_path} | awk -F'.' '{print $1}')
 echo "checkpoint name ${ckpt}"
 
@@ -35,7 +37,7 @@ 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 +47,7 @@ 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
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train
-  min_input_len: 0.0
+dev_manifest: data/manifest.dev
-  max_input_len: 27.0 # second
+test_manifest: data/manifest.test
-  min_output_len: 0.0
+min_input_len: 0.0
-  max_output_len: .inf
+max_input_len: 27.0 # second
-  min_output_input_ratio: 0.00
+min_output_len: 0.0
-  max_output_input_ratio: .inf
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
+###########################################
-  batch_size: 64 # one gpu
+#              Dataloader                 #
-  mean_std_filepath: data/mean_std.npz
+###########################################
-  unit_type: char
+batch_size: 64 # one gpu
-  vocab_filepath: data/vocab.txt 
+mean_std_filepath: data/mean_std.npz
-  augmentation_config: conf/augmentation.json
+unit_type: char
-  random_seed: 0
+vocab_filepath: data/vocab.txt 
-  spm_model_prefix: 
+augmentation_config: conf/augmentation.json
-  spectrum_type: linear
+random_seed: 0
-  feat_dim: 
+spm_model_prefix: 
-  delta_delta: False
+spectrum_type: linear
-  stride_ms: 10.0
+feat_dim: 
-  window_ms: 20.0
+delta_delta: False
-  n_fft: None
+stride_ms: 10.0
-  max_freq: None
+window_ms: 20.0
-  target_sample_rate: 16000
+n_fft: None
-  use_dB_normalization: True
+max_freq: None
-  target_dB: -20
+target_sample_rate: 16000
-  dither: 1.0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True
+dither: 1.0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 2
+sortagrad: True
+shuffle_method: batch_shuffle
+num_workers: 2
 
-model:
+############################################
-  num_conv_layers: 2
+#           Network Architecture           #
-  num_rnn_layers: 3
+############################################
-  rnn_layer_size: 1024
+num_conv_layers: 2
-  use_gru: True 
+num_rnn_layers: 3
-  share_rnn_weights: False
+rnn_layer_size: 1024
-  blank_id: 4333
+use_gru: True 
+share_rnn_weights: False
+blank_id: 4333
 
-training:
+###########################################
-  n_epoch: 80
+#                Training                 #
-  accum_grad: 1
+###########################################
-  lr: 2e-3
+n_epoch: 80
-  lr_decay: 0.83
+accum_grad: 1
-  weight_decay: 1e-06
+lr: 2e-3
-  global_grad_clip: 3.0
+lr_decay: 0.83
-  log_interval: 100
+weight_decay: 1e-06
-  checkpoint:
+global_grad_clip: 3.0
-    kbest_n: 50
+log_interval: 100
-    latest_n: 5
+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
+if [ $# != 4 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+    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
+decode_config_path=$2
-model_type=$3
+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
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test-clean
+train_manifest: data/manifest.train
-  min_input_len: 0.0
+dev_manifest: data/manifest.dev
-  max_input_len: .inf # second
+test_manifest: data/manifest.test-clean
-  min_output_len: 0.0
+min_input_len: 0.0
-  max_output_len: .inf
+max_input_len: .inf # second
-  min_output_input_ratio: 0.00
+min_output_len: 0.0
-  max_output_input_ratio: .inf
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
+###########################################
-  batch_size: 64 # one gpu
+#              Dataloader                 #
-  mean_std_filepath: data/mean_std.npz
+###########################################
-  unit_type: char
+batch_size: 64 # one gpu
-  vocab_filepath: data/vocab.txt 
+mean_std_filepath: data/mean_std.npz
-  augmentation_config: conf/augmentation.json
+unit_type: char
-  random_seed: 0
+vocab_filepath: data/vocab.txt 
-  spm_model_prefix: 
+augmentation_config: conf/augmentation.json
-  spectrum_type: linear
+random_seed: 0
-  feat_dim: 
+spm_model_prefix: 
-  delta_delta: False
+spectrum_type: linear
-  stride_ms: 10.0
+feat_dim: 
-  window_ms: 20.0
+delta_delta: False
-  n_fft: None
+stride_ms: 10.0
-  max_freq: None
+window_ms: 20.0
-  target_sample_rate: 16000
+n_fft: None
-  use_dB_normalization: True
+max_freq: None
-  target_dB: -20
+target_sample_rate: 16000
-  dither: 1.0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True
+dither: 1.0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 2
+sortagrad: True
+shuffle_method: batch_shuffle
+num_workers: 2
 
-model:
+############################################
-  num_conv_layers: 2
+#           Network Architecture           #
-  num_rnn_layers: 3
+############################################
-  rnn_layer_size: 1024
+num_conv_layers: 2
-  use_gru: True
+num_rnn_layers: 3
-  share_rnn_weights: False
+rnn_layer_size: 1024
-  blank_id: 28
+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
+if [ $# != 4 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+    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
+decode_config_path=$2
-model_type=$3
+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
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test-clean
+train_manifest: data/manifest.train
-  min_input_len: 0.0
+dev_manifest: data/manifest.dev
-  max_input_len: 1000.0 # second
+test_manifest: data/manifest.test-clean
-  min_output_len: 0.0
+min_input_len: 0.0
-  max_output_len: .inf
+max_input_len: 1000.0 # second
-  min_output_input_ratio: 0.00
+min_output_len: 0.0
-  max_output_input_ratio: .inf
+max_output_len: .inf
+min_output_input_ratio: 0.00
+max_output_input_ratio: .inf
 
-collator:
+###########################################
-  batch_size: 64 # one gpu
+#              Dataloader                 #
-  mean_std_filepath: data/mean_std.npz
+###########################################
-  unit_type: char
+batch_size: 64 # one gpu
-  vocab_filepath: data/vocab.txt 
+mean_std_filepath: data/mean_std.npz
-  augmentation_config: conf/augmentation.json
+unit_type: char
-  random_seed: 0
+vocab_filepath: data/vocab.txt 
-  spm_model_prefix: 
+augmentation_config: conf/augmentation.json
-  spectrum_type: linear
+random_seed: 0
-  feat_dim: 
+spm_model_prefix: 
-  delta_delta: False
+spectrum_type: linear
-  stride_ms: 10.0
+feat_dim: 
-  window_ms: 20.0
+delta_delta: False
-  n_fft: None
+stride_ms: 10.0
-  max_freq: None
+window_ms: 20.0
-  target_sample_rate: 16000
+n_fft: None
-  use_dB_normalization: True
+max_freq: None
-  target_dB: -20
+target_sample_rate: 16000
-  dither: 1.0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True
+dither: 1.0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 2
+sortagrad: True
+shuffle_method: batch_shuffle
+num_workers: 2
 
-model:
+############################################
-  num_conv_layers: 2
+#           Network Architecture           #
-  num_rnn_layers: 3
+############################################
-  rnn_layer_size: 2048
+num_conv_layers: 2
-  use_gru: False
+num_rnn_layers: 3
-  share_rnn_weights: True
+rnn_layer_size: 2048
-  blank_id: 28
+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
+if [ $# != 4 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+    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
+decode_config_path=$2
-model_type=$3
+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

@@ -13,6 +13,7 @@
 # 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
@@ -44,6 +45,10 @@ if __name__ == "__main__":
     config = get_cfg_defaults(args.model_type)
     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

@@ -233,11 +233,11 @@ 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_conv_layers=config.num_conv_layers,
-                    num_rnn_layers=config.model.num_rnn_layers,
+                    num_rnn_layers=config.num_rnn_layers,
-                    rnn_size=config.model.rnn_layer_size,
+                    rnn_size=config.rnn_layer_size,
-                    use_gru=config.model.use_gru,
+                    use_gru=config.use_gru,
-                    share_rnn_weights=config.model.share_rnn_weights)
+                    share_rnn_weights=config.share_rnn_weights)
         infos = Checkpoint().load_parameters(
             model, checkpoint_path=checkpoint_path)
         logger.info(f"checkpoint info: {infos}")
@@ -250,7 +250,7 @@ class DeepSpeech2Model(nn.Layer):
         Parameters
 
         config: yacs.config.CfgNode
-            config.model
+            config
         Returns
         -------
         DeepSpeech2Model

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

@@ -64,7 +64,7 @@ class DeepSpeech2Trainer(Trainer):
         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 +98,7 @@ 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 +126,7 @@ 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 +146,15 @@ 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.feat_size = self.train_loader.collate_fn.feature_size
-        #config.model.dict_size = self.train_loader.collate_fn.vocab_size
+        #config.dict_size = self.train_loader.collate_fn.vocab_size
-        config.model.dict_size = len(self.train_loader.collate_fn.vocab_list)
+        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 +163,13 @@ class DeepSpeech2Trainer(Trainer):
         logger.info(f"{model}")
         layer_tools.print_params(model, logger.info)
 
-        grad_clip = ClipGradByGlobalNormWithLog(
+        grad_clip = ClipGradByGlobalNormWithLog(config.global_grad_clip)
-            config.training.global_grad_clip)
         lr_scheduler = paddle.optimizer.lr.ExponentialDecay(
-            learning_rate=config.training.lr,
+            learning_rate=config.lr, gamma=config.lr_decay, verbose=True)
-            gamma=config.training.lr_decay,
-            verbose=True)
         optimizer = paddle.optimizer.Adam(
             learning_rate=lr_scheduler,
             parameters=model.parameters(),
-            weight_decay=paddle.regularizer.L2Decay(
+            weight_decay=paddle.regularizer.L2Decay(config.weight_decay),
-                config.training.weight_decay),
             grad_clip=grad_clip)
 
         self.model = model
@@ -184,59 +180,59 @@ 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,
+                sortagrad=config.sortagrad,
-                shuffle_method=config.collator.shuffle_method)
+                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,
+                sortagrad=config.sortagrad,
-                shuffle_method=config.collator.shuffle_method)
+                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.keep_transcription_text = True
-        config.collator.augmentation_config = ""
+        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)
@@ -274,7 +270,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
     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 +289,7 @@ 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 +395,3 @@ 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/transformer.yaml

@@ -1,109 +1,96 @@
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.train.tiny
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train.tiny
-  min_input_len: 0.05  # second
+dev_manifest: data/manifest.dev
-  max_input_len: 30.0 # second
+test_manifest: data/manifest.test
-  min_output_len: 0.0 # tokens
+min_input_len: 0.05  # second
-  max_output_len: 400.0 # tokens
+max_input_len: 30.0 # second
-  min_output_input_ratio: 0.01
+min_output_len: 0.0 # tokens
-  max_output_input_ratio: 20.0
+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
+#              Dataloader                 #
-  unit_type: 'spm'
+###########################################
-  spm_model_prefix: data/lang_char/bpe_unigram_8000
+vocab_filepath: data/lang_char/vocab.txt
-  mean_std_filepath: ""
+unit_type: 'spm'
-  # augmentation_config: conf/augmentation.json
+spm_model_prefix: data/lang_char/bpe_unigram_8000
-  batch_size: 10
+mean_std_filepath: ""
-  raw_wav: True  # use raw_wav or kaldi feature
+# augmentation_config: conf/augmentation.json
-  spectrum_type: fbank #linear, mfcc, fbank
+batch_size: 10
-  feat_dim: 80
+raw_wav: True  # use raw_wav or kaldi feature
-  delta_delta: False
+spectrum_type: fbank #linear, mfcc, fbank
-  dither: 1.0
+feat_dim: 80
-  target_sample_rate: 16000
+delta_delta: False
-  max_freq: None
+dither: 1.0
-  n_fft: None
+target_sample_rate: 16000
-  stride_ms: 10.0
+max_freq: None
-  window_ms: 25.0
+n_fft: None
-  use_dB_normalization: True
+stride_ms: 10.0
-  target_dB: -20
+window_ms: 25.0
-  random_seed: 0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True 
+random_seed: 0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 2
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 2
 
 
-# network architecture
+############################################
-model:
+#           Network Architecture           #
-    cmvn_file: "data/mean_std.json"
+############################################
-    cmvn_file_type: "json"
+cmvn_file: "data/mean_std.json"
-    # encoder related
+cmvn_file_type: "json"
-    encoder: transformer
+# encoder related
-    encoder_conf:
+encoder: transformer
-        output_size: 256    # dimension of attention
+encoder_conf:
-        attention_heads: 4
+    output_size: 256    # dimension of attention
-        linear_units: 2048  # the number of units of position-wise feed forward
+    attention_heads: 4
-        num_blocks: 12      # the number of encoder blocks
+    linear_units: 2048  # the number of units of position-wise feed forward
-        dropout_rate: 0.1
+    num_blocks: 12      # the number of encoder blocks
-        positional_dropout_rate: 0.1
+    dropout_rate: 0.1
-        attention_dropout_rate: 0.0
+    positional_dropout_rate: 0.1
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    attention_dropout_rate: 0.0
-        normalize_before: true
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
+# decoder related
-    decoder: transformer
+decoder: transformer
-    decoder_conf:
+decoder_conf:
-        attention_heads: 4
+    attention_heads: 4
-        linear_units: 2048
+    linear_units: 2048
-        num_blocks: 6
+    num_blocks: 6
-        dropout_rate: 0.1
+    dropout_rate: 0.1
-        positional_dropout_rate: 0.1
+    positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
+    self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
+# hybrid CTC/attention
-    model_conf:
+model_conf:
-        asr_weight: 0.0
+    asr_weight: 0.0
-        ctc_weight: 0.0
+    ctc_weight: 0.0
-        lsm_weight: 0.1     # label smoothing option
+    lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+    length_normalized_loss: false
 
-
+###########################################
-training:
+#                Training                 #
-  n_epoch: 120
+###########################################
-  accum_grad: 2
+n_epoch: 120
-  global_grad_clip: 5.0
+accum_grad: 2
-  optim: adam
+global_grad_clip: 5.0
-  optim_conf:
+optim: adam
-    lr: 0.004
+optim_conf:
-    weight_decay: 1e-06
+  lr: 0.004
-  scheduler: warmuplr     
+  weight_decay: 1.0e-06
-  scheduler_conf:
+scheduler: warmuplr     
-    warmup_steps: 25000
+scheduler_conf:
-    lr_decay: 1.0
+  warmup_steps: 25000
-  log_interval: 5
+  lr_decay: 1.0
-  checkpoint:
+log_interval: 5
-    kbest_n: 50
+checkpoint:
-    latest_n: 5
+  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
-  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/transformer_mtl_noam.yaml

@@ -1,112 +1,100 @@
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.train
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train
-  min_input_len: 0.05  # second
+dev_manifest: data/manifest.dev
-  max_input_len: 30.0 # second
+test_manifest: data/manifest.test
-  min_output_len: 0.0 # tokens
+min_input_len: 0.05  # second
-  max_output_len: 400.0 # tokens
+max_input_len: 30.0 # second
-  min_output_input_ratio: 0.01
+min_output_len: 0.0 # tokens
-  max_output_input_ratio: 20.0
+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
+#              Dataloader                 #
-  unit_type: 'spm'
+###########################################
-  spm_model_prefix: data/lang_char/bpe_unigram_8000
+vocab_filepath: data/lang_char/vocab.txt
-  mean_std_filepath: ""
+unit_type: 'spm'
-  # augmentation_config: conf/augmentation.json
+spm_model_prefix: data/lang_char/bpe_unigram_8000
-  batch_size: 10
+mean_std_filepath: ""
-  raw_wav: True  # use raw_wav or kaldi feature
+# augmentation_config: conf/augmentation.json
-  spectrum_type: fbank #linear, mfcc, fbank
+batch_size: 10
-  feat_dim: 80
+raw_wav: True  # use raw_wav or kaldi feature
-  delta_delta: False
+spectrum_type: fbank #linear, mfcc, fbank
-  dither: 1.0
+feat_dim: 80
-  target_sample_rate: 16000
+delta_delta: False
-  max_freq: None
+dither: 1.0
-  n_fft: None
+target_sample_rate: 16000
-  stride_ms: 10.0
+max_freq: None
-  window_ms: 25.0
+n_fft: None
-  use_dB_normalization: True
+stride_ms: 10.0
-  target_dB: -20
+window_ms: 25.0
-  random_seed: 0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True 
+random_seed: 0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 2
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 2
 
 
-# network architecture
+############################################
-model:
+#           Network Architecture           #
-    cmvn_file: "data/mean_std.json"
+############################################
-    cmvn_file_type: "json"
+cmvn_file: "data/mean_std.json"
-    # encoder related
+cmvn_file_type: "json"
-    encoder: transformer
+# encoder related
-    encoder_conf:
+encoder: transformer
-        output_size: 256    # dimension of attention
+encoder_conf:
-        attention_heads: 4
+    output_size: 256    # dimension of attention
-        linear_units: 2048  # the number of units of position-wise feed forward
+    attention_heads: 4
-        num_blocks: 12      # the number of encoder blocks
+    linear_units: 2048  # the number of units of position-wise feed forward
-        dropout_rate: 0.1
+    num_blocks: 12      # the number of encoder blocks
-        positional_dropout_rate: 0.1
+    dropout_rate: 0.1
-        attention_dropout_rate: 0.0
+    positional_dropout_rate: 0.1
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    attention_dropout_rate: 0.0
-        normalize_before: true
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
+# decoder related
-    decoder: transformer
+decoder: transformer
-    decoder_conf:
+decoder_conf:
-        attention_heads: 4
+    attention_heads: 4
-        linear_units: 2048
+    linear_units: 2048
-        num_blocks: 6
+    num_blocks: 6
-        dropout_rate: 0.1
+    dropout_rate: 0.1
-        positional_dropout_rate: 0.1
+    positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
+    self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
+# hybrid CTC/attention
-    model_conf:
+model_conf:
-        asr_weight: 0.5
+    asr_weight: 0.5
-        ctc_weight: 0.3
+    ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
+    lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+    length_normalized_loss: false
 
 
-training:
+###########################################
-  n_epoch: 120
+#                Training                 #
-  accum_grad: 2
+###########################################
-  global_grad_clip: 5.0
+n_epoch: 120
-  optim: adam
+accum_grad: 2
-  optim_conf:
+global_grad_clip: 5.0
-    lr: 2.5
+optim: adam
-    weight_decay: 1e-06
+optim_conf:
-  scheduler: noam    
+  lr: 2.5
-  scheduler_conf:
+  weight_decay: 1.0e-06
-    warmup_steps: 25000
+scheduler: noam    
-    lr_decay: 1.0
+scheduler_conf:
-  log_interval: 50
+  warmup_steps: 25000
-  checkpoint:
+  lr_decay: 1.0
-    kbest_n: 50
+log_interval: 50
-    latest_n: 5
+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
+if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+    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
 
 for type in fullsentence; do
     echo "decoding ${type}"
@@ -17,10 +18,11 @@ for type in fullsentence; 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 decode.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decode_batch_size ${batch_size}
 
     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/transformer.yaml

@@ -1,110 +1,97 @@
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.train.tiny
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train.tiny
-  min_input_len: 5.0  # frame
+dev_manifest: data/manifest.dev
-  max_input_len: 3000.0 # frame
+test_manifest: data/manifest.test
-  min_output_len: 0.0 # tokens
+min_input_len: 5.0  # frame
-  max_output_len: 400.0 # tokens
+max_input_len: 3000.0 # frame
-  min_output_input_ratio: 0.01
+min_output_len: 0.0 # tokens
-  max_output_input_ratio: 20.0
+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
+#              Dataloader                 #
-  unit_type: 'spm'
+###########################################
-  spm_model_prefix: data/lang_char/bpe_unigram_8000
+vocab_filepath: data/lang_char/vocab.txt
-  mean_std_filepath: ""
+unit_type: 'spm'
-  # augmentation_config: conf/augmentation.json
+spm_model_prefix: data/lang_char/bpe_unigram_8000
-  batch_size: 10
+mean_std_filepath: ""
-  raw_wav: True  # use raw_wav or kaldi feature
+# augmentation_config: conf/augmentation.json
-  spectrum_type: fbank #linear, mfcc, fbank
+batch_size: 10
-  feat_dim: 83
+raw_wav: True  # use raw_wav or kaldi feature
-  delta_delta: False
+spectrum_type: fbank #linear, mfcc, fbank
-  dither: 1.0
+feat_dim: 83
-  target_sample_rate: 16000
+delta_delta: False
-  max_freq: None
+dither: 1.0
-  n_fft: None
+target_sample_rate: 16000
-  stride_ms: 10.0
+max_freq: None
-  window_ms: 25.0
+n_fft: None
-  use_dB_normalization: True
+stride_ms: 10.0
-  target_dB: -20
+window_ms: 25.0
-  random_seed: 0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True 
+random_seed: 0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 2
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 2
 
 
-# network architecture
+############################################
-model:
+#           Network Architecture           #
-    cmvn_file: None
+############################################
-    cmvn_file_type: "json"
+cmvn_file: None
-    # encoder related
+cmvn_file_type: "json"
-    encoder: transformer
+# encoder related
-    encoder_conf:
+encoder: transformer
-        output_size: 256    # dimension of attention
+encoder_conf:
-        attention_heads: 4
+    output_size: 256    # dimension of attention
-        linear_units: 2048  # the number of units of position-wise feed forward
+    attention_heads: 4
-        num_blocks: 12      # the number of encoder blocks
+    linear_units: 2048  # the number of units of position-wise feed forward
-        dropout_rate: 0.1
+    num_blocks: 12      # the number of encoder blocks
-        positional_dropout_rate: 0.1
+    dropout_rate: 0.1
-        attention_dropout_rate: 0.0
+    positional_dropout_rate: 0.1
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    attention_dropout_rate: 0.0
-        normalize_before: true
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
+# decoder related
-    decoder: transformer
+decoder: transformer
-    decoder_conf:
+decoder_conf:
-        attention_heads: 4
+    attention_heads: 4
-        linear_units: 2048
+    linear_units: 2048
-        num_blocks: 6
+    num_blocks: 6
-        dropout_rate: 0.1
+    dropout_rate: 0.1
-        positional_dropout_rate: 0.1
+    positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
+    self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
+# hybrid CTC/attention
-    model_conf:
+model_conf:
-        asr_weight: 0.0
+    asr_weight: 0.0
-        ctc_weight: 0.0
+    ctc_weight: 0.0
-        lsm_weight: 0.1     # label smoothing option
+    lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+    length_normalized_loss: false
 
 
-training:
+###########################################
-  n_epoch: 20
+#                Training                 #
-  accum_grad: 2
+###########################################
-  global_grad_clip: 5.0
+n_epoch: 20
-  optim: adam
+accum_grad: 2
-  optim_conf:
+global_grad_clip: 5.0
-    lr: 0.004
+optim: adam
-    weight_decay: 1e-06
+optim_conf:
-  scheduler: warmuplr     
+  lr: 0.004
-  scheduler_conf:
+  weight_decay: 1.0e-06
-    warmup_steps: 25000
+scheduler: warmuplr     
-    lr_decay: 1.0
+scheduler_conf:
-  log_interval: 5
+  warmup_steps: 25000
-  checkpoint:
+  lr_decay: 1.0
-    kbest_n: 50
+log_interval: 5
-    latest_n: 5
+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/st1/conf/transformer_mtl_noam.yaml

@@ -1,110 +1,97 @@
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.train
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train
-  min_input_len: 5.0  # frame
+dev_manifest: data/manifest.dev
-  max_input_len: 3000.0 # frame
+test_manifest: data/manifest.test
-  min_output_len: 0.0 # tokens
+min_input_len: 5.0  # frame
-  max_output_len: 400.0 # tokens
+max_input_len: 3000.0 # frame
-  min_output_input_ratio: 0.01
+min_output_len: 0.0 # tokens
-  max_output_input_ratio: 20.0
+max_output_len: 400.0 # tokens
+min_output_input_ratio: 0.01
+max_output_input_ratio: 20.0
 
-collator:
+###########################################
-  vocab_filepath: data/lang_char/ted_en_zh_bpe8000.txt
+#              Dataloader                 #
-  unit_type: 'spm'
+###########################################
-  spm_model_prefix: data/lang_char/ted_en_zh_bpe8000
+vocab_filepath: data/lang_char/ted_en_zh_bpe8000.txt
-  mean_std_filepath: ""
+unit_type: 'spm'
-  # augmentation_config: conf/augmentation.json
+spm_model_prefix: data/lang_char/ted_en_zh_bpe8000
-  batch_size: 10
+mean_std_filepath: ""
-  raw_wav: True  # use raw_wav or kaldi feature
+# augmentation_config: conf/augmentation.json
-  spectrum_type: fbank #linear, mfcc, fbank
+batch_size: 10
-  feat_dim: 83
+raw_wav: True  # use raw_wav or kaldi feature
-  delta_delta: False
+spectrum_type: fbank #linear, mfcc, fbank
-  dither: 1.0
+feat_dim: 83
-  target_sample_rate: 16000
+delta_delta: False
-  max_freq: None
+dither: 1.0
-  n_fft: None
+target_sample_rate: 16000
-  stride_ms: 10.0
+max_freq: None
-  window_ms: 25.0
+n_fft: None
-  use_dB_normalization: True
+stride_ms: 10.0
-  target_dB: -20
+window_ms: 25.0
-  random_seed: 0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True 
+random_seed: 0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 2
+sortagrad: True 
+shuffle_method: batch_shuffle
+num_workers: 2
 
 
-# network architecture
+############################################
-model:
+#           Network Architecture           #
-    cmvn_file: None
+############################################
-    cmvn_file_type: "json"
+cmvn_file: None
-    # encoder related
+cmvn_file_type: "json"
-    encoder: transformer
+# encoder related
-    encoder_conf:
+encoder: transformer
-        output_size: 256    # dimension of attention
+encoder_conf:
-        attention_heads: 4
+    output_size: 256    # dimension of attention
-        linear_units: 2048  # the number of units of position-wise feed forward
+    attention_heads: 4
-        num_blocks: 12      # the number of encoder blocks
+    linear_units: 2048  # the number of units of position-wise feed forward
-        dropout_rate: 0.1
+    num_blocks: 12      # the number of encoder blocks
-        positional_dropout_rate: 0.1
+    dropout_rate: 0.1
-        attention_dropout_rate: 0.0
+    positional_dropout_rate: 0.1
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    attention_dropout_rate: 0.0
-        normalize_before: true
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
+# decoder related
-    decoder: transformer
+decoder: transformer
-    decoder_conf:
+decoder_conf:
-        attention_heads: 4
+    attention_heads: 4
-        linear_units: 2048
+    linear_units: 2048
-        num_blocks: 6
+    num_blocks: 6
-        dropout_rate: 0.1
+    dropout_rate: 0.1
-        positional_dropout_rate: 0.1
+    positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
+    self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
+# hybrid CTC/attention
-    model_conf:
+model_conf:
-        asr_weight: 0.5
+    asr_weight: 0.5
-        ctc_weight: 0.3
+    ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
+    lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+    length_normalized_loss: false
 
 
-training:
+###########################################
-  n_epoch: 20
+#                Training                 #
-  accum_grad: 2
+###########################################
-  global_grad_clip: 5.0
+n_epoch: 20
-  optim: adam
+accum_grad: 2
-  optim_conf:
+global_grad_clip: 5.0
-    lr: 2.5
+optim: adam
-    weight_decay: 1e-06
+optim_conf:
-  scheduler: noam    
+  lr: 2.5
-  scheduler_conf:
+  weight_decay: 1.0e-06
-    warmup_steps: 25000
+scheduler: noam    
-    lr_decay: 1.0
+scheduler_conf:
-  log_interval: 5
+  warmup_steps: 25000
-  checkpoint:
+  lr_decay: 1.0
-    kbest_n: 50
+log_interval: 5
-    latest_n: 5
+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/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
+if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+    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
 
 for type in fullsentence; do
     echo "decoding ${type}"
@@ -17,10 +18,11 @@ for type in fullsentence; 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 decode.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --opts decode.decode_batch_size ${batch_size}
 
     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
@@ -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
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_pat} exp/${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
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train
-  min_input_len: 0.0  # second
+dev_manifest: data/manifest.dev
-  max_input_len: 10.0 # second
+test_manifest: data/manifest.test
-  min_output_len: 0.0 # tokens
-  max_output_len: 150.0 # tokens
-  min_output_input_ratio: 0.005
-  max_output_input_ratio: 1000.0
 
-collator:
+###########################################
-  vocab_filepath: data/lang_char/vocab.txt
+#              Dataloader                 #
-  unit_type: "word"
+###########################################
-  mean_std_filepath: ""
+vocab_filepath: data/lang_char/vocab.txt
-  augmentation_config: conf/preprocess.yaml
+spm_model_prefix: ''
-  batch_size: 64
+unit_type: "word"
-  raw_wav: True  # use raw_wav or kaldi feature
+mean_std_filepath: ""
-  spectrum_type: fbank #linear, mfcc, fbank
+preprocess_config: conf/preprocess.yaml
-  feat_dim: 80
+feat_dim: 80
-  delta_delta: False
+stride_ms: 10.0
-  dither: 1.0
+window_ms: 25.0
-  target_sample_rate: 16000
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  max_freq: None
+batch_size: 64
-  n_fft: None
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  stride_ms: 10.0
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  window_ms: 25.0
+minibatches: 0 # for debug
-  use_dB_normalization: True
+batch_count: auto
-  target_dB: -20
+batch_bins: 0 
-  random_seed: 0
+batch_frames_in: 0
-  keep_transcription_text: False
+batch_frames_out: 0
-  sortagrad: True 
+batch_frames_inout: 0
-  shuffle_method: batch_shuffle
+num_workers: 0
-  num_workers: 2
+subsampling_factor: 1
+num_encs: 1
 
 
-# network architecture
+############################################
-model:
+#           Network Architecture           #
-    cmvn_file: 
+############################################
-    cmvn_file_type: "json"
+cmvn_file: 
-    # encoder related
+cmvn_file_type: "json"
-    encoder: transformer
+# encoder related
-    encoder_conf:
+encoder: transformer
-        output_size: 128    # dimension of attention
+encoder_conf:
-        attention_heads: 4
+    output_size: 128    # dimension of attention
-        linear_units: 1024  # the number of units of position-wise feed forward
+    attention_heads: 4
-        num_blocks: 6      # the number of encoder blocks
+    linear_units: 1024  # the number of units of position-wise feed forward
-        dropout_rate: 0.1
+    num_blocks: 6      # the number of encoder blocks
-        positional_dropout_rate: 0.1
+    dropout_rate: 0.1
-        attention_dropout_rate: 0.0
+    positional_dropout_rate: 0.1
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    attention_dropout_rate: 0.0
-        normalize_before: true
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    normalize_before: true
 
-    # decoder related
+# decoder related
-    decoder: transformer
+decoder: transformer
-    decoder_conf:
+decoder_conf:
-        attention_heads: 4
+    attention_heads: 4
-        linear_units: 1024
+    linear_units: 1024
-        num_blocks: 6
+    num_blocks: 6
-        dropout_rate: 0.1
+    dropout_rate: 0.1
-        positional_dropout_rate: 0.1
+    positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
+    self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
+# hybrid CTC/attention
-    model_conf:
+model_conf:
-        ctc_weight: 0.5
+    ctc_weight: 0.5
-        lsm_weight: 0.1     # label smoothing option
+    lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+    length_normalized_loss: false
 
 
-training:
+###########################################
-  n_epoch: 50
+#                Training                 #
-  accum_grad: 1
+###########################################
-  global_grad_clip: 5.0
+n_epoch: 50
-  optim: adam
+accum_grad: 1
-  optim_conf:
+global_grad_clip: 5.0
-    lr: 0.004
+optim: adam
-    weight_decay: 1e-06
+optim_conf:
-  scheduler: warmuplr     
+  lr: 0.004
-  scheduler_conf:
+  weight_decay: 1.0e-6
-    warmup_steps: 1200
+scheduler: warmuplr     
-    lr_decay: 1.0
+scheduler_conf:
-  log_interval: 10
+  warmup_steps: 1200
-  checkpoint:
+  lr_decay: 1.0
-    kbest_n: 50
+log_interval: 10
-    latest_n: 5
+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.
-
 

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
+if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+    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
+if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+    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 decode.decoding_method ${type} \
-        --opts decoding.batch_size ${batch_size}
+        --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 decode.decoding_method ${type} \
-        --opts decoding.batch_size ${batch_size}
+        --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 decode.decoding_method ${type} \
-        --opts decoding.batch_size ${batch_size}
+        --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
+if [ ${stage} -le 51 ] && [ ${stop_stage} -ge 51 ]; then
-#     # export ckpt avg_n
+     # export ckpt avg_n
-#     CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
+     CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
-# fi
+fi

examples/tiny/asr0/conf/deepspeech2.yaml

@@ -1,70 +1,67 @@
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.tiny
+#                   Data                  #
-  dev_manifest: data/manifest.tiny
+###########################################
-  test_manifest: data/manifest.tiny 
+train_manifest: data/manifest.tiny
-  min_input_len: 0.0
+dev_manifest: data/manifest.tiny
-  max_input_len: 30.0
+test_manifest: data/manifest.tiny 
-  min_output_len: 0.0
+min_input_len: 0.0
-  max_output_len: 400.0
+max_input_len: 30.0
-  min_output_input_ratio: 0.05
+min_output_len: 0.0
-  max_output_input_ratio: 10.0
+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
+#              Dataloader                 #
-  unit_type: char
+###########################################
-  vocab_filepath: data/lang_char/vocab.txt
+mean_std_filepath: data/mean_std.json
-  augmentation_config: conf/augmentation.json
+unit_type: char
-  random_seed: 0
+vocab_filepath: data/lang_char/vocab.txt
-  spm_model_prefix: 
+augmentation_config: conf/augmentation.json
-  spectrum_type: linear
+random_seed: 0
-  feat_dim: 
+spm_model_prefix: 
-  delta_delta: False
+spectrum_type: linear
-  stride_ms: 10.0
+feat_dim: 
-  window_ms: 20.0
+delta_delta: False
-  n_fft: None
+stride_ms: 10.0
-  max_freq: None
+window_ms: 20.0
-  target_sample_rate: 16000
+n_fft: None
-  use_dB_normalization: True
+max_freq: None
-  target_dB: -20
+target_sample_rate: 16000
-  dither: 1.0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True 
+dither: 1.0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 2
+sortagrad: True 
-  batch_size: 4
+shuffle_method: batch_shuffle
+num_workers: 2
+batch_size: 4
   
-model:
+############################################
-  num_conv_layers: 2
+#           Network Architecture           #
-  num_rnn_layers: 3
+############################################
-  rnn_layer_size: 2048
+num_conv_layers: 2
-  use_gru: False 
+num_rnn_layers: 3
-  share_rnn_weights: True 
+rnn_layer_size: 2048
-  blank_id: 0
+use_gru: False 
+share_rnn_weights: True 
+blank_id: 0
   
 
-training:
+###########################################
-  n_epoch: 5
+#                Training                 #
-  accum_grad: 1
+###########################################
-  lr: 1e-5 
+n_epoch: 5
-  lr_decay: 0.8 
+accum_grad: 1
-  weight_decay: 1e-06
+lr: 1e-5 
-  global_grad_clip: 5.0
+lr_decay: 0.8 
-  log_interval: 1
+weight_decay: 1e-06
-  checkpoint:
+global_grad_clip: 5.0
-    kbest_n: 3
+log_interval: 1
-    latest_n: 2
+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
+#                   Data                  #
-  dev_manifest: data/manifest.tiny
+###########################################
-  test_manifest: data/manifest.tiny 
+train_manifest: data/manifest.tiny
-  min_input_len: 0.0
+dev_manifest: data/manifest.tiny
-  max_input_len: 30.0
+test_manifest: data/manifest.tiny 
-  min_output_len: 0.0
+min_input_len: 0.0
-  max_output_len: 400.0
+max_input_len: 30.0
-  min_output_input_ratio: 0.05
+min_output_len: 0.0
-  max_output_input_ratio: 10.0
+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
+#              Dataloader                 #
-  unit_type: char
+###########################################
-  vocab_filepath: data/lang_char/vocab.txt
+mean_std_filepath: data/mean_std.json
-  augmentation_config: conf/augmentation.json
+unit_type: char
-  random_seed: 0
+vocab_filepath: data/lang_char/vocab.txt
-  spm_model_prefix: 
+augmentation_config: conf/augmentation.json
-  spectrum_type: linear
+random_seed: 0
-  feat_dim: 
+spm_model_prefix: 
-  delta_delta: False
+spectrum_type: linear
-  stride_ms: 10.0
+feat_dim: 
-  window_ms: 20.0
+delta_delta: False
-  n_fft: None
+stride_ms: 10.0
-  max_freq: None
+window_ms: 20.0
-  target_sample_rate: 16000
+n_fft: None
-  use_dB_normalization: True
+max_freq: None
-  target_dB: -20
+target_sample_rate: 16000
-  dither: 1.0
+use_dB_normalization: True
-  keep_transcription_text: False
+target_dB: -20
-  sortagrad: True 
+dither: 1.0
-  shuffle_method: batch_shuffle
+keep_transcription_text: False
-  num_workers: 0
+sortagrad: True 
-  batch_size: 4
+shuffle_method: batch_shuffle
+num_workers: 0
+batch_size: 4
   
-model:
+############################################
-  num_conv_layers: 2
+#           Network Architecture           #
-  num_rnn_layers: 4
+############################################
-  rnn_layer_size: 2048
+num_conv_layers: 2
-  rnn_direction: forward
+num_rnn_layers: 4
-  num_fc_layers: 2
+rnn_layer_size: 2048
-  fc_layers_size_list: 512, 256
+rnn_direction: forward
-  use_gru: True 
+num_fc_layers: 2
-  blank_id: 0
+fc_layers_size_list: 512, 256
+use_gru: True 
+blank_id: 0
   
 
-training:
+###########################################
-  n_epoch: 5
+#                Training                 #
-  accum_grad: 1
+###########################################
-  lr: 1e-5 
+n_epoch: 5
-  lr_decay: 1.0 
+accum_grad: 1
-  weight_decay: 1e-06
+lr: 1e-5 
-  global_grad_clip: 5.0
+lr_decay: 1.0 
-  log_interval: 1
+weight_decay: 1e-06
-  checkpoint:
+global_grad_clip: 5.0
-    kbest_n: 3
+log_interval: 1
-    latest_n: 2
+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
+if [ $# != 4 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix model_type"
+    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
+decode_config_path=$2
-model_type=$3
+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 +1,98 @@
-# https://yaml.org/type/float.html
+############################################
-data:
+#           Network Architecture           #
-  train_manifest: data/manifest.tiny
+############################################
-  dev_manifest: data/manifest.tiny
+cmvn_file: "data/mean_std.json"
-  test_manifest: data/manifest.tiny
+cmvn_file_type: "json"
-  min_input_len: 0.5  # second
+# encoder related
-  max_input_len: 30.0 # second
+encoder: conformer
-  min_output_len: 0.0 # tokens
+encoder_conf:
-  max_output_len: 400.0 # tokens
+    output_size: 256    # dimension of attention
-  min_output_input_ratio: 0.05
+    attention_heads: 4
-  max_output_input_ratio: 10.0
+    linear_units: 2048  # the number of units of position-wise feed forward
-  
+    num_blocks: 12      # the number of encoder blocks
-collator:
+    dropout_rate: 0.1
-  mean_std_filepath: ""
+    positional_dropout_rate: 0.1
-  vocab_filepath: data/lang_char/vocab.txt 
+    attention_dropout_rate: 0.0
-  unit_type: 'spm'
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-  spm_model_prefix: 'data/lang_char/bpe_unigram_200'
+    normalize_before: True
-  augmentation_config: conf/preprocess.yaml
+    use_cnn_module: True
-  batch_size: 4
+    cnn_module_kernel: 15
-  raw_wav: True  # use raw_wav or kaldi feature
+    activation_type: 'swish'
-  spectrum_type: fbank #linear, mfcc, fbank
+    pos_enc_layer_type: 'rel_pos'
-  feat_dim: 80
+    selfattention_layer_type: 'rel_selfattn'
-  delta_delta: False
+    causal: True
-  dither: 1.0
+    use_dynamic_chunk: True
-  target_sample_rate: 16000
+    cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
-  max_freq: None
+    use_dynamic_left_chunk: false
-  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
 
-    # hybrid CTC/attention
+# decoder related
-    model_conf:
+decoder: transformer
-        ctc_weight: 0.3
+decoder_conf:
-        lsm_weight: 0.1     # label smoothing option
+    attention_heads: 4
-        length_normalized_loss: false
+    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
+#              Dataloader                 #
-  decoding_method: attention  # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
+###########################################
-  lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
+mean_std_filepath: ""
-  alpha: 2.5
+vocab_filepath: data/lang_char/vocab.txt 
-  beta: 0.3
+unit_type: 'spm'
-  beam_size: 10
+spm_model_prefix: 'data/lang_char/bpe_unigram_200'
-  cutoff_prob: 1.0
+preprocess_config: conf/preprocess.yaml
-  cutoff_top_n: 0
+feat_dim: 80
-  num_proc_bsearch: 8
+stride_ms: 10.0
-  ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
+window_ms: 25.0
-  decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-      # <0: for decoding, use full chunk.
+batch_size: 4
-      # >0: for decoding, use fixed chunk size as set.
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-      # 0: used for training, it's prohibited here. 
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  num_decoding_left_chunks: -1  # number of left chunks for decoding. Defaults to -1.
+minibatches: 0 # for debug
-  simulate_streaming: False  # simulate streaming inference. Defaults to False.
+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: 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:
+#           Network Architecture           #
-  train_manifest: data/manifest.tiny
+############################################
-  dev_manifest: data/manifest.tiny
+cmvn_file: "data/mean_std.json"
-  test_manifest: data/manifest.tiny
+cmvn_file_type: "json"
-  min_input_len: 0.5  # second
+# encoder related
-  max_input_len: 20.0 # second
+encoder: transformer
-  min_output_len: 0.0 # tokens
+encoder_conf:
-  max_output_len: 400.0 # tokens
+    output_size: 256    # dimension of attention
-  min_output_input_ratio: 0.05
+    attention_heads: 4
-  max_output_input_ratio: 10.0
+    linear_units: 2048  # the number of units of position-wise feed forward
-  
+    num_blocks: 12      # the number of encoder blocks
-collator:
+    dropout_rate: 0.1
-  mean_std_filepath: ""
+    positional_dropout_rate: 0.1
-  vocab_filepath: data/lang_char/vocab.txt 
+    attention_dropout_rate: 0.0
-  unit_type: 'spm'
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-  spm_model_prefix: 'data/lang_char/bpe_unigram_200'
+    normalize_before: true
-  augmentation_config: conf/preprocess.yaml
+    use_dynamic_chunk: true
-  batch_size: 4
+    use_dynamic_left_chunk: false
-  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
 
-    # decoder related
+# decoder related
-    decoder: transformer
+decoder: transformer
-    decoder_conf:
+decoder_conf:
-        attention_heads: 4
+    attention_heads: 4
-        linear_units: 2048
+    linear_units: 2048
-        num_blocks: 6
+    num_blocks: 6
-        dropout_rate: 0.1
+    dropout_rate: 0.1
-        positional_dropout_rate: 0.1
+    positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
+    self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
+# hybrid CTC/attention
-    model_conf:
+model_conf:
-        ctc_weight: 0.3
+    ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
+    lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+    length_normalized_loss: false
 
 
-training:
+# https://yaml.org/type/float.html
-  n_epoch: 5
+###########################################
-  accum_grad: 1
+#                   Data                  #
-  global_grad_clip: 5.0
+###########################################
-  optim: adam
+train_manifest: data/manifest.tiny
-  optim_conf:
+dev_manifest: data/manifest.tiny
-    lr: 0.002
+test_manifest: data/manifest.tiny
-    weight_decay: 1e-06
+  
-  scheduler: warmuplr     
+###########################################
-  scheduler_conf:
+#              Dataloader                 #
-    warmup_steps: 25000
+###########################################
-    lr_decay: 1.0
+mean_std_filepath: ""
-  log_interval: 1
+vocab_filepath: data/lang_char/vocab.txt 
-  checkpoint:
+unit_type: 'spm'
-    kbest_n: 10
+spm_model_prefix: 'data/lang_char/bpe_unigram_200'
-    latest_n: 1
+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

@@ -1,46 +1,4 @@
 # 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
-  
-
-###########################################
-#              Dataloader                 #
-###########################################
-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           #
 ############################################
@@ -83,7 +41,41 @@ model_conf:
 
 
 ###########################################
-#                 training                #
+#                   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
@@ -91,7 +83,7 @@ global_grad_clip: 5.0
 optim: adam
 optim_conf:
   lr: 0.002
-  weight_decay: 1e-06
+  weight_decay: 1.0e-06
 scheduler: warmuplr     
 scheduler_conf:
   warmup_steps: 25000

examples/tiny/asr1/conf/transformer.yaml

@@ -1,44 +1,4 @@
 # 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
-  
-###########################################
-#              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'
-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           #
 ############################################
@@ -74,9 +34,41 @@ model_conf:
     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                #
+#                 Training                #
 ###########################################
 n_epoch: 5
 accum_grad: 1
@@ -84,7 +76,7 @@ global_grad_clip: 5.0
 optim: adam
 optim_conf:
   lr: 0.002
-  weight_decay: 1e-06
+  weight_decay: 1.0e-06
 scheduler: warmuplr     
 scheduler_conf:
   warmup_steps: 25000

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
+if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+    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
+if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+    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 decode.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --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 decode.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --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:
+#           Network Architecture           #
-    # encoder related
+############################################
-    encoder: conformer
+cmvn_file: 
-    encoder_conf:
+cmvn_file_type: "json"
-        output_size: 512    # dimension of attention
+# encoder related
-        attention_heads: 8
+encoder: conformer
-        linear_units: 2048  # the number of units of position-wise feed forward
+encoder_conf:
-        num_blocks: 12      # the number of encoder blocks
+    output_size: 512    # dimension of attention
-        dropout_rate: 0.1
+    attention_heads: 8
-        positional_dropout_rate: 0.1
+    linear_units: 2048  # the number of units of position-wise feed forward
-        attention_dropout_rate: 0.0
+    num_blocks: 12      # the number of encoder blocks
-        input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
+    dropout_rate: 0.1
-        normalize_before: True
+    positional_dropout_rate: 0.1
-        use_cnn_module: True
+    attention_dropout_rate: 0.0
-        cnn_module_kernel: 15
+    input_layer: conv2d # encoder input type, you can chose conv2d, conv2d6 and conv2d8
-        cnn_module_norm: layer_norm
+    normalize_before: True
-        activation_type: swish
+    use_cnn_module: True
-        pos_enc_layer_type: rel_pos
+    cnn_module_kernel: 15
-        selfattention_layer_type: rel_selfattn
+    cnn_module_norm: layer_norm
+    activation_type: swish
+    pos_enc_layer_type: rel_pos
+    selfattention_layer_type: rel_selfattn
 
-    # decoder related
+# decoder related
-    decoder: transformer
+decoder: transformer
-    decoder_conf:
+decoder_conf:
-        attention_heads: 8
+    attention_heads: 8
-        linear_units: 2048
+    linear_units: 2048
-        num_blocks: 6
+    num_blocks: 6
-        dropout_rate: 0.1
+    dropout_rate: 0.1
-        positional_dropout_rate: 0.1
+    positional_dropout_rate: 0.1
-        self_attention_dropout_rate: 0.0
+    self_attention_dropout_rate: 0.0
-        src_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
 
-    # hybrid CTC/attention
+# hybrid CTC/attention
-    model_conf:
+model_conf:
-        ctc_weight: 0.3
+    ctc_weight: 0.3
-        lsm_weight: 0.1     # label smoothing option
+    lsm_weight: 0.1     # label smoothing option
-        length_normalized_loss: false
+    length_normalized_loss: false
 
 # https://yaml.org/type/float.html
-data:
+###########################################
-  train_manifest: data/manifest.train
+#                   Data                  #
-  dev_manifest: data/manifest.dev
+###########################################
-  test_manifest: data/manifest.test
+train_manifest: data/manifest.train
-  min_input_len: 0.1 # second
+dev_manifest: data/manifest.dev
-  max_input_len: 12.0 # second
+test_manifest: data/manifest.test
-  min_output_len: 1.0
-  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 
+#              Dataloader                 #
-  unit_type: 'char'
+###########################################
-  spm_model_prefix: ''
+vocab_filepath: data/lang_char/vocab.txt 
-  augmentation_config: conf/preprocess.yaml
+unit_type: 'char'
-  batch_size: 64
+preprocess_config: conf/preprocess.yaml
-  raw_wav: True  # use raw_wav or kaldi feature
+spm_model_prefix: ''
-  spectrum_type: fbank #linear, mfcc, fbank
+feat_dim: 80
-  feat_dim: 80
+stride_ms: 10.0
-  delta_delta: False
+window_ms: 25.0
-  dither: 1.0
+sortagrad: 0 # Feed samples from shortest to longest ; -1: enabled for all epochs, 0: disabled, other: enabled for 'other' epochs 
-  target_sample_rate: 16000
+batch_size: 64
-  max_freq: None
+maxlen_in: 512  # if input length  > maxlen-in, batchsize is automatically reduced
-  n_fft: None
+maxlen_out: 150  # if output length > maxlen-out, batchsize is automatically reduced
-  stride_ms: 10.0
+minibatches: 0 # for debug
-  window_ms: 25.0
+batch_count: auto
-  use_dB_normalization: True 
+batch_bins: 0 
-  target_dB: -20
+batch_frames_in: 0
-  random_seed: 0
+batch_frames_out: 0
-  keep_transcription_text: False
+batch_frames_inout: 0
-  sortagrad: True 
+num_workers: 0
-  shuffle_method: batch_shuffle
+subsampling_factor: 1
-  num_workers: 2
+num_encs: 1
 
 
-training:
+###########################################
-  n_epoch: 240 
+#                 Training                #
-  accum_grad: 16
+###########################################
-  global_grad_clip: 5.0
+n_epoch: 240 
-  log_interval: 100
+accum_grad: 16
-  checkpoint:
+global_grad_clip: 5.0
-    kbest_n: 50
+log_interval: 100
-    latest_n: 5
+checkpoint:
-  optim: adam
+  kbest_n: 50
-  optim_conf:
+  latest_n: 5
-    lr: 0.001
+optim: adam
-    weight_decay: 1e-6
+optim_conf:
-  scheduler: warmuplr     
+  lr: 0.001
-  scheduler_conf:
+  weight_decay: 1.0e-6
-    warmup_steps: 5000
+scheduler: warmuplr     
-    lr_decay: 1.0
+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.

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
+if [ $# != 3 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix"
+    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 decode.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --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 decode.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size}
+    --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
+if [ $# != 4 ];then
-    echo "usage: ${0} config_path ckpt_path_prefix audio_file"
+    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
+decode_config_path=$2
-audio_file=$3
+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 decode.decoding_method ${type} \
-    --opts decoding.batch_size ${batch_size} \
+    --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/exps/deepspeech2/bin/deploy/runtime.py

@@ -80,13 +80,13 @@ def inference(config, args):
 def start_server(config, args):
     """Start the ASR server"""
     config.defrost()
-    config.data.manifest = config.data.test_manifest
+    config.manifest = config.test_manifest
     dataset = ManifestDataset.from_config(config)
 
-    config.collator.augmentation_config = ""
+    config.augmentation_config = ""
-    config.collator.keep_transcription_text = True
+    config.keep_transcription_text = True
-    config.collator.batch_size = 1
+    config.batch_size = 1
-    config.collator.num_workers = 0
+    config.num_workers = 0
     collate_fn = SpeechCollator.from_config(config)
     test_loader = DataLoader(dataset, collate_fn=collate_fn, num_workers=0)
 
@@ -105,14 +105,14 @@ def start_server(config, args):
             paddle.to_tensor(audio),
             paddle.to_tensor(audio_len),
             vocab_list=test_loader.collate_fn.vocab_list,
-            decoding_method=config.decoding.decoding_method,
+            decoding_method=config.decode.decoding_method,
-            lang_model_path=config.decoding.lang_model_path,
+            lang_model_path=config.decode.lang_model_path,
-            beam_alpha=config.decoding.alpha,
+            beam_alpha=config.decode.alpha,
-            beam_beta=config.decoding.beta,
+            beam_beta=config.decode.beta,
-            beam_size=config.decoding.beam_size,
+            beam_size=config.decode.beam_size,
-            cutoff_prob=config.decoding.cutoff_prob,
+            cutoff_prob=config.decode.cutoff_prob,
-            cutoff_top_n=config.decoding.cutoff_top_n,
+            cutoff_top_n=config.decode.cutoff_top_n,
-            num_processes=config.decoding.num_proc_bsearch)
+            num_processes=config.decode.num_proc_bsearch)
         return result_transcript[0]
 
     # warming up with utterrances sampled from Librispeech
@@ -179,12 +179,16 @@ if __name__ == "__main__":
     config = get_cfg_defaults()
     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()
     print(config)
 
-    args.warmup_manifest = config.data.test_manifest
+    args.warmup_manifest = config.test_manifest
     print_arguments(args, globals())
 
     if args.dump_config:

paddlespeech/s2t/exps/deepspeech2/bin/deploy/server.py

@@ -33,13 +33,13 @@ from paddlespeech.s2t.utils.utility import print_arguments
 def start_server(config, args):
     """Start the ASR server"""
     config.defrost()
-    config.data.manifest = config.data.test_manifest
+    config.manifest = config.test_manifest
     dataset = ManifestDataset.from_config(config)
 
-    config.collator.augmentation_config = ""
+    config.augmentation_config = ""
-    config.collator.keep_transcription_text = True
+    config.keep_transcription_text = True
-    config.collator.batch_size = 1
+    config.batch_size = 1
-    config.collator.num_workers = 0
+    config.num_workers = 0
     collate_fn = SpeechCollator.from_config(config)
     test_loader = DataLoader(dataset, collate_fn=collate_fn, num_workers=0)
 
@@ -62,14 +62,14 @@ def start_server(config, args):
             paddle.to_tensor(audio),
             paddle.to_tensor(audio_len),
             vocab_list=test_loader.collate_fn.vocab_list,
-            decoding_method=config.decoding.decoding_method,
+            decoding_method=config.decode.decoding_method,
-            lang_model_path=config.decoding.lang_model_path,
+            lang_model_path=config.decode.lang_model_path,
-            beam_alpha=config.decoding.alpha,
+            beam_alpha=config.decode.alpha,
-            beam_beta=config.decoding.beta,
+            beam_beta=config.decode.beta,
-            beam_size=config.decoding.beam_size,
+            beam_size=config.decode.beam_size,
-            cutoff_prob=config.decoding.cutoff_prob,
+            cutoff_prob=config.decode.cutoff_prob,
-            cutoff_top_n=config.decoding.cutoff_top_n,
+            cutoff_top_n=config.decode.cutoff_top_n,
-            num_processes=config.decoding.num_proc_bsearch)
+            num_processes=config.decode.num_proc_bsearch)
         return result_transcript[0]
 
     # warming up with utterrances sampled from Librispeech
@@ -114,12 +114,16 @@ if __name__ == "__main__":
     config = get_cfg_defaults()
     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()
     print(config)
 
-    args.warmup_manifest = config.data.test_manifest
+    args.warmup_manifest = config.test_manifest
     print_arguments(args, globals())
 
     if args.dump_config:

paddlespeech/s2t/exps/deepspeech2/bin/test.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Evaluation for DeepSpeech2 model."""
+from yacs.config import CfgNode
+
 from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults
 from paddlespeech.s2t.exps.deepspeech2.model import DeepSpeech2Tester as Tester
 from paddlespeech.s2t.training.cli import default_argument_parser
@@ -44,6 +46,10 @@ if __name__ == "__main__":
     config = get_cfg_defaults(args.model_type)
     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()

paddlespeech/s2t/exps/deepspeech2/bin/test_export.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Evaluation for DeepSpeech2 model."""
+from yacs.config import CfgNode
+
 from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults
 from paddlespeech.s2t.exps.deepspeech2.model import DeepSpeech2ExportTester as ExportTester
 from paddlespeech.s2t.training.cli import default_argument_parser
@@ -49,6 +51,10 @@ if __name__ == "__main__":
     config = get_cfg_defaults(args.model_type)
     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()

paddlespeech/s2t/exps/deepspeech2/bin/test_wav.py

@@ -18,6 +18,7 @@ from pathlib import Path
 
 import paddle
 import soundfile
+from yacs.config import CfgNode
 
 from paddlespeech.s2t.exps.deepspeech2.config import get_cfg_defaults
 from paddlespeech.s2t.frontend.featurizer.text_featurizer import TextFeaturizer
@@ -41,7 +42,7 @@ class DeepSpeech2Tester_hub():
         self.audio_file = args.audio_file
         self.collate_fn_test = SpeechCollator.from_config(config)
         self._text_featurizer = TextFeaturizer(
-            unit_type=config.collator.unit_type, vocab=None)
+            unit_type=config.unit_type, vocab=None)
 
     def compute_result_transcripts(self, audio, audio_len, vocab_list, cfg):
         result_transcripts = self.model.decode(
@@ -74,7 +75,7 @@ class DeepSpeech2Tester_hub():
         audio = paddle.unsqueeze(audio, axis=0)
         vocab_list = collate_fn_test.vocab_list
         result_transcripts = self.compute_result_transcripts(
-            audio, audio_len, vocab_list, cfg.decoding)
+            audio, audio_len, vocab_list, cfg.decode)
         logger.info("result_transcripts: " + result_transcripts[0])
 
     def run_test(self):
@@ -110,13 +111,13 @@ class DeepSpeech2Tester_hub():
     def setup_model(self):
         config = self.config.clone()
         with UpdateConfig(config):
-            config.model.input_dim = self.collate_fn_test.feature_size
+            config.input_dim = self.collate_fn_test.feature_size
-            config.model.output_dim = self.collate_fn_test.vocab_size
+            config.output_dim = self.collate_fn_test.vocab_size
 
         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")
 
@@ -134,8 +135,8 @@ class DeepSpeech2Tester_hub():
         self.checkpoint_dir = checkpoint_dir
 
         self.checkpoint = Checkpoint(
-            kbest_n=self.config.training.checkpoint.kbest_n,
+            kbest_n=self.config.checkpoint.kbest_n,
-            latest_n=self.config.training.checkpoint.latest_n)
+            latest_n=self.config.checkpoint.latest_n)
 
     def resume(self):
         """Resume from the checkpoint at checkpoints in the output
@@ -190,6 +191,10 @@ if __name__ == "__main__":
     config = get_cfg_defaults(args.model_type)
     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()

paddlespeech/s2t/exps/deepspeech2/config.py

@@ -23,17 +23,6 @@ from paddlespeech.s2t.models.ds2_online import DeepSpeech2ModelOnline
 
 def get_cfg_defaults(model_type='offline'):
     _C = CfgNode()
-    _C.data = ManifestDataset.params()
-    _C.collator = SpeechCollator.params()
-    _C.training = DeepSpeech2Trainer.params()
-    _C.decoding = DeepSpeech2Tester.params()
-    if model_type == 'offline':
-        _C.model = DeepSpeech2Model.params()
-    else:
-        _C.model = DeepSpeech2ModelOnline.params()
-    """Get a yacs CfgNode object with default values for my_project."""
-    # Return a clone so that the defaults will not be altered
-    # This is for the "local variable" use pattern
     config = _C.clone()
     config.set_new_allowed(True)
     return config

paddlespeech/s2t/exps/deepspeech2/model.py

@@ -69,8 +69,8 @@ class DeepSpeech2Trainer(Trainer):
         super().__init__(config, args)
 
     def train_batch(self, batch_index, batch_data, msg):
-        batch_size = self.config.collator.batch_size
+        batch_size = self.config.batch_size
-        accum_grad = self.config.training.accum_grad
+        accum_grad = self.config.accum_grad
 
         start = time.time()
 
@@ -133,7 +133,7 @@ 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
 
@@ -154,16 +154,16 @@ class DeepSpeech2Trainer(Trainer):
         config = self.config.clone()
         with UpdateConfig(config):
             if self.train:
-                config.model.input_dim = self.train_loader.collate_fn.feature_size
+                config.input_dim = self.train_loader.collate_fn.feature_size
-                config.model.output_dim = self.train_loader.collate_fn.vocab_size
+                config.output_dim = self.train_loader.collate_fn.vocab_size
             else:
-                config.model.input_dim = self.test_loader.collate_fn.feature_size
+                config.input_dim = self.test_loader.collate_fn.feature_size
-                config.model.output_dim = self.test_loader.collate_fn.vocab_size
+                config.output_dim = self.test_loader.collate_fn.vocab_size
 
         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:
@@ -177,17 +177,13 @@ class DeepSpeech2Trainer(Trainer):
         if not self.train:
             return
 
-        grad_clip = ClipGradByGlobalNormWithLog(
+        grad_clip = ClipGradByGlobalNormWithLog(config.global_grad_clip)
-            config.training.global_grad_clip)
         lr_scheduler = paddle.optimizer.lr.ExponentialDecay(
-            learning_rate=config.training.lr,
+            learning_rate=config.lr, gamma=config.lr_decay, verbose=True)
-            gamma=config.training.lr_decay,
-            verbose=True)
         optimizer = paddle.optimizer.Adam(
             learning_rate=lr_scheduler,
             parameters=model.parameters(),
-            weight_decay=paddle.regularizer.L2Decay(
+            weight_decay=paddle.regularizer.L2Decay(config.weight_decay),
-                config.training.weight_decay),
             grad_clip=grad_clip)
         self.optimizer = optimizer
         self.lr_scheduler = lr_scheduler
@@ -198,66 +194,67 @@ class DeepSpeech2Trainer(Trainer):
         config.defrost()
         if self.train:
             # train
-            config.data.manifest = config.data.train_manifest
+            config.manifest = config.train_manifest
             train_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,
+                    sortagrad=config.sortagrad,
-                    shuffle_method=config.collator.shuffle_method)
+                    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,
+                    sortagrad=config.sortagrad,
-                    shuffle_method=config.collator.shuffle_method)
+                    shuffle_method=config.shuffle_method)
 
-            config.collator.keep_transcription_text = False
+            config.keep_transcription_text = False
             collate_fn_train = 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)
 
             # dev
-            config.data.manifest = config.data.dev_manifest
+            config.manifest = config.dev_manifest
             dev_dataset = ManifestDataset.from_config(config)
 
-            config.collator.augmentation_config = ""
+            config.augmentation_config = ""
-            config.collator.keep_transcription_text = False
+            config.keep_transcription_text = False
             collate_fn_dev = SpeechCollator.from_config(config)
             self.valid_loader = DataLoader(
                 dev_dataset,
-                batch_size=int(config.collator.batch_size),
+                batch_size=int(config.batch_size),
                 shuffle=False,
                 drop_last=False,
                 collate_fn=collate_fn_dev,
-                num_workers=config.collator.num_workers)
+                num_workers=config.num_workers)
             logger.info("Setup train/valid  Dataloader!")
         else:
             # test
-            config.data.manifest = config.data.test_manifest
+            config.manifest = config.test_manifest
             test_dataset = ManifestDataset.from_config(config)
 
-            config.collator.augmentation_config = ""
+            config.augmentation_config = ""
-            config.collator.keep_transcription_text = True
+            config.keep_transcription_text = True
             collate_fn_test = SpeechCollator.from_config(config)
-
+            decode_batch_size = config.get('decode', dict()).get(
+                'decode_batch_size', 1)
             self.test_loader = DataLoader(
                 test_dataset,
-                batch_size=config.decoding.batch_size,
+                batch_size=decode_batch_size,
                 shuffle=False,
                 drop_last=False,
                 collate_fn=collate_fn_test,
-                num_workers=config.collator.num_workers)
+                num_workers=config.num_workers)
             logger.info("Setup test  Dataloader!")
 
 
@@ -286,7 +283,7 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
     def __init__(self, config, args):
         super().__init__(config, args)
         self._text_featurizer = TextFeaturizer(
-            unit_type=config.collator.unit_type, vocab=None)
+            unit_type=config.unit_type, vocab=None)
 
     def ordid2token(self, texts, texts_len):
         """ ord() id to chr() chr """
@@ -304,17 +301,17 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
                         texts,
                         texts_len,
                         fout=None):
-        cfg = self.config.decoding
+        decode_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
+        errors_func = error_rate.char_errors if decode_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
+        error_rate_func = error_rate.cer if decode_cfg.error_rate_type == 'cer' else error_rate.wer
 
         vocab_list = self.test_loader.collate_fn.vocab_list
 
         target_transcripts = self.ordid2token(texts, texts_len)
 
-        result_transcripts = self.compute_result_transcripts(audio, audio_len,
+        result_transcripts = self.compute_result_transcripts(
-                                                             vocab_list, cfg)
+            audio, audio_len, vocab_list, decode_cfg)
 
         for utt, target, result in zip(utts, target_transcripts,
                                        result_transcripts):
@@ -327,29 +324,31 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
             logger.info(f"Utt: {utt}")
             logger.info(f"Ref: {target}")
             logger.info(f"Hyp: {result}")
-            logger.info("Current error rate [%s] = %f" %
+            logger.info(
-                        (cfg.error_rate_type, error_rate_func(target, result)))
+                "Current error rate [%s] = %f" %
+                (decode_cfg.error_rate_type, error_rate_func(target, result)))
 
         return dict(
             errors_sum=errors_sum,
             len_refs=len_refs,
             num_ins=num_ins,
             error_rate=errors_sum / len_refs,
-            error_rate_type=cfg.error_rate_type)
+            error_rate_type=decode_cfg.error_rate_type)
 
-    def compute_result_transcripts(self, audio, audio_len, vocab_list, cfg):
+    def compute_result_transcripts(self, audio, audio_len, vocab_list,
+                                   decode_cfg):
         result_transcripts = self.model.decode(
             audio,
             audio_len,
             vocab_list,
-            decoding_method=cfg.decoding_method,
+            decoding_method=decode_cfg.decoding_method,
-            lang_model_path=cfg.lang_model_path,
+            lang_model_path=decode_cfg.lang_model_path,
-            beam_alpha=cfg.alpha,
+            beam_alpha=decode_cfg.alpha,
-            beam_beta=cfg.beta,
+            beam_beta=decode_cfg.beta,
-            beam_size=cfg.beam_size,
+            beam_size=decode_cfg.beam_size,
-            cutoff_prob=cfg.cutoff_prob,
+            cutoff_prob=decode_cfg.cutoff_prob,
-            cutoff_top_n=cfg.cutoff_top_n,
+            cutoff_top_n=decode_cfg.cutoff_top_n,
-            num_processes=cfg.num_proc_bsearch)
+            num_processes=decode_cfg.num_proc_bsearch)
 
         return result_transcripts
 
@@ -358,7 +357,6 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
     def test(self):
         logger.info(f"Test Total Examples: {len(self.test_loader.dataset)}")
         self.model.eval()
-        cfg = self.config
         error_rate_type = None
         errors_sum, len_refs, num_ins = 0.0, 0, 0
         with jsonlines.open(self.args.result_file, 'w') as fout:
@@ -412,11 +410,10 @@ class DeepSpeech2ExportTester(DeepSpeech2Tester):
         if self.args.enable_auto_log is True:
             from paddlespeech.s2t.utils.log import Autolog
             self.autolog = Autolog(
-                batch_size=self.config.decoding.batch_size,
+                batch_size=self.config.decode.decode_batch_size,
                 model_name="deepspeech2",
                 model_precision="fp32").getlog()
         self.model.eval()
-        cfg = self.config
         error_rate_type = None
         errors_sum, len_refs, num_ins = 0.0, 0, 0
         with jsonlines.open(self.args.result_file, 'w') as fout:
@@ -441,7 +438,8 @@ class DeepSpeech2ExportTester(DeepSpeech2Tester):
         if self.args.enable_auto_log is True:
             self.autolog.report()
 
-    def compute_result_transcripts(self, audio, audio_len, vocab_list, cfg):
+    def compute_result_transcripts(self, audio, audio_len, vocab_list,
+                                   decode_cfg):
         if self.args.model_type == "online":
             output_probs, output_lens = self.static_forward_online(audio,
                                                                    audio_len)
@@ -454,13 +452,15 @@ class DeepSpeech2ExportTester(DeepSpeech2Tester):
         self.predictor.clear_intermediate_tensor()
         self.predictor.try_shrink_memory()
 
-        self.model.decoder.init_decode(cfg.alpha, cfg.beta, cfg.lang_model_path,
+        self.model.decoder.init_decode(decode_cfg.alpha, decode_cfg.beta,
-                                       vocab_list, cfg.decoding_method)
+                                       decode_cfg.lang_model_path, vocab_list,
+                                       decode_cfg.decoding_method)
 
         result_transcripts = self.model.decoder.decode_probs(
-            output_probs, output_lens, vocab_list, cfg.decoding_method,
+            output_probs, output_lens, vocab_list, decode_cfg.decoding_method,
-            cfg.lang_model_path, cfg.alpha, cfg.beta, cfg.beam_size,
+            decode_cfg.lang_model_path, decode_cfg.alpha, decode_cfg.beta,
-            cfg.cutoff_prob, cfg.cutoff_top_n, cfg.num_proc_bsearch)
+            decode_cfg.beam_size, decode_cfg.cutoff_prob,
+            decode_cfg.cutoff_top_n, decode_cfg.num_proc_bsearch)
         #replace the <space> with ' '
         result_transcripts = [
             self._text_featurizer.detokenize(sentence)
@@ -531,12 +531,10 @@ class DeepSpeech2ExportTester(DeepSpeech2Tester):
             num_chunk = int(num_chunk)
 
             chunk_state_h_box = np.zeros(
-                (self.config.model.num_rnn_layers, 1,
+                (self.config.num_rnn_layers, 1, self.config.rnn_layer_size),
-                 self.config.model.rnn_layer_size),
                 dtype=x.dtype)
             chunk_state_c_box = np.zeros(
-                (self.config.model.num_rnn_layers, 1,
+                (self.config.num_rnn_layers, 1, self.config.rnn_layer_size),
-                 self.config.model.rnn_layer_size),
                 dtype=x.dtype)
 
             input_names = self.predictor.get_input_names()

paddlespeech/s2t/exps/u2/bin/alignment.py

@@ -43,9 +43,9 @@ if __name__ == "__main__":
     config = get_cfg_defaults()
     if args.config:
         config.merge_from_file(args.config)
-    if args.decode_config:
+    if args.decode_cfg:
         decode_confs = CfgNode(new_allowed=True)
-        decode_confs.merge_from_file(args.decode_config)
+        decode_confs.merge_from_file(args.decode_cfg)
         config.decode = decode_confs
     if args.opts:
         config.merge_from_list(args.opts)