I optimized the implementation of S2T.

It must be said that some problems still confuse me:
1. Whether to scale in the input layer (I try to replace it with layer specification);
2. The detailed setting of weight sharing between output projection matrix and embedding matrix in the adapter (I notice that inconsistent variance will lead to bad results);
3. The biggest confusion is that the variance increases with the calculation layer by layer (I am not sure if this phenomenon is reasonable, I will compare the behavior on the latest code).
Finally, the detailed implementation is so important to the final performance, even if it is a subtle difference.

egs/aishell/asr/conf/inter.yaml

@@ -4,9 +4,9 @@ interleaved-ctc-layers: 6,9
 interleaved-ctc-temperature: 1.0
 interleaved-ctc-drop-prob: 0
 
-sae-adapter: league
-sae-drop-prob: 0.2
-sae-distribution-cutoff: 10
+sae-adapter: inter_league
+sae-drop-prob: 0.0
+sae-distribution-cutoff: 0
 share-ctc-and-sae: False
 
 ctc-self-distill-weight: 0

egs/aishell/asr/conf/mixup.yaml

@@ -2,3 +2,4 @@ inter_mixup: True
 inter_mixup_layer: -1
 inter_mixup_prob: 1.0
 inter_mixup_ratio: 0.2
+inter_mixup_beta: 0.2

egs/aishell/asr/conf/purectc.yaml

+arch: s2t_ctc
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam_betas: (0.9,0.98)
+
+criterion: ctc
+zero_infinity: True
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.1
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+encoder-attention-heads: 4
+
+#load-pretrained-encoder-from:
\ No newline at end of file

egs/iwslt14/mt/conf/inter.yaml

 #ctc-layer:
-ctc-weight: 0.2
-interleaved-ctc-weight: 0.1
-interleaved-ctc-layers: 6,9
+#ctc-weight: 0.2
+interleaved-ctc-weight: 0.3
+interleaved-ctc-layers: 2,4
 interleaved-ctc-temperature: 1.0
 interleaved-ctc-drop-prob: 0
 interleaved_ctc_upsampling_ratio: 2
 
-sae-adapter: league
+sae-adapter: inter_league
 sae-drop-prob: 0.0
 #sae-distribution-cutoff: 10
-share-ctc-and-sae: False
+#share-ctc-and-sae: True
 
 ctc-self-distill-weight: 0
\ No newline at end of file

egs/iwslt14/mt/train.sh

@@ -8,6 +8,7 @@ max_tokens=8192
 
 exp_tag=baseline
 config_list=(base)
+config_list=(base_ctc inter)
 
 # exp full name
 exp_name=

egs/libri_trans/asr/conf/debug.yaml

-arch: s2t_transformer
+arch: s2t_transformer_s
 share-decoder-input-output-embed: True
 optimizer: adam
 clip-norm: 10.0
@@ -34,28 +34,19 @@ decoder-attention-heads: 4
 attention-dropout: 0.1
 activation-dropout: 0.1
 
-#load-pretrained-encoder-from:
-#load-pretrained-decoder-from:
-
-inter_mixup: True
-inter_mixup_layer: -1
-inter_mixup_ratio: 0.2
-
 ctc-weight: 0.2
 interleaved-ctc-weight: 0.1
 interleaved-ctc-layers: 6,9
-interleaved-temperature: 2
-
-#target-ctc-weight: 0.3
-#target-ctc-layer: 6
-#target-interleaved-ctc-weight: 0.1
-#target-interleaved-ctc-layers: 2,4
+interleaved-ctc-temperature: 1.0
+interleaved-ctc-drop-prob: 0
 
 sae-adapter: league
-share-ctc-and-sae: False
-sae-drop-prob: 0.2
-interleaved-ctc-drop-prob: 0.2
+sae-drop-prob: 0.0
 sae-distribution-cutoff: 10
+share-ctc-and-sae: False
 
 ctc-self-distill-weight: 0
-post-process: sentencepiece
\ No newline at end of file
+
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/asr/conf/purectc.yaml

@@ -9,7 +9,6 @@ adam_betas: (0.9,0.98)
 
 criterion: ctc
 zero_infinity: True
-post-process: sentencepiece
 
 subsampling-type: conv1d
 subsampling-layers: 2

egs/mustc/mt/conf/inter.yaml

 #ctc-layer:
 #ctc-weight: 0.2
 interleaved-ctc-weight: 0.3
-interleaved-ctc-layers: 6,9
-interleaved-ctc-temperature: 1.0
+interleaved-ctc-layers: 8
+interleaved-ctc-temperature: 1
 interleaved-ctc-drop-prob: 0
 interleaved_ctc_upsampling_ratio: 2
 
-sae-adapter: league
+sae-adapter: inter_league
 sae-drop-prob: 0.0
 #sae-distribution-cutoff: 10
-share-ctc-and-sae: False
+#share-ctc-and-sae: True
 
 ctc-self-distill-weight: 0

egs/mustc/st/conf/basis.yaml

@@ -8,6 +8,7 @@ best_checkpoint_metric: loss
 maximize_best_checkpoint_metric: False
 post-process: sentencepiece
 
+#fp16-scale-tolerance: 0.25
 no-epoch-checkpoints: True
 #keep-last-epochs: 10
 keep-best-checkpoints: 10

egs/mustc/st/conf/inter.yaml

-ctc-weight: 0.2
-interleaved-ctc-weight: 0.1
+ctc-weight: 0.3
+interleaved-ctc-weight: 0.2
 interleaved-ctc-layers: 6,9
 interleaved-ctc-temperature: 1.0
 interleaved-ctc-drop-prob: 0
 
 #target-ctc-weight: 0.3
 #target-ctc-layer: 6
-target-interleaved-ctc-weight: 0.1
-target-interleaved-ctc-layers: 2,4
+#target-interleaved-ctc-weight: 0.1
+#target-interleaved-ctc-layers: 2,4
 
-sae-adapter: league
+sae-adapter: inter_league
 sae-drop-prob: 0.0
-#sae-distribution-cutoff: 10
-share-ctc-and-sae: False
-share-target-ctc-and-sae: False
+#sae-distribution-cutoff: 0
+#share-ctc-and-sae: True
+#share-target-ctc-and-sae: True
 
 ctc-self-distill-weight: 0
\ No newline at end of file

egs/mustc/st/conf/sate.yaml

@@ -37,6 +37,11 @@ decoder-attention-heads: 4
 
 acoustic-encoder: transformer
 adapter: 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:

egs/mustc/st/run.sh

@@ -21,6 +21,7 @@ stop_stage=0
 ######## hardware ########
 # devices
 #device=()
+use_auto=0
 gpu_num=8
 update_freq=1
 
@@ -215,6 +216,9 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     echo "stage 1: ST Network Training"
     [[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1;
 
+    if [[ ${use_auto} -eq 1 ]]; then
+        device=(-1)
+    fi
     if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
 		if [[ ${gpu_num} -eq 0 ]]; then
 			device=""
@@ -330,14 +334,21 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     echo "${time} | ${device} | ${data_dir} | ${exp_name} | ${model_dir} " >> $log
     tail -n 50 ${log} > tmp.log
     mv tmp.log $log
-    export CUDA_VISIBLE_DEVICES=${device}
 
     cmd="nohup ${cmd} >> ${model_dir}/train.log 2>&1 &"
     if [[ $eval -eq 1 ]]; then
+        if [[ ${use_auto} -eq 1 ]]; then
+            cmd=$(echo ${cmd} | tr -d "\n")
+            auto_run -c "${cmd}" -n ${gpu_num}
+        else
+            export CUDA_VISIBLE_DEVICES=${device}
 	    	eval $cmd
-		sleep 2s
+        fi
+		sleep 5s
+		if [[ -f ${model_dir}/train.log ]]; then
 		    tail -n "$(wc -l ${model_dir}/train.log | awk '{print $1+1}')" -f ${model_dir}/train.log
         fi
+	fi
 fi
 wait
 
@@ -359,6 +370,9 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
 		dec_model=${dec_model}
 	fi
 
+    if [[ ${use_auto} -eq 1 ]]; then
+        device=(-1)
+    fi
     if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
 		if [[ ${gpu_num} -eq 0 ]]; then
 			device=""
@@ -367,7 +381,6 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
         	device=$(get_devices $gpu_num 0)
 		fi
     fi
-    export CUDA_VISIBLE_DEVICES=${device}
 
     suffix=beam${beam_size}_alpha${len_penalty}_tokens${max_tokens}
     if [[ ${n_average} -ne 1 ]]; then
@@ -409,7 +422,13 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     	echo -e "\033[34mRun command: \n${cmd} \033[0m"
 
         if [[ $eval -eq 1 ]]; then
+            if [[ ${use_auto} -eq 1 ]]; then
+                cmd=$(echo ${cmd} | tr -d "\n")
+                auto_run -c ${cmd} -n ${gpu_num}
+            else
+                export CUDA_VISIBLE_DEVICES=${device}
         	    eval $cmd
+            fi
     	    tail -n 1 ${model_dir}/generate-${subset}.txt >> ${result_file}
             mv ${model_dir}/generate-${subset}.txt ${model_dir}/generate-${subset}-${suffix}.txt
             mv ${model_dir}/translation-${subset}.txt ${model_dir}/translation-${subset}-${suffix}.txt

egs/tibetan/asr/binary.sh

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

egs/tibetan/asr/conf/base.yaml

+arch: s2t_transformer_s
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 1e-3
+#lr: 5e-4
+adam_betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+subsampling-type: conv1d
+subsampling-layers: 2
+#subsampling-filter: 2048
+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: /home/xuchen/st/checkpoints/aishell/asr/0506_sp_char_base_ctc_sample1024/avg_10_checkpoint.pt
+load-pretrained-decoder-from: /home/xuchen/st/checkpoints/aishell/asr/0506_sp_char_base_ctc_sample1024/avg_10_checkpoint.pt
+load-pretrained-encoder-from: /home/xuchen/st/checkpoints/librispeech/asr/base_baseline/avg_10_checkpoint.pt
+load-pretrained-decoder-from: /home/xuchen/st/checkpoints/librispeech/asr/base_baseline/avg_10_checkpoint.pt
+#load-pretrained-encoder-from: /home/xuchen/st/checkpoints/librispeech/asr/base_conformer_baseline_batch50k_16/avg_10_checkpoint.pt
+#load-pretrained-decoder-from: /home/xuchen/st/checkpoints/librispeech/asr/base_conformer_baseline_batch50k_16/avg_10_checkpoint.pt

egs/tibetan/asr/conf/basis.yaml

+train-subset: train
+valid-subset: dev
+
+max-epoch: 100
+max-update: 100000
+patience: 20
+best_checkpoint_metric: loss
+maximize_best_checkpoint_metric: False
+
+no-epoch-checkpoints: True
+#keep-last-epochs: 10
+keep-best-checkpoints: 10
+
+num-workers: 8
+no-progress-bar: True
+log-interval: 100
+seed: 1
+report-accuracy: True
+skip-invalid-size-inputs-valid-test: True

egs/tibetan/asr/conf/big.yaml

+arch: s2t_transformer_m
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam_betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+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/tibetan/asr/conf/big_wenet.yaml

+arch: s2t_transformer_m
+share-decoder-input-output-embed: True
+share-ctc-and-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
+label_smoothing: 0.1
+
+subsampling-type: conv2d
+subsmapling-layers: 2
+subsampling-filter: 512
+subsampling-kernel: 3
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: relu
+
+dropout: 0.15
+activation-fn: relu
+encoder-embed-dim: 512
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+cnn-module-norm: layer_norm
+
+load-pretrained-encoder-from: /home/xuchen/after.pt
+load-pretrained-decoder-from: /home/xuchen/after.pt
+#load-pretrained-decoder-from:

egs/tibetan/asr/conf/conformer.yaml

+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 15
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish

egs/tibetan/asr/conf/ctc.yaml

+ctc-weight: 0.3
+post-process: sentencepiece

egs/tibetan/asr/conf/dlcl.yaml

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

egs/tibetan/asr/conf/inter.yaml

+ctc-weight: 0.2
+intermedia-ctc-layers: 6,9
+intermedia-adapter: league
+intermedia-ctc-weight: 0.1
+ctc-self-distill-weight: 0
+post-process: sentencepiece
\ No newline at end of file

egs/tibetan/asr/conf/local_attn.yaml

+encoder-attention-type: local
+hard-mask-window: 0
+gauss-mask-sigma: 3
+init-mask-weight: 0
\ No newline at end of file

egs/tibetan/asr/conf/pds_base.yaml

+arch: pdss2t_transformer_s_8
+pds-fusion: True
+ctc-layer: 12
+
+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
+label_smoothing: 0.1
+
+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/tibetan/asr/conf/pds_base_16.yaml

+arch: pdss2t_transformer_s_16
+
+encoder-embed-dim: 256
+pds-stages: 4
+ctc-layer: 12
+pds-layers: 2_2_6_2
+pds-ratios: 2_2_2_2
+pds-fusion: True
+pds-fusion-method: all_conv
+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
+label_smoothing: 0.1
+
+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/tibetan/asr/conf/pds_base_32.yaml

+arch: pdss2t_transformer_s_32
+
+encoder-embed-dim: 256
+pds-stages: 5
+ctc-layer: 12
+pds-layers: 2_2_3_3_2
+pds-ratios: 2_2_2_2_2
+pds-fusion: True
+pds-fusion-method: all_conv
+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
+label_smoothing: 0.1
+
+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/tibetan/asr/conf/pds_base_4.yaml

+arch: pdss2t_transformer_s_8
+
+#pds-ctc: 0_1_1_0
+#intermedia-adapter: league
+#intermedia-ctc-weight: 0.1
+
+#encoder-attention-type: reduced
+#pds-attn-ds-ratios: 4_2_2_1
+#attention-reduced-method: pool
+#attention-reduced-q: True
+
+encoder-embed-dim: 256
+pds-stages: 4
+#ctc-layer: 12
+pds-layers: 4_2_2_4
+pds-ratios: 2_2_1_1
+pds-fusion: True
+pds-fusion-method: all_conv
+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: 1e-3
+adam_betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+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:

egs/tibetan/asr/conf/pds_base_8.yaml

+arch: pdss2t_transformer_s_8
+
+#pds-ctc: 0_1_1_0
+#intermedia-adapter: league
+#intermedia-ctc-weight: 0.1
+
+#encoder-attention-type: reduced
+#pds-attn-ds-ratios: 4_2_2_1
+#attention-reduced-method: pool
+#attention-reduced-q: True
+
+encoder-embed-dim: 256
+pds-stages: 4
+#ctc-layer: 12
+pds-layers: 4_2_2_4
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+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: 1e-3
+adam_betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+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:

egs/tibetan/asr/conf/pds_base_8_grow.yaml

+arch: pdss2t_transformer_s_8
+
+#pds-ctc: 0_1_1_0
+#intermedia-adapter: league
+#intermedia-ctc-weight: 0.1
+
+#encoder-attention-type: reduced
+#pds-attn-ds-ratios: 4_2_2_1
+#attention-reduced-method: pool
+#attention-reduced-q: True
+
+encoder-embed-dim: 512
+pds-stages: 4
+#ctc-layer: 15
+encoder-layers: 18
+pds-layers: 6_3_3_6
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 256_384_384_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_4_4_4
+pds-attn-heads: 4_6_6_8
+
+share-decoder-input-output-embed: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 2e-3
+adam_betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+dropout: 0.1
+activation-fn: relu
+
+decoder-layers: 6
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:

egs/tibetan/asr/conf/pds_big_8.yaml

+arch: pdss2t_transformer_m_8
+
+encoder-embed-dim: 512
+pds-stages: 4
+ctc-layer: 12
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+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: 2e-3
+adam_betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+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:

egs/tibetan/asr/conf/purectc.yaml

+arch: s2t_ctc
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 1e-3
+adam_betas: (0.9,0.98)
+
+criterion: ctc
+zero_infinity: True
+post-process: sentencepiece
+
+subsampling-type: conv1d
+subsampling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.1
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+encoder-attention-heads: 4
+
+load-pretrained-encoder-from: /home/xuchen/st/checkpoints/aishell/asr/0506_sp_char_base_ctc_sample1024/avg_10_checkpoint.pt
+#load-pretrained-encoder-from: /home/xuchen/st/checkpoints/librispeech/asr/base_baseline/avg_10_checkpoint.pt

egs/tibetan/asr/conf/purectc_pds_base.yaml

+arch: s2t_ctc
+encode-type: pds
+
+encoder-embed-dim: 256
+pds-stages: 4
+ctc-layer: 12
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+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
+label_smoothing: 0.1
+
+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/tibetan/asr/conf/purectc_pds_base_8.yaml

+arch: s2t_ctc
+encoder-type: pds
+
+#pds-ctc: 0_1_1_0
+#intermedia-adapter: league
+#intermedia-ctc-weight: 1
+
+#encoder-attention-type: reduced
+#pds-attn-ds-ratios: 8_4_2_1
+#attention-reduced-method: pool
+#attention-reduced-q: True
+
+encoder-embed-dim: 256
+pds-stages: 4
+#ctc-layer: 12
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+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: ctc
+post-process: sentencepiece
+
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+encoder-attention-heads: 4
+
+#load-pretrained-encoder-from:

egs/tibetan/asr/conf/purectc_pds_base_8_grow.yaml

+arch: s2t_ctc
+encoder-type: pds
+
+#pds-ctc: 0_1_1_0
+#intermedia-adapter: league
+#intermedia-ctc-weight: 1
+#intermedia-temperature: 5
+
+encoder-attention-type: rel_pos
+#encoder-attention-type: reduced_rel_pos
+#pds-attn-ds-ratios: 4_2_2_1
+#attention-reduced-method: pool
+#attention-reduced-q: True
+
+encoder-embed-dim: 512
+pds-stages: 4
+#ctc-layer: 15
+encoder-layers: 10
+pds-layers: 3_2_2_3
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 256_384_384_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_4_4_4
+pds-attn-heads: 4_6_6_8
+
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 10000
+lr: 0.002
+adam_betas: (0.9,0.98)
+
+criterion: ctc
+post-process: sentencepiece
+
+dropout: 0.1
+activation-fn: relu
+
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 31
+encoder-activation-fn: swish
+
+#load-pretrained-encoder-from:

egs/tibetan/asr/conf/rpr.yaml

+encoder-attention-type: rel_pos
+#encoder-attention-type: relative
+#max-encoder-relative-length: 100

egs/tibetan/asr/decode.sh

+#! /bin/bash
+
+gpu_num=1
+
+data_dir=
+test_subset=(dev test)
+
+exp_name=
+if [ "$#" -eq 1 ]; then
+    exp_name=$1
+fi
+
+cer=1
+n_average=10
+beam_size=5
+len_penalty=1.0
+max_tokens=10000
+dec_model=checkpoint_best.pt
+
+cmd="./run.sh
+    --stage 2
+    --stop_stage 2
+    --gpu_num ${gpu_num}
+    --exp_name ${exp_name}
+    --n_average ${n_average}
+    --cer ${cer}
+    --beam_size ${beam_size}
+    --len_penalty ${len_penalty}
+    --max_tokens ${max_tokens}
+    --dec_model ${dec_model}
+    "
+
+if [[ -n ${data_dir} ]]; then
+    cmd="$cmd --data_dir ${data_dir}"
+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/tibetan/asr/local/monitor.sh

+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/tibetan/asr/local/parse_options.sh

+#!/usr/bin/env bash
+
+# Copyright 2012  Johns Hopkins University (Author: Daniel Povey);
+#                 Arnab Ghoshal, Karel Vesely
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Parse command-line options.
+# To be sourced by another script (as in ". parse_options.sh").
+# Option format is: --option-name arg
+# and shell variable "option_name" gets set to value "arg."
+# The exception is --help, which takes no arguments, but prints the
+# $help_message variable (if defined).
+
+
+###
+### The --config file options have lower priority to command line
+### options, so we need to import them first...
+###
+
+# Now import all the configs specified by command-line, in left-to-right order
+for ((argpos=1; argpos<$#; argpos++)); do
+  if [ "${!argpos}" == "--config" ]; then
+    argpos_plus1=$((argpos+1))
+    config=${!argpos_plus1}
+    [ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
+    . $config  # source the config file.
+  fi
+done
+
+
+###
+### Now we process the command line options
+###
+while true; do
+  [ -z "${1:-}" ] && break;  # break if there are no arguments
+  case "$1" 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 '$1'"
+      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 "$1" | 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"
+      eval $name=\"$2\";
+
+      # Check that Boolean-valued arguments are really Boolean.
+      if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then
+        echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2
+        exit 1;
+      fi
+      shift 2;
+      ;;
+  *) break;
+  esac
+done
+
+
+# Check for an empty argument to the --cmd option, which can easily occur as a
+# result of scripting errors.
+[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
+
+
+true; # so this script returns exit code 0.

egs/tibetan/asr/local/utils.sh

+
+get_devices(){
+    gpu_num=$1
+    use_cpu=$2
+    device=()
+    while :
+    do
+        record=$(mktemp -t temp.record.XXXXXX)
+        gpustat > $record
+        all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
+
+        count=0
+        for dev in ${all_devices[@]}
+        do
+            line=$((dev + 2))
+            use=$(head -n $line ${record} | tail -1 | cut -d '|' -f3 | cut -d '/' -f1)
+            if [[ $use -lt 100 ]]; then
+                device[$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/tibetan/asr/run.sh

+#! /bin/bash
+
+# Processing MuST-C Datasets
+
+# Copyright 2021 Natural Language Processing Laboratory 
+# Xu Chen (xuchenneu@163.com)
+
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+#set -u
+set -o pipefail
+export PYTHONIOENCODING=UTF-8
+
+eval=1
+time=$(date "+%m%d_%H%M")
+
+stage=0
+stop_stage=0
+
+######## hardware ########
+# devices
+#device=()
+gpu_num=8
+update_freq=1
+
+root_dir=~/st
+code_dir=${root_dir}/Fairseq-S2T
+pwd_dir=$PWD
+
+# dataset
+src_lang=ti
+tgt_lang=de
+lang=${src_lang}-${tgt_lang}
+
+dataset=tibetan
+task=speech_to_text
+vocab_type=unigram
+vocab_type=char
+#vocab_type=word
+vocab_size=1700
+speed_perturb=0
+lcrm=0
+tokenizer=1
+use_raw_audio=0
+
+use_specific_dict=0
+specific_prefix=st
+specific_dir=${root_dir}/data/mustc/st
+asr_vocab_prefix=spm_unigram10000_st_share
+
+org_data_dir=${root_dir}/data/${dataset}
+data_dir=${root_dir}/data/${dataset}/asr_char
+data_dir=${root_dir}/data/${dataset}/asr_word
+#data_dir=${root_dir}/data/${dataset}/asr
+train_split=train
+valid_split=dev
+test_split=test
+test_subset=test
+
+# exp
+exp_prefix=$(date "+%m%d")
+extra_tag=
+extra_parameter=
+exp_tag=baseline
+exp_name=
+
+# config
+train_config=base
+data_config=config.yaml
+
+# training setting
+fp16=1
+max_tokens=40000
+step_valid=0
+
+# decoding setting
+cer=0
+dec_model=checkpoint_best.pt
+n_average=10
+beam_size=5
+len_penalty=1.0
+
+if [[ ${speed_perturb} -eq 1 ]]; then
+    data_dir=${data_dir}_sp
+    exp_prefix=${exp_prefix}_sp
+fi
+if [[ ${lcrm} -eq 1 ]]; then
+    data_dir=${data_dir}_lcrm
+    exp_prefix=${exp_prefix}_lcrm
+fi
+if [[ ${use_specific_dict} -eq 1 ]]; then
+    data_dir=${data_dir}_${specific_prefix}
+    exp_prefix=${exp_prefix}_${specific_prefix}
+fi
+if [[ ${tokenizer} -eq 1 ]]; then
+    data_dir=${data_dir}_tok
+    exp_prefix=${exp_prefix}_tok
+fi
+if [[ ${use_raw_audio} -eq 1 ]]; then
+    data_dir=${data_dir}_raw
+    exp_prefix=${exp_prefix}_raw
+fi
+
+. ./local/parse_options.sh || exit 1;
+
+if [[ -z ${exp_name} ]]; then
+    config_string=${train_config//,/_}
+    exp_name=${exp_prefix}_${config_string}_${exp_tag}
+    if [[ -n ${extra_tag} ]]; then
+        exp_name=${exp_name}_${extra_tag}
+    fi
+fi
+model_dir=${root_dir}/checkpoints/${dataset}/asr/${exp_name}
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+    echo "stage -1: Data Download"
+    # pass
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    ### Task dependent. You have to make data the following preparation part by yourself.
+    ### But you can utilize Kaldi recipes in most cases
+    echo "stage 0: ASR Data Preparation"
+    if [[ ! -e ${data_dir} ]]; then
+        mkdir -p ${data_dir}
+    fi
+    feature_zip=fbank80.zip
+    if [[ ${speed_perturb} -eq 1 ]]; then
+        feature_zip=fbank80_sp.zip
+    fi
+    if [[ ! -f ${data_dir}/${feature_zip} && -f ${data_dir}/../${feature_zip} ]]; then
+        ln -s ${data_dir}/../${feature_zip} ${data_dir}
+    fi
+
+    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}
+        --splits ${valid_split},${test_split},${train_split}
+        --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}
+
+    if [[ -f ${data_dir}/${feature_zip} && ! -f ${data_dir}/../${feature_zip} ]]; then
+        mv ${data_dir}/${feature_zip} ${data_dir}/..
+        ln -s ${data_dir}/../${feature_zip} ${data_dir}
+    fi
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    echo "stage 1: ASR Network Training"
+    [[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1;
+
+    if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
+		if [[ ${gpu_num} -eq 0 ]]; then
+			device=""
+		else
+        	source ./local/utils.sh
+        	device=$(get_devices $gpu_num 0)
+		fi
+    fi
+
+    echo -e "dev=${device} data=${data_dir} model=${model_dir}"
+
+    if [[ ! -d ${model_dir} ]]; then
+        mkdir -p ${model_dir}
+    else
+        echo "${model_dir} exists."
+    fi
+
+    cp ${BASH_SOURCE[0]} ${model_dir}
+    cp ${PWD}/train.sh ${model_dir}
+
+    extra_parameter="${extra_parameter}
+        --train-config ${pwd_dir}/conf/basis.yaml"
+    cp ${pwd_dir}/conf/basis.yaml ${model_dir}
+    config_list="${train_config//,/ }"
+    idx=1
+    for config in ${config_list[@]}
+    do
+        config_path=${pwd_dir}/conf/${config}.yaml
+        if [[ ! -f ${config_path} ]]; then
+            echo "No config file ${config_path}"
+            exit
+        fi
+        cp ${config_path} ${model_dir}
+
+        extra_parameter="${extra_parameter}
+        --train-config${idx} ${config_path}"
+        idx=$((idx + 1))
+    done
+
+    cmd="python3 -u ${code_dir}/fairseq_cli/train.py
+        ${data_dir}
+        --config-yaml ${data_config}
+        --task ${task}
+        --max-tokens ${max_tokens}
+        --skip-invalid-size-inputs-valid-test
+        --update-freq ${update_freq}
+        --log-interval 100
+        --save-dir ${model_dir}
+        --tensorboard-logdir ${model_dir}"
+
+	if [[ -n ${extra_parameter} ]]; then
+        cmd="${cmd}
+        ${extra_parameter}"
+    fi
+	if [[ ${gpu_num} -gt 0 ]]; then
+		cmd="${cmd}
+        --distributed-world-size $gpu_num
+        --ddp-backend no_c10d"
+	fi
+    if [[ $fp16 -eq 1 ]]; then
+        cmd="${cmd}
+        --fp16"
+    fi
+    if [[ $step_valid -eq 1 ]]; then
+        validate_interval=1
+        save_interval=1
+        keep_last_epochs=10
+        no_epoch_checkpoints=0
+        save_interval_updates=500
+        keep_interval_updates=10
+    else
+        validate_interval=1
+        keep_last_epochs=10
+    fi
+    if [[ -n $no_epoch_checkpoints && $no_epoch_checkpoints -eq 1 ]]; then
+        cmd="$cmd
+        --no-epoch-checkpoints"
+    fi
+    if [[ -n $validate_interval ]]; then
+        cmd="${cmd}
+        --validate-interval $validate_interval "
+    fi
+    if [[ -n $save_interval ]]; then
+        cmd="${cmd}
+        --save-interval $save_interval "
+    fi
+    if [[ -n $keep_last_epochs ]]; then
+        cmd="${cmd}
+        --keep-last-epochs $keep_last_epochs "
+    fi
+    if [[ -n $save_interval_updates ]]; then
+        cmd="${cmd}
+        --save-interval-updates $save_interval_updates"
+        if [[ -n $keep_interval_updates ]]; then
+        cmd="${cmd}
+        --keep-interval-updates $keep_interval_updates"
+        fi
+    fi
+
+    echo -e "\033[34mRun command: \n${cmd} \033[0m"
+
+    # save info
+    log=./history.log
+    echo "${time} | ${device} | ${data_dir} | ${exp_name} | ${model_dir} " >> $log
+    tail -n 50 ${log} > tmp.log
+    mv tmp.log $log
+    export CUDA_VISIBLE_DEVICES=${device}
+
+    cmd="nohup ${cmd} >> ${model_dir}/train.log 2>&1 &"
+    if [[ $eval -eq 1 ]]; then
+		eval $cmd
+		sleep 2s
+		tail -n "$(wc -l ${model_dir}/train.log | awk '{print $1+1}')" -f ${model_dir}/train.log
+	fi
+fi
+wait
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    echo "stage 2: ASR Decoding"
+    if [[ ${n_average} -ne 1 ]]; then
+        # Average models
+		dec_model=avg_${n_average}_checkpoint.pt
+
+        if [[ ! -f ${model_dir}/${dec_model} ]]; then
+            cmd="python ${code_dir}/scripts/average_checkpoints.py
+            --inputs ${model_dir}
+            --num-best-checkpoints ${n_average}
+            --output ${model_dir}/${dec_model}"
+            echo -e "\033[34mRun command: \n${cmd} \033[0m"
+            [[ $eval -eq 1 ]] && eval $cmd
+        fi
+	else
+		dec_model=${dec_model}
+	fi
+
+    if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
+		if [[ ${gpu_num} -eq 0 ]]; then
+			device=""
+		else
+        	source ./local/utils.sh
+        	device=$(get_devices $gpu_num 0)
+		fi
+    fi
+    export CUDA_VISIBLE_DEVICES=${device}
+
+	result_file=${model_dir}/decode_result
+	[[ -f ${result_file} ]] && rm ${result_file}
+
+    test_subset=${test_subset//,/ }
+	for subset in ${test_subset[@]}; do
+        subset=${subset}
+  		cmd="python ${code_dir}/fairseq_cli/generate.py
+        ${data_dir}
+        --config-yaml ${data_config}
+        --gen-subset ${subset}
+        --task speech_to_text
+        --path ${model_dir}/${dec_model}
+        --results-path ${model_dir}
+        --max-tokens ${max_tokens}
+        --beam ${beam_size}
+        --lenpen ${len_penalty}
+        --scoring wer
+        "
+
+        if [[ ${cer} -eq 1 ]]; then
+            cmd="${cmd}
+        --wer-char-level"
+        fi
+
+    	echo -e "\033[34mRun command: \n${cmd} \033[0m"
+
+        if [[ $eval -eq 1 ]]; then
+    	    eval $cmd
+    	    tail -n 1 ${model_dir}/generate-${subset}.txt >> ${result_file}
+        fi
+	done
+
+    cat ${result_file}
+fi

egs/tibetan/asr/train.sh

+#! /bin/bash
+
+# training the model
+
+gpu_num=1
+update_freq=1
+max_tokens=20000
+
+#extra_tag=lr0.0005
+#extra_tag=lr0.001
+#extra_tag=char
+extra_tag=word
+extra_parameter=
+#extra_tag="${extra_tag}"
+#extra_parameter="${extra_parameter} "
+
+exp_tag=batch5w
+exp_tag=pretrain
+#exp_tag=batch5w_pre_libri
+
+config_list=(purectc)
+config_list=(base)
+#config_list=(base ctc)
+#config_list=(base conformer)
+config_list=(big_wenet conformer ctc)
+
+#config_list=(pds_base_4 ctc)
+#config_list=(pds_base_8 ctc)
+
+# exp full name
+exp_name=
+
+train_config=$(echo ${config_list[*]} | sed 's/ /,/g')
+
+cmd="./run.sh
+    --stage 1
+    --stop_stage 1
+    --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/wmt16/mt/conf/deep_ctc.yaml

+arch: transformer_ctc
+share-all-embeddings: True
+optimizer: adam
+clip-norm: 10.0
+lr-scheduler: inverse_sqrt
+warmup-init-lr: 1e-7
+warmup-updates: 16000
+lr: 2e-3
+adam_betas: (0.9,0.997)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+dropout: 0.1
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+activation-fn: relu
+encoder-normalize-before: True
+decoder-normalize-before: True
+encoder-embed-dim: 512
+encoder-ffn-embed-dim: 2048
+encoder-layers: 20
+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/wmt16/mt/conf/inter.yaml

+#ctc-weight: 0.2
+interleaved-ctc-weight: 0.3
+interleaved-ctc-layers: 10,15
+interleaved-ctc-temperature: 1.0
+interleaved-ctc-drop-prob: 0
+interleaved_ctc_upsampling_ratio: 2
+
+sae-adapter: inter_league
+sae-drop-prob: 0.0
+#sae-distribution-cutoff: 10
+share-ctc-and-sae: True
+share-ctc-and-embed: True
+
+ctc-self-distill-weight: 0

fairseq/criterions/ctc.py

@@ -14,6 +14,8 @@ import logging
 
 import torch
 import torch.nn.functional as F
+from torch.distributions import Categorical
+
 from fairseq import metrics, utils
 from fairseq.criterions import FairseqCriterion, register_criterion
 from fairseq.dataclass import FairseqDataclass
@@ -43,6 +45,10 @@ class CtcCriterionConfig(FairseqDataclass):
         default=0.0,
         metadata={"help": "weight of CTC entropy"},
     )
+    ctc_entropy_cutoff: int = field(
+        default=0,
+        metadata={"help": "cutoff for CTC entropy computation"},
+    )
     interleaved_ctc_weight: float = field(
         default=0.0,
         metadata={"help": "weight of interleaved CTC loss"},
@@ -132,6 +138,7 @@ class CtcCriterion(FairseqCriterion):
         self.target_interleaved_ctc_weight = cfg.target_interleaved_ctc_weight
         self.ctc_self_distill_weight = cfg.ctc_self_distill_weight
         self.ctc_entropy = cfg.ctc_entropy
+        self.ctc_entropy_cutoff = cfg.ctc_entropy_cutoff
         self.all_ctc_weight = self.ctc_weight + self.interleaved_ctc_weight + \
                               self.target_ctc_weight + self.target_interleaved_ctc_weight + \
                               self.ctc_self_distill_weight + self.ctc_entropy
@@ -218,13 +225,16 @@ class CtcCriterion(FairseqCriterion):
                 ctc_loss += self.get_loss(lprobs, flat, input_lengths, lengths) * coef
 
             if self.ctc_entropy > 0:
-                from torch.distributions import Categorical
+
+                if self.ctc_entropy_cutoff != 0:
                     # ctc_logit = ctc_logit.sort(dim=-1, descending=True)[0][:, :, 0:100]
                     # ctc_logit = ctc_logit / ctc_logit.sum(dim=-1, keepdim=True)
-                # cut_ctc_logit = ctc_logit.sort(dim=-1, descending=True)[0][:, :, 0:100]
-                # ctc_entropy = Categorical(logits=cut_ctc_logit).entropy().sum()
-
+                    cut_ctc_logit = ctc_logit.sort(dim=-1, descending=True)[0][:, :, 0:self.ctc_entropy_cutoff]
+                    cut_ctc_logit = cut_ctc_logit / cut_ctc_logit.sum(dim=-1, keepdim=True)
+                    ctc_entropy = Categorical(logits=cut_ctc_logit).entropy().sum()
+                else:
                     ctc_entropy = Categorical(logits=ctc_logit).entropy().sum()
+
                 logging_output["ctc_entropy"] = utils.item(ctc_entropy.data)
             logging_output["ctc_loss"] = utils.item(ctc_loss.data)
 
@@ -328,14 +338,13 @@ class CtcCriterion(FairseqCriterion):
                 out = net_output["interleaved_ctc_logits"][i]
                 if type(out) == list:
                     inter_ctc_logit = out[0]
-                    padding = ~out[1]
+                    non_padding_mask = ~out[1]
                 else:
                     inter_ctc_logit = out
 
                 if inter_ctc_logit.size() != ctc_logit.size():
                     continue
 
-                ctc_self_distill_num += 1
                 loss = F.kl_div(
                     F.log_softmax(inter_ctc_logit, dim=-1, dtype=torch.float32),
                     F.softmax(ctc_logit, dim=-1, dtype=torch.float32),
@@ -344,7 +353,9 @@ class CtcCriterion(FairseqCriterion):
                 loss = loss.sum(-1).transpose(0, 1).masked_fill_(~non_padding_mask, 0.0)
                 loss = loss.sum()
                 ctc_self_distill_loss += loss
+                ctc_self_distill_num += 1
 
+            if ctc_self_distill_num != 0:
                 ctc_self_distill_loss /= ctc_self_distill_num
                 logging_output["ctc_self_distill_loss"] = utils.item(ctc_self_distill_loss.data)
 

fairseq/criterions/label_smoothed_cross_entropy_with_ctc.py

@@ -52,6 +52,16 @@ class LabelSmoothedCrossEntropyCriterionWithCTC(
         3) logging outputs to display while training
         """
         src_tokens, src_lengths, prev_output_tokens = sample["net_input"].values()
+
+        from fairseq.models.speech_to_text import S2TDualModel
+        if isinstance(model, S2TDualModel):
+            assert "transcript" in sample
+            text_src_tokens = sample["transcript"]["tokens"]
+            text_src_lengths = sample["transcript"]["lengths"]
+
+            encoder_out = model.encoder(src_tokens, src_lengths,
+                                        text_src_tokens, text_src_lengths)
+        else:
             encoder_out = model.encoder(src_tokens=src_tokens, src_lengths=src_lengths)
 
         use_mixup = False

fairseq/dataclass/configs.py

@@ -854,6 +854,10 @@ class GenerationConfig(FairseqDataclass):
         default=False,
         metadata={"help": "if set, dont use seed for initializing random generators"},
     )
+    ctc_infer: bool = field(
+        default=False,
+        metadata={"help": "generate CTC decoding results during inference"}
+    )
 
 
 @dataclass

fairseq/models/speech_to_text/pdss2t_transformer.py

@@ -595,7 +595,7 @@ class PDSS2TTransformerEncoder(FairseqEncoder):
 
         pad1 = encoder_padding_mask[idx1]
         pad2 = encoder_padding_mask[idx2]
-        encoder_padding_mask = pad1 + pad2
+        encoder_padding_mask = pad1 & pad2
         input_lengths = (~encoder_padding_mask).sum(-1)
 
         mixup = {

fairseq/models/speech_to_text/s2t_dual.py

@@ -26,6 +26,7 @@ from fairseq.models.speech_to_text import (
     S2TTransformerEncoder,
     PDSS2TTransformerModel,
     PDSS2TTransformerEncoder,
+    S2TSATEModel,
     S2TSATEEncoder,
 )
 from fairseq.modules.speech_to_text import Adapter, CTC
@@ -58,73 +59,18 @@ class S2TDualModel(FairseqEncoderDecoderModel):
     @staticmethod
     def add_args(parser):
         """Add model-specific arguments to the parser."""
-        PDSS2TTransformerModel.add_args(parser)
-        # SATE setting
-        parser.add_argument(
-            "--text-encoder-layers",
-            default=6,
-            type=int,
-            help="layers of the text encoder",
-        )
-        parser.add_argument(
-            "--text-attention-type",
-            default="selfattn",
-            type=str,
-            help="attention type of the textual encoder",
-        )
-        parser.add_argument(
-            "--adapter",
-            default="league",
-            type=str,
-            help="adapter type",
-        )
-        parser.add_argument(
-            "--ctc-compress-strategy",
-            default="avg",
-            type=str,
-            help="compress strategy, such as avg, weighted, and softmax",
-        )
-        parser.add_argument(
-            "--share-ctc-and-adapter",
-            default=False,
-            action="store_true",
-            help="share the projection weights of the ctc and adapter",
-        )
-        parser.add_argument(
-            "--temperature",
-            default=1.0,
-            type=float,
-            help="temperature of the CTC softmax",
-        )
-        parser.add_argument(
-            "--acoustic-encoder",
-            default="transformer",
-            type=str,
-            help="the architecture of the acoustic encoder",
-        )
-        parser.add_argument(
-            "--target-ctc-layers",
-            default=None,
-            type=str,
-            help="ctc layers for target sentence",
-        )
-        parser.add_argument(
-            "--load-pretrained-acoustic-encoder-from",
-            type=str,
-            metavar="STR",
-            help="model to take acoustic encoder weights from (for initialization)",
-        )
-        parser.add_argument(
-            "--load-pretrained-text-encoder-from",
-            type=str,
-            metavar="STR",
-            help="model to take text encoder weights from (for initialization)",
-        )
+        S2TTransformerModel.add_args(parser)
+        PDSS2TTransformerModel.add_specific_args(parser)
+        S2TSATEModel.add_specific_args(parser)
+        S2TDualModel.add_specific_args(parser)
 
+    @staticmethod
+    def add_specific_args(parser):
         # multi-encoder
         parser.add_argument(
             "--asr-encoder",
             default="transformer",
+            choices=["transformer", "pds", "sate", "wav2vec"],
             type=str,
             help="the architecture of the ASR encoder",
         )
@@ -134,12 +80,11 @@ class S2TDualModel(FairseqEncoderDecoderModel):
             type=str,
             help="the architecture of the MT encoder",
         )
-        # parser.add_argument(
-        #     "--mt-encoder-dim",
-        #     default="transformer",
-        #     type=str,
-        #     help="the dimension of the MT encoder",
-        # )
+        parser.add_argument(
+            "--mt-encoder-dim",
+            type=int,
+            help="the dimension of the MT encoder",
+        )
         parser.add_argument(
             "--mt-encoder-layers",
             default=6,
@@ -148,13 +93,13 @@ class S2TDualModel(FairseqEncoderDecoderModel):
         )
         parser.add_argument(
             "--encoder-asr-ratio",
-            default=1,
+            default=0.5,
             type=float,
             help="the ratio of the asr representation",
         )
         parser.add_argument(
             "--encoder-mt-ratio",
-            default=1,
+            default=0.5,
             type=float,
             help="the ratio of the mt representation",
         )
@@ -165,9 +110,9 @@ class S2TDualModel(FairseqEncoderDecoderModel):
         )
         parser.add_argument(
             "--encoder-drop-net-prob",
-            default=0.5,
+            default=0.2,
             type=float,
-            help="the probability of dropping",
+            help="probability of dropping one of the representations",
         )
         parser.add_argument(
             "--encoder-drop-net-mix",
@@ -268,9 +213,6 @@ class S2TDualModel(FairseqEncoderDecoderModel):
                 tgt_dict, args.decoder_embed_dim
             )
 
-        setattr(args, "encoder_s1_ratio", args.encoder_asr_ratio)
-        setattr(args, "encoder_s2_ratio", args.encoder_mt_ratio)
-
         encoder = cls.build_encoder(args, task, encoder_embed_tokens)
         if getattr(args, "encoder_freeze_module", None):
             utils.freeze_parameters(encoder, args.encoder_freeze_module)
@@ -326,8 +268,11 @@ class S2TDualEncoder(FairseqEncoder):
         else:
             logger.error("Unsupported ASR architecture: %s." % asr_encoder_type)
 
-        attn_type = args.encoder_attention_type
+        setattr(args, "encoder_s1_ratio", args.encoder_asr_ratio)
+        setattr(args, "encoder_s2_ratio", args.encoder_mt_ratio)
         setattr(args, "encoder_layers", args.mt_encoder_layers)
+
+        attn_type = args.encoder_attention_type
         setattr(args, "encoder_attention_type", "selfattn")
         self.mt_encoder = TransformerS2Encoder(args, task.source_dictionary, embed_tokens)
         setattr(args, "encoder_attention_type", attn_type)

fairseq/models/speech_to_text/s2t_transformer.py

@@ -5,7 +5,6 @@ from typing import Dict, List, Optional, Tuple
 import numpy as np
 import torch
 import torch.nn as nn
-import torch.nn.functional as F
 
 from fairseq import checkpoint_utils, utils
 from fairseq.data.data_utils import lengths_to_padding_mask
@@ -397,6 +396,11 @@ class S2TTransformerModel(FairseqEncoderDecoderModel):
             action="store_true",
             help="use linear transform after down-sampling",
         )
+        parser.add_argument(
+            "--embed-norm",
+            action="store_true",
+            help="use layer norm after down-sampling",
+        )
 
         # interleaved CTC layers
         parser.add_argument(
@@ -438,10 +442,22 @@ class S2TTransformerModel(FairseqEncoderDecoderModel):
             help="cutoff of the distribution in sae",
         )
         parser.add_argument(
-            "--share-ctc-and-sae",
+            "--share-sae-and-ctc",
             action="store_true",
             help="share the weight of ctc and sae",
         )
+        parser.add_argument(
+            "--sae-embed-norm",
+            default=False,
+            action="store_true",
+            help="use the layer norm for embed output",
+        )
+        parser.add_argument(
+            "--sae-out-norm",
+            default=False,
+            action="store_true",
+            help="use the layer norm for final output",
+        )
 
         # Mixup
         parser.add_argument(
@@ -580,8 +596,11 @@ class S2TTransformerEncoder(FairseqEncoder):
 
         self.subsample = subsampling(args)
         self.embed_linear = getattr(args, "embed_linear", False)
+        self.embed_norm = getattr(args, "embed_norm", False)
         if self.embed_linear:
             self.linear = nn.Linear(dim, dim)
+        if self.embed_norm:
+            self.embed_ln = LayerNorm(dim)
 
         self.attn_type = getattr(args, "encoder_attention_type", "selfattn")
         if self.attn_type == "rel_pos":
@@ -645,13 +664,16 @@ class S2TTransformerEncoder(FairseqEncoder):
             if not self.use_ctc:
                 self.ctc = CTC(dim,
                                dictionary_size=len(task.source_dictionary),
-                               dropout=args.dropout)
+                               dropout=args.dropout,
+                               )
                 if getattr(args, "share_ctc_and_embed", False) and \
                         task.source_dictionary == task.target_dictionary and \
                         embed_tokens is not None and dim == embed_tokens.embedding_dim:
                     self.ctc.ctc_projection.weight = embed_tokens.weight
 
             strategy = {
+                "embed_norm": getattr(args, "sae_embed_norm", False),
+                "out_norm": getattr(args, "sae_out_norm", False),
                 "ctc_compress_strategy": getattr(args, "ctc_compress_strategy", None),
                 "distribution_cutoff": getattr(args, "sae_distribution_cutoff", None),
                 "drop_prob": getattr(args, "sae_drop_prob", 0),
@@ -661,7 +683,7 @@ class S2TTransformerEncoder(FairseqEncoder):
                                len(task.source_dictionary),
                                strategy=strategy,
                                )
-            if args.share_ctc_and_sae and hasattr(self.sae, "embed_adapter"):
+            if args.share_sae_and_ctc and hasattr(self.sae, "embed_adapter"):
                 self.sae.embed_adapter.weight = self.ctc.ctc_projection.weight
 
         # mixup
@@ -682,6 +704,17 @@ class S2TTransformerEncoder(FairseqEncoder):
         self.dis = 2
         self.cos_sim = dict()
 
+        # debug the variance
+        self.debug_var = False
+
+    def set_ctc_infer(self, ctc_infer, post_process, src_dict=None, tgt_dict=None):
+        if hasattr(self, "ctc"):
+            assert src_dict is not None
+            self.ctc.set_infer(ctc_infer, post_process, src_dict)
+
+    def set_debug_var(self, debug_var_flag):
+        self.debug_var = debug_var_flag
+
     @staticmethod
     def pooling_ratio():
         return 4
@@ -730,7 +763,7 @@ class S2TTransformerEncoder(FairseqEncoder):
 
         pad1 = encoder_padding_mask[idx1]
         pad2 = encoder_padding_mask[idx2]
-        encoder_padding_mask = pad1 + pad2
+        encoder_padding_mask = pad1 & pad2
         input_lengths = (~encoder_padding_mask).sum(-1)
 
         mixup = {
@@ -741,7 +774,15 @@ class S2TTransformerEncoder(FairseqEncoder):
         }
         return x, encoder_padding_mask, input_lengths, mixup
 
-    def forward(self, src_tokens, src_lengths):
+    def show_debug(self, x, text=None):
+        if not self.debug_var:
+            return
+        if text:
+            logger.info("--- Variance of %s: %f." % (text, x.var()))
+        else:
+            logger.info("--- Variance: %f." % (x.var()))
+
+    def forward(self, src_tokens, src_lengths=None, **kwargs):
 
         layer_idx = -1
         mixup = None
@@ -753,6 +794,7 @@ class S2TTransformerEncoder(FairseqEncoder):
         x = src_tokens.transpose(0, 1)
         input_lengths = src_lengths
 
+        self.show_debug(x, "input x")
         # gather cosine similarity
         cos_sim_idx = -1
         dis = self.dis
@@ -766,11 +808,19 @@ class S2TTransformerEncoder(FairseqEncoder):
 
         # down-sampling
         x, input_lengths = self.subsample(x, input_lengths)
+        self.show_debug(x, "x after subsampling")
 
         encoder_padding_mask = lengths_to_padding_mask(input_lengths)
+        if encoder_padding_mask is not None:
+            x = x * (1 - encoder_padding_mask.transpose(0, 1).unsqueeze(-1).type_as(x))
+
+        if self.embed_norm:
+            x = self.embed_ln(x)
+            self.show_debug(x, "x after embed norm")
 
         # embedding scaling
         x = self.embed_scale * x
+        self.show_debug(x, "x after scale")
 
         # position embedding
         if self.attn_type in ["rel_pos", "rel_pos_legacy", "rel_selfattn"]:
@@ -783,9 +833,12 @@ class S2TTransformerEncoder(FairseqEncoder):
             positions = self.embed_positions(encoder_padding_mask).transpose(0, 1)
             x += positions
             positions = None
+        self.show_debug(x, "x after position embedding")
 
         if self.embed_linear:
             x = self.linear(x)
+            self.show_debug(x, "x after embed linear")
+
         x = self.dropout_module(x)
 
         # add emb into history
@@ -806,6 +859,7 @@ class S2TTransformerEncoder(FairseqEncoder):
             if torch.rand(1) < self.mixup_prob:
                 x, encoder_padding_mask, input_lengths, mixup = self.apply_mixup(x, encoder_padding_mask)
 
+        self.show_debug(x, "x before encoding")
         for layer in self.layers:
             if self.history is not None:
                 x = self.history.pop()
@@ -813,13 +867,14 @@ class S2TTransformerEncoder(FairseqEncoder):
             # encoder layer
             x = layer(x, encoder_padding_mask, pos_emb=positions)
             layer_idx += 1
+            self.show_debug(x, "x after layer %d" % layer_idx)
 
             if self.training and self.mixup and layer_idx == self.mixup_layer:
                 if torch.rand(1) < self.mixup_prob:
                     x, encoder_padding_mask, input_lengths, mixup = self.apply_mixup(x, encoder_padding_mask)
 
             if self.use_ctc and self.inter_ctc and self.ctc_layer == layer_idx:
-                ctc_logit = self.ctc(x.clone())
+                ctc_logit = self.ctc(x.clone(), encoder_padding_mask)
 
             # interleaved CTC
             if layer_idx in self.interleaved_ctc_layers:
@@ -829,7 +884,7 @@ class S2TTransformerEncoder(FairseqEncoder):
                         break
 
                 norm_x = self.layer_norm(x)
-                logit = self.ctc(norm_x)
+                logit = self.ctc(norm_x, encoder_padding_mask)
 
                 interleaved_ctc_logits.append(logit)
 
@@ -837,7 +892,8 @@ class S2TTransformerEncoder(FairseqEncoder):
                                     max=1e8 if logit.dtype == torch.float32 else 1e4)
 
                 prob = utils.softmax(logit / self.interleaved_ctc_temperature, dim=-1)
-                x, encoder_padding_mask = self.sae([x, prob], encoder_padding_mask)
+                x, encoder_padding_mask = self.sae([norm_x, prob], encoder_padding_mask)
+                self.show_debug(x, "x after sae")
 
             # gather cosine similarity
             if self.gather_cos_sim:
@@ -850,11 +906,14 @@ class S2TTransformerEncoder(FairseqEncoder):
         if self.history is not None:
             x = self.history.pop()
 
+        self.show_debug(x, "x after encoding")
         if self.layer_norm is not None:
             x = self.layer_norm(x)
+        self.show_debug(x, "x after encoding layer norm")
 
         if self.use_ctc and ctc_logit is None:
-            ctc_logit = self.ctc(x)
+            ctc_logit = self.ctc(x, encoder_padding_mask)
+            self.show_debug(x, "x after ctc")
 
         return {
             "encoder_out": [x],  # T x B x C
@@ -1005,6 +1064,7 @@ def base_architecture(args):
     args.quant_noise_pq = getattr(args, "quant_noise_pq", 0)
 
     args.embed_linear = getattr(args, "embed_linear", False)
+    args.embed_norm = getattr(args, "embed_norm", False)
 
     # CTC
     args.ctc_layer = getattr(args, "ctc_layer", 0)
@@ -1045,7 +1105,9 @@ def base_architecture(args):
 
     # Semantics-augmented Encoding (sae)
     args.sae_adapter = getattr(args, "sae_adapter", "none")
-    args.share_ctc_and_sae = getattr(args, "share_ctc_and_sae", False)
+    args.share_sae_and_ctc = getattr(args, "share_sae_and_ctc", False)
+    args.sae_embed_norm = getattr(args, "sae_embed_norm", False)
+    args.sae_out_norm = getattr(args, "sae_out_norm", False)
     args.sae_drop_prob = getattr(args, "sae_drop_prob", 0)
     args.sae_distribution_cutoff = getattr(args, "sae_distribution_cutoff", None)
 

fairseq/models/transformer.py

@@ -1063,7 +1063,7 @@ class TransformerDecoder(FairseqIncrementalDecoder):
             if self_attn_padding_mask is not None:
                 pad1 = self_attn_padding_mask[idx1]
                 pad2 = self_attn_padding_mask[idx2]
-                self_attn_padding_mask = pad1 + pad2
+                self_attn_padding_mask = pad1 & pad2
 
         # decoder layers
         avg_attn = None

fairseq/models/transformer_ctc.py

@@ -606,6 +606,7 @@ class TransformerCTCEncoder(FairseqEncoder):
             self.ctc = CTC(embed_dim,
                            dictionary_size=decoder_embed_tokens.num_embeddings,
                            dropout=args.dropout,
+                           dictionary=dictionary,
                            need_layernorm=True if self.inter_ctc else False)
 
             self.ctc.ctc_projection.weight = decoder_embed_tokens.weight
@@ -627,6 +628,7 @@ class TransformerCTCEncoder(FairseqEncoder):
             if not self.use_ctc:
                 self.ctc = CTC(embed_dim,
                                dictionary_size=decoder_embed_tokens.num_embeddings,
+                               dictionary=dictionary,
                                dropout=args.dropout)
 
                 self.ctc.ctc_projection.weight = decoder_embed_tokens.weight
@@ -715,6 +717,10 @@ class TransformerCTCEncoder(FairseqEncoder):
         x = x.unsqueeze(1).expand(-1, ratio, -1, -1).reshape(-1, bsz, dim)
         return x
 
+    def set_ctc_infer(self, ctc_infer, post_process):
+        if hasattr(self, "ctc"):
+            self.ctc.set_infer(ctc_infer, post_process)
+
     # TorchScript doesn't support super() method so that the scriptable Subclass
     # can't access the base class model in Torchscript.
     # Current workaround is to add a helper function with different name and
@@ -764,12 +770,19 @@ class TransformerCTCEncoder(FairseqEncoder):
 
         # B x T x C -> T x B x C
         x = x.transpose(0, 1)
+        bsz = x.size(1)
 
         encoder_states = []
 
         if return_all_hiddens:
             encoder_states.append(x)
 
+        org_encoder_padding_mask = encoder_padding_mask
+        ctc_padding_mask = encoder_padding_mask
+        if self.use_ctc or len(self.interleaved_ctc_layers) != 0:
+            ctc_padding_mask = encoder_padding_mask.unsqueeze(-1). \
+                expand(-1, -1, self.interleaved_ctc_upsampling_ratio).reshape(bsz, -1)
+
         # add emb into history
         if self.history is not None:
             self.history.push(x)
@@ -782,6 +795,10 @@ class TransformerCTCEncoder(FairseqEncoder):
             if self.history is not None:
                 x = self.history.pop()
 
+            if layer_idx + 1 in self.interleaved_ctc_layers:
+                x = self.upsampling(x)
+                encoder_padding_mask = ctc_padding_mask
+
             x = layer(
                 x, encoder_padding_mask=encoder_padding_mask if has_pads else None
             )
@@ -792,7 +809,7 @@ class TransformerCTCEncoder(FairseqEncoder):
 
             # CTC
             if self.use_ctc and self.inter_ctc and self.ctc_layer == layer_idx:
-                ctc_logit = self.ctc(self.upsampling(x.clone()))
+                ctc_logit = self.ctc(self.upsampling(x.clone()), ctc_padding_mask)
 
             # Interleaved CTC
             if layer_idx in self.interleaved_ctc_layers:
@@ -802,17 +819,17 @@ class TransformerCTCEncoder(FairseqEncoder):
                         break
 
                 norm_x = self.layer_norm(x)
-                up_x = self.upsampling(norm_x)
-                up_logit = self.ctc(up_x)
+                logit = self.ctc(norm_x, ctc_padding_mask)
 
-                interleaved_ctc_logits.append(up_logit)
-                up_prob = utils.softmax(up_logit / self.interleaved_ctc_temperature, dim=-1)
+                interleaved_ctc_logits.append(logit)
+                prob = utils.softmax(logit / self.interleaved_ctc_temperature, dim=-1)
 
-                up_prob = up_prob.permute(1, 2, 0)
-                prob = self.pool(up_prob)
-                prob = prob.permute(2, 0, 1)
+                x, _ = self.sae([norm_x, prob])
 
-                x, _ = self.sae([x, prob])
+                x = x.permute(1, 2, 0)
+                x = self.pool(x)
+                x = x.permute(2, 0, 1)
+                encoder_padding_mask = org_encoder_padding_mask
 
             if self.history is not None:
                 self.history.push(x)
@@ -824,13 +841,7 @@ class TransformerCTCEncoder(FairseqEncoder):
             x = self.layer_norm(x)
 
         if self.use_ctc and ctc_logit is None:
-            ctc_logit = self.ctc(self.upsampling(x))
-
-        ctc_padding_mask = encoder_padding_mask
-        if ctc_logit is not None or len(interleaved_ctc_logits) != 0:
-            bsz = encoder_padding_mask.size(0)
-            ctc_padding_mask = encoder_padding_mask.unsqueeze(-1). \
-                expand(-1, -1, self.interleaved_ctc_upsampling_ratio).reshape(bsz, -1)
+            ctc_logit = self.ctc(self.upsampling(x), ctc_padding_mask)
 
         # The Pytorch Mobile lite interpreter does not supports returning NamedTuple in
         # `forward` so we use a dictionary instead.
@@ -1226,7 +1237,7 @@ class TransformerDecoder(FairseqIncrementalDecoder):
             if self_attn_padding_mask is not None:
                 pad1 = self_attn_padding_mask[idx1]
                 pad2 = self_attn_padding_mask[idx2]
-                self_attn_padding_mask = pad1 + pad2
+                self_attn_padding_mask = pad1 & pad2
 
         # decoder layers
         avg_attn = None

fairseq/models/transformer_s2.py

@@ -317,7 +317,7 @@ class TransformerS2Decoder(TransformerDecoder):
             if self_attn_padding_mask is not None:
                 pad1 = self_attn_padding_mask[idx1]
                 pad2 = self_attn_padding_mask[idx2]
-                self_attn_padding_mask = pad1 + pad2
+                self_attn_padding_mask = pad1 & pad2
 
         # decoder layers
         avg_attn = None

fairseq/modules/speech_to_text/adapter.py

@@ -69,15 +69,19 @@ class Adapter(nn.Module):
         if self.adapter_type in ["linear", "league", "gated_league", "gated_league2"]:
             self.cal_linear = True
             self.linear_adapter = nn.Sequential(
-                nn.Linear(dim, dim),
-                LayerNorm(dim),
+                nn.Linear(dim, 2 * dim),
                 nn.ReLU(),
+                nn.Linear(2 * dim, dim),
+                LayerNorm(dim),
             )
 
         if self.adapter_type in ["context", "league", "gated_league", "gated_league2", "inter_league"]:
             self.cal_context = True
             self.embed_adapter = nn.Linear(dim, dictionary_size, bias=False)    # reverse for initialization
             nn.init.normal_(self.embed_adapter.weight, mean=0, std=dim ** -0.5)
+            self.embed_norm = strategy.get("embed_norm", False)
+            if self.embed_norm:
+                self.embed_ln = LayerNorm(dim)
             if embed_tokens is not None:
                 self.embed_adapter.weight = embed_tokens.weight
 
@@ -90,7 +94,7 @@ class Adapter(nn.Module):
         # additional strategy
         if self.adapter_type == "shrink":
             assert strategy is not None
-            ctc_compress_strategy = getattr(strategy, "ctc_compress_strategy", "avg")
+            ctc_compress_strategy = strategy.get("ctc_compress_strategy", "avg")
             self.ctc_compress = getattr(CTCCompressStrategy, ctc_compress_strategy)
             logger.info("CTC Compress Strategy: %s" % ctc_compress_strategy)
 
@@ -103,6 +107,9 @@ class Adapter(nn.Module):
             self.drop_prob = strategy.get("drop_prob", 0)
             if self.drop_prob != 0:
                 logger.info("Adapter drop probability: %f" % self.drop_prob)
+        self.out_norm = strategy.get("out_norm", False)
+        if self.out_norm:
+            self.out_ln = LayerNorm(dim)
 
     def forward(self, x, padding=None):
 
@@ -119,13 +126,19 @@ class Adapter(nn.Module):
         if self.cal_context:
             if self.distribution_cutoff is not None:
                 cutoff = min(int(self.distribution_cutoff), org_distribution.size(-1) - 1)
-                threshold = org_distribution.sort(dim=-1, descending=True)[0][:, :, cutoff:cutoff+1]
+                # threshold = org_distribution.sort(dim=-1, descending=True)[0][:, :, cutoff:cutoff+1]
+                # distribution = torch.where(
+                #     org_distribution > threshold, org_distribution, torch.zeros_like(org_distribution)
+                # )
+                threshold = org_distribution.sort(dim=-1, descending=True)[0][:, :, :cutoff].sum(-1, keepdim=True)
                 distribution = torch.where(
-                    org_distribution > threshold, org_distribution, torch.zeros_like(org_distribution)
+                    threshold > 0.9, org_distribution, torch.zeros_like(org_distribution)
                 )
                 distribution = distribution.view(-1, distribution.size(-1))
 
             soft_out = torch.mm(distribution, self.embed_adapter.weight).view(seq_len, bsz, -1)
+            if self.embed_norm:
+                soft_out = self.embed_ln(soft_out)
 
         if self.adapter_type == "linear":
             out = linear_out
@@ -134,7 +147,7 @@ class Adapter(nn.Module):
             out = soft_out
 
         elif self.adapter_type == "league":
-            if self.drop_prob > 0 and torch.rand(1).uniform_() < self.drop_prob:
+            if self.training and self.drop_prob > 0 and torch.rand(1).uniform_() < self.drop_prob:
                 if torch.rand(1).uniform_() < 0.5:
                     out = linear_out
                 else:
@@ -178,4 +191,7 @@ class Adapter(nn.Module):
             out = None
             logging.error("Unsupported adapter type: {}.".format(self.adapter_type))
 
+        if self.out_norm:
+            out = self.out_ln(out)
+
         return out, padding

fairseq/modules/speech_to_text/ctc.py

 import logging
 
+import editdistance
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -8,13 +9,15 @@ from fairseq.modules import (
     FairseqDropout,
     LayerNorm,
 )
+from fairseq.data.data_utils import post_process
 
 logger = logging.getLogger(__name__)
 
 
 class CTC(nn.Module):
 
-    def __init__(self, embed_dim, dictionary_size, dropout, need_layernorm=False):
+    def __init__(self, embed_dim, dictionary_size, dropout,
+                 need_layernorm=False, dictionary=None):
         super(CTC, self).__init__()
 
         self.embed_dim = embed_dim
@@ -26,15 +29,31 @@ class CTC(nn.Module):
         self.ctc_dropout_module = FairseqDropout(
             p=dropout, module_name=self.__class__.__name__
         )
+
         self.need_layernorm = need_layernorm
         if self.need_layernorm:
             self.LayerNorm = LayerNorm(embed_dim)
 
-    def forward(self, x):
+        self.dictionary = dictionary
+        self.infer_decoding = False
+        self.post_process = "sentencepiece"
+
+    def set_infer(self, is_infer, text_post_process, dictionary):
+        self.infer_decoding = is_infer
+        self.post_process = text_post_process
+        self.dictionary = dictionary
+
+    def forward(self, x, padding=None):
         if self.need_layernorm:
             x = self.LayerNorm(x)
 
         x = self.ctc_projection(self.ctc_dropout_module(x))
+
+        if not self.training and self.infer_decoding:
+            assert self.dictionary is not None
+            input_lengths = (~padding).sum(-1)
+            self.infer(x.transpose(0, 1).float().contiguous().cpu(), input_lengths)
+
         return x
 
     def softmax(self, x, temperature=1.0):
@@ -45,3 +64,55 @@ class CTC(nn.Module):
 
     def argmax(self, x):
         return torch.argmax(self.ctc_projection(x), dim=-1)
+
+    def infer(self, logits_or_probs, lengths):
+        for lp, inp_l in zip(
+                logits_or_probs,
+                lengths,
+        ):
+            lp = lp[:inp_l].unsqueeze(0)
+
+            toks = lp.argmax(dim=-1).unique_consecutive()
+            pred_units_arr = toks[toks != self.dictionary.bos()].tolist()
+
+            pred_units = self.dictionary.string(pred_units_arr)
+            pred_words_raw = post_process(pred_units, self.post_process).split()
+
+            print(pred_words_raw)
+
+    def valid(self, logits_or_probs, target, lengths):
+
+        c_err = 0
+        c_len = 0
+        w_errs = 0
+        w_len = 0
+
+        for lp, t, inp_l in zip(
+                logits_or_probs,
+                target,
+                lengths,
+        ):
+            lp = lp[:inp_l].unsqueeze(0)
+
+            p = (t != self.task.target_dictionary.pad()) & (
+                    t != self.task.target_dictionary.eos()
+            )
+            targ = t[p]
+            targ_units = self.task.target_dictionary.string(targ)
+            targ_units_arr = targ.tolist()
+
+            toks = lp.argmax(dim=-1).unique_consecutive()
+            pred_units_arr = toks[toks != self.blank_idx].tolist()
+
+            c_err += editdistance.eval(pred_units_arr, targ_units_arr)
+            c_len += len(targ_units_arr)
+
+            targ_words = post_process(targ_units, self.post_process).split()
+
+            pred_units = self.task.target_dictionary.string(pred_units_arr)
+            pred_words_raw = post_process(pred_units, self.post_process).split()
+
+            dist = editdistance.eval(pred_words_raw, targ_words)
+            w_errs += dist
+
+            w_len += len(targ_words)
\ No newline at end of file

fairseq/modules/transformer_s2_layer.py

@@ -242,7 +242,11 @@ class TransformerS2EncoderLayer(nn.Module):
         x = self.dropout_module(x)
 
         if x2 is not None:
-            x2, _ = self.s2_attn(x, x2, x2, x2_encoder_padding_mask)
+            x2, _ = self.s2_attn(
+                query=x,
+                key=x2,
+                value=x2,
+                key_padding_mask=x2_encoder_padding_mask)
             x2 = self.dropout_module(x2)
             ratio = self.get_ratio()
             x = x * ratio[0] + x2 * ratio[1]
@@ -557,6 +561,8 @@ class TransformerS2DecoderLayer(nn.Module):
                     saved_state["prev_key_padding_mask"] = prev_attn_state[2]
                 assert incremental_state is not None
                 self.encoder_attn._set_input_buffer(incremental_state, saved_state)
+                # notice here
+                # self.s2_attn._set_input_buffer(incremental_state, saved_state)
 
             x, attn = self.encoder_attn(
                 query=x,

fairseq/search.py

@@ -137,7 +137,8 @@ class BeamSearch(Search):
         scores_buf = top_prediction[0]
         indices_buf = top_prediction[1]
         # Project back into relative indices and beams
-        beams_buf = indices_buf // vocab_size
+        # beams_buf = indices_buf // vocab_size
+        beams_buf = torch.div(indices_buf, vocab_size, rounding_mode='trunc')
         indices_buf = indices_buf.fmod(vocab_size)
 
         # At this point, beams_buf and indices_buf are single-dim and contain relative indices

fairseq/sequence_generator.py

@@ -14,6 +14,8 @@ from fairseq.models import FairseqIncrementalDecoder
 from torch import Tensor
 from fairseq.ngram_repeat_block import NGramRepeatBlock
 
+from fairseq.models.speech_to_text import S2TDualModel
+
 
 class SequenceGenerator(nn.Module):
     def __init__(
@@ -197,11 +199,9 @@ class SequenceGenerator(nn.Module):
         )
         net_input = sample["net_input"]
 
-        # if "transcript" in sample:
-        #     text_src_tokens = sample["transcript"]["tokens"]
-        #     text_src_lengths = sample["transcript"]["lengths"]
-        #     net_input["text_src_tokens"] = text_src_tokens
-        #     net_input["text_src_lengths"] = text_src_lengths
+        if "transcript" in sample:
+            net_input["text_src_tokens"] = sample["transcript"]["tokens"]
+            net_input["text_src_lengths"] = sample["transcript"]["lengths"]
 
         if "src_tokens" in net_input:
             src_tokens = net_input["src_tokens"]
@@ -662,7 +662,8 @@ class SequenceGenerator(nn.Module):
             idx = bbsz_idx[i]
             score = eos_scores[i]
             # sentence index in the current (possibly reduced) batch
-            unfin_idx = idx // beam_size
+            # unfin_idx = idx // beam_size
+            unfin_idx = torch.div(idx, beam_size, rounding_mode='trunc')
             # sentence index in the original (unreduced) batch
             sent = unfin_idx + cum_unfin[unfin_idx]
             # Cannot create dict for key type '(int, int)' in torchscript.
@@ -760,7 +761,18 @@ class EnsembleModel(nn.Module):
     def forward_encoder(self, net_input: Dict[str, Tensor]):
         if not self.has_encoder():
             return None
-        return [model.encoder.forward_torchscript(net_input) for model in self.models]
+
+        encoder_outs = []
+        for model in self.models:
+            if not isinstance(model, S2TDualModel):
+                if "text_src_tokens" in net_input:
+                    net_input.pop("text_src_tokens", None)
+                if "text_src_lengths" in net_input:
+                    net_input.pop("text_src_lengths", None)
+            encoder_outs.append(model.encoder.forward_torchscript(net_input))
+        return encoder_outs
+
+        # return [model.encoder.forward_torchscript(net_input) for model in self.models]
 
     @torch.jit.export
     def forward_decoder(

fairseq_cli/generate.py

@@ -105,6 +105,10 @@ def _main(cfg: DictConfig, output_file, translation_path=None):
         strict=(cfg.checkpoint.checkpoint_shard_count == 1),
         num_shards=cfg.checkpoint.checkpoint_shard_count,
     )
+    for model in models:
+        if hasattr(model, "encoder") and hasattr(model.encoder, "set_ctc_infer"):
+            model.encoder.set_ctc_infer(cfg.generation.ctc_infer, cfg.common_eval.post_process,
+                                        src_dict, tgt_dict)
 
     # loading the dataset should happen after the checkpoint has been loaded so we can give it the saved task config
     task.load_dataset(cfg.dataset.gen_subset, task_cfg=saved_cfg.task)