Ignore padding frames during RNN-T decoding. (k2-fsa#358)

* Ignore padding frames during RNN-T decoding. * Fix outdated decoding code. * Minor fixes.
goodatlas · May 12, 2022 · aeb8986 · aeb8986
1 parent bc284e8
commit aeb8986
Show file tree

Hide file tree

Showing 36 changed files with 2,205 additions and 764 deletions.
diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless-2022-03-12.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless-2022-03-12.sh
@@ -33,7 +33,7 @@ for sym in 1 2 3; do
     $repo/test_wavs/1221-135766-0002.wav
 done
 
-for method in modified_beam_search beam_search; do
+for method in fast_beam_search modified_beam_search beam_search; do
   log "$method"
 
   ./pruned_transducer_stateless/pretrained.py \
@@ -47,7 +47,8 @@ for method in modified_beam_search beam_search; do
 done
 
 echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
-if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
   mkdir -p pruned_transducer_stateless/exp
   ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless/exp/epoch-999.pt
   ln -s $PWD/$repo/data/lang_bpe_500 data/
@@ -58,9 +59,9 @@ if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
   log "Decoding test-clean and test-other"
 
   # use a small value for decoding with CPU
-  max_duration=50
+  max_duration=100
 
-  for method in greedy_search fast_beam_search; do
+  for method in greedy_search fast_beam_search modified_beam_search; do
     log "Decoding with $method"
 
     ./pruned_transducer_stateless/decode.py \

diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh
@@ -51,7 +51,8 @@ for method in modified_beam_search beam_search fast_beam_search; do
 done
 
 echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
-if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
   mkdir -p pruned_transducer_stateless2/exp
   ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless2/exp/epoch-999.pt
   ln -s $PWD/$repo/data/lang_bpe_500 data/
@@ -62,9 +63,9 @@ if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
   log "Decoding test-clean and test-other"
 
   # use a small value for decoding with CPU
-  max_duration=50
+  max_duration=100
 
-  for method in greedy_search fast_beam_search; do
+  for method in greedy_search fast_beam_search modified_beam_search; do
     log "Decoding with $method"
 
     ./pruned_transducer_stateless2/decode.py \

diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh
@@ -51,7 +51,8 @@ for method in modified_beam_search beam_search fast_beam_search; do
 done
 
 echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
-if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
   mkdir -p pruned_transducer_stateless3/exp
   ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless3/exp/epoch-999.pt
   ln -s $PWD/$repo/data/lang_bpe_500 data/
@@ -62,9 +63,9 @@ if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
   log "Decoding test-clean and test-other"
 
   # use a small value for decoding with CPU
-  max_duration=50
+  max_duration=100
 
-  for method in greedy_search fast_beam_search; do
+  for method in greedy_search fast_beam_search modified_beam_search; do
     log "Decoding with $method"
 
     ./pruned_transducer_stateless3/decode.py \

diff --git a/.github/scripts/run-librispeech-transducer-stateless2-2022-04-19.sh b/.github/scripts/run-librispeech-transducer-stateless2-2022-04-19.sh
@@ -33,7 +33,7 @@ for sym in 1 2 3; do
     $repo/test_wavs/1221-135766-0002.wav
 done
 
-for method in modified_beam_search beam_search; do
+for method in fast_beam_search modified_beam_search beam_search; do
   log "$method"
 
   ./transducer_stateless2/pretrained.py \
@@ -47,7 +47,8 @@ for method in modified_beam_search beam_search; do
 done
 
 echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
-if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
   mkdir -p transducer_stateless2/exp
   ln -s $PWD/$repo/exp/pretrained.pt transducer_stateless2/exp/epoch-999.pt
   ln -s $PWD/$repo/data/lang_bpe_500 data/
@@ -58,9 +59,9 @@ if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
   log "Decoding test-clean and test-other"
 
   # use a small value for decoding with CPU
-  max_duration=50
+  max_duration=100
 
-  for method in greedy_search modified_beam_search; do
+  for method in greedy_search fast_beam_search modified_beam_search; do
     log "Decoding with $method"
 
     ./transducer_stateless2/decode.py \

diff --git a/.github/scripts/run-pre-trained-transducer-stateless-librispeech-100h.sh b/.github/scripts/run-pre-trained-transducer-stateless-librispeech-100h.sh
@@ -33,7 +33,7 @@ for sym in 1 2 3; do
     $repo/test_wavs/1221-135766-0002.wav
 done
 
-for method in modified_beam_search beam_search; do
+for method in modified_beam_search beam_search fast_beam_search; do
   log "$method"
 
   ./transducer_stateless_multi_datasets/pretrained.py \
@@ -45,3 +45,32 @@ for method in modified_beam_search beam_search; do
     $repo/test_wavs/1221-135766-0001.wav \
     $repo/test_wavs/1221-135766-0002.wav
 done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p transducer_stateless_multi_datasets/exp
+  ln -s $PWD/$repo/exp/pretrained.pt transducer_stateless_multi_datasets/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh transducer_stateless_multi_datasets/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./transducer_stateless_multi_datasets/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir transducer_stateless_multi_datasets/exp
+  done
+
+  rm transducer_stateless_multi_datasets/exp/*.pt
+fi
diff --git a/.github/scripts/run-pre-trained-transducer-stateless-librispeech-960h.sh b/.github/scripts/run-pre-trained-transducer-stateless-librispeech-960h.sh
@@ -33,7 +33,7 @@ for sym in 1 2 3; do
     $repo/test_wavs/1221-135766-0002.wav
 done
 
-for method in modified_beam_search beam_search; do
+for method in modified_beam_search beam_search fast_beam_search; do
   log "$method"
 
   ./transducer_stateless_multi_datasets/pretrained.py \
@@ -45,3 +45,32 @@ for method in modified_beam_search beam_search; do
     $repo/test_wavs/1221-135766-0001.wav \
     $repo/test_wavs/1221-135766-0002.wav
 done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p transducer_stateless_multi_datasets/exp
+  ln -s $PWD/$repo/exp/pretrained.pt transducer_stateless_multi_datasets/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh transducer_stateless_multi_datasets/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./transducer_stateless_multi_datasets/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir transducer_stateless_multi_datasets/exp
+  done
+
+  rm transducer_stateless_multi_datasets/exp/*.pt
+fi
diff --git a/.github/scripts/run-pre-trained-transducer-stateless.sh b/.github/scripts/run-pre-trained-transducer-stateless.sh
@@ -33,7 +33,7 @@ for sym in 1 2 3; do
     $repo/test_wavs/1221-135766-0002.wav
 done
 
-for method in modified_beam_search beam_search; do
+for method in fast_beam_search modified_beam_search beam_search; do
   log "$method"
 
   ./transducer_stateless/pretrained.py \
@@ -46,15 +46,31 @@ for method in modified_beam_search beam_search; do
     $repo/test_wavs/1221-135766-0002.wav
 done
 
-for method in modified_beam_search beam_search; do
-  log "$method"
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p transducer_stateless/exp
+  ln -s $PWD/$repo/exp/pretrained.pt transducer_stateless/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
 
-  ./transducer_stateless_multi_datasets/pretrained.py \
-    --method $method \
-    --beam-size 4 \
-    --checkpoint $repo/exp/pretrained.pt \
-    --bpe-model $repo/data/lang_bpe_500/bpe.model \
-    $repo/test_wavs/1089-134686-0001.wav \
-    $repo/test_wavs/1221-135766-0001.wav \
-    $repo/test_wavs/1221-135766-0002.wav
-done
+  ls -lh data
+  ls -lh transducer_stateless/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./transducer_stateless/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir transducer_stateless/exp
+  done
+
+  rm transducer_stateless/exp/*.pt
+fi
diff --git a/.github/workflows/run-librispeech-2022-03-12.yml b/.github/workflows/run-librispeech-2022-03-12.yml
@@ -35,7 +35,7 @@ on:
 
 jobs:
   run_librispeech_2022_03_12:
-    if: github.event.label.name == 'ready' || github.event_name == 'push' || github.event_name == 'schedule'
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
@@ -107,11 +107,11 @@ jobs:
         run: |
           .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
 
-
       - name: Inference with pre-trained model
         shell: bash
         env:
           GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
         run: |
           mkdir -p egs/librispeech/ASR/data
           ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
@@ -124,8 +124,8 @@ jobs:
 
           .github/scripts/run-librispeech-pruned-transducer-stateless-2022-03-12.sh
 
-      - name: Display decoding results
-        if: github.event_name == 'schedule'
+      - name: Display decoding results for pruned_transducer_stateless
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
         shell: bash
         run: |
           cd egs/librispeech/ASR/
@@ -141,9 +141,13 @@ jobs:
           find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
           find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
 
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
       - name: Upload decoding results for pruned_transducer_stateless
         uses: actions/upload-artifact@v2
-        if: github.event_name == 'schedule'
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
         with:
           name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless-2022-03-12
           path: egs/librispeech/ASR/pruned_transducer_stateless/exp/
diff --git a/.github/workflows/run-librispeech-2022-04-29.yml b/.github/workflows/run-librispeech-2022-04-29.yml
@@ -35,7 +35,7 @@ on:
 
 jobs:
   run_librispeech_2022_04_29:
-    if: github.event.label.name == 'ready' || github.event_name == 'push' || github.event_name == 'schedule'
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
@@ -111,6 +111,7 @@ jobs:
         shell: bash
         env:
           GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
         run: |
           mkdir -p egs/librispeech/ASR/data
           ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
@@ -125,44 +126,54 @@ jobs:
 
           .github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh
 
-      - name: Display decoding results
-        if: github.event_name == 'schedule'
+      - name: Display decoding results for pruned_transducer_stateless2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
         shell: bash
         run: |
           cd egs/librispeech/ASR
           tree pruned_transducer_stateless2/exp
-          cd pruned_transducer_stateless2
-          echo "results for pruned_transducer_stateless2"
+          cd pruned_transducer_stateless2/exp
           echo "===greedy search==="
-          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
-          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
 
           echo "===fast_beam_search==="
-          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
-          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
 
-          cd ../
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Display decoding results for pruned_transducer_stateless3
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR
           tree pruned_transducer_stateless3/exp
-          cd pruned_transducer_stateless3
-          echo "results for pruned_transducer_stateless3"
+          cd pruned_transducer_stateless3/exp
           echo "===greedy search==="
-          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
-          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
 
           echo "===fast_beam_search==="
-          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
-          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
 
       - name: Upload decoding results for pruned_transducer_stateless2
         uses: actions/upload-artifact@v2
-        if: github.event_name == 'schedule'
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
         with:
           name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless2-2022-04-29
           path: egs/librispeech/ASR/pruned_transducer_stateless2/exp/
 
       - name: Upload decoding results for pruned_transducer_stateless3
         uses: actions/upload-artifact@v2
-        if: github.event_name == 'schedule'
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
         with:
           name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless3-2022-04-29
           path: egs/librispeech/ASR/pruned_transducer_stateless3/exp/