remove multi from librispeech

egs/librispeech/ASR/prepare_multidataset.sh

@@ -1,340 +0,0 @@
-#!/usr/bin/env bash
-
-# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
-export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
-
-set -eou pipefail
-
-nj=16
-stage=-1
-stop_stage=100
-
-# We assume dl_dir (download dir) contains the following
-# directories and files. If not, they will be downloaded
-# by this script automatically.
-#
-#  - $dl_dir/LibriSpeech
-#      You can find BOOKS.TXT, test-clean, train-clean-360, etc, inside it.
-#      You can download them from https://www.openslr.org/12
-#
-#  - $dl_dir/lm
-#      This directory contains the following files downloaded from
-#       http://www.openslr.org/resources/11
-#
-#        - 3-gram.pruned.1e-7.arpa.gz
-#        - 3-gram.pruned.1e-7.arpa
-#        - 4-gram.arpa.gz
-#        - 4-gram.arpa
-#        - librispeech-vocab.txt
-#        - librispeech-lexicon.txt
-#        - librispeech-lm-norm.txt.gz
-#
-#  - $dl_dir/musan
-#      This directory contains the following directories downloaded from
-#       http://www.openslr.org/17/
-#
-#     - music
-#     - noise
-#     - speech
-
-dl_dir=$PWD/download
-
-. shared/parse_options.sh || exit 1
-
-# vocab size for sentence piece models.
-# It will generate data/lang_bpe_xxx,
-# data/lang_bpe_yyy if the array contains xxx, yyy
-vocab_sizes=(
-  # 5000
-  # 2000
-  # 1000
-  500
-)
-
-# multidataset list.
-# LibriSpeech and musan are required.
-# The others are optional.
-multidataset=(
-  "gigaspeech",
-  "commonvoice",
-  "peoples_speech",
-)
-
-# All files generated by this script are saved in "data".
-# You can safely remove "data" and rerun this script to regenerate it.
-mkdir -p data
-
-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]}) $*"
-}
-
-log "dl_dir: $dl_dir"
-
-log "Dataset: LibriSpeech and musan"
-if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
-  log "Stage -1: Download LM"
-  mkdir -p $dl_dir/lm
-  if [ ! -e $dl_dir/lm/.done ]; then
-    ./local/download_lm.py --out-dir=$dl_dir/lm
-    touch $dl_dir/lm/.done
-  fi
-fi
-
-if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
-  log "Stage 0: Download data"
-
-  # If you have pre-downloaded it to /path/to/LibriSpeech,
-  # you can create a symlink
-  #
-  #   ln -sfv /path/to/LibriSpeech $dl_dir/LibriSpeech
-  #
-  if [ ! -d $dl_dir/LibriSpeech/train-other-500 ]; then
-    lhotse download librispeech --full $dl_dir
-  fi
-
-  # If you have pre-downloaded it to /path/to/musan,
-  # you can create a symlink
-  #
-  #   ln -sfv /path/to/musan $dl_dir/
-  #
-  if [ ! -d $dl_dir/musan ]; then
-    lhotse download musan $dl_dir
-  fi
-fi
-
-if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
-  log "Stage 1: Prepare LibriSpeech manifest"
-  # We assume that you have downloaded the LibriSpeech corpus
-  # to $dl_dir/LibriSpeech
-  mkdir -p data/manifests
-  if [ ! -e data/manifests/.librispeech.done ]; then
-    lhotse prepare librispeech -j $nj $dl_dir/LibriSpeech data/manifests
-    touch data/manifests/.librispeech.done
-  fi
-fi
-
-if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
-  log "Stage 2: Prepare musan manifest"
-  # We assume that you have downloaded the musan corpus
-  # to data/musan
-  mkdir -p data/manifests
-  if [ ! -e data/manifests/.musan.done ]; then
-    lhotse prepare musan $dl_dir/musan data/manifests
-    touch data/manifests/.musan.done
-  fi
-fi
-
-if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
-  log "Stage 3: Compute fbank for librispeech"
-  mkdir -p data/fbank
-  if [ ! -e data/fbank/.librispeech.done ]; then
-    ./local/compute_fbank_librispeech.py --perturb-speed False
-    touch data/fbank/.librispeech.done
-  fi
-
-  if [ ! -f data/fbank/librispeech_cuts_train-all-shuf.jsonl.gz ]; then
-    cat <(gunzip -c data/fbank/librispeech_cuts_train-clean-100.jsonl.gz) \
-      <(gunzip -c data/fbank/librispeech_cuts_train-clean-360.jsonl.gz) \
-      <(gunzip -c data/fbank/librispeech_cuts_train-other-500.jsonl.gz) | \
-      shuf | gzip -c > data/fbank/librispeech_cuts_train-all-shuf.jsonl.gz
-  fi
-
-  if [ ! -e data/fbank/.librispeech-validated.done ]; then
-    log "Validating data/fbank for LibriSpeech"
-    parts=(
-      train-clean-100
-      train-clean-360
-      train-other-500
-      test-clean
-      test-other
-      dev-clean
-      dev-other
-    )
-    for part in ${parts[@]}; do
-      python3 ./local/validate_manifest.py \
-        data/fbank/librispeech_cuts_${part}.jsonl.gz
-    done
-    touch data/fbank/.librispeech-validated.done
-  fi
-fi
-
-if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
-  log "Stage 4: Compute fbank for musan"
-  mkdir -p data/fbank
-  if [ ! -e data/fbank/.musan.done ]; then
-    ./local/compute_fbank_musan.py
-    touch data/fbank/.musan.done
-  fi
-fi
-
-if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
-  log "Stage 5: Prepare phone based lang"
-  lang_dir=data/lang_phone
-  mkdir -p $lang_dir
-
-  (echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
-    cat - $dl_dir/lm/librispeech-lexicon.txt |
-    sort | uniq > $lang_dir/lexicon.txt
-
-  if [ ! -f $lang_dir/L_disambig.pt ]; then
-    ./local/prepare_lang.py --lang-dir $lang_dir
-  fi
-
-  if [ ! -f $lang_dir/L.fst ]; then
-    log "Converting L.pt to L.fst"
-    ./shared/convert-k2-to-openfst.py \
-      --olabels aux_labels \
-      $lang_dir/L.pt \
-      $lang_dir/L.fst
-  fi
-
-  if [ ! -f $lang_dir/L_disambig.fst ]; then
-    log "Converting L_disambig.pt to L_disambig.fst"
-    ./shared/convert-k2-to-openfst.py \
-      --olabels aux_labels \
-      $lang_dir/L_disambig.pt \
-      $lang_dir/L_disambig.fst
-  fi
-fi
-
-if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
-  log "Stage 6: Prepare BPE based lang"
-
-  for vocab_size in ${vocab_sizes[@]}; do
-    lang_dir=data/lang_bpe_${vocab_size}
-    mkdir -p $lang_dir
-    # We reuse words.txt from phone based lexicon
-    # so that the two can share G.pt later.
-    cp data/lang_phone/words.txt $lang_dir
-
-    if [ ! -f $lang_dir/transcript_words.txt ]; then
-      log "Generate data for BPE training"
-      files=$(
-        find "$dl_dir/LibriSpeech/train-clean-100" -name "*.trans.txt"
-        find "$dl_dir/LibriSpeech/train-clean-360" -name "*.trans.txt"
-        find "$dl_dir/LibriSpeech/train-other-500" -name "*.trans.txt"
-      )
-      for f in ${files[@]}; do
-        cat $f | cut -d " " -f 2-
-      done > $lang_dir/transcript_words.txt
-    fi
-
-    if [ ! -f $lang_dir/bpe.model ]; then
-      ./local/train_bpe_model.py \
-        --lang-dir $lang_dir \
-        --vocab-size $vocab_size \
-        --transcript $lang_dir/transcript_words.txt
-    fi
-
-    if [ ! -f $lang_dir/L_disambig.pt ]; then
-      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
-
-      log "Validating $lang_dir/lexicon.txt"
-      ./local/validate_bpe_lexicon.py \
-        --lexicon $lang_dir/lexicon.txt \
-        --bpe-model $lang_dir/bpe.model
-    fi
-
-    if [ ! -f $lang_dir/L.fst ]; then
-      log "Converting L.pt to L.fst"
-      ./shared/convert-k2-to-openfst.py \
-        --olabels aux_labels \
-        $lang_dir/L.pt \
-        $lang_dir/L.fst
-    fi
-
-    if [ ! -f $lang_dir/L_disambig.fst ]; then
-      log "Converting L_disambig.pt to L_disambig.fst"
-      ./shared/convert-k2-to-openfst.py \
-        --olabels aux_labels \
-        $lang_dir/L_disambig.pt \
-        $lang_dir/L_disambig.fst
-    fi
-  done
-fi
-
-if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
-  log "Stage 7: Prepare G"
-  # We assume you have install kaldilm, if not, please install
-  # it using: pip install kaldilm
-
-  mkdir -p data/lm
-  if [ ! -f data/lm/G_3_gram.fst.txt ]; then
-    # It is used in building HLG
-    python3 -m kaldilm \
-      --read-symbol-table="data/lang_phone/words.txt" \
-      --disambig-symbol='#0' \
-      --max-order=3 \
-      $dl_dir/lm/3-gram.pruned.1e-7.arpa > data/lm/G_3_gram.fst.txt
-  fi
-
-  if [ ! -f data/lm/G_4_gram.fst.txt ]; then
-    # It is used for LM rescoring
-    python3 -m kaldilm \
-      --read-symbol-table="data/lang_phone/words.txt" \
-      --disambig-symbol='#0' \
-      --max-order=4 \
-      $dl_dir/lm/4-gram.arpa > data/lm/G_4_gram.fst.txt
-  fi
-fi
-
-if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
-  log "Stage 8: Compile LG"
-  ./local/compile_lg.py --lang-dir data/lang_phone
-
-  for vocab_size in ${vocab_sizes[@]}; do
-    lang_dir=data/lang_bpe_${vocab_size}
-    ./local/compile_lg.py --lang-dir $lang_dir
-  done
-fi
-
-if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
-  log "Stage 9: Prepare the other datasets"
-  # GigaSpeech
-  if [[ "${multidataset[@]}" =~ "gigaspeech" ]] && [ ! -f data/fbank/.gigaspeech.done ]; then
-    log "Dataset: GigaSpeech"
-    cd data/fbank
-    if [ -f ../../../../gigaspeech/ASR/data/fbank/XL_split/.split_completed ]; then
-      ln -svf $(realpath ../../../../gigaspeech/ASR/data/fbank/XL_split) .
-    else
-      log "Abort! Please run ../../gigaspeech/ASR/prepare.sh --stage 5 --stop-stage 6"
-      exit 1
-    fi
-
-    touch .gigaspeech.done
-    cd ../..
-  fi
-
-  # CommonVoice
-  if [[ "${multidataset[@]}" =~ "commonvoice" ]] && [ ! -f data/fbank/.commonvoice.done ]; then
-    log "Dataset: CommonVoice"
-    cd data/fbank
-    if [ -f ../../../../commonvoice/ASR/data/en/fbank/.cv-en_train.done ]; then
-      ln -svf $(realpath ../../../../commonvoice/ASR/data/en/fbank/cv-en_train_split_1000) .
-      ln -svf $(realpath ../../../../commonvoice/ASR/data/en/fbank/cv-en_cuts_train.jsonl.gz) .
-    else
-      log "Abort! Please run ../../commonvoice/ASR/prepare.sh --stage 5 --stop-stage 6"
-      exit 1
-    fi
-
-    touch .commonvoice.done
-    cd ../..
-  fi
-
-  # People's Speech
-  if [[ "${multidataset[@]}" =~ "peoples_speech" ]] && [ ! -f data/fbank/.peoples_speech.done ]; then
-    log "Dataset: People's Speech"
-    cd data/fbank
-    if [ -f ../../../../peoples_speech/ASR/data/fbank/.peoples_speech_train.done ]; then
-      ln -svf $(realpath ../../../../peoples_speech/ASR/data/fbank/peoples_speech_train_split) .
-    else
-      log "Abort! Please run ../../peoples_speech/prepare.sh --stage 5 --stop-stage 6"
-      exit 1
-    fi
-
-    touch .peoples_speech.done
-    cd ../..
-  fi
-fi

egs/librispeech/ASR/pruned_transducer_stateless7/multidataset.py

@@ -1,100 +0,0 @@
-# Copyright      2023  Xiaomi Corp.        (authors: Yifan Yang)
-#
-# See ../../../../LICENSE for clarification regarding multiple authors
-#
-# 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
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-
-import glob
-import logging
-import re
-from pathlib import Path
-
-import lhotse
-from lhotse import CutSet, load_manifest_lazy
-
-
-class MultiDataset:
-    def __init__(self, manifest_dir: str):
-        """
-        Args:
-          manifest_dir:
-            It is expected to contain the following files:
-
-            - librispeech_cuts_train-all-shuf.jsonl.gz
-            - XL_split_2000/cuts_XL.*.jsonl.gz
-            - cv-en_cuts_train.jsonl.gz
-            - peoples_speech_train_split/peoples_speech_cuts_dirty.*.jsonl.gz
-            - peoples_speech_train_split/peoples_speech_cuts_dirty_sa.*.jsonl.gz
-            - peoples_speech_train_split/peoples_speech_cuts_clean.*.jsonl.gz
-            - peoples_speech_train_split/peoples_speech_cuts_clean_sa.*.jsonl.gz
-        """
-        self.manifest_dir = Path(manifest_dir)
-
-    def train_cuts(self) -> CutSet:
-        logging.info("About to get multidataset train cuts")
-
-        # LibriSpeech
-        logging.info("Loading LibriSpeech in lazy mode")
-        librispeech_cuts = load_manifest_lazy(
-            self.manifest_dir / "librispeech_cuts_train-all-shuf.jsonl.gz"
-        )
-
-        # GigaSpeech
-        filenames = glob.glob(f"{self.manifest_dir}/XL_split/cuts_XL.*.jsonl.gz")
-
-        pattern = re.compile(r"cuts_XL.([0-9]+).jsonl.gz")
-        idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames)
-        idx_filenames = sorted(idx_filenames, key=lambda x: x[0])
-
-        sorted_filenames = [f[1] for f in idx_filenames]
-
-        logging.info(f"Loading GigaSpeech {len(sorted_filenames)} splits in lazy mode")
-
-        gigaspeech_cuts = lhotse.combine(
-            lhotse.load_manifest_lazy(p) for p in sorted_filenames
-        )
-
-        # CommonVoice
-        logging.info("Loading CommonVoice in lazy mode")
-        commonvoice_cuts = load_manifest_lazy(
-            self.manifest_dir / f"cv-en_cuts_train.jsonl.gz"
-        )
-
-        # People's Speech
-        sorted_filenames = sorted(
-            glob.glob(
-                f"{self.manifest_dir}/peoples_speech_train_split/peoples_speech_cuts_*[yna].*.jsonl.gz"
-            )
-        )
-
-        logging.info(
-            f"Loading People's Speech {len(sorted_filenames)} splits in lazy mode"
-        )
-
-        peoples_speech_cuts = lhotse.combine(
-            lhotse.load_manifest_lazy(p) for p in sorted_filenames
-        )
-
-        return CutSet.mux(
-            librispeech_cuts,
-            gigaspeech_cuts,
-            commonvoice_cuts,
-            peoples_speech_cuts,
-            weights=[
-                len(librispeech_cuts),
-                len(gigaspeech_cuts),
-                len(commonvoice_cuts),
-                len(peoples_speech_cuts),
-            ],
-        )

egs/librispeech/ASR/pruned_transducer_stateless7/train.py

@@ -66,7 +66,6 @@ from lhotse.cut import Cut
 from lhotse.dataset.sampling.base import CutSampler
 from lhotse.utils import fix_random_seed
 from model import Transducer
-from multidataset import MultiDataset
 from optim import Eden, ScaledAdam
 from torch import Tensor
 from torch.cuda.amp import GradScaler
@@ -376,13 +375,6 @@ def get_parser():
         help="Whether to use half precision training.",
     )
 
-    parser.add_argument(
-        "--use-multidataset",
-        type=str2bool,
-        default=False,
-        help="Whether to use multidataset to train.",
-    )
-
     add_model_arguments(parser)
 
     return parser
@@ -1042,16 +1034,12 @@ def run(rank, world_size, args):
 
     librispeech = LibriSpeechAsrDataModule(args)
 
-    if params.use_multidataset:
+    if params.mini_libri:
-        multidataset = MultiDataset(params.manifest_dir)
+        train_cuts = librispeech.train_clean_5_cuts()
-        train_cuts = multidataset.train_cuts()
+    elif params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
     else:
-        if params.mini_libri:
+        train_cuts = librispeech.train_clean_100_cuts()
-            train_cuts = librispeech.train_clean_5_cuts()
-        elif params.full_libri:
-            train_cuts = librispeech.train_all_shuf_cuts()
-        else:
-            train_cuts = librispeech.train_clean_100_cuts()
 
     def remove_short_and_long_utt(c: Cut):
         # Keep only utterances with duration between 1 second and 20 seconds
@@ -1107,7 +1095,7 @@ def run(rank, world_size, args):
         valid_cuts += librispeech.dev_other_cuts()
     valid_dl = librispeech.valid_dataloaders(valid_cuts)
 
-    if not params.use_multidataset and not params.print_diagnostics:
+    if not params.print_diagnostics:
         scan_pessimistic_batches_for_oom(
             model=model,
             train_dl=train_dl,

egs/librispeech/ASR/zipformer/multidataset.py

@@ -1 +0,0 @@
-../pruned_transducer_stateless7/multidataset.py

egs/librispeech/ASR/zipformer/train.py

@@ -68,7 +68,6 @@ from lhotse.cut import Cut
 from lhotse.dataset.sampling.base import CutSampler
 from lhotse.utils import fix_random_seed
 from model import Transducer
-from multidataset import MultiDataset
 from optim import Eden, ScaledAdam
 from scaling import ScheduledFloat
 from subsampling import Conv2dSubsampling
@@ -444,13 +443,6 @@ def get_parser():
         help="Whether to use half precision training.",
     )
 
-    parser.add_argument(
-        "--use-multidataset",
-        type=str2bool,
-        default=False,
-        help="Whether to use multidataset to train.",
-    )
-
     add_model_arguments(parser)
 
     return parser
@@ -1134,14 +1126,10 @@ def run(rank, world_size, args):
 
     librispeech = LibriSpeechAsrDataModule(args)
 
-    if params.use_multidataset:
+    train_cuts = librispeech.train_clean_100_cuts()
-        multidataset = MultiDataset(params.manifest_dir)
+    if params.full_libri:
-        train_cuts = multidataset.train_cuts()
+        train_cuts += librispeech.train_clean_360_cuts()
-    else:
+        train_cuts += librispeech.train_other_500_cuts()
-        train_cuts = librispeech.train_clean_100_cuts()
-        if params.full_libri:
-            train_cuts += librispeech.train_clean_360_cuts()
-            train_cuts += librispeech.train_other_500_cuts()
 
     def remove_short_and_long_utt(c: Cut):
         # Keep only utterances with duration between 1 second and 20 seconds
@@ -1197,7 +1185,7 @@ def run(rank, world_size, args):
     valid_cuts += librispeech.dev_other_cuts()
     valid_dl = librispeech.valid_dataloaders(valid_cuts)
 
-    if not params.use_multidataset and not params.print_diagnostics:
+    if not params.print_diagnostics:
         scan_pessimistic_batches_for_oom(
             model=model,
             train_dl=train_dl,