Support computing RNN-T loss with torchaudio (#316)

k2-fsa · Apr 19, 2022 · fce7f3c · fce7f3c
1 parent 021c798
commit fce7f3c
Show file tree

Hide file tree

Showing 37 changed files with 2,536 additions and 680 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
@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless-2022-03-12
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --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 \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless/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 \
+    $repo/test_wavs/1221-135766-0002.wav
+done
diff --git a/.github/scripts/run-librispeech-transducer-stateless2-2022-04-19.sh b/.github/scripts/run-librispeech-transducer-stateless2-2022-04-19.sh
@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless2-torchaudio-2022-04-19
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless2/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --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 \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search; do
+  log "$method"
+
+  ./transducer_stateless2/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 \
+    $repo/test_wavs/1221-135766-0002.wav
+done
diff --git a/.github/scripts/run-pre-trained-conformer-ctc.sh b/.github/scripts/run-pre-trained-conformer-ctc.sh
@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://github.com/csukuangfj/icefall-asr-conformer-ctc-bpe-500
+git lfs install
+git clone $repo
+
+log "Downloading pre-trained model from $repo_url"
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.flac
+ls -lh $repo/test_wavs/*.flac
+
+log "CTC decoding"
+
+./conformer_ctc/pretrained.py \
+  --method ctc-decoding \
+  --num-classes 500 \
+  --checkpoint $repo/exp/pretrained.pt \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  $repo/test_wavs/1089-134686-0001.flac \
+  $repo/test_wavs/1221-135766-0001.flac \
+  $repo/test_wavs/1221-135766-0002.flac
+
+log "HLG decoding"
+
+./conformer_ctc/pretrained.py \
+  --method 1best \
+  --num-classes 500 \
+  --checkpoint $repo/exp/pretrained.pt \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --words-file $repo/data/lang_bpe_500/words.txt \
+  --HLG $repo/data/lang_bpe_500/HLG.pt \
+  $repo/test_wavs/1089-134686-0001.flac \
+  $repo/test_wavs/1221-135766-0001.flac \
+  $repo/test_wavs/1221-135766-0002.flac
diff --git a/.github/scripts/run-pre-trained-transducer-stateless-librispeech-100h.sh b/.github/scripts/run-pre-trained-transducer-stateless-librispeech-100h.sh
@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-100h-transducer-stateless-multi-datasets-bpe-500-2022-02-21
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless_multi_datasets/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --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 \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search; do
+  log "$method"
+
+  ./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 \
+    $repo/test_wavs/1221-135766-0002.wav
+done
diff --git a/.github/scripts/run-pre-trained-transducer-stateless-librispeech-960h.sh b/.github/scripts/run-pre-trained-transducer-stateless-librispeech-960h.sh
@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-multi-datasets-bpe-500-2022-03-01
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless_multi_datasets/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --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 \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search; do
+  log "$method"
+
+  ./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 \
+    $repo/test_wavs/1221-135766-0002.wav
+done
diff --git a/.github/scripts/run-pre-trained-transducer-stateless-modified-2-aishell.sh b/.github/scripts/run-pre-trained-transducer-stateless-modified-2-aishell.sh
@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/aishell/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2-2022-03-01
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless_modified-2/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
+
+for method in modified_beam_search beam_search; do
+  log "$method"
+
+  ./transducer_stateless_modified-2/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
diff --git a/.github/scripts/run-pre-trained-transducer-stateless-modified-aishell.sh b/.github/scripts/run-pre-trained-transducer-stateless-modified-aishell.sh
@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/aishell/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2022-03-01
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless_modified/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
+
+for method in modified_beam_search beam_search; do
+  log "$method"
+
+  ./transducer_stateless_modified/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $repo/test_wavs/BAC009S0764W0123.wav
+done
diff --git a/.github/scripts/run-pre-trained-transducer-stateless.sh b/.github/scripts/run-pre-trained-transducer-stateless.sh
@@ -0,0 +1,60 @@
+#!/usr/bin/env bash
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./transducer_stateless/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --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 \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search; do
+  log "$method"
+
+  ./transducer_stateless/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 \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search; do
+  log "$method"
+
+  ./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