accu update

egs/librispeech/asr/nohup.out

-./run.sh --stage 2 --stop_stage 2 --gpu_num 1 --exp_name 1123_purectc_conformer_norm_100h_layer36_interw1_x1 --n_average 10 --beam_size 1 --len_penalty 1.0 --batch_size 1 --max_tokens 50000 --dec_model checkpoint_best.pt --ctc_infer 0 --infer_ctc_weight 0 --infer_debug 0 --data_tag asr --test_subset dev-clean,dev-other,test-clean,test-other
-Start Stage: 2
-Stop  Stage: 2
-Stage 2: Decoding
-Run command: 
-python3  /xuchen/st/S2T/fairseq_cli/generate.py
-            /xuchen/st/data/librispeech/asr
-            --config-yaml config.yaml
-            --gen-subset dev-clean
-            --task speech_to_text
-            --path /xuchen/st/checkpoints/librispeech/asr//1123_purectc_conformer_norm_100h_layer36_interw1_x1/avg_best10_checkpoint.pt
-            --results-path /xuchen/st/checkpoints/librispeech/asr//1123_purectc_conformer_norm_100h_layer36_interw1_x1
-            --batch-size 1
-            --max-tokens 50000
-            --beam 1
-            --lenpen 1.0
-            --infer-ctc-weight 0
-            --scoring wer
-            --early-exit-count 6 
-Run command: 
-python3  /xuchen/st/S2T/fairseq_cli/generate.py
-            /xuchen/st/data/librispeech/asr
-            --config-yaml config.yaml
-            --gen-subset dev-other
-            --task speech_to_text
-            --path /xuchen/st/checkpoints/librispeech/asr//1123_purectc_conformer_norm_100h_layer36_interw1_x1/avg_best10_checkpoint.pt
-            --results-path /xuchen/st/checkpoints/librispeech/asr//1123_purectc_conformer_norm_100h_layer36_interw1_x1
-            --batch-size 1
-            --max-tokens 50000
-            --beam 1
-            --lenpen 1.0
-            --infer-ctc-weight 0
-            --scoring wer
-            --early-exit-count 6 
-Run command: 
-python3  /xuchen/st/S2T/fairseq_cli/generate.py
-            /xuchen/st/data/librispeech/asr
-            --config-yaml config.yaml
-            --gen-subset test-clean
-            --task speech_to_text
-            --path /xuchen/st/checkpoints/librispeech/asr//1123_purectc_conformer_norm_100h_layer36_interw1_x1/avg_best10_checkpoint.pt
-            --results-path /xuchen/st/checkpoints/librispeech/asr//1123_purectc_conformer_norm_100h_layer36_interw1_x1
-            --batch-size 1
-            --max-tokens 50000
-            --beam 1
-            --lenpen 1.0
-            --infer-ctc-weight 0
-            --scoring wer
-            --early-exit-count 6 
-Run command: 
-python3  /xuchen/st/S2T/fairseq_cli/generate.py
-            /xuchen/st/data/librispeech/asr
-            --config-yaml config.yaml
-            --gen-subset test-other
-            --task speech_to_text
-            --path /xuchen/st/checkpoints/librispeech/asr//1123_purectc_conformer_norm_100h_layer36_interw1_x1/avg_best10_checkpoint.pt
-            --results-path /xuchen/st/checkpoints/librispeech/asr//1123_purectc_conformer_norm_100h_layer36_interw1_x1
-            --batch-size 1
-            --max-tokens 50000
-            --beam 1
-            --lenpen 1.0
-            --infer-ctc-weight 0
-            --scoring wer
-            --early-exit-count 6 
-
-
-2023-11-26 00:02:34 | INFO | fairseq_cli.generate | Translated 2,703 sentences (59,825 tokens) in 169.3s (15.96 sentences/s, 353.34 tokens/s)
-Generate dev-clean with beam=1: WER: 8.96
-
-2023-11-26 00:07:24 | INFO | fairseq_cli.generate | Translated 2,864 sentences (55,375 tokens) in 219.3s (13.06 sentences/s, 252.51 tokens/s)
-Generate dev-other with beam=1: WER: 21.21
-
-2023-11-26 00:10:57 | INFO | fairseq_cli.generate | Translated 2,620 sentences (57,617 tokens) in 158.0s (16.59 sentences/s, 364.77 tokens/s)
-Generate test-clean with beam=1: WER: 9.34
-
-2023-11-26 00:15:40 | INFO | fairseq_cli.generate | Translated 2,939 sentences (56,865 tokens) in 226.7s (12.96 sentences/s, 250.80 tokens/s)
-Generate test-other with beam=1: WER: 21.57

egs/mustc/joint/binary.sh

+set -e
+
+eval=1
+
+lcrm=1
+tokenizer=0
+
+vocab_type=unigram
+vocab_size=5000
+use_raw_audio=0
+speed_perturb=0
+
+dataset=mustc
+root_dir=~/st
+code_dir=${root_dir}/Fairseq-S2T
+org_data_dir=${root_dir}/data/${dataset}
+data_dir=${root_dir}/data/${dataset}/st
+
+use_specific_dict=0
+specific_prefix=st
+specific_dir=${root_dir}/data/mustc/st
+asr_vocab_prefix=spm_unigram10000_st_share
+
+src_lang=en
+tgt_lang=zh
+subsets=(2019)
+
+splits=$(echo ${subsets[*]} | sed 's/ /_/g')
+cmd="python ${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}
+    --tgt-lang ${tgt_lang}
+    --splits ${splits}
+    --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}

egs/mustc/joint/conf/aipa.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
+
+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/mustc/joint/conf/aipa_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
+
+ctc-mixup-consistent-weight: 0.15
+inter-ctc-mixup-consistent-weight: 0.1
+mixup-consistent-weight: 0.5
+
+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/mustc/joint/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
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: 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: 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:

egs/mustc/joint/conf/basis.yaml

+train-subset: train_mix_1,train_mix_2
+valid-subset: dev_en,dev_en2de
+ignore-prefix-size: 1
+use-additional-ctc-text: true
+sharded-data-load: true
+
+max-epoch: 100
+max-update: 300000
+patience: 20
+post-process: sentencepiece
+
+best_checkpoint_metric: loss
+maximize_best_checkpoint_metric: False
+
+#eval-bleu: True
+eval-bleu-args: {"beam": 5, "lenpen": 1.0, "prefix_size": 1}
+eval-bleu-detok: moses
+eval-bleu-remove-bpe: sentencepiece
+eval-bleu-print-samples: True
+#best_checkpoint_metric: bleu
+#maximize_best_checkpoint_metric: True
+
+# no-epoch-checkpoints: True
+validate-interval: 1
+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/mustc/joint/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: 1e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+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: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:

egs/mustc/joint/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/mustc/joint/conf/ctc.yaml

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

egs/mustc/joint/conf/ctc_aug_base.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-encoder-no-scale-embedding: 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: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: transformer
+adapter: none
+
+# load-pretrained-encoder-from: 
+# load-pretrained-acoustic-encoder-from:
+# load-pretrained-text-encoder-from:
+# load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/ctc_aug_big.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-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: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+acoustic-encoder: transformer
+adapter: none
+
+# load-pretrained-encoder-from: 
+# load-pretrained-acoustic-encoder-from:
+# load-pretrained-text-encoder-from:
+# load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/ctc_aug_pds_big.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-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-layers: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+acoustic-encoder: pds
+adapter: none
+
+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
+
+# load-pretrained-encoder-from: 
+# load-pretrained-acoustic-encoder-from:
+# load-pretrained-text-encoder-from:
+# load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/dlcl.yaml

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

egs/mustc/joint/conf/dual.yaml

+arch: s2t_dual
+
+asr-encoder: transformer
+mt-encoder-layers: 6
+mt-encoder: transformer
+
+encoder-collaboration-mode: parallel
+decoder-collaboration-mode: parallel
+
+encoder-league-s1-ratio: 0.5
+encoder-league-s2-ratio: 0.5
+encoder-league-drop-net: False
+encoder-league-drop-net-prob: 0.2
+encoder-league-drop-net-mix: False
+
+decoder-league-s1-ratio: 0.5
+decoder-league-s2-ratio: 0.5
+decoder-league-drop-net: False
+decoder-league-drop-net-prob: 0.0
+decoder-league-drop-net-mix: False
+
+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: join_speech_and_text_loss
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-asr-encoder-from:
+#load-pretrained-mt-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/dynamic.yaml

+compression-metric: threshold
+compression-mode: create
+compression-layers: 6,9
+compression-threshold: 0.99
+compression-norm: True
+compression-pos: True
\ No newline at end of file

egs/mustc/joint/conf/inter.yaml

+inter-ctc-weight: 0.2
+inter-ctc-layers: 6,9
+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/mustc/joint/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/mustc/joint/conf/mlo.yaml

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

egs/mustc/joint/conf/multibranch.yaml

+arch: s2t_multibranch
+
+junior-acoustic-encoder: transformer
+acoustic-adapter: none
+textual-adapter: none
+junior-acoustic-encoder-layers: 12
+senior-acoustic-encoder-layers: 6
+textual-encoder-layers: 6
+
+# collaboration-direction: none
+# collaboration-direction: acoustic
+# collaboration-direction: textual
+collaboration-direction: both
+collaboration-start: 0:0
+collaboration-step: 1:1
+# encoder-collaboration-mode: serial
+# decoder-collaboration-mode: serial
+encoder-collaboration-mode: parallel
+decoder-collaboration-mode: parallel
+
+use-raw-text: False
+modality-switch: False
+text-to-hidden-progress: 1:0:500
+
+encoder-league-s1-ratio: 0.5
+encoder-league-s2-ratio: 0.5
+encoder-league-out-norm: False
+encoder-league-gated: False
+encoder-league-drop-net: False
+encoder-league-drop-net-prob: 0.2
+encoder-league-drop-net-mix: False
+
+decoder-league-s1-ratio: 0.5
+decoder-league-s2-ratio: 0.5
+decoder-league-out-norm: False
+decoder-league-gated: False
+decoder-league-drop-net: False
+decoder-league-drop-net-prob: 0.0
+decoder-league-drop-net-mix: False
+
+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-attention-heads: 4
+
+decoder-layers: 6
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+#adapter: inter_league
+#adapter: none
+#adapter-embed-norm: True
+#adapter-out-norm: True
+#share-adapter-and-ctc: True
+#share-adapter-and-embed: True
+
+#load-pretrained-junior-acoustic-encoder-from:
+#load-pretrained-senior-acoustic-encoder-from:
+#load-pretrained-textual-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/nast_base.yaml

+arch: s2t_ctc
+encoder-type: sate
+
+criterion: ctc
+zero_infinity: True
+xctc-weight: 1.0
+ctc-weight: 1.0
+
+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)
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-encoder-no-scale-embedding: 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: 12
+text-encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: transformer
+adapter: none
\ No newline at end of file

egs/mustc/joint/conf/nast_big.yaml

+arch: s2t_ctc
+encoder-type: sate
+
+criterion: ctc
+zero_infinity: True
+xctc-weight: 1.0
+ctc-weight: 1.0
+
+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
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-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: 12
+text-encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+acoustic-encoder: transformer
+adapter: none
\ No newline at end of file

egs/mustc/joint/conf/nast_pds_big.yaml

+arch: s2t_ctc
+encoder-type: sate
+
+criterion: ctc
+zero_infinity: True
+xctc-weight: 1.0
+ctc-weight: 1.0
+
+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
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-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-layers: 12
+text-encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+acoustic-encoder: pds
+adapter: none
+
+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
\ No newline at end of file

egs/mustc/joint/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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/pds_base_4_444.yaml

+arch: pdss2t_transformer_s_8
+
+encoder-embed-dim: 256
+pds-stages: 3
+pds-layers: 4_4_4
+pds-ratios: 2_2_1
+pds-fusion: False
+pds-fusion-method: all_conv2
+pds-fusion-layers: 1_1_1 
+pds-fusion-weight: 0.2_0.3_0.5
+pds-embed-dims: 256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1
+pds-kernel-sizes: 5_5_5
+pds-ffn-ratios: 8_8_8
+pds-attn-heads: 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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/pds_base_8_444.yaml

+arch: pdss2t_transformer_s_8
+
+encoder-embed-dim: 256
+pds-stages: 3
+pds-layers: 4_4_4
+pds-ratios: 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
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1
+pds-kernel-sizes: 5_5_5
+pds-ffn-ratios: 8_8_8
+pds-attn-heads: 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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/pds_big_16.yaml

+arch: pdss2t_transformer_m_16
+
+encoder-embed-dim: 512
+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: 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: 1e-3
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/pds_big_32.yaml

+arch: pdss2t_transformer_m_32
+
+encoder-embed-dim: 512
+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: 512_512_512_512_512
+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: 4_4_4_4_4
+pds-attn-heads: 8_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: 1e-3
+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
\ No newline at end of file

egs/mustc/joint/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: 1e-3
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/pds_big_deep_16.yaml

+arch: pdss2t_transformer_s_16
+
+encoder-embed-dim: 512
+pds-stages: 4
+pds-layers: 3_3_8_4
+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: 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: 8_8_8_8
+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: 1e-3
+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: 18
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/pds_big_deep_32.yaml

+arch: pdss2t_transformer_s_16
+
+encoder-embed-dim: 512
+pds-stages: 5
+pds-layers: 3_3_4_4_4
+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: 512_512_512_512_512
+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: 8_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: 1e-3
+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: 18
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/pds_big_deep_8.yaml

+arch: pdss2t_transformer_s_8
+
+encoder-embed-dim: 512
+pds-stages: 4
+pds-layers: 4_5_5_4
+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: 8_8_8_8
+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: 1e-3
+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: 18
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/pds_deep_16.yaml

+arch: pdss2t_transformer_sd_16
+
+encoder-embed-dim: 256
+pds-stages: 4
+pds-layers: 3_3_8_4
+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: 18
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/pds_deep_32.yaml

+arch: pdss2t_transformer_sd_32
+
+encoder-embed-dim: 256
+pds-stages: 5
+pds-layers: 3_3_4_4_4
+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: 18
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/pds_deep_8.yaml

+arch: pdss2t_transformer_sd_8
+
+encoder-embed-dim: 256
+pds-stages: 4
+pds-layers: 4_5_5_4
+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: 18
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4

egs/mustc/joint/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/joint/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/joint/conf/reproduction_aipa_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
+
+# 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/mustc/joint/conf/reproduction_aipa_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
+
+# MTL
+ctc-weight: 0.3
+share-ctc-and-embed: True
+inter-ctc-weight: 0.2
+inter-ctc-layers: 6,9
+share-inter-ctc: 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/mustc/joint/conf/reproduction_bil_ctc_progressive.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-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: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+acoustic-encoder: transformer
+adapter: inter_league
+
+# 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: 0.2
+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: 4
+
+# 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

egs/mustc/joint/conf/reproduction_bil_ctc_progressive2.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
+
+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
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+# 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: 0.2
+ctc-layer: 12
+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: 16
+
+# 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

egs/mustc/joint/conf/reproduction_bil_ctc_progressive2_256.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: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+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: 0.1
+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
+
+# 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: 0.2
+ctc-layer: 12
+#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: 16
+
+# 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

egs/mustc/joint/conf/reproduction_bil_ctc_synchronous.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
+
+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
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+# 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: 0.2
+xctc-weight: 0.1
+
+# InterCTC
+inter-ctc-weight: 0.1
+inter-ctc-layers: 6,9,12,15
+share-inter-ctc: True
+inter-xctc-weight: 0.05
+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

egs/mustc/joint/conf/reproduction_bil_ctc_synchronous_256.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: 2e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+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: 0.1
+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
+
+# 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: 0.2
+xctc-weight: 0.1
+
+# InterCTC
+inter-ctc-weight: 0.1
+inter-ctc-layers: 6,9,12,15
+share-inter-ctc: True
+inter-xctc-weight: 0.05
+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

egs/mustc/joint/conf/reproduction_ctc_aug.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-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: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+acoustic-encoder: transformer
+adapter: inter_league
+
+# Conformer 
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
+
+# CTC & XCTC
+share-ctc-and-embed: True
+share-xctc-and-embed: True
+ctc-weight: 0.2
+xctc-weight: 0.2
+
+# InterCTC
+share-inter-ctc: True
+inter-ctc-weight: 0.1
+inter-ctc-layers: 6,9
+inter-xctc-weight: 0.1
+inter-xctc-layers: 4
+
+# 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: 3
+cross-attn-layer: 2
+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: 0.5
+xctc-pae-ground-truth-only-mistake: True 
+pae-oracle-smooth: True

egs/mustc/joint/conf/reproduction_nast.yaml

+arch: s2t_ctc
+encoder-type: sate
+
+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
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-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: 12
+text-encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+acoustic-encoder: transformer
+adapter: inter_league
+
+# 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
+inter-xctc-weight: 1.0
+inter-xctc-layers: 6,9
+
+# Prediction-aware encoding
+ctc-pae: inter_league
+xctc-pae: inter_league
+
+# 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/joint/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/joint/conf/reproduction_sate.yaml

+arch: s2t_sate
+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)
+
+ctc-weight: 0.3
+share-ctc-and-embed: True
+criterion: label_smoothed_cross_entropy_with_ctc
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-encoder-no-scale-embedding: 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: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: transformer
+adapter: inter_league
+# adapter: none
+# adapter-embed-norm: True
+# adapter-out-norm: True
+
+# load-pretrained-encoder-from:
+# load-pretrained-acoustic-encoder-from:
+# load-pretrained-text-encoder-from:
+# load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/rpr.yaml

+encoder-attention-type: rel_pos
+
+# encoder-attention-type: relative
+# decoder-attention-type: relative
+# max-encoder-relative-length: 100
+# max-decoder-relative-length: 20
\ No newline at end of file

egs/mustc/joint/conf/sate.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-encoder-no-scale-embedding: 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: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: transformer
+adapter: inter_league
+#adapter: none
+#adapter-embed-norm: True
+#adapter-out-norm: True
+#share-adapter-and-ctc: True
+#share-adapter-and-embed: True
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:

egs/mustc/joint/conf/sate_big.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-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: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+acoustic-encoder: transformer
+adapter: inter_league
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/sate_big_pds.yaml

+arch: s2t_sate
+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
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-encoder-no-scale-embedding: True
+
+encoder-normalize-before: True
+decoder-normalize-before: 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-ffn-embed-dim: 2048
+encoder-layers: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+acoustic-encoder: pds
+adapter: inter_league
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/sate_deep.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-encoder-no-scale-embedding: 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
+text-encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: transformer
+adapter: inter_league
+#adapter-embed-norm: True
+#adapter-out-norm: True
+#share-adapter-and-ctc: True
+#share-adapter-and-embed: True
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:

egs/mustc/joint/conf/sate_pds_16.yaml

+arch: s2t_sate
+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
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-encoder-no-scale-embedding: True
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: pds
+adapter: inter_league
+
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/sate_pds_8.yaml

+arch: s2t_sate
+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
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-encoder-no-scale-embedding: True
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: pds
+adapter: inter_league
+
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/sate_pds_8_444.yaml

+arch: s2t_sate
+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
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-encoder-no-scale-embedding: True
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: pds
+adapter: inter_league
+
+
+encoder-embed-dim: 256
+pds-stages: 3
+pds-layers: 4_4_4
+pds-ratios: 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
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1
+pds-kernel-sizes: 5_5_5
+pds-ffn-ratios: 8_8_8
+pds-attn-heads: 4_4_4
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/sate_pds_base_8.yaml

+arch: s2t_sate
+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
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-encoder-no-scale-embedding: True
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: pds
+adapter: inter_league
+
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/sate_pds_deep_8.yaml

+arch: s2t_sate
+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
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+textual-encoder-embed-norm: True
+textual-encoder-no-scale-embedding: True
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 18
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: pds
+adapter: inter_league
+
+encoder-embed-dim: 256
+pds-stages: 4
+pds-layers: 4_5_5_4
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/w2v2.yaml

+arch: s2t_w2v2_transformer
+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
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: 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: 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
+
+w2v2-model-path: /home/xuchen/st/models/w2v2/wav2vec_small.pt
+freeze-w2v: False
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/joint/conf/xctc.yaml

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

egs/mustc/joint/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/mustc/joint/decode.sh

+#!/usr/bin/env bash
+
+gpu_num=1
+
+data_tag=
+test_subset=(tst-COMMON_en-de tst-COMMON_en-fr tst-COMMON_en-es tst-COMMON_en-it tst-COMMON_en-nl tst-COMMON_en-pt tst-COMMON_en-ro tst-COMMON_en-ru)
+#test_subset=(tst-COMMON_en-de)
+#test_subset=(test_en-fr_1k)
+
+exp_name=
+if [ "$#" -eq 1 ]; then
+    exp_name=$1
+fi
+
+sacrebleu=1
+ctc_infer=0
+n_average=10
+beam_size=5
+infer_ctc_weight=0
+len_penalty=1.0
+max_tokens=10000
+batch_size=0
+infer_debug=0
+dec_model=checkpoint_last.pt
+
+cmd="./run.sh
+    --stage 2
+    --stop_stage 2
+    --gpu_num ${gpu_num}
+    --exp_name ${exp_name}
+    --sacrebleu ${sacrebleu}
+    --n_average ${n_average}
+    --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/mustc/joint/ensemble.sh

+set -e
+
+gpu_num=1
+root_dir=/home/xuchen/st/Fairseq-S2T
+ckpt=/home/xuchen/st/checkpoints/mustc-v2/st
+
+model_txt=$1
+set=$2
+test_subset=$3
+
+#data_dir=/home/xuchen/st/data/mustc-v2/st_lcrm/en-de
+#test_subset=(tst-COMMON)
+
+data_dir=/media/data/tst/$set/en-de
+#test_subset=(office)
+#test_subset=(webrtc1)
+#test_subset=(adap2)
+
+data_config=config_st_share.yaml
+result_file=./result
+
+beam_size=5
+lenpen=0.6
+max_tokens=10000
+
+models=()
+i=0
+for line in `cat $model_txt`; do
+    i=`expr $i + 1`
+    
+    model_dir=$ckpt/$line
+    [[ ! -d $model_dir ]] && echo $model_dir && exit 1;
+
+    if [[ -f $model_dir/avg_10_checkpoint.pt ]]; then
+        model=$model_dir/avg_10_checkpoint.pt
+    else
+        model=$model_dir/checkpoint_best.pt
+    fi
+    [[ ! -f $model ]] && echo $model && exit 1;
+
+    models[$i]=$model
+done
+
+models=`echo ${models[*]} | sed 's/ /:/g'`
+
+res_dir=$ckpt/ensemble/$set
+i=0
+while : 
+do
+    if [[ -d $res_dir/$i ]]; then
+        i=`expr $i + 1`
+    else
+        res_dir=$res_dir/$i
+        break
+    fi 
+done
+
+mkdir -p $res_dir
+cp $model_txt $res_dir
+
+
+if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
+    if [[ ${gpu_num} -eq 0 ]]; then
+        device=()
+    else
+        source ./local/utils.sh
+        device=$(get_devices $gpu_num 0)
+    fi
+fi
+export CUDA_VISIBLE_DEVICES=${device}
+
+for subset in ${test_subset[@]}; do
+    subset=${subset}_st
+    cmd="python ${root_dir}/fairseq_cli/generate.py
+    ${data_dir}
+    --config-yaml ${data_config}
+    --gen-subset ${subset}
+    --task speech_to_text
+    --path ${models}
+    --results-path ${res_dir}
+    --skip-invalid-size-inputs-valid-test
+    --max-tokens ${max_tokens}
+    --beam ${beam_size}
+    --lenpen ${lenpen}
+    --scoring sacrebleu"
+    echo -e "\033[34mRun command: \n${cmd} \033[0m"
+
+    eval $cmd
+    tail -n 1 ${res_dir}/generate-${subset}.txt
+
+    cd $res_dir
+    evaluate.sh translation-${subset}.txt $set
+    cd -
+done
+

egs/mustc/joint/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/mustc/joint/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/mustc/joint/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.Strip(),
+        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/mustc/joint/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/mustc/joint/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/mustc/joint/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/mustc/joint/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/mustc/joint/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/mustc/joint/train.sh

+#!/usr/bin/env bash
+
+# training the model
+
+gpu_num=1
+update_freq=1
+max_tokens=40000
+
+extra_tag=
+extra_parameter=
+#extra_tag="${extra_tag}"
+#extra_parameter="${extra_parameter} "
+
+exp_tag=
+
+# Base
+config_list=(base conformer ctc)
+
+# SATE
+#config_list=(sate conformer ctc)
+
+# PDS
+#config_list=(pds_base_8 conformer ctc)
+
+# 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/mustc/multi-st/binary.sh

+set -e
+
+eval=1
+
+lcrm=1
+tokenizer=0
+
+vocab_type=unigram
+vocab_size=5000
+use_raw_audio=0
+speed_perturb=0
+
+dataset=mustc
+root_dir=~/st
+code_dir=${root_dir}/Fairseq-S2T
+org_data_dir=${root_dir}/data/${dataset}
+data_dir=${root_dir}/data/${dataset}/st
+
+use_specific_dict=0
+specific_prefix=st
+specific_dir=${root_dir}/data/mustc/st
+asr_vocab_prefix=spm_unigram10000_st_share
+
+src_lang=en
+tgt_lang=zh
+subsets=(2019)
+
+splits=$(echo ${subsets[*]} | sed 's/ /_/g')
+cmd="python ${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}
+    --tgt-lang ${tgt_lang}
+    --splits ${splits}
+    --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}

egs/mustc/multi-st/conf/aipa.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
+
+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/mustc/multi-st/conf/aipa_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
+
+ctc-mixup-consistent-weight: 0.15
+inter-ctc-mixup-consistent-weight: 0.1
+mixup-consistent-weight: 0.5
+
+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/mustc/multi-st/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
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: 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: 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:

egs/mustc/multi-st/conf/basis.yaml

+#train-subset: train_en-de,train_en-fr,train_en-es,train_en-it,train_en-nl,train_en-pt,train_en-ro,train_en-ru
+train-subset: train_mix_1,train_mix_2,train_mix_3,train_mix_4,train_mix_5,train_mix_6,train_mix_7,train_mix_8
+valid-subset: dev_en-de,dev_en-fr,dev_en-es,dev_en-it,dev_en-nl,dev_en-pt,dev_en-ro,dev_en-ru
+#valid-subset: dev_en-de
+#train-subset: train_en-de,train_en-fr,train_en-es,train_en-it
+#valid-subset: dev_en-de,dev_en-fr,dev_en-es,dev_en-it
+ignore-prefix-size: 1
+prefix-size: 1
+
+sharded-data-load: True
+max-epoch: 100
+max-update: 300000
+patience: 20
+post-process: sentencepiece
+
+best_checkpoint_metric: loss
+maximize_best_checkpoint_metric: False
+
+#eval-bleu: True
+eval-bleu-args: {"beam": 5, "lenpen": 1.0}
+eval-bleu-detok: moses
+eval-bleu-remove-bpe: sentencepiece
+eval-bleu-print-samples: True
+#best_checkpoint_metric: bleu
+#maximize_best_checkpoint_metric: True
+
+# no-epoch-checkpoints: True
+validate-interval: 5
+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/mustc/multi-st/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: 1e-3
+adam-betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+
+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: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:

egs/mustc/multi-st/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/mustc/multi-st/conf/ctc.yaml

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

egs/mustc/multi-st/conf/ctc_aug_base.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-encoder-no-scale-embedding: 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: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: transformer
+adapter: none
+
+# load-pretrained-encoder-from: 
+# load-pretrained-acoustic-encoder-from:
+# load-pretrained-text-encoder-from:
+# load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/multi-st/conf/ctc_aug_big.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-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: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+acoustic-encoder: transformer
+adapter: none
+
+# load-pretrained-encoder-from: 
+# load-pretrained-acoustic-encoder-from:
+# load-pretrained-text-encoder-from:
+# load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/multi-st/conf/ctc_aug_pds_big.yaml

+arch: s2t_sate
+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
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-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-layers: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+acoustic-encoder: pds
+adapter: none
+
+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
+
+# load-pretrained-encoder-from: 
+# load-pretrained-acoustic-encoder-from:
+# load-pretrained-text-encoder-from:
+# load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/multi-st/conf/dlcl.yaml

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

egs/mustc/multi-st/conf/dual.yaml

+arch: s2t_dual
+
+asr-encoder: transformer
+mt-encoder-layers: 6
+mt-encoder: transformer
+
+encoder-collaboration-mode: parallel
+decoder-collaboration-mode: parallel
+
+encoder-league-s1-ratio: 0.5
+encoder-league-s2-ratio: 0.5
+encoder-league-drop-net: False
+encoder-league-drop-net-prob: 0.2
+encoder-league-drop-net-mix: False
+
+decoder-league-s1-ratio: 0.5
+decoder-league-s2-ratio: 0.5
+decoder-league-drop-net: False
+decoder-league-drop-net-prob: 0.0
+decoder-league-drop-net-mix: False
+
+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: join_speech_and_text_loss
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-asr-encoder-from:
+#load-pretrained-mt-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/multi-st/conf/dynamic.yaml

+compression-metric: threshold
+compression-mode: create
+compression-layers: 6,9
+compression-threshold: 0.99
+compression-norm: True
+compression-pos: True
\ No newline at end of file

egs/mustc/multi-st/conf/inter.yaml

+inter-ctc-weight: 0.2
+inter-ctc-layers: 6,9
+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/mustc/multi-st/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/mustc/multi-st/conf/mlo.yaml

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

egs/mustc/multi-st/conf/multibranch.yaml

+arch: s2t_multibranch
+
+junior-acoustic-encoder: transformer
+acoustic-adapter: none
+textual-adapter: none
+junior-acoustic-encoder-layers: 12
+senior-acoustic-encoder-layers: 6
+textual-encoder-layers: 6
+
+# collaboration-direction: none
+# collaboration-direction: acoustic
+# collaboration-direction: textual
+collaboration-direction: both
+collaboration-start: 0:0
+collaboration-step: 1:1
+# encoder-collaboration-mode: serial
+# decoder-collaboration-mode: serial
+encoder-collaboration-mode: parallel
+decoder-collaboration-mode: parallel
+
+use-raw-text: False
+modality-switch: False
+text-to-hidden-progress: 1:0:500
+
+encoder-league-s1-ratio: 0.5
+encoder-league-s2-ratio: 0.5
+encoder-league-out-norm: False
+encoder-league-gated: False
+encoder-league-drop-net: False
+encoder-league-drop-net-prob: 0.2
+encoder-league-drop-net-mix: False
+
+decoder-league-s1-ratio: 0.5
+decoder-league-s2-ratio: 0.5
+decoder-league-out-norm: False
+decoder-league-gated: False
+decoder-league-drop-net: False
+decoder-league-drop-net-prob: 0.0
+decoder-league-drop-net-mix: False
+
+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-attention-heads: 4
+
+decoder-layers: 6
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+#adapter: inter_league
+#adapter: none
+#adapter-embed-norm: True
+#adapter-out-norm: True
+#share-adapter-and-ctc: True
+#share-adapter-and-embed: True
+
+#load-pretrained-junior-acoustic-encoder-from:
+#load-pretrained-senior-acoustic-encoder-from:
+#load-pretrained-textual-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/multi-st/conf/nast_base.yaml

+arch: s2t_ctc
+encoder-type: sate
+
+criterion: ctc
+zero_infinity: True
+xctc-weight: 1.0
+ctc-weight: 1.0
+
+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)
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+encoder-embed-norm: True
+encoder-no-scale-embedding: True
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-encoder-no-scale-embedding: 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: 12
+text-encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: transformer
+adapter: none
\ No newline at end of file

egs/mustc/multi-st/conf/nast_big.yaml

+arch: s2t_ctc
+encoder-type: sate
+
+criterion: ctc
+zero_infinity: True
+xctc-weight: 1.0
+ctc-weight: 1.0
+
+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
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-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: 12
+text-encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+acoustic-encoder: transformer
+adapter: none
\ No newline at end of file

egs/mustc/multi-st/conf/nast_pds_big.yaml

+arch: s2t_ctc
+encoder-type: sate
+
+criterion: ctc
+zero_infinity: True
+xctc-weight: 1.0
+ctc-weight: 1.0
+
+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
+text-no-pos-emb: True
+textual-encoder-embed-norm: False
+textual-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-layers: 12
+text-encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+acoustic-encoder: pds
+adapter: none
+
+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
\ No newline at end of file

egs/mustc/multi-st/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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/multi-st/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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/multi-st/conf/pds_base_4_444.yaml

+arch: pdss2t_transformer_s_8
+
+encoder-embed-dim: 256
+pds-stages: 3
+pds-layers: 4_4_4
+pds-ratios: 2_2_1
+pds-fusion: False
+pds-fusion-method: all_conv2
+pds-fusion-layers: 1_1_1 
+pds-fusion-weight: 0.2_0.3_0.5
+pds-embed-dims: 256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1
+pds-kernel-sizes: 5_5_5
+pds-ffn-ratios: 8_8_8
+pds-attn-heads: 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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/multi-st/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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/multi-st/conf/pds_base_8_444.yaml

+arch: pdss2t_transformer_s_8
+
+encoder-embed-dim: 256
+pds-stages: 3
+pds-layers: 4_4_4
+pds-ratios: 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
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1
+pds-kernel-sizes: 5_5_5
+pds-ffn-ratios: 8_8_8
+pds-attn-heads: 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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file