accu update

egs/mustc/st/conf/aipa.yaml

@@ -14,7 +14,7 @@ ctc-mixup-consistent-weight: 0
 inter-ctc-mixup-consistent-weight: 0
 inter-ctc-mixup-consistent-weight: 0
 mixup-consistent-weight: 0
 mixup-consistent-weight: 0
 
 
-cal-mixup-loss: True
+mixup-no-hard-loss: False
 no-specaugment: False
 no-specaugment: False
 layer-out-norm: False
 layer-out-norm: False
 
 

egs/mustc/st/conf/aipa_kd.yaml

@@ -14,7 +14,7 @@ ctc-mixup-consistent-weight: 0.15
 inter-ctc-mixup-consistent-weight: 0.1
 inter-ctc-mixup-consistent-weight: 0.1
 mixup-consistent-weight: 0.5
 mixup-consistent-weight: 0.5
 
 
-cal-mixup-loss: True
+mixup-no-hard-loss: False
 no-specaugment: False
 no-specaugment: False
 layer-out-norm: False
 layer-out-norm: False
 
 

egs/mustc/st/conf/base.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/conf/basis.yaml

@@ -25,5 +25,6 @@ num-workers: 8
 no-progress-bar: True
 no-progress-bar: True
 log-interval: 100
 log-interval: 100
 seed: 1
 seed: 1
+label-smoothing: 0.1
 report-accuracy: True
 report-accuracy: True
 skip-invalid-size-inputs-valid-test: True
 skip-invalid-size-inputs-valid-test: True
\ No newline at end of file

egs/mustc/st/conf/big.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/conf/ctc_aug_base.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/ctc_aug_big.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/ctc_aug_pds_big.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/dual.yaml

@@ -26,10 +26,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: join_speech_and_text_loss
 criterion: join_speech_and_text_loss
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/ipa.yaml

@@ -13,7 +13,7 @@ ctc-mixup-consistent-weight: 0
 inter-ctc-mixup-consistent-weight: 0
 inter-ctc-mixup-consistent-weight: 0
 mixup-consistent-weight: 0
 mixup-consistent-weight: 0
 
 
-cal-mixup-loss: True
+mixup-no-hard-loss: False
 no-specaugment: False
 no-specaugment: False
 layer-out-norm: False
 layer-out-norm: False
 
 

egs/mustc/st/conf/multibranch.yaml

@@ -45,10 +45,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 subsampling-type: conv1d
 subsampling-type: conv1d
 subsampling-layers: 2
 subsampling-layers: 2

egs/mustc/st/conf/nast_base.yaml

@@ -13,7 +13,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/nast_big.yaml

@@ -13,7 +13,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/nast_pds_big.yaml

@@ -13,7 +13,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/pds_base_16.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_base_32.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_base_4_444.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_base_8.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_base_8_444.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_big_16.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_big_32.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_big_8.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_big_deep_16.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_big_deep_32.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_big_deep_8.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_deep_16.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_deep_32.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/pds_deep_8.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/mustc/st/conf/reproduction_aipa_bilctc_kd.yaml

+# Append-based Interpolation Augmentation
+inter-mixup: True
+
+inter-mixup-layer: -1
+inter-mixup-decoder-layer: 0
+inter-mixup-prob: 1.0
+inter-mixup-ratio: 1.0
+inter-mixup-beta: 0.2
+
+inter-mixup-keep-org: True
+inter-mixup-decoder-emb: True
+
+mixup-no-hard-loss: False
+no-specaugment: False
+layer-out-norm: False
+
+inter-mixup-ratio-decay: False
+inter-mixup-ratio-decay-params: 20000,40000,0
+
+# Bilingual CTC
+share-ctc-and-embed: True
+share-xctc-and-embed: True
+ctc-weight: 0.2
+xctc-weight: 0.1
+
+# InterCTC
+inter-ctc-weight: 0.1
+inter-ctc-layers: 6,9
+share-inter-ctc: True
+inter-xctc-weight: 0.05
+inter-xctc-layers: 6,9
+
+# Prediction-aware encoding
+ctc-pae: inter_league
+xctc-pae: inter_league
+pae-unnorm-input: True
+
+ctc-mixup-consistent-weight: 0.1
+inter-ctc-mixup-consistent-weight: 0.05
+xctc-mixup-consistent-weight: 0.05
+inter-xctc-mixup-consistent-weight: 0.25
+mixup-consistent-weight: 0.5
+
+# Conformer
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+layer-padding-mask: True
\ No newline at end of file

egs/mustc/st/conf/reproduction_aipa_bilctc_kd_nohardloss.yaml

+# Append-based Interpolation Augmentation
+inter-mixup: True
+
+inter-mixup-layer: -1
+inter-mixup-decoder-layer: 0
+inter-mixup-prob: 1.0
+inter-mixup-ratio: 1.0
+inter-mixup-beta: 0.2
+
+inter-mixup-keep-org: True
+inter-mixup-decoder-emb: True
+
+mixup-no-hard-loss: True
+no-specaugment: False
+layer-out-norm: False
+
+inter-mixup-ratio-decay: False
+inter-mixup-ratio-decay-params: 20000,40000,0
+
+# Bilingual CTC
+share-ctc-and-embed: True
+share-xctc-and-embed: True
+ctc-weight: 0.2
+xctc-weight: 0.1
+
+# InterCTC
+inter-ctc-weight: 0.1
+inter-ctc-layers: 6,9
+share-inter-ctc: True
+inter-xctc-weight: 0.05
+inter-xctc-layers: 6,9
+
+# Prediction-aware encoding
+ctc-pae: inter_league
+xctc-pae: inter_league
+pae-unnorm-input: True
+
+ctc-mixup-consistent-weight: 0.1
+inter-ctc-mixup-consistent-weight: 0.05
+xctc-mixup-consistent-weight: 0.05
+inter-xctc-mixup-consistent-weight: 0.25
+mixup-consistent-weight: 0.5
+
+# Conformer
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+layer-padding-mask: True
\ No newline at end of file

egs/mustc/st/conf/reproduction_aipa_kd.yaml

@@ -10,7 +10,7 @@ inter-mixup-beta: 0.2
 inter-mixup-keep-org: True
 inter-mixup-keep-org: True
 inter-mixup-decoder-emb: True
 inter-mixup-decoder-emb: True
 
 
-cal-mixup-loss: True
+mixup-no-hard-loss: False
 no-specaugment: False
 no-specaugment: False
 layer-out-norm: False
 layer-out-norm: False
 
 

egs/mustc/st/conf/reproduction_aipa_kd_womixuploss.yaml → egs/mustc/st/conf/reproduction_aipa_kd_nohardloss.yaml

@@ -10,7 +10,7 @@ inter-mixup-beta: 0.2
 inter-mixup-keep-org: True
 inter-mixup-keep-org: True
 inter-mixup-decoder-emb: True
 inter-mixup-decoder-emb: True
 
 
-cal-mixup-loss: False
+mixup-no-hard-loss: True
 no-specaugment: False
 no-specaugment: False
 layer-out-norm: False
 layer-out-norm: False
 
 

egs/mustc/st/conf/reproduction_bil_ctc_progressive.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/reproduction_bil_ctc_progressive2.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/conf/reproduction_bil_ctc_synchronous.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/conf/reproduction_ctc_aug.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/reproduction_nast.yaml

@@ -15,7 +15,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/reproduction_nast_bilctc_syn.yaml

+arch: s2t_ctc
+encoder-type: transformer
+
+criterion: ctc
+zero_infinity: True
+xctc-weight: 1.0
+ctc-weight: 1.0
+share-ctc-and-embed: True 
+share-xctc-and-embed: True
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 1e-3
+adam-betas: (0.9,0.98)
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 2048
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.15
+activation-fn: relu
+encoder-embed-dim: 512
+encoder-ffn-embed-dim: 2048
+encoder-layers: 18
+encoder-attention-heads: 8
+
+# Conformer 
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+
+# InterCTC
+share-inter-ctc: True
+inter-ctc-weight: 1.0
+inter-ctc-layers: 6,9,12,15
+inter-xctc-weight: 1.0
+inter-xctc-layers: 6,9,12,15
+
+# Prediction-aware encoding
+ctc-pae: inter_league
+xctc-pae: inter_league
+pae-unnorm-input: True
+
+# Cross-layer attn
+# xctc-cross-attn: True
+# cross-attn-start-layer: 4
+# cross-attn-layer: 3
+# cross-attn-collaboration-mode: serial
+# cross-attn-league-drop-net: True
+# cross-attn-league-drop-net-prob: 0.1
+
+# Curriculum learning mixing
+xctc-pae-ground-truth-ratio: 0.8
+xctc-pae-ground-truth-only-mistake: True 
+pae-oracle-smooth: True

egs/mustc/st/conf/reproduction_sate.yaml

@@ -6,12 +6,11 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 ctc-weight: 0.3
 ctc-weight: 0.3
 share-ctc-and-embed: True
 share-ctc-and-embed: True
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/sate.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/sate_big.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/sate_big_pds.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/conf/sate_deep.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-normalize-before: True
 encoder-normalize-before: True
 decoder-normalize-before: True
 decoder-normalize-before: True

egs/mustc/st/conf/sate_pds_16.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/conf/sate_pds_8.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/conf/sate_pds_8_444.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/conf/sate_pds_base_8.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/conf/sate_pds_deep_8.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/conf/w2v2.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/mustc/st/decode.sh

@@ -11,7 +11,7 @@ if [ "$#" -eq 1 ]; then
 fi
 fi
 
 
 sacrebleu=1
 sacrebleu=1
-ctc_infer=0
+ctc_infer=1
 n_average=10
 n_average=10
 beam_size=5
 beam_size=5
 infer_ctc_weight=0
 infer_ctc_weight=0

egs/mustc/st/run.sh

@@ -354,8 +354,8 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     mv tmp.log $log
     mv tmp.log $log
 
 
     log=${model_dir}/train.log
     log=${model_dir}/train.log
-     # cmd="${cmd} 2>&1 | tee -a ${log}"
-    cmd="${cmd} >> ${log} 2>&1 "
+    cmd="${cmd} 2>&1 | tee -a ${log}"
+    #cmd="${cmd} >> ${log} 2>&1 "
     if [[ $eval -eq 1 ]]; then
     if [[ $eval -eq 1 ]]; then
         # tensorboard
         # tensorboard
         port=6666
         port=6666

egs/talcs/asr/binary.sh

+set -e
+
+eval=1
+
+lcrm=0
+tokenizer=0
+
+root_dir=~/st/Fairseq-S2T
+data_dir=~/st/data/test
+vocab_dir=~/st/data/mustc/st/en-de
+asr_vocab_prefix=spm_unigram10000_st_share
+
+src_lang=en
+tgt_lang=de
+subsets=(2019)
+
+cp -r ${vocab_dir}/${asr_vocab_prefix}.* ${data_dir}/${src_lang}-${tgt_lang}
+rm -rf ${data_dir}/${src_lang}-${tgt_lang}/fbank80.zip
+
+splits=$(echo ${subsets[*]} | sed 's/ /,/g')
+cmd="python ${root_dir}/examples/speech_to_text/prep_st_data.py
+    --data-root ${data_dir}
+    --output-root ${data_dir}
+    --splits ${splits}
+    --task asr
+    --src-lang ${src_lang}
+    --tgt-lang ${tgt_lang}
+    --add-src
+    --share
+    --asr-prefix ${asr_vocab_prefix}
+    --cmvn-type utterance"
+
+    if [[ ${lcrm} -eq 1 ]]; then
+        cmd="$cmd
+    --lowercase-src
+    --rm-punc-src"
+    fi
+    if [[ ${tokenizer} -eq 1 ]]; then
+        cmd="$cmd
+    --tokenizer"
+    fi
+
+echo -e "\033[34mRun command: \n${cmd} \033[0m"
+[[ $eval -eq 1 ]] && eval ${cmd}

egs/talcs/asr/conf/base.yaml

+arch: s2t_transformer_s
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/talcs/asr/conf/base_nonorm.yaml

+arch: s2t_transformer_s
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/talcs/asr/conf/basis.yaml

+train-subset: train
+valid-subset: dev
+
+max-epoch: 100
+max-update: 100000
+patience: 20
+post-process: sentencepiece
+
+# best-checkpoint-metric: loss
+# maximize-best-checkpoint-metric: False
+
+eval-wer: True
+eval-wer-args: {"beam": 1, "lenpen": 1.0}
+eval-wer-tok-args: {"wer_remove_punct": true, "wer_lowercase": true, "wer_char_level": true}
+eval-wer-remove-bpe: sentencepiece
+eval-wer-print-samples: True
+best_checkpoint_metric: dec_wer
+maximize_best_checkpoint_metric: False
+
+no-epoch-checkpoints: True
+# keep-last-epochs: 10
+keep-best-checkpoints: 10
+
+num-workers: 8
+no-progress-bar: True
+log-interval: 100
+seed: 1
+label-smoothing: 0.1
+report-accuracy: True
+skip-invalid-size-inputs-valid-test: True
\ No newline at end of file

egs/talcs/asr/conf/big.yaml

+arch: s2t_transformer_m
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 0.0014
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 2048
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.15
+activation-fn: relu
+encoder-embed-dim: 512
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
\ No newline at end of file

egs/talcs/asr/conf/big_wenet.yaml

+arch: s2t_transformer_m
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 1e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv2d
+subsampling-layers: 2
+subsampling-filter: 512
+subsampling-kernel: 3
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: relu
+
+dropout: 0.15
+activation-fn: relu
+encoder-embed-dim: 512
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+cnn-module-norm: layer_norm
+
+load-pretrained-encoder-from: /home/xuchen/after.pt
+load-pretrained-decoder-from: /home/xuchen/after.pt
+#load-pretrained-decoder-from:

egs/talcs/asr/conf/conformer.yaml

+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+layer-padding-mask: True
\ No newline at end of file

egs/talcs/asr/conf/ctc.yaml

+ctc-weight: 0.3
\ No newline at end of file

egs/talcs/asr/conf/dlcl.yaml

+use-enc-dlcl: True
+use-dec-dlcl: True

egs/talcs/asr/conf/dynamic.yaml

+compression-metric: threshold
+compression-mode: create
+compression-layers: 6,9
+compression-threshold: 0.95
+compression-norm: True
+compression-pos: True
\ No newline at end of file

egs/talcs/asr/conf/inter.yaml

+inter-ctc-weight: 0.2
+inter-ctc-layers: 6,9
+inter-ctc-drop-prob: 0
+share-inter-ctc: True
+
+ctc-pae: none
+# ctc-pae: inter_league
+
+# ctc-pae-ground-truth-ratio: 0.1
+# pae-gumbel: True
+# pae-distribution-hard: True
+# pae-drop-prob: 0.0
+# pae-distribution-cutoff: 10
+# share-pae-and-ctc: True
+# pae-embed-norm: True
+# pae-out-norm: True
+
+# ctc-self-distill-weight: 1
+# target-ctc-self-distill-weight: 1
+# ctc-self-distill-prob: 0.1
+# cal-all-ctc: True
\ No newline at end of file

egs/talcs/asr/conf/ipa.yaml

+inter-mixup: True
+inter-mixup-layer: -1
+inter-mixup-decoder-layer: 0
+inter-mixup-prob: 1.0
+inter-mixup-ratio: 1.0
+inter-mixup-beta: 0.2
+inter-mixup-keep-org: False
+inter-mixup-decoder-emb: False
+ctc-mixup-consistent-weight: 0
+inter-ctc-mixup-consistent-weight: 0
+mixup-consistent-weight: 0
+mixup-no-hard-loss: False
+no-specaugment: False
+layer-out-norm: False
+
+inter-mixup-ratio-decay: False
+inter-mixup-ratio-decay-params: 20000,40000,0
\ No newline at end of file

egs/talcs/asr/conf/mlo.yaml

+inter-ctc-mlo: 1:2:3
\ No newline at end of file

egs/talcs/asr/conf/norm.yaml

+encoder-embed-norm: True
+encoder-no-scale-embedding: True
\ No newline at end of file

egs/talcs/asr/conf/pds_base.yaml

+arch: pdss2t_transformer_s_8
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
\ No newline at end of file

egs/talcs/asr/conf/pds_base_16.yaml

+arch: pdss2t_transformer_s_16
+
+encoder-embed-dim: 256
+pds-stages: 4
+pds-layers: 2_2_6_2
+pds-ratios: 2_2_2_2
+pds-fusion: False
+pds-fusion-method: all_conv2
+pds-fusion-layers: 0_1_1_1 
+pds-fusion-weight: 0.2_0.3_0.5
+pds-embed-dims: 256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 8_8_8_8
+pds-attn-heads: 4_4_4_4
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
\ No newline at end of file

egs/talcs/asr/conf/pds_base_32.yaml

+arch: pdss2t_transformer_s_32
+
+encoder-embed-dim: 256
+pds-stages: 5
+pds-layers: 2_2_3_3_2
+pds-ratios: 2_2_2_2_2
+pds-fusion: False
+pds-fusion-method: all_conv2
+pds-fusion-layers: 0_0_1_1_1 
+pds-fusion-weight: 0.2_0.3_0.5
+pds-embed-dims: 256_256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1_1
+pds-kernel-sizes: 5_5_5_5_5
+pds-ffn-ratios: 8_8_8_8_8
+pds-attn-heads: 4_4_4_4_4
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
\ No newline at end of file

egs/talcs/asr/conf/pds_base_8.yaml

+arch: pdss2t_transformer_s_8
+
+encoder-embed-dim: 256
+pds-stages: 4
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: False
+pds-fusion-method: all_conv2
+pds-fusion-layers: 0_1_1_1 
+pds-fusion-weight: 0.2_0.3_0.5
+pds-embed-dims: 256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 8_8_8_8
+pds-attn-heads: 4_4_4_4
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
\ No newline at end of file

egs/talcs/asr/conf/pds_big_8.yaml

+arch: pdss2t_transformer_m_8
+
+encoder-embed-dim: 512
+pds-stages: 4
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: False
+pds-fusion-method: all_conv2
+pds-fusion-layers: 0_1_1_1 
+pds-fusion-weight: 0.2_0.3_0.5
+pds-embed-dims: 512_512_512_512
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 4_4_4_4
+pds-attn-heads: 8_8_8_8
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 0.0014
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+dropout: 0.15
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8

egs/talcs/asr/conf/purectc.yaml

+arch: s2t_ctc
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: ctc
+zero_infinity: True
+ctc-weight: 1.0
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 18
+encoder-attention-heads: 4
\ No newline at end of file

egs/talcs/asr/conf/rpr.yaml

+encoder-attention-type: rel_pos
+
+#encoder-attention-type: rel_pos_legacy
+#encoder-attention-type: rel_selfattn
+#encoder-attention-type: relative
+#decoder-attention-type: relative
+#max-encoder-relative-length: 100
+#max-decoder-relative-length: 20

egs/talcs/asr/conf/xctc.yaml

+xctc-weight: 0.3
+share-xctc-and-embed: True
\ No newline at end of file

egs/talcs/asr/conf/xinter.yaml

+inter-xctc-weight: 0.2
+inter-xctc-layers: 6,9
+
+xctc-pae: none
+# xctc-pae: inter_league
+
+xctc-cross-attn: False
+cross-attn-start-layer: 7
+cross-attn-layer: 6
+cross-attn-collaboration-mode: parallel
+cross-attn-league-s1-ratio: 0.5
+cross-attn-league-s2-ratio: 0.5
+cross-attn-league-out-norm: False
+cross-attn-league-gated: False
+cross-attn-league-drop-net: False
+cross-attn-league-drop-net-prob: 0.2
+cross-attn-league-drop-net-mix: False
+
+# xctc-pae-ground-truth-ratio: 0.1
+# xctc-pae-ground-truth-ratio-adaptive: True 
+# xctc-pae-ground-truth-only-mistake: True 
+# pae-oracle-smooth: True
+# pae-gumbel: True
+# pae-distribution-hard: True
+# pae-drop-prob: 0.0
+# pae-distribution-cutoff: 10
+# share-pae-and-xctc: True
+# pae-embed-norm: True
+# pae-out-norm: True
+
+# ctc-self-distill-weight: 1
+# target-ctc-self-distill-weight: 1
+# ctc-self-distill-prob: 0.1
+# cal-all-ctc: True
\ No newline at end of file

egs/talcs/asr/decode.sh

+#!/usr/bin/env bash
+
+gpu_num=1
+
+data_tag=asr
+test_subset=(dev test)
+
+exp_name=
+if [ "$#" -eq 1 ]; then
+    exp_name=$1
+fi
+
+cer=1
+ctc_infer=1
+n_average=10
+beam_size=5
+infer_ctc_weight=0.1
+len_penalty=1.0
+max_tokens=50000
+batch_size=1
+infer_debug=0
+dec_model=checkpoint_best.pt
+
+cmd="./run.sh
+    --stage 2
+    --stop_stage 2
+    --gpu_num ${gpu_num}
+    --exp_name ${exp_name}
+    --n_average ${n_average}
+    --cer ${cer}
+    --ctc_infer ${ctc_infer}
+    --beam_size ${beam_size}
+    --len_penalty ${len_penalty}
+    --batch_size ${batch_size}
+    --max_tokens ${max_tokens}
+    --dec_model ${dec_model}
+    --ctc_infer ${ctc_infer}
+    --infer_ctc_weight ${infer_ctc_weight}
+    --infer_debug ${infer_debug}
+    "
+
+if [[ -n ${data_tag} ]]; then
+    cmd="$cmd --data_tag ${data_tag}"
+fi
+if [[ ${#test_subset[@]} -ne 0 ]]; then
+    subsets=$(echo ${test_subset[*]} | sed 's/ /,/g')
+    cmd="$cmd --test_subset ${subsets}"
+fi
+
+echo $cmd
+eval $cmd

egs/talcs/asr/local/cal_bleu.sh

+#!/usr/bin/env bash
+
+set -e
+
+ref=$1
+gen=$2
+tokenizer=$3
+lang=$4
+lang_pair=en-${lang}
+
+record=$(mktemp -t temp.record.XXXXXX)
+if [[ ${tokenizer} -eq 1 ]]; then
+    echo "MultiBLEU" > ${record}
+    cmd="multi-bleu.perl ${ref} < ${gen}"
+    eval $cmd | head -n 1 >> ${record}
+
+    cmd="detokenizer.perl -q -l ${lang} --threads 32 < ${ref} > ${ref}.detok"
+    eval $cmd
+    cmd="detokenizer.perl -q -l ${lang} --threads 32 < ${gen} > ${gen}.detok"
+    eval $cmd
+    ref=${ref}.detok
+    gen=${gen}.detok
+fi
+
+echo "SacreBLEU" >> ${record}
+cmd="cat ${gen} | sacrebleu ${ref} -m bleu -w 4 -l ${lang_pair}"
+eval $cmd >> ${record}
+cat ${record}
+rm ${record}
\ No newline at end of file

egs/talcs/asr/local/cal_ctc_bleu.sh

+#!/usr/bin/env bash
+
+set -e
+
+infer_dir=$1
+tag=$2
+s2s_infer_file=${infer_dir}/$3
+org_ctc_infer_file=${infer_dir}/$4
+ref=$5
+tokenizer=$6
+lang=$7
+
+idx=${infer_dir}/${tag}_idx
+ctc_infer=${infer_dir}/${tag}_ctc_infer
+ctc_infer_sort=${infer_dir}/${tag}_ctc_infer_sort
+
+if [[ ! -f ${ctc_infer_sort} ]]; then
+    cut -f1 ${s2s_infer_file} > ${idx}
+    paste ${idx} ${org_ctc_infer_file} > ${ctc_infer}
+    sort -n -t $'\t' ${ctc_infer} | cut -f2 > ${ctc_infer_sort}
+fi
+
+gen=${ctc_infer_sort}
+./cal_bleu.sh ${ref} ${gen} ${tokenizer} ${lang}
\ No newline at end of file

egs/talcs/asr/local/cal_wer.py

+import unicodedata
+import re
+import jiwer
+import jiwer.transforms as tr
+import sys
+
+ref_file = sys.argv[1]
+hyp_file = sys.argv[2]
+
+
+wer_standardize = tr.Compose(
+    [
+        tr.SubstituteRegexes({r"<<unk>>": r"@"}),
+        tr.ToLowerCase(),
+        tr.RemovePunctuation(),
+        tr.ExpandCommonEnglishContractions(),
+        tr.RemoveKaldiNonWords(),
+        tr.RemoveWhiteSpace(replace_by_space=True),
+        tr.ReduceToListOfListOfWords(),
+    ]
+)
+cer_standardize = tr.Compose(
+    [
+        tr.SubstituteRegexes({r"<<unk>>": r"@"}),
+        tr.ToLowerCase(),
+        tr.RemovePunctuation(),
+        tr.Strip(),
+        tr.ReduceToListOfListOfChars(),
+    ]
+)
+
+def process_text(text):
+    # 将中文字符和英文字符间加空格
+    text = re.sub(r'([\u4e00-\u9fa5])([a-zA-Z0-9])', r'\1 \2', text)
+    text = re.sub(r'([a-zA-Z0-9])([\u4e00-\u9fa5])', r'\1 \2', text)
+
+    # 将中文字符间加空格
+    text = re.sub(r'([\u4e00-\u9fa5])', r'\1 ', text)
+
+    # 去掉多余的空格
+    text = re.sub(r'\s+', ' ', text).strip()
+
+    return text
+
+ref_lines = open(ref_file, "r").readlines()
+hyp_lines = open(hyp_file, "r").readlines()
+ref_lines = [process_text(line) for line in ref_lines]
+hyp_lines = [process_text(line) for line in hyp_lines]
+
+print(hyp_lines[:10])
+
+wer = jiwer.wer(ref_lines, hyp_lines,
+                truth_transform=wer_standardize,
+                hypothesis_transform=wer_standardize,
+                )
+cer = jiwer.cer(ref_lines, hyp_lines,
+                truth_transform=cer_standardize,
+                hypothesis_transform=cer_standardize,
+                )
+
+print("WER: %.4f" % wer)
+print("CER: %.4f" % cer)

egs/talcs/asr/local/cal_wer.sh

+#!/usr/bin/env bash
+
+set -e
+
+infer_dir=$1
+tag=$2
+s2s_infer_file=${infer_dir}/$3
+org_ctc_infer_file=${infer_dir}/$4
+ref=$5
+
+idx=${infer_dir}/${tag}_idx
+ctc_infer=${infer_dir}/${tag}_ctc_infer
+ctc_infer_sort=${infer_dir}/${tag}_ctc_infer_sort
+
+cut -f1 ${s2s_infer_file} > ${idx}
+paste ${idx} ${org_ctc_infer_file} > ${ctc_infer}
+sort -n -t $'\t' ${ctc_infer} | cut -f2 > ${ctc_infer_sort}
+python3 ./cal_wer.py ${ref} ${ctc_infer_sort}
\ No newline at end of file

egs/talcs/asr/local/extract_txt_from_tsv.py

+import sys
+import csv
+
+tsv_file = sys.argv[1]
+out_file = sys.argv[2]
+extract_item = sys.argv[3]
+
+with open(tsv_file) as f:
+    reader = csv.DictReader(
+        f,
+        delimiter="\t",
+        quotechar=None,
+        doublequote=False,
+        lineterminator="\n",
+        quoting=csv.QUOTE_NONE,
+    )
+    samples = [dict(e) for e in reader]
+
+fw = open(out_file, "w", encoding="utf-8")
+for s in samples:
+    if extract_item in s:
+        fw.write("%s\n" % s[extract_item])
+    else:
+        print("Error in sample: ")
+        print(s)
+        exit()
+

egs/talcs/asr/local/monitor.sh

+#!/usr/bin/env bash
+
+gpu_num=4
+cmd="sh train.sh"
+
+while :
+do
+    record=$(mktemp -t temp.record.XXXXXX)
+    gpustat > $record
+    all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
+
+    count=0
+    for dev in ${all_devices[@]}
+    do
+        line=$((dev + 2))
+        use=$(head -n $line ${record} | tail -1 | cut -d '|' -f3 | cut -d '/' -f1)
+
+        if [[ $use -lt 100 ]]; then
+            device[$count]=$dev
+            count=$((count + 1))
+            if [[ $count -eq $gpu_num ]]; then
+                break
+            fi
+        fi
+    done
+    if [[ ${#device[@]} -lt $gpu_num ]]; then
+        sleep 60s
+    else
+        echo "Run $cmd"
+        eval $cmd
+        sleep 10s
+        exit
+    fi
+done

egs/talcs/asr/local/parse_options.sh

+#!/usr/bin/env bash
+
+# Copyright 2012  Johns Hopkins University (Author: Daniel Povey);
+#                 Arnab Ghoshal, Karel Vesely
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Parse command-line options.
+# To be sourced by another script (as in ". parse_options.sh").
+# Option format is: --option-name arg
+# and shell variable "option_name" gets set to value "arg."
+# The exception is --help, which takes no arguments, but prints the
+# $help_message variable (if defined).
+
+
+###
+### The --config file options have lower priority to command line
+### options, so we need to import them first...
+###
+
+# Now import all the configs specified by command-line, in left-to-right order
+for ((argpos=1; argpos<$#; argpos++)); do
+  if [ "${!argpos}" == "--config" ]; then
+    argpos_plus1=$((argpos+1))
+    config=${!argpos_plus1}
+    [ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
+    . $config  # source the config file.
+  fi
+done
+
+
+###
+### Now we process the command line options
+###
+i=1
+argv="$@"
+while true; do
+  key=${!i}
+  j=$(($i + 1))
+  value=${!j}
+
+  [ -z "${!i:-}" ] && break;  # break if there are no arguments
+  case "${key}" in
+    # If the enclosing script is called with --help option, print the help
+    # message and exit.  Scripts should put help messages in $help_message
+    --help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2;
+      else printf "$help_message\n" 1>&2 ; fi;
+      exit 0 ;;
+    --*=*) echo "$0: options to scripts must be of the form --name value, got '${key}}'"
+      exit 1 ;;
+    # If the first command-line argument begins with "--" (e.g. --foo-bar),
+    # then work out the variable name as $name, which will equal "foo_bar".
+    --*) name=`echo "${key}" | sed s/^--// | sed s/-/_/g`;
+      # Next we test whether the variable in question is undefned-- if so it's
+      # an invalid option and we die.  Note: $0 evaluates to the name of the
+      # enclosing script.
+      # The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar
+      # is undefined.  We then have to wrap this test inside "eval" because
+      # foo_bar is itself inside a variable ($name).
+      #eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
+
+      oldval="`eval echo \\$$name`";
+      # Work out whether we seem to be expecting a Boolean argument.
+      if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then
+        was_bool=true;
+      else
+        was_bool=false;
+      fi
+
+      # Set the variable to the right value-- the escaped quotes make it work if
+      # the option had spaces, like --cmd "queue.pl -sync y"
+      # echo $name
+      eval $name=\"${value}\";
+
+      # Check that Boolean-valued arguments are really Boolean.
+      if $was_bool && [[ "${value}" != "true" && "${value}" != "false" ]]; then
+        echo "$0: expected \"true\" or \"false\": ${key} ${value}" 1>&2
+        exit 1;
+      fi
+      # shift 2;
+      i=$(($i + 2))
+      ;;
+  *) break;
+  esac
+done
+
+
+# Check for an empty argument to the --cmd option, which can easily occur as a
+# result of scripting errors.
+[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
+
+
+true; # so this script returns exit code 0.

egs/talcs/asr/local/utils.sh

+
+get_devices(){
+    gpu_num=$1
+    use_cpu=$2
+    device=()
+    while :
+    do
+        record=$(mktemp -t temp.record.XXXXXX)
+        gpustat > $record
+        all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
+
+        count=0
+        for dev in ${all_devices[@]}
+        do
+            line=$((dev + 2))
+            use=$(head -n $line ${record} | tail -1 | cut -d '|' -f3 | cut -d '/' -f1)
+            if [[ $use -lt 1000 ]]; then
+                device[$count]=$dev
+                count=$((count + 1))
+                if [[ $count -eq $gpu_num ]]; then
+                    break
+                fi
+            fi
+        done
+        if [[ ${#device[@]} -lt $gpu_num ]]; then
+            if [[ $use_cpu -eq 1 ]]; then
+                device=(-1)
+            else
+                sleep 60s
+            fi
+        else
+            break
+        fi
+    done
+
+    echo ${device[*]} | sed 's/ /,/g'
+    return $?
+}
+
+

egs/talcs/asr/run.sh

+#!/usr/bin/env bash
+
+# Processing AIShell ASR Datasets
+
+# Copyright 2021 Chen Xu (xuchennlp@outlook.com)
+
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+#set -u
+set -o pipefail
+export PYTHONIOENCODING=UTF-8
+
+eval=1
+time=$(date "+%m%d_%H%M")
+
+stage=1
+stop_stage=2
+
+######## Hardware ########
+# Devices
+device=(0)
+gpu_num=2
+update_freq=1
+max_tokens=100000
+
+pwd_dir=$PWD
+root_dir=${ST_ROOT}
+data_root_dir=${root_dir}
+
+code_dir=${root_dir}/S2T
+
+# Dataset
+src_lang=zh_en.tok
+lang=${src_lang}
+dataset=talcs
+data_tag=asr
+
+task=speech_to_text
+vocab_type=unigram
+vocab_size=10000
+speed_perturb=0
+lcrm=0
+tokenizer=0
+use_raw_audio=0
+
+. ./local/parse_options.sh || exit 1;
+
+use_specific_dict=0
+specific_prefix=st
+specific_dir=${root_dir}/data/mustc/st
+asr_vocab_prefix=spm_unigram10000_st_share
+
+data_model_subfix=${dataset}/${data_tag}
+org_data_dir=${data_root_dir}/data/${dataset}
+data_dir=${data_root_dir}/data/${data_model_subfix}
+train_split=train
+valid_split=dev
+test_split=test
+test_subset=dev,test
+
+# exp
+sub_tag=
+exp_prefix=$(date "+%m%d")
+extra_tag=
+extra_parameter=
+exp_tag=baseline
+exp_name=
+
+# Training Settings
+train_config=base,ctc
+fp16=1
+step_valid=0
+
+# Decoding Settings
+dec_model=checkpoint_best.pt
+cer=1
+ctc_infer=0
+infer_ctc_weight=0
+ctc_self_ensemble=0
+ctc_inter_logit=0
+n_average=10
+batch_size=0
+beam_size=5
+len_penalty=1.0
+single=0
+epoch_ensemble=0
+best_ensemble=1
+infer_debug=0
+infer_score=0
+# infer_parameters="--cal-monotonic-cross-attn-weights --cal-localness --localness-window 0.1 --cal-topk-cross-attn-weights --topk-cross-attn-weights 15 --cal-entropy"
+
+data_config=config.yaml
+
+# Parsing Options
+if [[ ${speed_perturb} -eq 1 ]]; then
+    data_dir=${data_dir}_sp
+    exp_prefix=${exp_prefix}_sp
+fi
+if [[ ${lcrm} -eq 1 ]]; then
+    data_dir=${data_dir}_lcrm
+    exp_prefix=${exp_prefix}_lcrm
+fi
+if [[ ${use_specific_dict} -eq 1 ]]; then
+    data_dir=${data_dir}_${specific_prefix}
+    exp_prefix=${exp_prefix}_${specific_prefix}
+fi
+if [[ ${tokenizer} -eq 1 ]]; then
+    data_dir=${data_dir}_tok
+    exp_prefix=${exp_prefix}_tok
+fi
+if [[ ${use_raw_audio} -eq 1 ]]; then
+    data_dir=${data_dir}_raw
+    exp_prefix=${exp_prefix}_raw
+fi
+if [[ "${vocab_type}" == "char" ]]; then
+    data_dir=${data_dir}_char
+    exp_prefix=${exp_prefix}_char
+fi
+
+# setup nccl envs
+export NCCL_IB_DISABLE=0
+export NCCL_IB_HCA=$ARNOLD_RDMA_DEVICE:1
+export NCCL_IB_GID_INDEX=3
+export NCCL_SOCKET_IFNAME=eth0
+
+export PATH=$PATH:${code_dir}/scripts
+. ./local/parse_options.sh || exit 1;
+
+if [[ -z ${exp_name} ]]; then
+    config_string=${train_config//,/_}
+    exp_name=${exp_prefix}_${config_string}_${exp_tag}
+    if [[ -n ${extra_tag} ]]; then
+        exp_name=${exp_name}_${extra_tag}
+    fi
+    if [[ -n ${exp_subfix} ]]; then
+        exp_name=${exp_name}_${exp_subfix}
+    fi
+fi
+
+ckpt_dir=${root_dir}/checkpoints/
+model_dir=${root_dir}/checkpoints/${data_model_subfix}/${sub_tag}/${exp_name}
+
+# Start
+cd ${code_dir}
+echo "Start Stage: $stage"
+echo "Stop  Stage: $stop_stage"
+
+if [[ `pip list | grep fairseq | wc -l` -eq 0 ]]; then 
+    echo "Default Stage: env configure"
+    pip3 install -e ${code_dir}
+fi
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+    echo "Stage -1: Data Download"
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    ### Task dependent. You have to make data the following preparation part by yourself.
+    echo "Stage 0: Data Preparation"
+
+    if [[ ! -e ${data_dir} ]]; then
+        mkdir -p ${data_dir}
+    fi
+
+    cmd="python3 ${code_dir}/examples/speech_to_text/prep_audio_data.py
+        --data-root ${org_data_dir}
+        --output-root ${data_dir}
+        --task asr
+        --src-lang ${src_lang}
+        --splits ${valid_split},${test_split},${train_split}
+	    --add-src
+	    --share
+        --vocab-type ${vocab_type}
+        --vocab-size ${vocab_size}"
+
+    if [[ ${use_raw_audio} -eq 1 ]]; then
+        cmd="$cmd
+        --raw"
+    fi
+    if [[ ${use_specific_dict} -eq 1 ]]; then
+        cp -r ${specific_dir}/${asr_vocab_prefix}.* ${data_dir}
+        cmd="$cmd
+        --asr-prefix ${asr_vocab_prefix}"
+    fi
+    if [[ ${speed_perturb} -eq 1 ]]; then
+        cmd="$cmd
+        --speed-perturb"
+    fi
+    if [[ ${lcrm} -eq 1 ]]; then
+        cmd="$cmd
+        --lowercase-src
+        --rm-punc-src"
+    fi
+    if [[ ${tokenizer} -eq 1 ]]; then
+        cmd="$cmd
+        --tokenizer"
+    fi
+
+    echo -e "\033[34mRun command: \n${cmd} \033[0m"
+    [[ $eval -eq 1 ]] && eval ${cmd}
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    echo "Stage 1: Network Training"
+    [[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1;
+
+    if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
+        if [[ ${gpu_num} -eq 0 ]]; then
+            device=""
+        else
+            source ./local/utils.sh
+            device=$(get_devices $gpu_num 0)
+        fi
+        export CUDA_VISIBLE_DEVICES=${device}
+    fi
+
+    echo -e "data=${data_dir} model=${model_dir}"
+
+    if [[ ! -d ${model_dir} ]]; then
+        mkdir -p ${model_dir}
+    else
+        echo "${model_dir} exists."
+    fi
+
+    cp -f ${pwd_dir}/`basename ${BASH_SOURCE[0]}` ${model_dir}
+    cp -f ${pwd_dir}/train.sh ${model_dir}
+
+    extra_parameter="${extra_parameter}
+        --train-config ${pwd_dir}/conf/basis.yaml"
+    cp -f ${pwd_dir}/conf/basis.yaml ${model_dir}
+    config_list="${train_config//,/ }"
+    idx=1
+    for config in ${config_list[@]}
+    do
+        config_path=${pwd_dir}/conf/${config}.yaml
+        if [[ ! -f ${config_path} ]]; then
+            echo "No config file ${config_path}"
+            exit
+        fi
+        cp -f ${config_path} ${model_dir}
+
+        extra_parameter="${extra_parameter}
+        --train-config${idx} ${config_path}"
+        idx=$((idx + 1))
+    done
+
+    cmd="python3 -u ${code_dir}/fairseq_cli/train.py
+        ${data_dir}
+        --config-yaml ${data_config}
+        --task ${task}
+        --max-tokens ${max_tokens}
+        --skip-invalid-size-inputs-valid-test
+        --update-freq ${update_freq}
+        --log-interval 100
+        --save-dir ${model_dir}
+        --tensorboard-logdir ${model_dir}"
+
+    if [[ -n ${extra_parameter} ]]; then
+        cmd="${cmd}
+        ${extra_parameter}"
+    fi
+    if [[ ${gpu_num} -gt 0 ]]; then
+        cmd="${cmd}
+        --distributed-world-size $gpu_num
+        --ddp-backend no_c10d"
+    fi
+    if [[ $fp16 -eq 1 ]]; then
+        cmd="${cmd}
+        --fp16"
+    fi
+    if [[ $step_valid -eq 1 ]]; then
+        validate_interval=1
+        save_interval=1
+        no_epoch_checkpoints=0
+        save_interval_updates=500
+        keep_interval_updates=10
+    fi
+    if [[ -n $no_epoch_checkpoints && $no_epoch_checkpoints -eq 1 ]]; then
+        cmd="$cmd
+        --no-epoch-checkpoints"
+    fi
+    if [[ -n $validate_interval ]]; then
+        cmd="${cmd}
+        --validate-interval $validate_interval "
+    fi
+    if [[ -n $save_interval ]]; then
+        cmd="${cmd}
+        --save-interval $save_interval "
+    fi
+    if [[ -n $save_interval_updates ]]; then
+        cmd="${cmd}
+        --save-interval-updates $save_interval_updates"
+        if [[ -n $keep_interval_updates ]]; then
+        cmd="${cmd}
+        --keep-interval-updates $keep_interval_updates"
+        fi
+    fi
+
+    echo -e "\033[34mRun command: \n${cmd} \033[0m"
+
+    # save info
+    log=${ckpt_dir}/history.log
+    echo "${time} | ${data_dir} | ${exp_name} | ${model_dir} " >> $log
+    tail -n 50 ${log} > tmp.log
+    mv tmp.log $log
+
+    log=${model_dir}/train.log
+    cmd="${cmd} 2>&1 | tee -a ${log}"
+    #cmd="${cmd} >> ${log} 2>&1 "
+    if [[ $eval -eq 1 ]]; then
+        # tensorboard
+        port=6666
+        tensorboard --logdir ${model_dir} --port ${port} --bind_all &
+    
+        echo "${cmd}" > ${model_dir}/cmd
+        eval $cmd
+        #sleep 2s
+        #tail -n "$(wc -l ${log} | awk '{print $1+1}')" -f ${log}
+    fi
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    echo "Stage 2: Decoding"
+    dec_models=
+    if [[ ${n_average} -eq 1 ]]; then
+        dec_models=${dec_model}
+    fi
+    if [[ ${n_average} -ne 1 ]]; then
+        # Average models
+        if [[ ${epoch_ensemble} -eq 1 ]]; then
+            avg_model=avg_epoch${n_average}_checkpoint.pt
+
+            if [[ ! -f ${model_dir}/${avg_model} ]]; then
+                cmd="python3 ${code_dir}/scripts/average_checkpoints.py
+                --inputs ${model_dir}
+                --num-epoch-checkpoints ${n_average}
+                --output ${model_dir}/${avg_model}"
+                echo -e "\033[34mRun command: \n${cmd} \033[0m"
+                [[ $eval -eq 1 ]] && eval $cmd
+            fi
+            dec_models+=(${avg_model})
+        fi
+        if [[ ${best_ensemble} -eq 1 ]]; then
+            avg_model=avg_best${n_average}_checkpoint.pt
+
+            if [[ ! -f ${model_dir}/${avg_model} ]]; then
+                cmd="python3 ${code_dir}/scripts/average_checkpoints.py
+                --inputs ${model_dir}
+                --num-best-checkpoints ${n_average}
+                --output ${model_dir}/${avg_model}"
+                echo -e "\033[34mRun command: \n${cmd} \033[0m"
+                [[ $eval -eq 1 ]] && eval $cmd
+            fi
+            dec_models+=(${avg_model})
+        fi
+    fi
+
+    if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
+        if [[ ${gpu_num} -eq 0 ]]; then
+            device=""
+        else
+            source ./local/utils.sh
+            device=$(get_devices $gpu_num 0)
+        fi
+        export CUDA_VISIBLE_DEVICES=${device}
+    fi
+
+    for dec_model in ${dec_models[@]}; do
+        suffix=alpha${len_penalty}
+        model_str=`echo $dec_model | sed -e "s#checkpoint##" | sed "s#.pt##"`
+        suffix=${suffix}_${model_str}
+        if [[ -n ${cer} && ${cer} -eq 1 ]]; then
+            suffix=${suffix}_cer
+        else
+            suffix=${suffix}_wer
+        fi
+
+        suffix=${suffix}_beam${beam_size}
+        if [[ ${batch_size} -ne 0 ]]; then
+            suffix=${suffix}_batch${batch_size}
+        else
+            suffix=${suffix}_tokens${max_tokens}
+        fi
+        if [[ ${ctc_infer} -eq 1 ]]; then
+            suffix=${suffix}_ctc
+        fi
+        if [[ ${ctc_self_ensemble} -eq 1 ]]; then
+            suffix=${suffix}_ensemble
+        fi
+        if [[ ${ctc_inter_logit} -ne 0 ]]; then
+            suffix=${suffix}_logit${ctc_inter_logit}
+        fi
+        if (( $(echo "${infer_ctc_weight} > 0" | bc -l) )); then
+            suffix=${suffix}_ctc${infer_ctc_weight}
+        fi
+        if [[ ${infer_score} -eq 1 ]]; then
+            suffix=${suffix}_score
+        fi
+
+        suffix=`echo $suffix | sed -e "s#__#_#"`
+        result_file=${model_dir}/decode_result_${suffix}
+        [[ -f ${result_file} ]] && rm ${result_file}
+
+        test_subset=${test_subset//,/ }
+        for subset in ${test_subset[@]}; do
+            subset=${subset}
+            if [[ ${infer_debug} -ne 0 ]]; then
+                cmd="python3 -m debugpy --listen 0.0.0.0:5678 --wait-for-client"
+            else
+                cmd="python3 "
+            fi
+            cmd="$cmd ${code_dir}/fairseq_cli/generate.py
+            ${data_dir}
+            --config-yaml ${data_config}
+            --gen-subset ${subset}
+            --task speech_to_text
+            --path ${model_dir}/${dec_model}
+            --results-path ${model_dir}
+            --batch-size ${batch_size}
+            --max-tokens ${max_tokens}
+            --beam ${beam_size}
+            --lenpen ${len_penalty}
+            --infer-ctc-weight ${infer_ctc_weight}
+            --scoring wer"
+
+            if [[ ${cer} -eq 1 ]]; then
+                cmd="${cmd}
+            --wer-char-level"
+            fi
+            if [[ ${ctc_infer} -eq 1 ]]; then
+                cmd="${cmd}
+            --ctc-infer"
+            fi
+            if [[ ${ctc_self_ensemble} -eq 1 ]]; then
+                cmd="${cmd}
+            --ctc-self-ensemble"
+            fi
+            if [[ ${ctc_inter_logit} -ne 0 ]]; then
+                cmd="${cmd}
+            --ctc-inter-logit ${ctc_inter_logit}"
+            fi
+            if [[ ${infer_score} -eq 1 ]]; then
+                cmd="${cmd}
+            --score-reference"
+            fi
+            if [[ -n ${infer_parameters} ]]; then
+                cmd="${cmd}
+            ${infer_parameters}"
+            fi                        
+
+            echo -e "\033[34mRun command: \n${cmd} \033[0m"
+
+            cd ${code_dir}
+            if [[ $eval -eq 1 ]]; then
+                ctc_file=translation-${subset}.ctc
+                if [[ -f ${model_dir}/${ctc_file} ]]; then
+                    rm ${model_dir}/${ctc_file}
+                fi
+
+                eval $cmd
+                echo "" >> ${result_file}
+                tail -n 2 ${model_dir}/generate-${subset}.txt >> ${result_file}
+                mv ${model_dir}/generate-${subset}.txt ${model_dir}/generate-${subset}-${suffix}.txt
+                mv ${model_dir}/translation-${subset}.txt ${model_dir}/translation-${subset}-${suffix}.txt
+
+                cd ${pwd_dir}
+                if [[ -f ${model_dir}/enc_dump ]]; then
+                    mv ${model_dir}/enc_dump ${model_dir}/dump-${subset}-enc-${suffix}
+                fi
+                if [[ -f ${model_dir}/dec_dump ]]; then
+                    mv ${model_dir}/dec_dump ${model_dir}/dump-${subset}-dec-${suffix}
+                fi
+
+                trans_file=translation-${subset}-${suffix}.txt
+                if [[ ${ctc_infer} -eq 1 && -f ${model_dir}/${ctc_file} ]]; then
+                    ref_file=${model_dir}/${subset}.${src_lang}
+                    if [[ ! -f ${ref_file} ]]; then
+                        python3 ./local/extract_txt_from_tsv.py ${data_dir}/${subset}.tsv ${ref_file} "tgt_text"
+                    fi
+                    if [[ -f ${ref_file} ]]; then
+                        ctc=$(mktemp -t temp.record.XXXXXX)
+                        cd ./local
+                        cmd="./cal_wer.sh ${model_dir} ${subset} ${trans_file} ${ctc_file} ${ref_file} > ${ctc}"
+                        #echo $cmd
+                        eval $cmd
+                        cd ..
+
+                        echo "CTC WER" >> ${result_file}
+                        tail -n 2 ${ctc} >> ${result_file}
+
+                        src_bleu=$(mktemp -t temp.record.XXXXXX)
+                        cd local
+                        ./cal_ctc_bleu.sh ${model_dir} ${subset} ${trans_file} ${ctc_file} ${ref_file} ${tokenizer} ${src_lang} > ${src_bleu}
+                        cd ..
+                        cat ${src_bleu} >> ${result_file}
+
+                        rm ${ctc} ${src_bleu}
+                    else
+                        echo "No reference for source language."
+                    fi
+                fi
+            fi
+        done
+        echo
+        cat ${result_file}
+    done
+fi

egs/talcs/asr/train.sh

+#!/usr/bin/env bash
+
+# training the model
+
+gpu_num=1
+update_freq=1
+max_tokens=100000
+
+extra_tag=
+extra_parameter=
+#extra_tag="${extra_tag}"
+#extra_parameter="${extra_parameter} "
+
+exp_tag=
+
+# CTC
+config_list=(purectc)
+
+# Transformer
+config_list=(base ctc)
+
+# Conformer
+#config_list=(base conformer ctc)
+
+# PDS
+config_list=(purectc_pds_base_8)
+config_list=(pds_base_8)
+
+# exp full name
+exp_name=
+
+train_config=$(echo ${config_list[*]} | sed 's/ /,/g')
+
+cmd="./run.sh
+    --stage 1
+    --stop_stage 2
+    --gpu_num ${gpu_num}
+    --update_freq ${update_freq}
+    --train_config ${train_config}
+    --max_tokens ${max_tokens}
+    "
+
+if [[ -n ${exp_name} ]]; then
+    cmd="$cmd --exp_name ${exp_name}"
+fi
+if [[ -n ${exp_tag} ]]; then
+    cmd="$cmd --exp_tag ${exp_tag}"
+fi
+if [[ -n ${extra_tag} ]]; then
+    cmd="$cmd --extra_tag ${extra_tag}"
+fi
+if [[ -n ${extra_parameter} ]]; then
+    cmd="$cmd --extra_parameter \"${extra_parameter}\""
+fi
+
+echo ${cmd}
+eval ${cmd}

egs/tedlium2/asr/conf/base.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/tedlium2/asr/conf/base_nonorm.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 subsampling-type: conv1d
 subsampling-type: conv1d
 subsampling-layers: 2
 subsampling-layers: 2

egs/tedlium2/asr/conf/basis.yaml

@@ -25,5 +25,6 @@ num-workers: 8
 no-progress-bar: True
 no-progress-bar: True
 log-interval: 100
 log-interval: 100
 seed: 1
 seed: 1
+label-smoothing: 0.1
 report-accuracy: True
 report-accuracy: True
 skip-invalid-size-inputs-valid-test: True
 skip-invalid-size-inputs-valid-test: True
\ No newline at end of file

egs/tedlium2/asr/conf/big.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 encoder-embed-norm: True
 encoder-embed-norm: True
 encoder-no-scale-embedding: True
 encoder-no-scale-embedding: True

egs/tedlium2/asr/conf/big_wenet.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 subsampling-type: conv2d
 subsampling-type: conv2d
 subsampling-layers: 2
 subsampling-layers: 2

egs/tedlium2/asr/conf/mixup.yaml

@@ -8,6 +8,6 @@ inter-mixup-keep-org: False
 inter-mixup-decoder-emb: False
 inter-mixup-decoder-emb: False
 ctc-mixup-consistent-weight: 0
 ctc-mixup-consistent-weight: 0
 mixup-consistent-weight: 0
 mixup-consistent-weight: 0
-cal-mixup-loss: True
+mixup-no-hard-loss: False
 no-specaugment: False
 no-specaugment: False
 layer-out-norm: False
 layer-out-norm: False
\ No newline at end of file

egs/tedlium2/asr/conf/pds_base.yaml

@@ -10,10 +10,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_base_16.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_base_16_growth.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_base_16_growth_fusion320.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_base_32.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_base_4.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_base_8.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_base_8_444.yaml

@@ -23,10 +23,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_base_8_growth.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_base_8_growth_fusion256.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_big.yaml

@@ -12,10 +12,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_big_16.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 0.0014
 lr: 0.0014
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_big_32.yaml

@@ -21,10 +21,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 0.0014
 lr: 0.0014
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_big_8.yaml

@@ -21,10 +21,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 0.0014
 lr: 0.0014
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep.yaml

@@ -12,10 +12,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep18_16.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep18_32.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep18_8.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep18_big_8.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 0.0014
 lr: 0.0014
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.15
 dropout: 0.15
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep30_16.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep30_32.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep30_8.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep_16.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep_32.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/pds_deep_8.yaml

@@ -22,10 +22,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 activation-fn: relu
 activation-fn: relu

egs/tedlium2/asr/conf/purectc.yaml

@@ -5,7 +5,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 zero_infinity: True
 zero_infinity: True

egs/tedlium2/asr/conf/purectc_base.yaml

@@ -7,7 +7,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 0.002
 lr: 0.002
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 ctc-weight: 1.0
 ctc-weight: 1.0

egs/tedlium2/asr/conf/purectc_pds_base_16.yaml

@@ -29,7 +29,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 zero_infinity: True
 zero_infinity: True

egs/tedlium2/asr/conf/purectc_pds_base_16_growth.yaml

@@ -23,7 +23,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 ctc-weight: 1.0
 ctc-weight: 1.0

egs/tedlium2/asr/conf/purectc_pds_base_16_growth_fusion256.yaml

@@ -23,7 +23,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 ctc-weight: 1.0
 ctc-weight: 1.0

egs/tedlium2/asr/conf/purectc_pds_base_16_growth_fusion320.yaml

@@ -22,7 +22,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 ctc-weight: 1.0
 ctc-weight: 1.0

egs/tedlium2/asr/conf/purectc_pds_base_16_growth_fusion360.yaml

@@ -22,7 +22,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 ctc-weight: 1.0
 ctc-weight: 1.0

egs/tedlium2/asr/conf/purectc_pds_base_8.yaml

@@ -30,7 +30,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 zero_infinity: True
 zero_infinity: True

egs/tedlium2/asr/conf/purectc_pds_base_8_666.yaml

@@ -29,7 +29,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 zero_infinity: True
 zero_infinity: True

egs/tedlium2/asr/conf/purectc_pds_base_8_666_growth360.yaml

@@ -29,7 +29,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 zero_infinity: True
 zero_infinity: True

egs/tedlium2/asr/conf/purectc_pds_base_8_growth.yaml

@@ -23,7 +23,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 ctc-weight: 1.0
 ctc-weight: 1.0

egs/tedlium2/asr/conf/purectc_pds_base_8_growth360.yaml

@@ -22,7 +22,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 ctc-weight: 1.0
 ctc-weight: 1.0

egs/tedlium2/asr/conf/purectc_pds_base_8_growth_fusion256.yaml

@@ -22,7 +22,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 ctc-weight: 1.0
 ctc-weight: 1.0

egs/tedlium2/asr/conf/purectc_pds_big_16.yaml

@@ -29,7 +29,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 0.0014
 lr: 0.0014
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 zero_infinity: True
 zero_infinity: True

egs/tedlium2/asr/conf/purectc_pds_big_8.yaml

@@ -23,7 +23,7 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 warmup-updates: 10000
 lr: 0.0014
 lr: 0.0014
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: ctc
 criterion: ctc
 zero_infinity: True
 zero_infinity: True

egs/tedlium2/asr/run.sh

@@ -2,8 +2,7 @@
 
 
 # Processing TED-LIUM2 ASR Datasets
 # Processing TED-LIUM2 ASR Datasets
 
 
-# Copyright 2023 Natural Language Processing Laboratory 
-# Xu Chen (xuchenneu@163.com)
+# Copyright 2021 Chen Xu (xuchennlp@outlook.com)
 
 
 # Set bash to 'debug' mode, it will exit on :
 # Set bash to 'debug' mode, it will exit on :
 # -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
 # -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
@@ -16,22 +15,22 @@ eval=1
 time=$(date "+%m%d_%H%M")
 time=$(date "+%m%d_%H%M")
 
 
 stage=1
 stage=1
-stop_stage=4
+stop_stage=2
 
 
-######## hardware ########
-# devices
+######## Hardware ########
+# Devices
 device=(0)
 device=(0)
 gpu_num=2
 gpu_num=2
 update_freq=1
 update_freq=1
+max_tokens=100000
 
 
-root_dir=/opt/tiger
-#  data_root_dir=/mnt/bd/data-model
-data_root_dir=/mnt/bn/nas-xc-1
-
-code_dir=${root_dir}/s2t
 pwd_dir=$PWD
 pwd_dir=$PWD
+root_dir=${ST_ROOT}
+data_root_dir=${root_dir}
 
 
-# dataset
+code_dir=${root_dir}/S2T
+
+# Dataset
 src_lang=en
 src_lang=en
 lang=${src_lang}
 lang=${src_lang}
 dataset=tedlium2
 dataset=tedlium2
@@ -40,7 +39,7 @@ data_tag=asr
 task=speech_to_text
 task=speech_to_text
 vocab_type=unigram
 vocab_type=unigram
 vocab_size=10000
 vocab_size=10000
-speed_perturb=1
+speed_perturb=0
 lcrm=0
 lcrm=0
 tokenizer=0
 tokenizer=0
 use_raw_audio=0
 use_raw_audio=0
@@ -49,7 +48,7 @@ use_raw_audio=0
 
 
 use_specific_dict=0
 use_specific_dict=0
 specific_prefix=st
 specific_prefix=st
-specific_dir=${root_dir}/data/mustc/st
+specific_dir=${data_root_dir}/data/must_c/en-de/st
 asr_vocab_prefix=spm_unigram10000_st_share
 asr_vocab_prefix=spm_unigram10000_st_share
 
 
 data_model_subfix=${dataset}/${data_tag}
 data_model_subfix=${dataset}/${data_tag}
@@ -63,35 +62,37 @@ test_subset=dev,test
 # exp
 # exp
 sub_tag=
 sub_tag=
 exp_prefix=$(date "+%m%d")
 exp_prefix=$(date "+%m%d")
-# exp_subfix=${ARNOLD_JOB_ID}_${ARNOLD_TASK_ID}_${ARNOLD_TRIAL_ID}
 extra_tag=
 extra_tag=
 extra_parameter=
 extra_parameter=
 exp_tag=baseline
 exp_tag=baseline
 exp_name=
 exp_name=
 
 
-# config
-train_config=base,ctc
-data_config=config.yaml
-
-# training setting
+# Training Settings
+train_config=base
 fp16=1
 fp16=1
-max_tokens=100000
 step_valid=0
 step_valid=0
 
 
-# decoding setting
+# Decoding Settings
 dec_model=checkpoint_best.pt
 dec_model=checkpoint_best.pt
 cer=0
 cer=0
-ctc_infer=0
+ctc_infer=1
+infer_ctc_weight=0
 ctc_self_ensemble=0
 ctc_self_ensemble=0
 ctc_inter_logit=0
 ctc_inter_logit=0
 n_average=10
 n_average=10
+batch_size=0
 beam_size=5
 beam_size=5
 len_penalty=1.0
 len_penalty=1.0
 single=0
 single=0
 epoch_ensemble=0
 epoch_ensemble=0
 best_ensemble=1
 best_ensemble=1
-infer_parameters=
+infer_debug=0
+infer_score=0
+#infer_parameters="--cal-monotonic-cross-attn-weights --cal-localness --localness-window 0.1 --cal-topk-cross-attn-weights --topk-cross-attn-weights 15 --cal-entropy"
+
+data_config=config.yaml
 
 
+# Parsing Options
 if [[ ${speed_perturb} -eq 1 ]]; then
 if [[ ${speed_perturb} -eq 1 ]]; then
     data_dir=${data_dir}_sp
     data_dir=${data_dir}_sp
     exp_prefix=${exp_prefix}_sp
     exp_prefix=${exp_prefix}_sp
@@ -116,13 +117,6 @@ if [[ "${vocab_type}" == "char" ]]; then
     data_dir=${data_dir}_char
     data_dir=${data_dir}_char
     exp_prefix=${exp_prefix}_char
     exp_prefix=${exp_prefix}_char
 fi
 fi
-if [[ ! -d /mnt/bd/data-model && -d /mnt/bd/data-model2 ]]; then
-    sudo ln -s /mnt/bd/data-model2/ /mnt/bd/data-model
-fi
-if [[ ! -d ${data_dir} ]]; then
-    echo "No feature dir ${data_dir}"
-    exit
-fi
 
 
 # setup nccl envs
 # setup nccl envs
 export NCCL_IB_DISABLE=0
 export NCCL_IB_DISABLE=0
@@ -130,13 +124,6 @@ export NCCL_IB_HCA=$ARNOLD_RDMA_DEVICE:1
 export NCCL_IB_GID_INDEX=3
 export NCCL_IB_GID_INDEX=3
 export NCCL_SOCKET_IFNAME=eth0
 export NCCL_SOCKET_IFNAME=eth0
 
 
-HOSTS=$ARNOLD_WORKER_HOSTS
-HOST=(${HOSTS//,/ })
-HOST_SPLIT=(${HOST//:/ })
-PORT=${HOST_SPLIT[1]}
-INIT_METHOD="tcp://${ARNOLD_WORKER_0_HOST}:${ARNOLD_WORKER_0_PORT}"
-DIST_RANK=$((ARNOLD_ID * ARNOLD_WORKER_GPU))
-
 export PATH=$PATH:${code_dir}/scripts
 export PATH=$PATH:${code_dir}/scripts
 . ./local/parse_options.sh || exit 1;
 . ./local/parse_options.sh || exit 1;
 
 
@@ -150,21 +137,28 @@ if [[ -z ${exp_name} ]]; then
         exp_name=${exp_name}_${exp_subfix}
         exp_name=${exp_name}_${exp_subfix}
     fi
     fi
 fi
 fi
-model_dir=${code_dir}/checkpoints/${data_model_subfix}/${sub_tag}/${exp_name}
 
 
-echo "stage: $stage"
-echo "stop_stage: $stop_stage"
+ckpt_dir=${root_dir}/checkpoints/
+model_dir=${root_dir}/checkpoints/${data_model_subfix}/${sub_tag}/${exp_name}
+
+# Start
 cd ${code_dir}
 cd ${code_dir}
+echo "Start Stage: $stage"
+echo "Stop  Stage: $stop_stage"
+
+if [[ `pip list | grep fairseq | wc -l` -eq 0 ]]; then 
+    echo "Default Stage: env configure"
+    pip3 install -e ${code_dir}
+fi
 
 
 if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
 if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
-    echo "stage -1: Data Download"
-    # pass
+    echo "Stage -1: Data Download"
 fi
 fi
 
 
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
 if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     ### Task dependent. You have to make data the following preparation part by yourself.
     ### Task dependent. You have to make data the following preparation part by yourself.
-    ### But you can utilize Kaldi recipes in most cases
-    echo "stage 0: ASR Data Preparation"
+    echo "Stage 0: Data Preparation"
+
     if [[ ! -e ${data_dir} ]]; then
     if [[ ! -e ${data_dir} ]]; then
         mkdir -p ${data_dir}
         mkdir -p ${data_dir}
     fi
     fi
@@ -205,40 +199,18 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     [[ $eval -eq 1 ]] && eval ${cmd}
     [[ $eval -eq 1 ]] && eval ${cmd}
 fi
 fi
 
 
-echo "stage 1: env configure"
-if [[ `pip list | grep fairseq | wc -l` -eq 0 ]]; then 
-    pip3 install -e ${code_dir} -i https://bytedpypi.byted.org/simple  --no-build-isolation --default-timeout=10000
-fi
-
-if [[ -d /mnt/bn/nas-xc-1/checkpoints && ! -d ${code_dir}/checkpoints ]]; then
-    ln -s /mnt/bn/nas-xc-1/checkpoints ${code_dir}
-fi
-
-# if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
-#     ln_data_dir=`echo ${data_dir} | sed -e "s#${data_root_dir}#${code_dir}#"`
-#     echo ${ln_data_dir}
-#     mkdir -p ${ln_data_dir}
-#     ln -s ${data_dir}/../* ${ln_data_dir}
-#     rm -r ${ln_data_dir}
-
-#     hdfs_path=`echo ${data_dir} | sed -e "s#${data_root_dir}#hdfs://haruna/home/byte_arnold_lq_mlnlc/user/xuchen/#"`
-#     hdfs dfs -get ${hdfs_path} ${ln_data_dir}
-#     sed -i -e "s#${data_root_dir}#${code_dir}#" ${ln_data_dir}/config*
-#     data_dir=${ln_data_dir}
-
-# fi
-
-if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
-    echo "stage 2: ASR Network Training"
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    echo "Stage 1: Network Training"
     [[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1;
     [[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1;
 
 
     if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
     if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
-		if [[ ${gpu_num} -eq 0 ]]; then
-			device=""
-		else
+        if [[ ${gpu_num} -eq 0 ]]; then
+            device=""
+        else
             source ./local/utils.sh
             source ./local/utils.sh
             device=$(get_devices $gpu_num 0)
             device=$(get_devices $gpu_num 0)
-		fi
+        fi
+        export CUDA_VISIBLE_DEVICES=${device}
     fi
     fi
 
 
     echo -e "data=${data_dir} model=${model_dir}"
     echo -e "data=${data_dir} model=${model_dir}"
@@ -252,11 +224,9 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     cp -f ${pwd_dir}/`basename ${BASH_SOURCE[0]}` ${model_dir}
     cp -f ${pwd_dir}/`basename ${BASH_SOURCE[0]}` ${model_dir}
     cp -f ${pwd_dir}/train.sh ${model_dir}
     cp -f ${pwd_dir}/train.sh ${model_dir}
 
 
-    extra_parameter="${extra_parameter}
-        --train-config ${pwd_dir}/conf/basis.yaml"
-    cp -f ${pwd_dir}/conf/basis.yaml ${model_dir}
+    train_config=basis,${train_config}
     config_list="${train_config//,/ }"
     config_list="${train_config//,/ }"
-    idx=1
+    idx=0
     for config in ${config_list[@]}
     for config in ${config_list[@]}
     do
     do
         config_path=${pwd_dir}/conf/${config}.yaml
         config_path=${pwd_dir}/conf/${config}.yaml
@@ -266,8 +236,13 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
         fi
         fi
         cp -f ${config_path} ${model_dir}
         cp -f ${config_path} ${model_dir}
 
 
+        if [[ $idx -eq 0 ]]; then
+        extra_parameter="${extra_parameter}
+        --train-config ${config_path}"
+        else
         extra_parameter="${extra_parameter}
         extra_parameter="${extra_parameter}
         --train-config${idx} ${config_path}"
         --train-config${idx} ${config_path}"
+        fi
         idx=$((idx + 1))
         idx=$((idx + 1))
     done
     done
 
 
@@ -282,17 +257,15 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
         --save-dir ${model_dir}
         --save-dir ${model_dir}
         --tensorboard-logdir ${model_dir}"
         --tensorboard-logdir ${model_dir}"
 
 
-	if [[ -n ${extra_parameter} ]]; then
+    if [[ -n ${extra_parameter} ]]; then
         cmd="${cmd}
         cmd="${cmd}
         ${extra_parameter}"
         ${extra_parameter}"
     fi
     fi
-	if [[ ${gpu_num} -gt 0 ]]; then
-		cmd="${cmd}
+    if [[ ${gpu_num} -gt 0 ]]; then
+        cmd="${cmd}
         --distributed-world-size $gpu_num
         --distributed-world-size $gpu_num
         --ddp-backend no_c10d"
         --ddp-backend no_c10d"
-        #--distributed-init-method ${INIT_METHOD}
-        #--distributed-rank ${DIST_RANK}"
-	fi
+    fi
     if [[ $fp16 -eq 1 ]]; then
     if [[ $fp16 -eq 1 ]]; then
         cmd="${cmd}
         cmd="${cmd}
         --fp16"
         --fp16"
@@ -328,35 +301,28 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     echo -e "\033[34mRun command: \n${cmd} \033[0m"
     echo -e "\033[34mRun command: \n${cmd} \033[0m"
 
 
     # save info
     # save info
-    log=./history.log
+    log=${ckpt_dir}/history.log
     echo "${time} | ${data_dir} | ${exp_name} | ${model_dir} " >> $log
     echo "${time} | ${data_dir} | ${exp_name} | ${model_dir} " >> $log
     tail -n 50 ${log} > tmp.log
     tail -n 50 ${log} > tmp.log
     mv tmp.log $log
     mv tmp.log $log
-    # export CUDA_VISIBLE_DEVICES=${device}
 
 
     log=${model_dir}/train.log
     log=${model_dir}/train.log
     cmd="${cmd} 2>&1 | tee -a ${log}"
     cmd="${cmd} 2>&1 | tee -a ${log}"
-    #cmd="nohup ${cmd} >> ${log} 2>&1 &"
+    #cmd="${cmd} >> ${log} 2>&1 "
     if [[ $eval -eq 1 ]]; then
     if [[ $eval -eq 1 ]]; then
         # tensorboard
         # tensorboard
-        if [[ -z ${ARNOLD_TENSORBOARD_CURRENT_PORT} ]]; then
-            port=6666
-        else
-            port=${ARNOLD_TENSORBOARD_CURRENT_PORT}
-        fi
+        port=6666
         tensorboard --logdir ${model_dir} --port ${port} --bind_all &
         tensorboard --logdir ${model_dir} --port ${port} --bind_all &
     
     
         echo "${cmd}" > ${model_dir}/cmd
         echo "${cmd}" > ${model_dir}/cmd
-		eval $cmd
-		#sleep 2s
-		#tail -n "$(wc -l ${log} | awk '{print $1+1}')" -f ${log}
-	fi
+        eval $cmd
+    fi
 fi
 fi
 
 
-if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
-    echo "stage 3: ASR Decoding"
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    echo "Stage 2: Decoding"
     dec_models=
     dec_models=
-    if [[ ${single} -eq 1 ]]; then
+    if [[ ${n_average} -eq 1 ]]; then
         dec_models=${dec_model}
         dec_models=${dec_model}
     fi
     fi
     if [[ ${n_average} -ne 1 ]]; then
     if [[ ${n_average} -ne 1 ]]; then
@@ -387,20 +353,20 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
             fi
             fi
             dec_models+=(${avg_model})
             dec_models+=(${avg_model})
         fi
         fi
-	fi
+    fi
 
 
     if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
     if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
-		if [[ ${gpu_num} -eq 0 ]]; then
-			device=""
-		else
+        if [[ ${gpu_num} -eq 0 ]]; then
+            device=""
+        else
             source ./local/utils.sh
             source ./local/utils.sh
             device=$(get_devices $gpu_num 0)
             device=$(get_devices $gpu_num 0)
-		fi
+        fi
+        export CUDA_VISIBLE_DEVICES=${device}
     fi
     fi
-    # export CUDA_VISIBLE_DEVICES=${device}
 
 
     for dec_model in ${dec_models[@]}; do
     for dec_model in ${dec_models[@]}; do
-        suffix=beam${beam_size}_alpha${len_penalty}_tokens${max_tokens}
+        suffix=alpha${len_penalty}
         model_str=`echo $dec_model | sed -e "s#checkpoint##" | sed "s#.pt##"`
         model_str=`echo $dec_model | sed -e "s#checkpoint##" | sed "s#.pt##"`
         suffix=${suffix}_${model_str}
         suffix=${suffix}_${model_str}
         if [[ -n ${cer} && ${cer} -eq 1 ]]; then
         if [[ -n ${cer} && ${cer} -eq 1 ]]; then
@@ -408,6 +374,13 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
         else
         else
             suffix=${suffix}_wer
             suffix=${suffix}_wer
         fi
         fi
+
+        suffix=${suffix}_beam${beam_size}
+        if [[ ${batch_size} -ne 0 ]]; then
+            suffix=${suffix}_batch${batch_size}
+        else
+            suffix=${suffix}_tokens${max_tokens}
+        fi
         if [[ ${ctc_infer} -eq 1 ]]; then
         if [[ ${ctc_infer} -eq 1 ]]; then
             suffix=${suffix}_ctc
             suffix=${suffix}_ctc
         fi
         fi
@@ -417,6 +390,12 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
         if [[ ${ctc_inter_logit} -ne 0 ]]; then
         if [[ ${ctc_inter_logit} -ne 0 ]]; then
             suffix=${suffix}_logit${ctc_inter_logit}
             suffix=${suffix}_logit${ctc_inter_logit}
         fi
         fi
+        if (( $(echo "${infer_ctc_weight} > 0" | bc -l) )); then
+            suffix=${suffix}_ctc${infer_ctc_weight}
+        fi
+        if [[ ${infer_score} -eq 1 ]]; then
+            suffix=${suffix}_score
+        fi
 
 
         suffix=`echo $suffix | sed -e "s#__#_#"`
         suffix=`echo $suffix | sed -e "s#__#_#"`
         result_file=${model_dir}/decode_result_${suffix}
         result_file=${model_dir}/decode_result_${suffix}
@@ -425,16 +404,23 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
         test_subset=${test_subset//,/ }
         test_subset=${test_subset//,/ }
         for subset in ${test_subset[@]}; do
         for subset in ${test_subset[@]}; do
             subset=${subset}
             subset=${subset}
-            cmd="python3 ${code_dir}/fairseq_cli/generate.py
+            if [[ ${infer_debug} -ne 0 ]]; then
+                cmd="python3 -m debugpy --listen 0.0.0.0:5678 --wait-for-client"
+            else
+                cmd="python3 "
+            fi
+            cmd="$cmd ${code_dir}/fairseq_cli/generate.py
             ${data_dir}
             ${data_dir}
             --config-yaml ${data_config}
             --config-yaml ${data_config}
             --gen-subset ${subset}
             --gen-subset ${subset}
             --task speech_to_text
             --task speech_to_text
             --path ${model_dir}/${dec_model}
             --path ${model_dir}/${dec_model}
             --results-path ${model_dir}
             --results-path ${model_dir}
+            --batch-size ${batch_size}
             --max-tokens ${max_tokens}
             --max-tokens ${max_tokens}
             --beam ${beam_size}
             --beam ${beam_size}
             --lenpen ${len_penalty}
             --lenpen ${len_penalty}
+            --infer-ctc-weight ${infer_ctc_weight}
             --scoring wer"
             --scoring wer"
 
 
             if [[ ${cer} -eq 1 ]]; then
             if [[ ${cer} -eq 1 ]]; then
@@ -453,20 +439,28 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
                 cmd="${cmd}
                 cmd="${cmd}
             --ctc-inter-logit ${ctc_inter_logit}"
             --ctc-inter-logit ${ctc_inter_logit}"
             fi
             fi
+            if [[ ${infer_score} -eq 1 ]]; then
+                cmd="${cmd}
+            --score-reference"
+            fi
             if [[ -n ${infer_parameters} ]]; then
             if [[ -n ${infer_parameters} ]]; then
                 cmd="${cmd}
                 cmd="${cmd}
             ${infer_parameters}"
             ${infer_parameters}"
-            fi
+            fi                        
 
 
             echo -e "\033[34mRun command: \n${cmd} \033[0m"
             echo -e "\033[34mRun command: \n${cmd} \033[0m"
 
 
+            cd ${code_dir}
             if [[ $eval -eq 1 ]]; then
             if [[ $eval -eq 1 ]]; then
-                src_ctc_file=translation-${subset}.txt.ctc
-                if [[ -f ${model_dir}/${src_ctc_file} ]]; then
-                    rm ${model_dir}/${src_ctc_file}
+                ctc_file=translation-${subset}.ctc
+                xctc_file=translation-${subset}.xctc
+                if [[ -f ${model_dir}/${ctc_file} ]]; then
+                    rm ${model_dir}/${ctc_file}
+                fi
+                if [[ -f ${model_dir}/${xctc_file} ]]; then
+                    rm ${model_dir}/${xctc_file}
                 fi
                 fi
 
 
-                cd ${code_dir}
                 eval $cmd
                 eval $cmd
                 echo "" >> ${result_file}
                 echo "" >> ${result_file}
                 tail -n 2 ${model_dir}/generate-${subset}.txt >> ${result_file}
                 tail -n 2 ${model_dir}/generate-${subset}.txt >> ${result_file}
@@ -474,28 +468,39 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
                 mv ${model_dir}/translation-${subset}.txt ${model_dir}/translation-${subset}-${suffix}.txt
                 mv ${model_dir}/translation-${subset}.txt ${model_dir}/translation-${subset}-${suffix}.txt
 
 
                 cd ${pwd_dir}
                 cd ${pwd_dir}
+                if [[ -f ${model_dir}/enc_dump ]]; then
+                    mv ${model_dir}/enc_dump ${model_dir}/dump-${subset}-enc-${suffix}
+                fi
+                if [[ -f ${model_dir}/dec_dump ]]; then
+                    mv ${model_dir}/dec_dump ${model_dir}/dump-${subset}-dec-${suffix}
+                fi
+
                 trans_file=translation-${subset}-${suffix}.txt
                 trans_file=translation-${subset}-${suffix}.txt
-                if [[ ${ctc_infer} -eq 1 && -f ${model_dir}/${src_ctc_file} ]]; then
+                if [[ ! -f ${model_dir}/{ctc_file} && -f ${model_dir}/${xctc_file} ]]; then
+                    ctc_file=${xctc_file}
+                fi
+                if [[ ${ctc_infer} -eq 1 && -f ${model_dir}/${ctc_file} ]]; then
                     ref_file=${model_dir}/${subset}.${src_lang}
                     ref_file=${model_dir}/${subset}.${src_lang}
                     if [[ ! -f ${ref_file} ]]; then
                     if [[ ! -f ${ref_file} ]]; then
-                        python3 ./local/extract_txt_from_tsv.py ${data_dir}/${subset}.tsv ${ref_file} "src_text"
+                        python3 ./local/extract_txt_from_tsv.py ${data_dir}/${subset}.tsv ${ref_file} "tgt_text"
                     fi
                     fi
                     if [[ -f ${ref_file} ]]; then
                     if [[ -f ${ref_file} ]]; then
-                        src_ctc=$(mktemp -t temp.record.XXXXXX)
+                        ctc=$(mktemp -t temp.record.XXXXXX)
                         cd ./local
                         cd ./local
-                        ./cal_wer.sh ${model_dir} ${subset} ${trans_file} ${src_ctc_file} ${ref_file} > ${src_ctc}
+                        cmd="./cal_wer.sh ${model_dir} ${subset} ${trans_file} ${ctc_file} ${ref_file} > ${ctc}"
+                        eval $cmd
                         cd ..
                         cd ..
 
 
                         echo "CTC WER" >> ${result_file}
                         echo "CTC WER" >> ${result_file}
-                        tail -n 2 ${src_ctc} >> ${result_file}
+                        tail -n 2 ${ctc} >> ${result_file}
 
 
                         src_bleu=$(mktemp -t temp.record.XXXXXX)
                         src_bleu=$(mktemp -t temp.record.XXXXXX)
                         cd local
                         cd local
-                        ./cal_ctc_bleu.sh ${model_dir} ${subset} ${trans_file} ${src_ctc_file} ${ref_file} ${tokenizer} ${src_lang} > ${src_bleu}
+                        ./cal_ctc_bleu.sh ${model_dir} ${subset} ${trans_file} ${ctc_file} ${ref_file} ${tokenizer} ${src_lang} > ${src_bleu}
                         cd ..
                         cd ..
                         cat ${src_bleu} >> ${result_file}
                         cat ${src_bleu} >> ${result_file}
 
 
-                        rm ${src_ctc} ${src_bleu}
+                        rm ${ctc} ${src_bleu}
                     else
                     else
                         echo "No reference for source language."
                         echo "No reference for source language."
                     fi
                     fi
@@ -506,10 +511,3 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
         cat ${result_file}
         cat ${result_file}
     done
     done
 fi
 fi
-
-# if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
-#     echo "Stage 4: Upload model and log"
-#     echo "Path: hdfs://haruna/home/byte_arnold_lq_mlnlc/user/xuchen/s2t/checkpoints/${data_model_subfix}/${exp_name}"
-#     hdfs dfs -mkdir -p hdfs://haruna/home/byte_arnold_lq_mlnlc/user/xuchen/s2t/checkpoints/${data_model_subfix}
-#     hdfs dfs -put -f ${model_dir} hdfs://haruna/home/byte_arnold_lq_mlnlc/user/xuchen/s2t/checkpoints/${data_model_subfix}
-# fi

egs/wav2vec/conf/basis.yaml

@@ -10,5 +10,6 @@ num-workers: 8
 no-progress-bar: True
 no-progress-bar: True
 log-interval: 100
 log-interval: 100
 seed: 1
 seed: 1
+label-smoothing: 0.1
 report-accuracy: True
 report-accuracy: True
 skip-invalid-size-inputs-valid-test: True
 skip-invalid-size-inputs-valid-test: True

egs/wav2vec/conf/wav2vec2.yaml

@@ -38,7 +38,7 @@ optimization:
 
 
 optimizer:
 optimizer:
   _name: adam
   _name: adam
-  adam_betas: (0.9,0.98)
+  adam-betas: (0.9,0.98)
   adam_eps: 1e-06
   adam_eps: 1e-06
   weight_decay: 0.01
   weight_decay: 0.01
 
 

egs/wmt16/mt/conf/base.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 8000
 warmup-updates: 8000
 lr: 1e-3
 lr: 1e-3
-adam_betas: (0.9,0.997)
+adam-betas: (0.9,0.997)
 
 
 criterion: label_smoothed_cross_entropy
 criterion: label_smoothed_cross_entropy
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 attention-dropout: 0.1
 attention-dropout: 0.1

egs/wmt16/mt/conf/base_postnorm.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 4000
 warmup-updates: 4000
 lr: 7e-4
 lr: 7e-4
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy
 criterion: label_smoothed_cross_entropy
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 attention-dropout: 0.1
 attention-dropout: 0.1

egs/wmt16/mt/conf/basis.yaml

@@ -25,5 +25,6 @@ num-workers: 8
 no-progress-bar: True
 no-progress-bar: True
 log-interval: 100
 log-interval: 100
 seed: 1
 seed: 1
+label-smoothing: 0.1
 report-accuracy: True
 report-accuracy: True
 skip-invalid-size-inputs-valid-test: True
 skip-invalid-size-inputs-valid-test: True
\ No newline at end of file

egs/wmt16/mt/conf/big.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 8000
 warmup-updates: 8000
 lr: 7e-4
 lr: 7e-4
-adam_betas: (0.9,0.997)
+adam-betas: (0.9,0.997)
 
 
 criterion: label_smoothed_cross_entropy
 criterion: label_smoothed_cross_entropy
-label_smoothing: 0.1
 
 
 dropout: 0.3
 dropout: 0.3
 attention-dropout: 0.1
 attention-dropout: 0.1

egs/wmt16/mt/conf/big_postnorm.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 4000
 warmup-updates: 4000
 lr: 5e-4
 lr: 5e-4
-adam_betas: (0.9,0.98)
+adam-betas: (0.9,0.98)
 
 
 criterion: label_smoothed_cross_entropy
 criterion: label_smoothed_cross_entropy
-label_smoothing: 0.1
 
 
 dropout: 0.3
 dropout: 0.3
 attention-dropout: 0.1
 attention-dropout: 0.1

egs/wmt16/mt/conf/deep.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 16000
 warmup-updates: 16000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.997)
+adam-betas: (0.9,0.997)
 
 
 criterion: label_smoothed_cross_entropy
 criterion: label_smoothed_cross_entropy
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 attention-dropout: 0.1
 attention-dropout: 0.1

egs/wmt16/mt/conf/deep_ctc.yaml

@@ -6,10 +6,9 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-init-lr: 1e-7
 warmup-updates: 16000
 warmup-updates: 16000
 lr: 2e-3
 lr: 2e-3
-adam_betas: (0.9,0.997)
+adam-betas: (0.9,0.997)
 
 
 criterion: label_smoothed_cross_entropy_with_ctc
 criterion: label_smoothed_cross_entropy_with_ctc
-label_smoothing: 0.1
 
 
 dropout: 0.1
 dropout: 0.1
 attention-dropout: 0.1
 attention-dropout: 0.1

egs/zerocs/asr/binary.sh

+set -e
+
+eval=1
+
+lcrm=0
+tokenizer=0
+
+root_dir=~/st/Fairseq-S2T
+data_dir=~/st/data/test
+vocab_dir=~/st/data/mustc/st/en-de
+asr_vocab_prefix=spm_unigram10000_st_share
+
+src_lang=en
+tgt_lang=de
+subsets=(2019)
+
+cp -r ${vocab_dir}/${asr_vocab_prefix}.* ${data_dir}/${src_lang}-${tgt_lang}
+rm -rf ${data_dir}/${src_lang}-${tgt_lang}/fbank80.zip
+
+splits=$(echo ${subsets[*]} | sed 's/ /,/g')
+cmd="python ${root_dir}/examples/speech_to_text/prep_st_data.py
+    --data-root ${data_dir}
+    --output-root ${data_dir}
+    --splits ${splits}
+    --task asr
+    --src-lang ${src_lang}
+    --tgt-lang ${tgt_lang}
+    --add-src
+    --share
+    --asr-prefix ${asr_vocab_prefix}
+    --cmvn-type utterance"
+
+    if [[ ${lcrm} -eq 1 ]]; then
+        cmd="$cmd
+    --lowercase-src
+    --rm-punc-src"
+    fi
+    if [[ ${tokenizer} -eq 1 ]]; then
+        cmd="$cmd
+    --tokenizer"
+    fi
+
+echo -e "\033[34mRun command: \n${cmd} \033[0m"
+[[ $eval -eq 1 ]] && eval ${cmd}

egs/zerocs/asr/conf/base.yaml

+arch: s2t_transformer_s
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/zerocs/asr/conf/base_nonorm.yaml

+arch: s2t_transformer_s
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/zerocs/asr/conf/basis.yaml

+#train-subset: train_zh_en,train_en_zh
+train-subset: train_zh,train_en,train_cs
+valid-subset: dev_zh,dev_en,dev_cs
+
+max-epoch: 100
+max-update: 100000
+patience: 20
+post-process: sentencepiece
+
+# best-checkpoint-metric: loss
+# maximize-best-checkpoint-metric: False
+
+eval-wer: True
+eval-wer-args: {"beam": 1, "lenpen": 1.0}
+eval-wer-tok-args: {"wer_remove_punct": true, "wer_lowercase": true, "wer_char_level": false}
+eval-wer-remove-bpe: sentencepiece
+eval-wer-print-samples: True
+best_checkpoint_metric: dec_wer
+maximize_best_checkpoint_metric: False
+
+no-epoch-checkpoints: True
+# keep-last-epochs: 10
+keep-best-checkpoints: 10
+
+num-workers: 8
+no-progress-bar: True
+log-interval: 100
+seed: 1
+label-smoothing: 0.1
+report-accuracy: True
+skip-invalid-size-inputs-valid-test: True

egs/zerocs/asr/conf/big.yaml

+arch: s2t_transformer_m
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 0.0014
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 2048
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.15
+activation-fn: relu
+encoder-embed-dim: 512
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
\ No newline at end of file

egs/zerocs/asr/conf/big_wenet.yaml

+arch: s2t_transformer_m
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 1e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv2d
+subsampling-layers: 2
+subsampling-filter: 512
+subsampling-kernel: 3
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: relu
+
+dropout: 0.15
+activation-fn: relu
+encoder-embed-dim: 512
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+cnn-module-norm: layer_norm
+
+load-pretrained-encoder-from: /home/xuchen/after.pt
+load-pretrained-decoder-from: /home/xuchen/after.pt
+#load-pretrained-decoder-from:

egs/zerocs/asr/conf/conformer.yaml

+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+layer-padding-mask: True
\ No newline at end of file

egs/zerocs/asr/conf/ctc.yaml

+ctc-weight: 0.3
\ No newline at end of file

egs/zerocs/asr/conf/dlcl.yaml

+use-enc-dlcl: True
+use-dec-dlcl: True

egs/zerocs/asr/conf/dynamic.yaml

+compression-metric: threshold
+compression-mode: create
+compression-layers: 6,9
+compression-threshold: 0.95
+compression-norm: True
+compression-pos: True
\ No newline at end of file

egs/zerocs/asr/conf/inter.yaml

+inter-ctc-weight: 0.2
+inter-ctc-layers: 6,9
+inter-ctc-drop-prob: 0
+share-inter-ctc: True
+
+ctc-pae: none
+# ctc-pae: inter_league
+
+# ctc-pae-ground-truth-ratio: 0.1
+# pae-gumbel: True
+# pae-distribution-hard: True
+# pae-drop-prob: 0.0
+# pae-distribution-cutoff: 10
+# share-pae-and-ctc: True
+# pae-embed-norm: True
+# pae-out-norm: True
+
+# ctc-self-distill-weight: 1
+# target-ctc-self-distill-weight: 1
+# ctc-self-distill-prob: 0.1
+# cal-all-ctc: True
\ No newline at end of file

egs/zerocs/asr/conf/ipa.yaml

+inter-mixup: True
+inter-mixup-layer: -1
+inter-mixup-decoder-layer: 0
+inter-mixup-prob: 1.0
+inter-mixup-ratio: 1.0
+inter-mixup-beta: 0.2
+inter-mixup-keep-org: False
+inter-mixup-decoder-emb: False
+ctc-mixup-consistent-weight: 0
+inter-ctc-mixup-consistent-weight: 0
+mixup-consistent-weight: 0
+mixup-no-hard-loss: False
+no-specaugment: False
+layer-out-norm: False
+
+inter-mixup-ratio-decay: False
+inter-mixup-ratio-decay-params: 20000,40000,0
\ No newline at end of file

egs/zerocs/asr/conf/mlo.yaml

+inter-ctc-mlo: 1:2:3
\ No newline at end of file

egs/zerocs/asr/conf/norm.yaml

+encoder-embed-norm: True
+encoder-no-scale-embedding: True
\ No newline at end of file

egs/zerocs/asr/conf/pds_base.yaml

+arch: pdss2t_transformer_s_8
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
\ No newline at end of file

egs/zerocs/asr/conf/pds_base_16.yaml

+arch: pdss2t_transformer_s_16
+
+encoder-embed-dim: 256
+pds-stages: 4
+pds-layers: 2_2_6_2
+pds-ratios: 2_2_2_2
+pds-fusion: False
+pds-fusion-method: all_conv2
+pds-fusion-layers: 0_1_1_1 
+pds-fusion-weight: 0.2_0.3_0.5
+pds-embed-dims: 256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 8_8_8_8
+pds-attn-heads: 4_4_4_4
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
\ No newline at end of file

egs/zerocs/asr/conf/pds_base_32.yaml

+arch: pdss2t_transformer_s_32
+
+encoder-embed-dim: 256
+pds-stages: 5
+pds-layers: 2_2_3_3_2
+pds-ratios: 2_2_2_2_2
+pds-fusion: False
+pds-fusion-method: all_conv2
+pds-fusion-layers: 0_0_1_1_1 
+pds-fusion-weight: 0.2_0.3_0.5
+pds-embed-dims: 256_256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1_1
+pds-kernel-sizes: 5_5_5_5_5
+pds-ffn-ratios: 8_8_8_8_8
+pds-attn-heads: 4_4_4_4_4
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
\ No newline at end of file

egs/zerocs/asr/conf/pds_base_8.yaml

+arch: pdss2t_transformer_s_8
+
+encoder-embed-dim: 256
+pds-stages: 4
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: False
+pds-fusion-method: all_conv2
+pds-fusion-layers: 0_1_1_1 
+pds-fusion-weight: 0.2_0.3_0.5
+pds-embed-dims: 256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 8_8_8_8
+pds-attn-heads: 4_4_4_4
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
\ No newline at end of file

egs/zerocs/asr/conf/pds_big_8.yaml

+arch: pdss2t_transformer_m_8
+
+encoder-embed-dim: 512
+pds-stages: 4
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: False
+pds-fusion-method: all_conv2
+pds-fusion-layers: 0_1_1_1 
+pds-fusion-weight: 0.2_0.3_0.5
+pds-embed-dims: 512_512_512_512
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 4_4_4_4
+pds-attn-heads: 8_8_8_8
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 0.0014
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+dropout: 0.15
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8

egs/zerocs/asr/conf/purectc.yaml

+arch: s2t_ctc
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: ctc
+zero_infinity: True
+ctc-weight: 1.0
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 18
+encoder-attention-heads: 4
\ No newline at end of file

egs/zerocs/asr/conf/reproduction_bil_ctc_syn.yaml

+arch: s2t_transformer_s
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+weight-decay: 1e-4
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 18
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
+
+# Conformer 
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+
+# Bilingual CTC
+share-ctc-and-embed: True
+share-xctc-and-embed: True
+ctc-weight: 0.05
+xctc-weight: 0.2
+
+# InterCTC
+inter-ctc-weight: 0.025
+inter-ctc-layers: 6,9,12,15
+share-inter-ctc: True
+inter-xctc-weight: 0.1
+inter-xctc-layers: 6,9,12,15
+
+# Prediction-aware encoding
+ctc-pae: inter_league
+xctc-pae: inter_league
+pae-unnorm-input: True
+
+# Curriculum learning mixing
+xctc-pae-ground-truth-ratio: 0.1
+xctc-pae-ground-truth-only-mistake: True 
+pae-oracle-smooth: True
\ No newline at end of file

egs/zerocs/asr/conf/reproduction_encdec_aipa_bilctc_kd.yaml

+arch: s2t_transformer_s
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
+
+# Append-based Interpolation Augmentation
+inter-mixup: True
+
+inter-mixup-layer: -1
+inter-mixup-decoder-layer: 0
+inter-mixup-prob: 1.0
+inter-mixup-ratio: 1.0
+inter-mixup-beta: 0.2
+
+inter-mixup-keep-org: True
+inter-mixup-decoder-emb: True
+
+mixup-no-hard-loss: False
+no-specaugment: False
+layer-out-norm: False
+
+inter-mixup-ratio-decay: False
+inter-mixup-ratio-decay-params: 20000,40000,0
+
+# Bilingual CTC
+share-ctc-and-embed: True
+share-xctc-and-embed: True
+ctc-weight: 0.05
+xctc-weight: 0.2
+
+# InterCTC
+inter-ctc-weight: 0.025
+inter-ctc-layers: 6,9
+share-inter-ctc: True
+inter-xctc-weight: 0.1
+inter-xctc-layers: 6,9
+
+# Prediction-aware encoding
+ctc-pae: inter_league
+xctc-pae: inter_league
+pae-unnorm-input: True
+
+ctc-mixup-consistent-weight: 0.1
+xctc-mixup-consistent-weight: 0.025
+inter-ctc-mixup-consistent-weight: 0.05
+inter-xctc-mixup-consistent-weight: 0.0125
+# mixup-consistent-weight: 0.5
+
+# Conformer
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+layer-padding-mask: True
\ No newline at end of file

egs/zerocs/asr/conf/reproduction_encdec_aipa_kd.yaml

+arch: s2t_transformer_s
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
+
+# Append-based Interpolation Augmentation
+inter-mixup: True
+
+inter-mixup-layer: -1
+inter-mixup-decoder-layer: 0
+inter-mixup-prob: 1.0
+inter-mixup-ratio: 1.0
+inter-mixup-beta: 0.2
+
+inter-mixup-keep-org: True
+inter-mixup-decoder-emb: True
+
+mixup-no-hard-loss: False
+no-specaugment: False
+layer-out-norm: False
+
+inter-mixup-ratio-decay: False
+inter-mixup-ratio-decay-params: 20000,40000,0
+
+# MTL
+ctc-weight: 0.3
+#inter-ctc-weight: 0.2
+#inter-ctc-layers: 6,9
+#share-inter-ctc: True
+share-ctc-and-embed: True
+
+ctc-pae: inter_league
+pae-unnorm-input: True
+
+ctc-mixup-consistent-weight: 0.15
+#inter-ctc-mixup-consistent-weight: 0.1
+# mixup-consistent-weight: 0.5
+
+# Conformer
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+layer-padding-mask: True

egs/zerocs/asr/conf/reproduction_encdec_aipa_kd_nohardloss.yaml

+arch: s2t_transformer_s
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
+
+# Append-based Interpolation Augmentation
+inter-mixup: True
+
+inter-mixup-layer: -1
+inter-mixup-decoder-layer: 0
+inter-mixup-prob: 1.0
+inter-mixup-ratio: 1.0
+inter-mixup-beta: 0.2
+
+inter-mixup-keep-org: True
+inter-mixup-decoder-emb: True
+
+mixup-no-hard-loss: True
+no-specaugment: False
+layer-out-norm: False
+
+inter-mixup-ratio-decay: False
+inter-mixup-ratio-decay-params: 20000,40000,0
+
+# MTL
+ctc-weight: 0.3
+inter-ctc-weight: 0.2
+inter-ctc-layers: 6,9
+share-inter-ctc: True
+share-ctc-and-embed: True
+
+ctc-pae: inter_league
+pae-unnorm-input: True
+
+ctc-mixup-consistent-weight: 0.15
+inter-ctc-mixup-consistent-weight: 0.1
+mixup-consistent-weight: 0.5
+
+# Conformer
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+layer-padding-mask: True
\ No newline at end of file

egs/zerocs/asr/conf/reproduction_purectc_aipa_bilctc_kd.yaml

+arch: s2t_ctc
+encoder-type: transformer
+
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: ctc
+zero_infinity: True
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 18
+encoder-attention-heads: 4
+
+# Append-based Interpolation Augmentation
+inter-mixup: True
+
+inter-mixup-layer: -1
+inter-mixup-decoder-layer: 0
+inter-mixup-prob: 1.0
+inter-mixup-ratio: 1.0
+inter-mixup-beta: 0.2
+
+inter-mixup-keep-org: True
+inter-mixup-decoder-emb: True
+
+mixup-no-hard-loss: False
+no-specaugment: False
+layer-out-norm: False
+
+inter-mixup-ratio-decay: False
+inter-mixup-ratio-decay-params: 20000,40000,0
+
+# Bilingual CTC
+share-ctc-and-embed: True
+share-xctc-and-embed: True
+ctc-weight: 0.2
+xctc-weight: 1
+
+# InterCTC
+inter-ctc-weight: 0.1
+inter-ctc-layers: 6,9,12,15
+share-inter-ctc: True
+inter-xctc-weight: 1.0
+inter-xctc-layers: 6,9,12,15
+
+# Prediction-aware encoding
+ctc-pae: inter_league
+xctc-pae: inter_league
+pae-unnorm-input: True
+
+ctc-mixup-consistent-weight: 0.1
+inter-ctc-mixup-consistent-weight: 0.05
+xctc-mixup-consistent-weight: 0.5
+xinter-ctc-mixup-consistent-weight: 0.5
+
+# Conformer
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+layer-padding-mask: True
\ No newline at end of file

egs/zerocs/asr/conf/reproduction_purectc_aipa_kd.yaml

+arch: s2t_ctc
+encoder-type: transformer
+
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: ctc
+zero_infinity: True
+ctc-weight: 1.0
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 18
+encoder-attention-heads: 4
+
+# Append-based Interpolation Augmentation
+inter-mixup: True
+
+inter-mixup-layer: -1
+inter-mixup-decoder-layer: 0
+inter-mixup-prob: 1.0
+inter-mixup-ratio: 1.0
+inter-mixup-beta: 0.2
+
+inter-mixup-keep-org: True
+inter-mixup-decoder-emb: True
+
+mixup-no-hard-loss: False
+no-specaugment: False
+layer-out-norm: False
+
+inter-mixup-ratio-decay: False
+inter-mixup-ratio-decay-params: 20000,40000,0
+
+# MTL
+inter-ctc-weight: 1.0
+inter-ctc-layers: 6,9,12,15
+share-inter-ctc: True
+share-ctc-and-embed: True
+
+ctc-pae: inter_league
+pae-unnorm-input: True
+
+ctc-mixup-consistent-weight: 0.5
+inter-ctc-mixup-consistent-weight: 0.5
+
+# Conformer
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+layer-padding-mask: True
\ No newline at end of file

egs/zerocs/asr/conf/reproduction_purectc_aipa_kd_nohardloss.yaml

+arch: s2t_ctc
+encoder-type: transformer
+
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: ctc
+zero_infinity: True
+ctc-weight: 1.0
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 18
+encoder-attention-heads: 4
+
+# Append-based Interpolation Augmentation
+inter-mixup: True
+
+inter-mixup-layer: -1
+inter-mixup-decoder-layer: 0
+inter-mixup-prob: 1.0
+inter-mixup-ratio: 1.0
+inter-mixup-beta: 0.2
+
+inter-mixup-keep-org: True
+inter-mixup-decoder-emb: True
+
+mixup-no-hard-loss: True
+no-specaugment: False
+layer-out-norm: False
+
+inter-mixup-ratio-decay: False
+inter-mixup-ratio-decay-params: 20000,40000,0
+
+# MTL
+inter-ctc-weight: 1.0
+inter-ctc-layers: 6,9,12,15
+share-inter-ctc: True
+share-ctc-and-embed: True
+
+ctc-pae: inter_league
+pae-unnorm-input: True
+
+ctc-mixup-consistent-weight: 0.5
+inter-ctc-mixup-consistent-weight: 0.5
+
+# Conformer
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+layer-padding-mask: True
\ No newline at end of file

egs/zerocs/asr/conf/rpr.yaml

+encoder-attention-type: rel_pos
+
+#encoder-attention-type: rel_pos_legacy
+#encoder-attention-type: rel_selfattn
+#encoder-attention-type: relative
+#decoder-attention-type: relative
+#max-encoder-relative-length: 100
+#max-decoder-relative-length: 20

egs/zerocs/asr/conf/xctc.yaml

+xctc-weight: 0.3
+share-xctc-and-embed: True
\ No newline at end of file

egs/zerocs/asr/conf/xinter.yaml

+inter-xctc-weight: 0.2
+inter-xctc-layers: 6,9
+
+xctc-pae: none
+# xctc-pae: inter_league
+
+xctc-cross-attn: False
+cross-attn-start-layer: 7
+cross-attn-layer: 6
+cross-attn-collaboration-mode: parallel
+cross-attn-league-s1-ratio: 0.5
+cross-attn-league-s2-ratio: 0.5
+cross-attn-league-out-norm: False
+cross-attn-league-gated: False
+cross-attn-league-drop-net: False
+cross-attn-league-drop-net-prob: 0.2
+cross-attn-league-drop-net-mix: False
+
+# xctc-pae-ground-truth-ratio: 0.1
+# xctc-pae-ground-truth-ratio-adaptive: True 
+# xctc-pae-ground-truth-only-mistake: True 
+# pae-oracle-smooth: True
+# pae-gumbel: True
+# pae-distribution-hard: True
+# pae-drop-prob: 0.0
+# pae-distribution-cutoff: 10
+# share-pae-and-xctc: True
+# pae-embed-norm: True
+# pae-out-norm: True
+
+# ctc-self-distill-weight: 1
+# target-ctc-self-distill-weight: 1
+# ctc-self-distill-prob: 0.1
+# cal-all-ctc: True
\ No newline at end of file

egs/zerocs/asr/decode.sh

+#!/usr/bin/env bash
+
+gpu_num=1
+
+data_tag=asr
+test_subset=(dev test)
+test_subset=(test)
+
+exp_name=
+if [ "$#" -eq 1 ]; then
+    exp_name=$1
+fi
+
+cer=1
+ctc_infer=1
+n_average=10
+beam_size=5
+infer_ctc_weight=0
+len_penalty=1.0
+max_tokens=50000
+batch_size=0
+infer_debug=0
+dec_model=checkpoint_best.pt
+
+cmd="./run.sh
+    --stage 2
+    --stop_stage 2
+    --gpu_num ${gpu_num}
+    --exp_name ${exp_name}
+    --n_average ${n_average}
+    --cer ${cer}
+    --ctc_infer ${ctc_infer}
+    --beam_size ${beam_size}
+    --len_penalty ${len_penalty}
+    --batch_size ${batch_size}
+    --max_tokens ${max_tokens}
+    --dec_model ${dec_model}
+    --ctc_infer ${ctc_infer}
+    --infer_ctc_weight ${infer_ctc_weight}
+    --infer_debug ${infer_debug}
+    "
+
+if [[ -n ${data_tag} ]]; then
+    cmd="$cmd --data_tag ${data_tag}"
+fi
+if [[ ${#test_subset[@]} -ne 0 ]]; then
+    subsets=$(echo ${test_subset[*]} | sed 's/ /,/g')
+    cmd="$cmd --test_subset ${subsets}"
+fi
+
+echo $cmd
+eval $cmd

egs/zerocs/asr/local/cal_bleu.sh

+#!/usr/bin/env bash
+
+set -e
+
+ref=$1
+gen=$2
+tokenizer=$3
+lang=$4
+lang_pair=en-${lang}
+
+record=$(mktemp -t temp.record.XXXXXX)
+if [[ ${tokenizer} -eq 1 ]]; then
+    echo "MultiBLEU" > ${record}
+    cmd="multi-bleu.perl ${ref} < ${gen}"
+    eval $cmd | head -n 1 >> ${record}
+
+    cmd="detokenizer.perl -q -l ${lang} --threads 32 < ${ref} > ${ref}.detok"
+    eval $cmd
+    cmd="detokenizer.perl -q -l ${lang} --threads 32 < ${gen} > ${gen}.detok"
+    eval $cmd
+    ref=${ref}.detok
+    gen=${gen}.detok
+fi
+
+echo "SacreBLEU" >> ${record}
+cmd="cat ${gen} | sacrebleu ${ref} -m bleu -w 4 -l ${lang_pair}"
+eval $cmd >> ${record}
+cat ${record}
+rm ${record}
\ No newline at end of file

egs/zerocs/asr/local/cal_ctc_bleu.sh

+#!/usr/bin/env bash
+
+set -e
+
+infer_dir=$1
+tag=$2
+s2s_infer_file=${infer_dir}/$3
+org_ctc_infer_file=${infer_dir}/$4
+ref=$5
+tokenizer=$6
+lang=$7
+
+idx=${infer_dir}/${tag}_idx
+ctc_infer=${infer_dir}/${tag}_ctc_infer
+ctc_infer_sort=${infer_dir}/${tag}_ctc_infer_sort
+
+if [[ ! -f ${ctc_infer_sort} ]]; then
+    cut -f1 ${s2s_infer_file} > ${idx}
+    paste ${idx} ${org_ctc_infer_file} > ${ctc_infer}
+    sort -n -t $'\t' ${ctc_infer} | cut -f2 > ${ctc_infer_sort}
+fi
+
+gen=${ctc_infer_sort}
+./cal_bleu.sh ${ref} ${gen} ${tokenizer} ${lang}
\ No newline at end of file

egs/zerocs/asr/local/cal_wer.py

+import unicodedata
+import jiwer
+import jiwer.transforms as tr
+import sys
+
+ref_file = sys.argv[1]
+hyp_file = sys.argv[2]
+
+
+wer_standardize = tr.Compose(
+    [
+        tr.SubstituteRegexes({r"<<unk>>": r"@"}),
+        tr.ToLowerCase(),
+        tr.RemovePunctuation(),
+        tr.ExpandCommonEnglishContractions(),
+        tr.RemoveKaldiNonWords(),
+        tr.RemoveWhiteSpace(replace_by_space=True),
+        tr.ReduceToListOfListOfWords(),
+    ]
+)
+cer_standardize = tr.Compose(
+    [
+        tr.SubstituteRegexes({r"<<unk>>": r"@"}),
+        tr.ToLowerCase(),
+        tr.RemovePunctuation(),
+        tr.Strip(),
+        tr.ReduceToListOfListOfChars(),
+    ]
+)
+
+
+ref_lines = open(ref_file, "r").readlines()
+hyp_lines = open(hyp_file, "r").readlines()
+
+wer = jiwer.wer(ref_lines, hyp_lines,
+                truth_transform=wer_standardize,
+                hypothesis_transform=wer_standardize,
+                )
+cer = jiwer.cer(ref_lines, hyp_lines,
+                truth_transform=cer_standardize,
+                hypothesis_transform=cer_standardize,
+                )
+
+print("WER: %.4f" % wer)
+print("CER: %.4f" % cer)

egs/zerocs/asr/local/cal_wer.sh

+#!/usr/bin/env bash
+
+set -e
+
+infer_dir=$1
+tag=$2
+s2s_infer_file=${infer_dir}/$3
+org_ctc_infer_file=${infer_dir}/$4
+ref=$5
+
+idx=${infer_dir}/${tag}_idx
+ctc_infer=${infer_dir}/${tag}_ctc_infer
+ctc_infer_sort=${infer_dir}/${tag}_ctc_infer_sort
+
+cut -f1 ${s2s_infer_file} > ${idx}
+paste ${idx} ${org_ctc_infer_file} > ${ctc_infer}
+sort -n -t $'\t' ${ctc_infer} | cut -f2 > ${ctc_infer_sort}
+python3 ./cal_wer.py ${ref} ${ctc_infer_sort}
\ No newline at end of file

egs/zerocs/asr/local/extract_txt_from_tsv.py

+import sys
+import csv
+
+tsv_file = sys.argv[1]
+out_file = sys.argv[2]
+extract_item = sys.argv[3]
+
+with open(tsv_file) as f:
+    reader = csv.DictReader(
+        f,
+        delimiter="\t",
+        quotechar=None,
+        doublequote=False,
+        lineterminator="\n",
+        quoting=csv.QUOTE_NONE,
+    )
+    samples = [dict(e) for e in reader]
+
+fw = open(out_file, "w", encoding="utf-8")
+for s in samples:
+    if extract_item in s:
+        fw.write("%s\n" % s[extract_item])
+    else:
+        print("Error in sample: ")
+        print(s)
+        exit()
+

egs/zerocs/asr/local/monitor.sh

+#!/usr/bin/env bash
+
+gpu_num=4
+cmd="sh train.sh"
+
+while :
+do
+    record=$(mktemp -t temp.record.XXXXXX)
+    gpustat > $record
+    all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
+
+    count=0
+    for dev in ${all_devices[@]}
+    do
+        line=$((dev + 2))
+        use=$(head -n $line ${record} | tail -1 | cut -d '|' -f3 | cut -d '/' -f1)
+
+        if [[ $use -lt 100 ]]; then
+            device[$count]=$dev
+            count=$((count + 1))
+            if [[ $count -eq $gpu_num ]]; then
+                break
+            fi
+        fi
+    done
+    if [[ ${#device[@]} -lt $gpu_num ]]; then
+        sleep 60s
+    else
+        echo "Run $cmd"
+        eval $cmd
+        sleep 10s
+        exit
+    fi
+done

egs/zerocs/asr/local/parse_options.sh

+#!/usr/bin/env bash
+
+# Copyright 2012  Johns Hopkins University (Author: Daniel Povey);
+#                 Arnab Ghoshal, Karel Vesely
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Parse command-line options.
+# To be sourced by another script (as in ". parse_options.sh").
+# Option format is: --option-name arg
+# and shell variable "option_name" gets set to value "arg."
+# The exception is --help, which takes no arguments, but prints the
+# $help_message variable (if defined).
+
+
+###
+### The --config file options have lower priority to command line
+### options, so we need to import them first...
+###
+
+# Now import all the configs specified by command-line, in left-to-right order
+for ((argpos=1; argpos<$#; argpos++)); do
+  if [ "${!argpos}" == "--config" ]; then
+    argpos_plus1=$((argpos+1))
+    config=${!argpos_plus1}
+    [ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
+    . $config  # source the config file.
+  fi
+done
+
+
+###
+### Now we process the command line options
+###
+i=1
+argv="$@"
+while true; do
+  key=${!i}
+  j=$(($i + 1))
+  value=${!j}
+
+  [ -z "${!i:-}" ] && break;  # break if there are no arguments
+  case "${key}" in
+    # If the enclosing script is called with --help option, print the help
+    # message and exit.  Scripts should put help messages in $help_message
+    --help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2;
+      else printf "$help_message\n" 1>&2 ; fi;
+      exit 0 ;;
+    --*=*) echo "$0: options to scripts must be of the form --name value, got '${key}}'"
+      exit 1 ;;
+    # If the first command-line argument begins with "--" (e.g. --foo-bar),
+    # then work out the variable name as $name, which will equal "foo_bar".
+    --*) name=`echo "${key}" | sed s/^--// | sed s/-/_/g`;
+      # Next we test whether the variable in question is undefned-- if so it's
+      # an invalid option and we die.  Note: $0 evaluates to the name of the
+      # enclosing script.
+      # The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar
+      # is undefined.  We then have to wrap this test inside "eval" because
+      # foo_bar is itself inside a variable ($name).
+      #eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
+
+      oldval="`eval echo \\$$name`";
+      # Work out whether we seem to be expecting a Boolean argument.
+      if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then
+        was_bool=true;
+      else
+        was_bool=false;
+      fi
+
+      # Set the variable to the right value-- the escaped quotes make it work if
+      # the option had spaces, like --cmd "queue.pl -sync y"
+      # echo $name
+      eval $name=\"${value}\";
+
+      # Check that Boolean-valued arguments are really Boolean.
+      if $was_bool && [[ "${value}" != "true" && "${value}" != "false" ]]; then
+        echo "$0: expected \"true\" or \"false\": ${key} ${value}" 1>&2
+        exit 1;
+      fi
+      # shift 2;
+      i=$(($i + 2))
+      ;;
+  *) break;
+  esac
+done
+
+
+# Check for an empty argument to the --cmd option, which can easily occur as a
+# result of scripting errors.
+[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
+
+
+true; # so this script returns exit code 0.

egs/zerocs/asr/local/utils.sh

+
+get_devices(){
+    gpu_num=$1
+    use_cpu=$2
+    device=()
+    while :
+    do
+        record=$(mktemp -t temp.record.XXXXXX)
+        gpustat > $record
+        all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
+
+        count=0
+        for dev in ${all_devices[@]}
+        do
+            line=$((dev + 2))
+            use=$(head -n $line ${record} | tail -1 | cut -d '|' -f3 | cut -d '/' -f1)
+            if [[ $use -lt 1000 ]]; then
+                device[$count]=$dev
+                count=$((count + 1))
+                if [[ $count -eq $gpu_num ]]; then
+                    break
+                fi
+            fi
+        done
+        if [[ ${#device[@]} -lt $gpu_num ]]; then
+            if [[ $use_cpu -eq 1 ]]; then
+                device=(-1)
+            else
+                sleep 60s
+            fi
+        else
+            break
+        fi
+    done
+
+    echo ${device[*]} | sed 's/ /,/g'
+    return $?
+}
+
+

egs/zerocs/asr/run.sh

+#!/usr/bin/env bash
+
+# Processing AIShell ASR Datasets
+
+# Copyright 2021 Chen Xu (xuchennlp@outlook.com)
+
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+#set -u
+set -o pipefail
+export PYTHONIOENCODING=UTF-8
+
+eval=1
+time=$(date "+%m%d_%H%M")
+
+stage=1
+stop_stage=2
+
+######## Hardware ########
+# Devices
+device=(0)
+gpu_num=2
+update_freq=1
+max_tokens=50000
+
+pwd_dir=$PWD
+root_dir=${ST_ROOT}
+data_root_dir=${root_dir}
+
+code_dir=${root_dir}/S2T
+
+# Dataset
+src_lang=zh
+lang=${src_lang}
+dataset=zerocs
+data_tag=asr
+
+task=speech_to_text
+vocab_type=unigram
+vocab_size=32000
+speed_perturb=0
+lcrm=0
+tokenizer=0
+use_raw_audio=0
+
+. ./local/parse_options.sh || exit 1;
+
+use_specific_dict=0
+specific_prefix=st
+specific_dir=${root_dir}/data/mustc/st
+asr_vocab_prefix=spm_unigram10000_st_share
+
+data_model_subfix=${dataset}/${data_tag}
+org_data_dir=${data_root_dir}/data/${dataset}
+data_dir=${data_root_dir}/data/${data_model_subfix}
+train_split=train_zh_en,train_en_zh,train_cs
+train_split=train_zh,train_en,train_cs
+valid_split=dev_zh,dev_en,dev_cs
+test_split=test_zh,test_en,test_cs
+test_subset=test_zh,test_en,test_cs,test_cs_zh
+tgt_langs=zh,en,cs
+joint=1
+
+# exp
+sub_tag=
+exp_prefix=$(date "+%m%d")
+extra_tag=
+extra_parameter=
+exp_tag=baseline
+exp_name=
+
+# Training Settings
+train_config=base,ctc
+fp16=1
+step_valid=0
+
+# Decoding Settings
+dec_model=checkpoint_best.pt
+cer=0
+ctc_infer=0
+infer_ctc_weight=0
+ctc_self_ensemble=0
+ctc_inter_logit=0
+n_average=10
+batch_size=0
+beam_size=5
+len_penalty=1.0
+single=0
+epoch_ensemble=0
+best_ensemble=1
+infer_debug=0
+infer_score=0
+# infer_parameters="--cal-monotonic-cross-attn-weights --cal-localness --localness-window 0.1 --cal-topk-cross-attn-weights --topk-cross-attn-weights 15 --cal-entropy"
+
+data_config=config.yaml
+
+# Parsing Options
+if [[ ${speed_perturb} -eq 1 ]]; then
+    data_dir=${data_dir}_sp
+    exp_prefix=${exp_prefix}_sp
+fi
+if [[ ${lcrm} -eq 1 ]]; then
+    data_dir=${data_dir}_lcrm
+    exp_prefix=${exp_prefix}_lcrm
+fi
+if [[ ${use_specific_dict} -eq 1 ]]; then
+    data_dir=${data_dir}_${specific_prefix}
+    exp_prefix=${exp_prefix}_${specific_prefix}
+fi
+if [[ ${tokenizer} -eq 1 ]]; then
+    data_dir=${data_dir}_tok
+    exp_prefix=${exp_prefix}_tok
+fi
+if [[ ${use_raw_audio} -eq 1 ]]; then
+    data_dir=${data_dir}_raw
+    exp_prefix=${exp_prefix}_raw
+fi
+
+export PATH=$PATH:${code_dir}/scripts
+. ./local/parse_options.sh || exit 1;
+
+if [[ -z ${exp_name} ]]; then
+    config_string=${train_config//,/_}
+    exp_name=${exp_prefix}_${config_string}_${exp_tag}
+    if [[ -n ${extra_tag} ]]; then
+        exp_name=${exp_name}_${extra_tag}
+    fi
+    if [[ -n ${exp_subfix} ]]; then
+        exp_name=${exp_name}_${exp_subfix}
+    fi
+fi
+
+ckpt_dir=${root_dir}/checkpoints/
+model_dir=${root_dir}/checkpoints/${data_model_subfix}/${sub_tag}/${exp_name}
+
+# Start
+cd ${code_dir}
+echo "Start Stage: $stage"
+echo "Stop  Stage: $stop_stage"
+
+if [[ `pip list | grep fairseq | wc -l` -eq 0 ]]; then 
+    echo "Default Stage: env configure"
+    pip3 install -e ${code_dir}
+fi
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+    echo "Stage -1: Data Download"
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    ### Task dependent. You have to make data the following preparation part by yourself.
+    echo "Stage 0: Data Preparation"
+
+    if [[ ! -e ${data_dir} ]]; then
+        mkdir -p ${data_dir}
+    fi
+
+    cmd="python3 ${code_dir}/examples/speech_to_text/prep_audio_data.py
+        --data-root ${org_data_dir}
+        --output-root ${data_dir}
+        --task asr
+        --src-lang ${src_lang}
+        --splits ${valid_split},${test_split},${train_split}
+        --add-src
+        --share
+        --vocab-type ${vocab_type}
+        --vocab-size ${vocab_size}"
+
+    if [[ ${joint} -eq 1 ]]; then
+        cmd="$cmd
+	--add-syms
+	--tgt-langs ${tgt_langs}"
+    fi
+    if [[ ${use_raw_audio} -eq 1 ]]; then
+        cmd="$cmd
+        --raw"
+    fi
+    if [[ ${use_specific_dict} -eq 1 ]]; then
+        cp -r ${specific_dir}/${asr_vocab_prefix}.* ${data_dir}
+        cmd="$cmd
+        --asr-prefix ${asr_vocab_prefix}"
+    fi
+    if [[ ${speed_perturb} -eq 1 ]]; then
+        cmd="$cmd
+        --speed-perturb"
+    fi
+    if [[ ${lcrm} -eq 1 ]]; then
+        cmd="$cmd
+        --lowercase-src
+        --rm-punc-src"
+    fi
+    if [[ ${tokenizer} -eq 1 ]]; then
+        cmd="$cmd
+        --tokenizer"
+    fi
+
+    echo -e "\033[34mRun command: \n${cmd} \033[0m"
+    [[ $eval -eq 1 ]] && eval ${cmd}
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    echo "Stage 1: Network Training"
+    [[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1;
+
+    if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
+        if [[ ${gpu_num} -eq 0 ]]; then
+            device=""
+        else
+            source ./local/utils.sh
+            device=$(get_devices $gpu_num 0)
+        fi
+        export CUDA_VISIBLE_DEVICES=${device}
+    fi
+
+    echo -e "data=${data_dir} model=${model_dir}"
+
+    if [[ ! -d ${model_dir} ]]; then
+        mkdir -p ${model_dir}
+    else
+        echo "${model_dir} exists."
+    fi
+
+    cp -f ${pwd_dir}/`basename ${BASH_SOURCE[0]}` ${model_dir}
+    cp -f ${pwd_dir}/train.sh ${model_dir}
+
+    train_config=basis,${train_config}
+    config_list="${train_config//,/ }"
+    idx=0
+    for config in ${config_list[@]}
+    do
+        config_path=${pwd_dir}/conf/${config}.yaml
+        if [[ ! -f ${config_path} ]]; then
+            echo "No config file ${config_path}"
+            exit
+        fi
+        cp -f ${config_path} ${model_dir}
+
+        if [[ $idx -eq 0 ]]; then
+        extra_parameter="${extra_parameter}
+        --train-config ${config_path}"
+        else
+        extra_parameter="${extra_parameter}
+        --train-config${idx} ${config_path}"
+        fi
+        idx=$((idx + 1))
+    done
+
+    cmd="python3 -u ${code_dir}/fairseq_cli/train.py
+        ${data_dir}
+        --config-yaml ${data_config}
+        --task ${task}
+        --max-tokens ${max_tokens}
+        --skip-invalid-size-inputs-valid-test
+        --update-freq ${update_freq}
+        --log-interval 100
+        --save-dir ${model_dir}
+        --tensorboard-logdir ${model_dir}"
+
+    if [[ -n ${extra_parameter} ]]; then
+        cmd="${cmd}
+        ${extra_parameter}"
+    fi
+    if [[ ${gpu_num} -gt 0 ]]; then
+        cmd="${cmd}
+        --distributed-world-size $gpu_num
+        --ddp-backend no_c10d"
+    fi
+    if [[ $joint -eq 1 ]]; then
+        cmd="${cmd}
+	--prefix-size 1
+        --ignore-prefix-size 1"
+    fi
+    if [[ $fp16 -eq 1 ]]; then
+        cmd="${cmd}
+        --fp16"
+    fi
+    if [[ $step_valid -eq 1 ]]; then
+        validate_interval=1
+        save_interval=1
+        no_epoch_checkpoints=0
+        save_interval_updates=500
+        keep_interval_updates=10
+    fi
+    if [[ -n $no_epoch_checkpoints && $no_epoch_checkpoints -eq 1 ]]; then
+        cmd="$cmd
+        --no-epoch-checkpoints"
+    fi
+    if [[ -n $validate_interval ]]; then
+        cmd="${cmd}
+        --validate-interval $validate_interval "
+    fi
+    if [[ -n $save_interval ]]; then
+        cmd="${cmd}
+        --save-interval $save_interval "
+    fi
+    if [[ -n $save_interval_updates ]]; then
+        cmd="${cmd}
+        --save-interval-updates $save_interval_updates"
+        if [[ -n $keep_interval_updates ]]; then
+        cmd="${cmd}
+        --keep-interval-updates $keep_interval_updates"
+        fi
+    fi
+
+    echo -e "\033[34mRun command: \n${cmd} \033[0m"
+
+    # save info
+    log=${ckpt_dir}/history.log
+    echo "${time} | ${data_dir} | ${exp_name} | ${model_dir} " >> $log
+    tail -n 50 ${log} > tmp.log
+    mv tmp.log $log
+
+    log=${model_dir}/train.log
+    cmd="${cmd} 2>&1 | tee -a ${log}"
+    #cmd="${cmd} >> ${log} 2>&1 "
+    if [[ $eval -eq 1 ]]; then
+        # tensorboard
+        port=6666
+        tensorboard --logdir ${model_dir} --port ${port} --bind_all &
+    
+        echo "${cmd}" > ${model_dir}/cmd
+        eval $cmd
+    fi
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    echo "Stage 2: Decoding"
+    dec_models=
+    if [[ ${n_average} -eq 1 ]]; then
+        dec_models=${dec_model}
+    fi
+    if [[ ${n_average} -ne 1 ]]; then
+        # Average models
+        if [[ ${epoch_ensemble} -eq 1 ]]; then
+            avg_model=avg_epoch${n_average}_checkpoint.pt
+
+            if [[ ! -f ${model_dir}/${avg_model} ]]; then
+                cmd="python3 ${code_dir}/scripts/average_checkpoints.py
+                --inputs ${model_dir}
+                --num-epoch-checkpoints ${n_average}
+                --output ${model_dir}/${avg_model}"
+                echo -e "\033[34mRun command: \n${cmd} \033[0m"
+                [[ $eval -eq 1 ]] && eval $cmd
+            fi
+            dec_models+=(${avg_model})
+        fi
+        if [[ ${best_ensemble} -eq 1 ]]; then
+            avg_model=avg_best${n_average}_checkpoint.pt
+
+            if [[ ! -f ${model_dir}/${avg_model} ]]; then
+                cmd="python3 ${code_dir}/scripts/average_checkpoints.py
+                --inputs ${model_dir}
+                --num-best-checkpoints ${n_average}
+                --output ${model_dir}/${avg_model}"
+                echo -e "\033[34mRun command: \n${cmd} \033[0m"
+                [[ $eval -eq 1 ]] && eval $cmd
+            fi
+            dec_models+=(${avg_model})
+        fi
+    fi
+
+    if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
+        if [[ ${gpu_num} -eq 0 ]]; then
+            device=""
+        else
+            source ./local/utils.sh
+            device=$(get_devices $gpu_num 0)
+        fi
+        export CUDA_VISIBLE_DEVICES=${device}
+    fi
+
+    for dec_model in ${dec_models[@]}; do
+        suffix=alpha${len_penalty}
+        model_str=`echo $dec_model | sed -e "s#checkpoint##" | sed "s#.pt##"`
+        suffix=${suffix}_${model_str}
+        if [[ -n ${cer} && ${cer} -eq 1 ]]; then
+            suffix=${suffix}_cer
+        else
+            suffix=${suffix}_wer
+        fi
+
+        suffix=${suffix}_beam${beam_size}
+        if [[ ${batch_size} -ne 0 ]]; then
+            suffix=${suffix}_batch${batch_size}
+        else
+            suffix=${suffix}_tokens${max_tokens}
+        fi
+        if [[ ${ctc_infer} -eq 1 ]]; then
+            suffix=${suffix}_ctc
+        fi
+        if [[ ${ctc_self_ensemble} -eq 1 ]]; then
+            suffix=${suffix}_ensemble
+        fi
+        if [[ ${ctc_inter_logit} -ne 0 ]]; then
+            suffix=${suffix}_logit${ctc_inter_logit}
+        fi
+        if (( $(echo "${infer_ctc_weight} > 0" | bc -l) )); then
+            suffix=${suffix}_ctc${infer_ctc_weight}
+        fi
+        if [[ ${infer_score} -eq 1 ]]; then
+            suffix=${suffix}_score
+        fi
+
+        suffix=`echo $suffix | sed -e "s#__#_#"`
+        result_file=${model_dir}/decode_result_${suffix}
+        [[ -f ${result_file} ]] && rm ${result_file}
+
+        test_subset=${test_subset//,/ }
+        for subset in ${test_subset[@]}; do
+            subset=${subset}
+            if [[ ${infer_debug} -ne 0 ]]; then
+                cmd="python3 -m debugpy --listen 0.0.0.0:5678 --wait-for-client"
+            else
+                cmd="python3 "
+            fi
+            cmd="$cmd ${code_dir}/fairseq_cli/generate.py
+            ${data_dir}
+            --config-yaml ${data_config}
+            --gen-subset ${subset}
+            --task speech_to_text
+            --path ${model_dir}/${dec_model}
+            --results-path ${model_dir}
+            --batch-size ${batch_size}
+            --max-tokens ${max_tokens}
+            --beam ${beam_size}
+            --lenpen ${len_penalty}
+            --infer-ctc-weight ${infer_ctc_weight}
+            --scoring wer"
+
+            if [[ ${cer} -eq 1 ]]; then
+                cmd="${cmd}
+            --wer-char-level"
+            fi
+            if [[ ${ctc_infer} -eq 1 ]]; then
+                cmd="${cmd}
+            --ctc-infer"
+            fi
+            if [[ ${ctc_self_ensemble} -eq 1 ]]; then
+                cmd="${cmd}
+            --ctc-self-ensemble"
+            fi
+            if [[ ${ctc_inter_logit} -ne 0 ]]; then
+                cmd="${cmd}
+            --ctc-inter-logit ${ctc_inter_logit}"
+            fi
+            if [[ ${infer_score} -eq 1 ]]; then
+                cmd="${cmd}
+            --score-reference"
+            fi
+            if [[ -n ${infer_parameters} ]]; then
+                cmd="${cmd}
+            ${infer_parameters}"
+            fi                        
+            if [[ $joint -eq 1 ]]; then
+                cmd="${cmd}
+            --prefix-size 1"
+            fi
+
+            echo -e "\033[34mRun command: \n${cmd} \033[0m"
+
+            cd ${code_dir}
+            if [[ $eval -eq 1 ]]; then
+                ctc_file=translation-${subset}.ctc
+                xctc_file=translation-${subset}.xctc
+                if [[ -f ${model_dir}/${ctc_file} ]]; then
+                    rm ${model_dir}/${ctc_file}
+                fi
+                if [[ -f ${model_dir}/${xctc_file} ]]; then
+                    rm ${model_dir}/${xctc_file}
+                fi
+
+                eval $cmd
+                echo "" >> ${result_file}
+                tail -n 2 ${model_dir}/generate-${subset}.txt >> ${result_file}
+                mv ${model_dir}/generate-${subset}.txt ${model_dir}/generate-${subset}-${suffix}.txt
+                mv ${model_dir}/translation-${subset}.txt ${model_dir}/translation-${subset}-${suffix}.txt
+
+                cd ${pwd_dir}
+                if [[ -f ${model_dir}/enc_dump ]]; then
+                    mv ${model_dir}/enc_dump ${model_dir}/dump-${subset}-enc-${suffix}
+                fi
+                if [[ -f ${model_dir}/dec_dump ]]; then
+                    mv ${model_dir}/dec_dump ${model_dir}/dump-${subset}-dec-${suffix}
+                fi
+
+                trans_file=translation-${subset}-${suffix}.txt
+                if [[ ! -f ${model_dir}/{ctc_file} && -f ${model_dir}/${xctc_file} ]]; then
+                    ctc_file=${xctc_file}
+                fi
+                if [[ ${ctc_infer} -eq 1 && -f ${model_dir}/${ctc_file} ]]; then
+                    ref_file=${model_dir}/${subset}.${src_lang}
+                    if [[ ! -f ${ref_file} ]]; then
+                        python3 ./local/extract_txt_from_tsv.py ${data_dir}/${subset}.tsv ${ref_file} "tgt_text"
+                    fi
+                    if [[ -f ${ref_file} ]]; then
+                        ctc=$(mktemp -t temp.record.XXXXXX)
+                        cd ./local
+                        cmd="./cal_wer.sh ${model_dir} ${subset} ${trans_file} ${ctc_file} ${ref_file} > ${ctc}"
+                        eval $cmd
+                        cd ..
+
+                        echo "CTC WER" >> ${result_file}
+                        tail -n 2 ${ctc} >> ${result_file}
+
+                        src_bleu=$(mktemp -t temp.record.XXXXXX)
+                        cd local
+                        ./cal_ctc_bleu.sh ${model_dir} ${subset} ${trans_file} ${ctc_file} ${ref_file} ${tokenizer} ${src_lang} > ${src_bleu}
+                        cd ..
+                        cat ${src_bleu} >> ${result_file}
+
+                        rm ${ctc} ${src_bleu}
+                    else
+                        echo "No reference for source language."
+                    fi
+                fi
+            fi
+        done
+        echo
+        cat ${result_file}
+    done
+fi

egs/zerocs/asr/train.sh

+#!/usr/bin/env bash
+
+# training the model
+
+gpu_num=2
+update_freq=1
+max_tokens=100000
+
+extra_tag=
+extra_parameter=
+#extra_tag="${extra_tag}"
+#extra_parameter="${extra_parameter} "
+
+exp_tag=
+
+# CTC
+config_list=(purectc)
+
+# Transformer
+config_list=(base ctc)
+
+# Conformer
+#config_list=(base conformer ctc)
+
+# PDS
+config_list=(purectc_pds_base_8)
+config_list=(pds_base_8)
+
+# exp full name
+exp_name=
+
+train_config=$(echo ${config_list[*]} | sed 's/ /,/g')
+
+cmd="./run.sh
+    --stage 1
+    --stop_stage 2
+    --gpu_num ${gpu_num}
+    --update_freq ${update_freq}
+    --train_config ${train_config}
+    --max_tokens ${max_tokens}
+    "
+
+if [[ -n ${exp_name} ]]; then
+    cmd="$cmd --exp_name ${exp_name}"
+fi
+if [[ -n ${exp_tag} ]]; then
+    cmd="$cmd --exp_tag ${exp_tag}"
+fi
+if [[ -n ${extra_tag} ]]; then
+    cmd="$cmd --extra_tag ${extra_tag}"
+fi
+if [[ -n ${extra_parameter} ]]; then
+    cmd="$cmd --extra_parameter \"${extra_parameter}\""
+fi
+
+echo ${cmd}
+eval ${cmd}