Add Zipformer-MMI (#746)

* Minor fix to conformer-mmi

* Minor fixes

* Fix decode.py

* add training files

* train with ctc warmup

* add pruned_transducer_stateless7_mmi

* add zipformer_mmi/mmi_decode.py, using HP as decoding graph

* add mmi_decode.py

* remove pruned_transducer_stateless7_mmi

* rename zipformer_mmi/train_with_ctc.py as zipformer_mmi/train.py

* remove unused method

* rename mmi_decode.py

* add export.py pretrained.py jit_pretrained.py ...

* add RESULTS.md

* add CI test

* add docs

* add README.md

Co-authored-by: pkufool <wkang.pku@gmail.com>

.flake8

@@ -1,7 +1,7 @@
 [flake8]
 show-source=true
 statistics=true
-max-line-length = 80
+max-line-length = 88
 per-file-ignores =
     # line too long
     icefall/diagnostics.py: E501,
@@ -12,6 +12,7 @@ per-file-ignores =
     egs/librispeech/ASR/lstm_transducer_stateless*/*.py: E501, E203
     egs/librispeech/ASR/conv_emformer_transducer_stateless*/*.py: E501, E203
     egs/librispeech/ASR/conformer_ctc*/*py: E501,
+    egs/librispeech/ASR/zipformer_mmi/*.py: E501, E203
     egs/librispeech/ASR/RESULTS.md: E999,
 
     # invalid escape sequence (cause by tex formular), W605

.github/scripts/run-librispeech-conformer-ctc3-2022-11-28.sh

@@ -13,7 +13,6 @@ cd egs/librispeech/ASR
 repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-conformer-ctc3-2022-11-27
 
 log "Downloading pre-trained model from $repo_url"
-git lfs install
 GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
 repo=$(basename $repo_url)
 
@@ -23,7 +22,12 @@ soxi $repo/test_wavs/*.wav
 ls -lh $repo/test_wavs/*.wav
 
 pushd $repo/exp
-git lfs pull --include "data/*"
+git lfs pull --include "data/lang_bpe_500/HLG.pt"
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/LG.pt"
+git lfs pull --include "data/lang_bpe_500/Linv.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lm/G_4_gram.pt"
 git lfs pull --include "exp/jit_trace.pt"
 git lfs pull --include "exp/pretrained.pt"
 ln -s pretrained.pt epoch-99.pt

.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-2022-12-01.sh

@@ -13,7 +13,6 @@ cd egs/librispeech/ASR
 repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01
 
 log "Downloading pre-trained model from $repo_url"
-git lfs install
 GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
 repo=$(basename $repo_url)
 
@@ -23,7 +22,12 @@ soxi $repo/test_wavs/*.wav
 ls -lh $repo/test_wavs/*.wav
 
 pushd $repo/exp
-git lfs pull --include "data/*"
+git lfs pull --include "data/lang_bpe_500/HLG.pt"
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/LG.pt"
+git lfs pull --include "data/lang_bpe_500/Linv.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lm/G_4_gram.pt"
 git lfs pull --include "exp/cpu_jit.pt"
 git lfs pull --include "exp/pretrained.pt"
 ln -s pretrained.pt epoch-99.pt

.github/scripts/run-librispeech-zipformer-mmi-2022-12-08.sh

@@ -0,0 +1,103 @@
+#!/usr/bin/env bash
+
+set -e
+
+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/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 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
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/3gram.pt"
+git lfs pull --include "data/lang_bpe_500/4gram.pt"
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/LG.pt"
+git lfs pull --include "data/lang_bpe_500/Linv.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./zipformer_mmi/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./zipformer_mmi/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  --lang-dir $repo/data/lang_bpe_500 \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for method in 1best nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+  log "$method"
+
+  ./zipformer_mmi/pretrained.py \
+    --method $method \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_bpe_500 \
+    --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
+
+
+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 zipformer_mmi/exp
+  ln -s $PWD/$repo/exp/pretrained.pt zipformer_mmi/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh zipformer_mmi/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in 1best nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+    log "Decoding with $method"
+
+    ./zipformer_mmi/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --nbest-scale 1.2 \
+      --hp-scale 1.0 \
+      --max-duration $max_duration \
+      --lang-dir $repo/data/lang_bpe_500 \
+      --exp-dir zipformer_mmi/exp
+  done
+
+  rm zipformer_mmi/exp/*.pt
+fi

.github/workflows/run-librispeech-2022-12-08-zipformer-mmi.yml

@@ -0,0 +1,167 @@
+# Copyright      2022  Zengwei Yao
+
+# 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.
+
+name: run-librispeech-2022-12-08-zipformer-mmi
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_librispeech_2022_12_08_zipformer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_12_08_zipformer:
+    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:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2022-09-25
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        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: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-zipformer-mmi-2022-12-08.sh
+
+      - name: Display decoding results for librispeech zipformer-mmi
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./zipformer-mmi/exp
+
+          cd zipformer-mmi
+          echo "results for zipformer-mmi"
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest==="
+          find exp/nbest -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest-rescoring-LG==="
+          find exp/nbest-rescoring-LG -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest-rescoring-LG -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest-rescoring-3-gram==="
+          find exp/nbest-rescoring-3-gram -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest-rescoring-3-gram -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest-rescoring-4-gram==="
+          find exp/nbest-rescoring-4-gram -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest-rescoring-4-gram -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech zipformer-mmi
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-zipformer_mmi-2022-12-08
+          path: egs/librispeech/ASR/zipformer_mmi/exp/

docs/source/recipes/librispeech/index.rst

@@ -7,3 +7,4 @@ LibriSpeech
    tdnn_lstm_ctc
    conformer_ctc
    lstm_pruned_stateless_transducer
+   zipformer_mmi

docs/source/recipes/librispeech/zipformer_mmi.rst

@@ -0,0 +1,422 @@
+Zipformer MMI
+===============
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+This tutorial shows you how to train an Zipformer MMI model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+We use LF-MMI to compute the loss.
+
+.. note::
+
+   You can find the document about LF-MMI training at the following address:
+
+   `<https://github.com/k2-fsa/next-gen-kaldi-wechat/blob/master/pdf/LF-MMI-training-and-decoding-in-k2-Part-I.pdf>`_
+
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+.. note::
+
+   We encourage you to read ``./prepare.sh``.
+
+The data preparation contains several stages. You can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. hint::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. note::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+Training
+--------
+
+For stability, it uses CTC loss for model warm-up and then switches to MMI loss.
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./zipformer_mmi/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./zipformer_mmi/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./zipformer_mmi/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./zipformer_mmi/train.py --start-epoch 10`` loads the
+    checkpoint ``./zipformer_mmi/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./zipformer_mmi/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./zipformer_mmi/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./zipformer_mmi/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., weight decay,
+number of warmup steps, results dir, etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`zipformer_mmi/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/zipformer_mmi/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./zipformer_mmi/train.py`` directly.
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``zipformer_mmi/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./zipformer_mmi/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./zipformer_mmi/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd zipformer_mmi/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "Zipformer MMI training for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/xyOZUKpEQm62HBIlUD4uPA/
+
+    Note there is a URL in the above output. Click it and you will see
+    tensorboard.
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd zipformer_mmi/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 8 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./zipformer_mmi/train.py \
+    --world-size 4 \
+    --num-epochs 30 \
+    --start-epoch 1 \
+    --full-libri 1 \
+    --exp-dir zipformer_mmi/exp \
+    --max-duration 500 \
+    --use-fp16 1 \
+    --num-workers 2
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``zipformer_mmi/decode.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``zipformer_mmi/decode.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./zipformer_mmi/decode.py --help
+
+shows the options for decoding.
+
+The following shows the example using ``epoch-*.pt``:
+
+.. code-block:: bash
+
+  for m in nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+    ./zipformer_mmi/decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir ./zipformer_mmi/exp/ \
+      --max-duration 100 \
+      --lang-dir data/lang_bpe_500 \
+      --nbest-scale 1.2 \
+      --hp-scale 1.0 \
+      --decoding-method $m
+  done
+
+
+Export models
+-------------
+
+`zipformer_mmi/export.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/zipformer_mmi/export.py>`_ supports exporting checkpoints from ``zipformer_mmi/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``zipformer_mmi/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  ./zipformer_mmi/export.py \
+    --exp-dir ./zipformer_mmi/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 9 \
+    --jit 0
+
+It will generate a file ``./zipformer_mmi/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``zipformer_mmi/decode.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd zipformer_mmi/exp
+      ln -s pretrained epoch-9999.pt
+
+   And then pass ``--epoch 9999 --avg 1 --use-averaged-model 0`` to
+   ``./zipformer_mmi/decode.py``.
+
+To use the exported model with ``./zipformer_mmi/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method 1best \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./zipformer_mmi/export.py \
+    --exp-dir ./zipformer_mmi/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 9 \
+    --jit 1
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+To use the generated files with ``./zipformer_mmi/jit_pretrained.py``:
+
+.. code-block:: bash
+
+  ./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method 1best \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - `<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08>`_
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models

egs/librispeech/ASR/RESULTS.md

@@ -1,5 +1,62 @@
 ## Results
 
+### zipformer_mmi (zipformer with mmi loss)
+
+See <https://github.com/k2-fsa/icefall/pull/746> for more details.
+
+[zipformer_mmi](./zipformer_mmi)
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/xyOZUKpEQm62HBIlUD4uPA/>
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08>
+
+Number of model parameters: 69136519, i.e., 69.14 M
+
+|                          | test-clean | test-other  | comment             |
+|--------------------------|------------|-------------|---------------------|
+| 1best                    | 2.54       | 5.65        | --epoch 30 --avg 10 |
+| nbest                    | 2.54       | 5.66        | --epoch 30 --avg 10 |
+| nbest-rescoring-LG       | 2.49       | 5.42        | --epoch 30 --avg 10 |
+| nbest-rescoring-3-gram   | 2.52       | 5.62        | --epoch 30 --avg 10 |
+| nbest-rescoring-4-gram   | 2.5        | 5.51        | --epoch 30 --avg 10 |
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./zipformer_mmi/train.py \
+  --world-size 4 \
+  --master-port 12345 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --lang-dir data/lang_bpe_500 \
+  --max-duration 500 \
+  --full-libri 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer_mmi/exp
+```
+
+The decoding commands for the transducer branch are:
+```bash
+export CUDA_VISIBLE_DEVICES="5"
+
+for m in nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+  ./zipformer_mmi/decode.py \
+    --epoch 30 \
+    --avg 10 \
+    --exp-dir ./zipformer_mmi/exp/ \
+    --max-duration 100 \
+    --lang-dir data/lang_bpe_500 \
+    --nbest-scale 1.2 \
+    --hp-scale 1.0 \
+    --decoding-method $m
+done
+```
+
+
 ### pruned_transducer_stateless7_ctc (zipformer with transducer loss and ctc loss)
 
 See <https://github.com/k2-fsa/icefall/pull/683> for more details.

egs/librispeech/ASR/conformer_ctc3/jit_pretrained.py

@@ -291,7 +291,10 @@ def main():
 
     batch_size = nnet_output.shape[0]
     supervision_segments = torch.tensor(
-        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
         dtype=torch.int32,
     )
 

egs/librispeech/ASR/conformer_ctc3/pretrained.py

@@ -339,7 +339,10 @@ def main():
 
     batch_size = nnet_output.shape[0]
     supervision_segments = torch.tensor(
-        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
         dtype=torch.int32,
     )
 

egs/librispeech/ASR/conformer_mmi/decode.py

@@ -660,14 +660,22 @@ def main():
     # we need cut ids to display recognition results.
     args.return_cuts = True
     librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
     # CAUTION: `test_sets` is for displaying only.
     # If you want to skip test-clean, you have to skip
     # it inside the for loop. That is, use
     #
     #   if test_set == 'test-clean': continue
-    #
     test_sets = ["test-clean", "test-other"]
-    for test_set, test_dl in zip(test_sets, librispeech.test_dataloaders()):
+    test_dls = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
         results_dict = decode_dataset(
             dl=test_dl,
             params=params,

egs/librispeech/ASR/conformer_mmi/train-with-attention.py

@@ -30,6 +30,8 @@ import torch.multiprocessing as mp
 import torch.nn as nn
 from asr_datamodule import LibriSpeechAsrDataModule
 from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
 from lhotse.utils import fix_random_seed
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.nn.utils import clip_grad_norm_
@@ -100,6 +102,41 @@ def get_parser():
         """,
     )
 
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_mmi/exp-attn",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--use-pruned-intersect",
+        type=str2bool,
+        default=False,
+        help="""Whether to use `intersect_dense_pruned` to get denominator
+        lattice.""",
+    )
+
     return parser
 
 
@@ -114,12 +151,6 @@ def get_params() -> AttributeDict:
 
     Explanation of options saved in `params`:
 
-        - exp_dir: It specifies the directory where all training related
-                   files, e.g., checkpoints, log, etc, are saved
-
-        - lang_dir: It contains language related input files such as
-                    "lexicon.txt"
-
         - best_train_loss: Best training loss so far. It is used to select
                            the model that has the lowest training loss. It is
                            updated during the training.
@@ -164,8 +195,6 @@ def get_params() -> AttributeDict:
     """
     params = AttributeDict(
         {
-            "exp_dir": Path("conformer_mmi/exp_500_with_attention"),
-            "lang_dir": Path("data/lang_bpe_500"),
             "best_train_loss": float("inf"),
             "best_valid_loss": float("inf"),
             "best_train_epoch": -1,
@@ -184,15 +213,12 @@ def get_params() -> AttributeDict:
             "beam_size": 6,  # will change it to 8 after some batches (see code)
             "reduction": "sum",
             "use_double_scores": True,
-            #  "att_rate": 0.0,
-            #  "num_decoder_layers": 0,
             "att_rate": 0.7,
             "num_decoder_layers": 6,
             # parameters for Noam
             "weight_decay": 1e-6,
             "lr_factor": 5.0,
             "warm_step": 80000,
-            "use_pruned_intersect": False,
             "den_scale": 1.0,
             # use alignments before this number of batches
             "use_ali_until": 13000,
@@ -661,7 +687,7 @@ def run(rank, world_size, args):
     params = get_params()
     params.update(vars(args))
 
-    fix_random_seed(42)
+    fix_random_seed(params.seed)
     if world_size > 1:
         setup_dist(rank, world_size, params.master_port)
 
@@ -745,8 +771,29 @@ def run(rank, world_size, args):
         valid_ali = None
 
     librispeech = LibriSpeechAsrDataModule(args)
-    train_dl = librispeech.train_dataloaders()
+    train_cuts = librispeech.train_clean_100_cuts()
-    valid_dl = librispeech.valid_dataloaders()
+    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
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
 
     for epoch in range(params.start_epoch, params.num_epochs):
         train_dl.sampler.set_epoch(epoch)
@@ -796,6 +843,7 @@ def main():
     parser = get_parser()
     LibriSpeechAsrDataModule.add_arguments(parser)
     args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
 
     world_size = args.world_size
     assert world_size >= 1

egs/librispeech/ASR/conformer_mmi/train.py

@@ -30,6 +30,8 @@ import torch.multiprocessing as mp
 import torch.nn as nn
 from asr_datamodule import LibriSpeechAsrDataModule
 from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
 from lhotse.utils import fix_random_seed
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.nn.utils import clip_grad_norm_
@@ -100,6 +102,26 @@ def get_parser():
         """,
     )
 
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_mmi/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
     parser.add_argument(
         "--seed",
         type=int,
@@ -107,6 +129,14 @@ def get_parser():
         help="The seed for random generators intended for reproducibility",
     )
 
+    parser.add_argument(
+        "--use-pruned-intersect",
+        type=str2bool,
+        default=False,
+        help="""Whether to use `intersect_dense_pruned` to get denominator
+        lattice.""",
+    )
+
     return parser
 
 
@@ -121,12 +151,6 @@ def get_params() -> AttributeDict:
 
     Explanation of options saved in `params`:
 
-        - exp_dir: It specifies the directory where all training related
-                   files, e.g., checkpoints, log, etc, are saved
-
-        - lang_dir: It contains language related input files such as
-                    "lexicon.txt"
-
         - best_train_loss: Best training loss so far. It is used to select
                            the model that has the lowest training loss. It is
                            updated during the training.
@@ -171,8 +195,6 @@ def get_params() -> AttributeDict:
     """
     params = AttributeDict(
         {
-            "exp_dir": Path("conformer_mmi/exp_500"),
-            "lang_dir": Path("data/lang_bpe_500"),
             "best_train_loss": float("inf"),
             "best_valid_loss": float("inf"),
             "best_train_epoch": -1,
@@ -193,13 +215,10 @@ def get_params() -> AttributeDict:
             "use_double_scores": True,
             "att_rate": 0.0,
             "num_decoder_layers": 0,
-            #  "att_rate": 0.7,
-            #  "num_decoder_layers": 6,
             # parameters for Noam
             "weight_decay": 1e-6,
             "lr_factor": 5.0,
             "warm_step": 80000,
-            "use_pruned_intersect": False,
             "den_scale": 1.0,
             # use alignments before this number of batches
             "use_ali_until": 13000,
@@ -752,8 +771,29 @@ def run(rank, world_size, args):
         valid_ali = None
 
     librispeech = LibriSpeechAsrDataModule(args)
-    train_dl = librispeech.train_dataloaders()
+    train_cuts = librispeech.train_clean_100_cuts()
-    valid_dl = librispeech.valid_dataloaders()
+    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
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
 
     for epoch in range(params.start_epoch, params.num_epochs):
         fix_random_seed(params.seed + epoch)
@@ -804,6 +844,7 @@ def main():
     parser = get_parser()
     LibriSpeechAsrDataModule.add_arguments(parser)
     args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
 
     world_size = args.world_size
     assert world_size >= 1

egs/librispeech/ASR/generate-lm.sh

@@ -2,7 +2,7 @@
 
 lang_dir=data/lang_bpe_500
 
-for ngram in 2 3 5; do
+for ngram in 2 3 4 5; do
   if [ ! -f $lang_dir/${ngram}gram.arpa ]; then
     ./shared/make_kn_lm.py \
       -ngram-order ${ngram} \

egs/librispeech/ASR/pruned_transducer_stateless7_ctc/export.py

@@ -72,14 +72,14 @@ Check ./pretrained.py for its usage.
 Note: If you don't want to train a model from scratch, we have
 provided one for you. You can get it at
 
-https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01
 
 with the following commands:
 
     sudo apt-get install git-lfs
     git lfs install
-    git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+    git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01
-    # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/exp
+    # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01/exp
 """
 
 import argparse

egs/librispeech/ASR/pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py

@@ -304,7 +304,10 @@ def main():
 
     batch_size = nnet_output.shape[0]
     supervision_segments = torch.tensor(
-        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
         dtype=torch.int32,
     )
 

egs/librispeech/ASR/pruned_transducer_stateless7_ctc/pretrained_ctc.py

@@ -322,7 +322,10 @@ def main():
 
     batch_size = nnet_output.shape[0]
     supervision_segments = torch.tensor(
-        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
         dtype=torch.int32,
     )
 

egs/librispeech/ASR/zipformer_mmi/README.md

@@ -0,0 +1,26 @@
+This recipe implements Zipformer-MMI model.
+
+See https://k2-fsa.github.io/icefall/recipes/librispeech/zipformer_mmi.html for detailed tutorials.
+
+It uses **CTC loss for warm-up** and then switches to MMI loss during training.
+
+For decoding, it uses HP (H is ctc_topo, P is token-level bi-gram) as decoding graph. Supported decoding methods are:
+- **1best**. Extract the best path from the decoding lattice as the decoding result.
+- **nbest**. Extract n paths from the decoding lattice; the path with the highest score is the decoding result.
+- **nbest-rescoring-LG**. Extract n paths from the decoding lattice, rescore them with an word-level 3-gram LM, the path with the highest score is the decoding result.
+- **nbest-rescoring-3-gram**. Extract n paths from the decoding lattice, rescore them with an token-level 3-gram LM, the path with the highest score is the decoding result.
+- **nbest-rescoring-4-gram**. Extract n paths from the decoding lattice, rescore them with an token-level 4-gram LM, the path with the highest score is the decoding result.
+
+Experimental results training on train-clean-100 (epoch-30-avg-10):
+- 1best. 6.43 & 17.44
+- nbest, nbest-scale=1.2, 6.43 & 17.45
+- nbest-rescoring-LG, nbest-scale=1.2, 5.87 & 16.35
+- nbest-rescoring-3-gram,  nbest-scale=1.2, 6.19 & 16.57
+- nbest-rescoring-4-gram,  nbest-scale=1.2, 5.87 & 16.07
+
+Experimental results training on full librispeech (epoch-30-avg-10):
+- 1best. 2.54 & 5.65
+- nbest, nbest-scale=1.2, 2.54 & 5.66
+- nbest-rescoring-LG, nbest-scale=1.2, 2.49 & 5.42
+- nbest-rescoring-3-gram,  nbest-scale=1.2, 2.52 & 5.62
+- nbest-rescoring-4-gram,  nbest-scale=1.2, 2.5 & 5.51

egs/librispeech/ASR/zipformer_mmi/asr_datamodule.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py

egs/librispeech/ASR/zipformer_mmi/decode.py

@@ -0,0 +1,736 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Liyong Guo,
+#                                                 Zengwei Yao)
+#
+# 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.
+"""
+Usage:
+(1) 1best
+./zipformer_mmi/mmi_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer_mmi/exp \
+    --max-duration 100 \
+    --decoding-method 1best
+(2) nbest
+./zipformer_mmi/mmi_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer_mmi/exp \
+    --max-duration 100 \
+    --nbest-scale 1.0 \
+    --decoding-method nbest
+(3) nbest-rescoring-LG
+./zipformer_mmi/mmi_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer_mmi/exp \
+    --max-duration 100 \
+    --nbest-scale 1.0 \
+    --decoding-method nbest-rescoring-LG
+(4) nbest-rescoring-3-gram
+./zipformer_mmi/mmi_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer_mmi/exp \
+    --max-duration 100 \
+    --nbest-scale 1.0 \
+    --decoding-method nbest-rescoring-3-gram
+(5) nbest-rescoring-4-gram
+./zipformer_mmi/mmi_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer_mmi/exp \
+    --max-duration 100 \
+    --nbest-scale 1.0 \
+    --decoding-method nbest-rescoring-4-gram
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_rescore_with_LM,
+    one_best_decoding,
+)
+from icefall.lexicon import Lexicon
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer_mmi/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="1best",
+        help="""Decoding method. Use HP as decoding graph, where H is
+        ctc_topo and P is token-level bi-gram lm.
+        Supported values are:
+        - (1) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (2) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring-LG. Extract n paths from the decoding lattice,
+          rescore them with an word-level 3-gram LM, the path with the
+          highest score is the decoding result.
+        - (5) nbest-rescoring-3-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 3-gram LM, the path with
+          the highest score is the decoding result.
+        - (6) nbest-rescoring-4-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 4-gram LM, the path with
+          the highest score is the decoding result.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--hp-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `ctc_topo_P.scores`.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HP: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    G: Optional[k2.Fsa] = None,
+    LG: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - value: It contains the decoding result. `len(value)` equals to
+             batch size. `value[i]` is the decoding result for the i-th
+             utterance in the given batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring-LG", it uses nbest rescoring with word-level 3-gram LM.
+        - params.decoding_method is "nbest-rescoring-3-gram", it uses nbest rescoring with token-level 3-gram LM.
+        - params.decoding_method is "nbest-rescoring-4-gram", it uses nbest rescoring with token-level 4-gram LM.
+
+      model:
+        The neural model.
+      HP:
+        The decoding graph. H is ctc_topo, P is token-level bi-gram LM.
+      bpe_model:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      LG:
+        An LM. L is the lexicon, G is a word-level 3-gram LM.
+        It is used when params.decoding_method is "nbest-rescoring-LG".
+      G:
+        An LM. L is the lexicon, G is a token-level 3-gram or 4-gram LM.
+        It is used when params.decoding_method is "nbest-rescoring-3-gram"
+        or "nbest-rescoring-4-gram".
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    device = HP.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3, feature.shape
+    feature = feature.to(device)
+
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    nnet_output, encoder_out_lens = model(x=feature, x_lens=feature_lens)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HP,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    method = params.decoding_method
+
+    if method in ["1best", "nbest"]:
+        if method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using HP, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        return {key: hyps}
+
+    assert method in [
+        "nbest-rescoring-LG",  # word-level 3-gram lm
+        "nbest-rescoring-3-gram",  # token-level 3-gram lm
+        "nbest-rescoring-4-gram",  # token-level 4-gram lm
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if method == "nbest-rescoring-LG":
+        assert LG is not None
+        LM = LG
+    else:
+        assert G is not None
+        LM = G
+    best_path_dict = nbest_rescore_with_LM(
+        lattice=lattice,
+        LM=LM,
+        num_paths=params.num_paths,
+        lm_scale_list=lm_scale_list,
+        nbest_scale=params.nbest_scale,
+    )
+
+    ans = dict()
+    suffix = f"-nbest-scale-{params.nbest_scale}-{params.num_paths}"
+    for lm_scale_str, best_path in best_path_dict.items():
+        token_ids = get_texts(best_path)
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        ans[lm_scale_str + suffix] = hyps
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HP: k2.Fsa,
+    bpe_model: spm.SentencePieceProcessor,
+    G: Optional[k2.Fsa] = None,
+    LG: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HP:
+        The decoding graph. H is ctc_topo, P is token-level bi-gram LM.
+      bpe_model:
+        The BPE model.
+      LG:
+        An LM. L is the lexicon, G is a word-level 3-gram LM.
+        It is used when params.decoding_method is "nbest-rescoring-LG".
+      G:
+        An LM. L is the lexicon, G is a token-level 3-gram or 4-gram LM.
+        It is used when params.decoding_method is "nbest-rescoring-3-gram"
+        or "nbest-rescoring-4-gram".
+
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HP=HP,
+            bpe_model=bpe_model,
+            batch=batch,
+            G=G,
+            LG=LG,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "1best",
+        "nbest",
+        "nbest-rescoring-LG",  # word-level 3-gram lm
+        "nbest-rescoring-3-gram",  # token-level 3-gram lm
+        "nbest-rescoring-4-gram",  # token-level 4-gram lm
+    ), params.decoding_method
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    bpe_model = spm.SentencePieceProcessor()
+    bpe_model.load(str(params.lang_dir / "bpe.model"))
+    mmi_graph_compiler = MmiTrainingGraphCompiler(
+        params.lang_dir,
+        uniq_filename="lexicon.txt",
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+    HP = mmi_graph_compiler.ctc_topo_P
+    HP.scores *= params.hp_scale
+    if not hasattr(HP, "lm_scores"):
+        HP.lm_scores = HP.scores.clone()
+
+    LG = None
+    G = None
+
+    if params.decoding_method == "nbest-rescoring-LG":
+        lg_filename = params.lang_dir / "LG.pt"
+        logging.info(f"Loading {lg_filename}")
+        LG = k2.Fsa.from_dict(torch.load(lg_filename, map_location=device))
+        LG = k2.Fsa.from_fsas([LG]).to(device)
+        LG.lm_scores = LG.scores.clone()
+
+    elif params.decoding_method in ["nbest-rescoring-3-gram", "nbest-rescoring-4-gram"]:
+        order = params.decoding_method[-6]
+        assert order in ("3", "4"), (params.decoding_method, order)
+        order = int(order)
+        if not (params.lang_dir / f"{order}gram.pt").is_file():
+            logging.info(f"Loading {order}gram.fst.txt")
+            logging.warning("It may take a few minutes.")
+            with open(params.lang_dir / f"{order}gram.fst.txt") as f:
+                first_token_disambig_id = lexicon.token_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_token_disambig_id] = 0
+                G = k2.Fsa.from_fsas([G]).to(device)
+                # G = k2.remove_epsilon(G)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lang_dir / f"{order}gram.pt")
+        else:
+            logging.info(f"Loading pre-compiled {order}gram.pt")
+            d = torch.load(params.lang_dir / f"{order}gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        G.lm_scores = G.scores.clone()
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HP=HP,
+            bpe_model=bpe_model,
+            G=G,
+            LG=LG,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/zipformer_mmi/encoder_interface.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py

egs/librispeech/ASR/zipformer_mmi/export.py

@@ -0,0 +1,307 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# 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.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+(1) Export to torchscript model using torch.jit.script()
+
+./zipformer_mmi/export.py \
+  --exp-dir ./zipformer_mmi/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("cpu_jit.pt")`.
+
+Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
+are on CPU. You can use `to("cuda")` to move them to a CUDA device.
+
+Check
+https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+./zipformer_mmi/export.py \
+  --exp-dir ./zipformer_mmi/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+To use the generated file with `zipformer_mmi/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./zipformer_mmi/decode.py \
+        --exp-dir ./zipformer_mmi/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08
+    # You will find the pre-trained model in icefall-asr-librispeech-zipformer-mmi-2022-12-08/exp
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer_mmi/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named cpu_jit.pt
+
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_ctc_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        logging.info("Using torch.jit.script()")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/librispeech/ASR/zipformer_mmi/jit_pretrained.py

@@ -0,0 +1,391 @@
+#!/usr/bin/env python3
+# Copyright      2021-2022  Xiaomi Corp.   (authors: Fangjun Kuang,
+#                                                    Zengwei)
+#
+# 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.
+"""
+This script loads torchscript models, exported by `torch.jit.script()`
+and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./zipformer_mmi/export.py \
+  --exp-dir ./zipformer_mmi/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --jit 1
+
+Usage of this script:
+
+(1) 1best
+./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method 1best \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(2) nbest
+./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(3) nbest-rescoring-LG
+./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-LG \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(4) nbest-rescoring-3-gram
+./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-3-gram \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(5) nbest-rescoring-4-gram
+./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-4-gram \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params
+
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_rescore_with_LM,
+    one_best_decoding,
+)
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model-filename",
+        type=str,
+        required=True,
+        help="Path to the torchscript model cpu_jit.pt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method. Use HP as decoding graph, where H is
+        ctc_topo and P is token-level bi-gram lm.
+        Supported values are:
+        - (1) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (2) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring-LG. Extract n paths from the decoding lattice,
+          rescore them with an word-level 3-gram LM, the path with the
+          highest score is the decoding result.
+        - (5) nbest-rescoring-3-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 3-gram LM, the path with
+          the highest score is the decoding result.
+        - (6) nbest-rescoring-4-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 4-gram LM, the path with
+          the highest score is the decoding result.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.2,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.1,
+        help="""
+        Used when method is nbest-rescoring-LG, nbest-rescoring-3-gram,
+        and nbest-rescoring-4-gram.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--hp-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `ctc_topo_P.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float = 16000
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = torch.jit.load(params.nn_model_filename)
+    model.eval()
+    model.to(device)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    bpe_model = spm.SentencePieceProcessor()
+    bpe_model.load(str(params.lang_dir / "bpe.model"))
+    mmi_graph_compiler = MmiTrainingGraphCompiler(
+        params.lang_dir,
+        uniq_filename="lexicon.txt",
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+    HP = mmi_graph_compiler.ctc_topo_P
+    HP.scores *= params.hp_scale
+    if not hasattr(HP, "lm_scores"):
+        HP.lm_scores = HP.scores.clone()
+
+    method = params.method
+    assert method in (
+        "1best",
+        "nbest",
+        "nbest-rescoring-LG",  # word-level 3-gram lm
+        "nbest-rescoring-3-gram",  # token-level 3-gram lm
+        "nbest-rescoring-4-gram",  # token-level 4-gram lm
+    )
+    # loading language model for rescoring
+    LM = None
+    if method == "nbest-rescoring-LG":
+        lg_filename = params.lang_dir / "LG.pt"
+        logging.info(f"Loading {lg_filename}")
+        LG = k2.Fsa.from_dict(torch.load(lg_filename, map_location=device))
+        LG = k2.Fsa.from_fsas([LG]).to(device)
+        LG.lm_scores = LG.scores.clone()
+        LM = LG
+    elif method in ["nbest-rescoring-3-gram", "nbest-rescoring-4-gram"]:
+        order = method[-6]
+        assert order in ("3", "4")
+        order = int(order)
+        logging.info(f"Loading pre-compiled {order}gram.pt")
+        d = torch.load(params.lang_dir / f"{order}gram.pt", map_location=device)
+        G = k2.Fsa.from_dict(d)
+        G.lm_scores = G.scores.clone()
+        LM = G
+
+    # Encoder forward
+    nnet_output, encoder_out_lens = model(x=features, x_lens=feature_lengths)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HP,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if method in ["1best", "nbest"]:
+        if method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+    else:
+        best_path_dict = nbest_rescore_with_LM(
+            lattice=lattice,
+            LM=LM,
+            num_paths=params.num_paths,
+            lm_scale_list=[params.ngram_lm_scale],
+            nbest_scale=params.nbest_scale,
+        )
+        best_path = next(iter(best_path_dict.values()))
+
+    # Note: `best_path.aux_labels` contains token IDs, not word IDs
+    # since we are using HP, not HLG here.
+    #
+    # token_ids is a lit-of-list of IDs
+    token_ids = get_texts(best_path)
+    # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+    hyps = bpe_model.decode(token_ids)
+    # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+    hyps = [s.split() for s in hyps]
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/librispeech/ASR/zipformer_mmi/model.py

@@ -0,0 +1,75 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Zengwei Yao)
+#
+# 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.
+
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+
+class CTCModel(nn.Module):
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        encoder_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+
+        self.encoder = encoder
+
+        self.ctc_output = nn.Sequential(
+            nn.Dropout(p=0.1),
+            nn.Linear(encoder_dim, vocab_size),
+            nn.LogSoftmax(dim=-1),
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+        Returns:
+          Return the ctc outputs and encoder output lengths.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # compute ctc log-probs
+        ctc_output = self.ctc_output(encoder_out)
+
+        return ctc_output, x_lens

egs/librispeech/ASR/zipformer_mmi/optim.py

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

egs/librispeech/ASR/zipformer_mmi/pretrained.py

@@ -0,0 +1,410 @@
+#!/usr/bin/env python3
+# Copyright      2021-2022  Xiaomi Corp.   (authors: Fangjun Kuang,
+#                                                    Zengwei)
+#
+# 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.
+"""
+This script loads a checkpoint and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./zipformer_mmi/export.py \
+  --exp-dir ./zipformer_mmi/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+Usage of this script:
+
+(1) 1best
+./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method 1best \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(2) nbest
+./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(3) nbest-rescoring-LG
+./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-LG \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(4) nbest-rescoring-3-gram
+./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-3-gram \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+(5) nbest-rescoring-4-gram
+./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --nbest-scale 1.2 \
+    --method nbest-rescoring-4-gram \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+
+You can also use `./zipformer_mmi/exp/epoch-xx.pt`.
+
+Note: ./zipformer_mmi/exp/pretrained.pt is generated by
+./zipformer_mmi/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from decode import get_decoding_params
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_ctc_model, get_params
+
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_rescore_with_LM,
+    one_best_decoding,
+)
+from icefall.mmi_graph_compiler import MmiTrainingGraphCompiler
+from icefall.utils import get_texts
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method. Use HP as decoding graph, where H is
+        ctc_topo and P is token-level bi-gram lm.
+        Supported values are:
+        - (1) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (2) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring-LG. Extract n paths from the decoding lattice,
+          rescore them with an word-level 3-gram LM, the path with the
+          highest score is the decoding result.
+        - (5) nbest-rescoring-3-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 3-gram LM, the path with
+          the highest score is the decoding result.
+        - (6) nbest-rescoring-4-gram. Extract n paths from the decoding
+          lattice, rescore them with an token-level 4-gram LM, the path with
+          the highest score is the decoding result.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.2,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.1,
+        help="""
+        Used when method is nbest-rescoring-LG, nbest-rescoring-3-gram,
+        and nbest-rescoring-4-gram.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--hp-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `ctc_topo_P.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_ctc_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    bpe_model = spm.SentencePieceProcessor()
+    bpe_model.load(str(params.lang_dir / "bpe.model"))
+    mmi_graph_compiler = MmiTrainingGraphCompiler(
+        params.lang_dir,
+        uniq_filename="lexicon.txt",
+        device=device,
+        oov="<UNK>",
+        sos_id=1,
+        eos_id=1,
+    )
+    HP = mmi_graph_compiler.ctc_topo_P
+    HP.scores *= params.hp_scale
+    if not hasattr(HP, "lm_scores"):
+        HP.lm_scores = HP.scores.clone()
+
+    method = params.method
+    assert method in (
+        "1best",
+        "nbest",
+        "nbest-rescoring-LG",  # word-level 3-gram lm
+        "nbest-rescoring-3-gram",  # token-level 3-gram lm
+        "nbest-rescoring-4-gram",  # token-level 4-gram lm
+    )
+    # loading language model for rescoring
+    LM = None
+    if method == "nbest-rescoring-LG":
+        lg_filename = params.lang_dir / "LG.pt"
+        logging.info(f"Loading {lg_filename}")
+        LG = k2.Fsa.from_dict(torch.load(lg_filename, map_location=device))
+        LG = k2.Fsa.from_fsas([LG]).to(device)
+        LG.lm_scores = LG.scores.clone()
+        LM = LG
+    elif method in ["nbest-rescoring-3-gram", "nbest-rescoring-4-gram"]:
+        order = method[-6]
+        assert order in ("3", "4")
+        order = int(order)
+        logging.info(f"Loading pre-compiled {order}gram.pt")
+        d = torch.load(params.lang_dir / f"{order}gram.pt", map_location=device)
+        G = k2.Fsa.from_dict(d)
+        G.lm_scores = G.scores.clone()
+        LM = G
+
+    # Encoder forward
+    nnet_output, encoder_out_lens = model(x=features, x_lens=feature_lengths)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [
+            [i, 0, feature_lengths[i] // params.subsampling_factor]
+            for i in range(batch_size)
+        ],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=HP,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if method in ["1best", "nbest"]:
+        if method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+    else:
+        best_path_dict = nbest_rescore_with_LM(
+            lattice=lattice,
+            LM=LM,
+            num_paths=params.num_paths,
+            lm_scale_list=[params.ngram_lm_scale],
+            nbest_scale=params.nbest_scale,
+        )
+        best_path = next(iter(best_path_dict.values()))
+
+    # Note: `best_path.aux_labels` contains token IDs, not word IDs
+    # since we are using HP, not HLG here.
+    #
+    # token_ids is a lit-of-list of IDs
+    token_ids = get_texts(best_path)
+    # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+    hyps = bpe_model.decode(token_ids)
+    # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+    hyps = [s.split() for s in hyps]
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/librispeech/ASR/zipformer_mmi/scaling.py

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

egs/librispeech/ASR/zipformer_mmi/scaling_converter.py

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

egs/librispeech/ASR/zipformer_mmi/test_model.py

@@ -0,0 +1,57 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# 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.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./zipformer_mmi/test_model.py
+"""
+
+import torch
+from train import get_ctc_model, get_params
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.num_encoder_layers = "2,4,3,2,4"
+    #  params.feedforward_dims = "1024,1024,1536,1536,1024"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,31,31,31"
+    model = get_ctc_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    features = torch.randn(2, 100, 80)
+    feature_lengths = torch.full((2,), 100)
+    model(x=features, x_lens=feature_lengths)
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/zipformer_mmi/zipformer.py

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

icefall/decode.py

@@ -717,6 +717,107 @@ def rescore_with_n_best_list(
     return ans
 
 
+def nbest_rescore_with_LM(
+    lattice: k2.Fsa,
+    LM: k2.Fsa,
+    num_paths: int,
+    lm_scale_list: List[float],
+    nbest_scale: float = 1.0,
+    use_double_scores: bool = True,
+) -> Dict[str, k2.Fsa]:
+    """Rescore an n-best list with an n-gram LM.
+    The path with the maximum score is used as the decoding output.
+
+    Args:
+      lattice:
+        An FsaVec with axes [utt][state][arc]. It must have the following
+        attributes: ``aux_labels`` and ``lm_scores``. They are both token
+        IDs.
+      LM:
+        An FsaVec containing only a single FSA. It is one of follows:
+        - LG, L is lexicon and G is word-level n-gram LM.
+        - G, token-level n-gram LM.
+      num_paths:
+        Size of nbest list.
+      lm_scale_list:
+        A list of floats representing LM score scales.
+      nbest_scale:
+        Scale to be applied to ``lattice.score`` when sampling paths
+        using ``k2.random_paths``.
+      use_double_scores:
+        True to use double precision during computation. False to use
+        single precision.
+    Returns:
+      A dict of FsaVec, whose key is an lm_scale and the value is the
+      best decoding path for each utterance in the lattice.
+    """
+    device = lattice.device
+
+    assert len(lattice.shape) == 3
+    assert hasattr(lattice, "aux_labels")
+    assert hasattr(lattice, "lm_scores")
+
+    assert LM.shape == (1, None, None)
+    assert LM.device == device
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+    # nbest.fsa.scores contains 0s
+
+    nbest = nbest.intersect(lattice)
+
+    # Now nbest.fsa has its scores set
+    assert hasattr(nbest.fsa, "lm_scores")
+
+    # am scores + bi-gram scores
+    hp_scores = nbest.tot_scores()
+
+    # Now start to intersect nbest with LG or G
+    inv_fsa = k2.invert(nbest.fsa)
+    if hasattr(LM, "aux_labels"):
+        # LM is LG here
+        # delete token IDs as it is not needed
+        del inv_fsa.aux_labels
+    inv_fsa.scores.zero_()
+    inv_fsa_with_epsilon_loops = k2.linear_fsa_with_self_loops(inv_fsa)
+    path_to_utt_map = nbest.shape.row_ids(1)
+
+    LM = k2.arc_sort(LM)
+    path_lattice = k2.intersect_device(
+        LM,
+        inv_fsa_with_epsilon_loops,
+        b_to_a_map=torch.zeros_like(path_to_utt_map),
+        sorted_match_a=True,
+    )
+
+    # Its labels are token IDs.
+    # If LM is G, its aux_labels are tokens IDs;
+    # If LM is LG, its aux_labels are words IDs.
+    path_lattice = k2.top_sort(k2.connect(path_lattice))
+    one_best = k2.shortest_path(path_lattice, use_double_scores=use_double_scores)
+
+    lm_scores = one_best.get_tot_scores(
+        use_double_scores=use_double_scores,
+        log_semiring=True,  # Note: we always use True
+    )
+    # If LM is LG, we might get empty paths
+    lm_scores[lm_scores == float("-inf")] = -1e9
+
+    ans = dict()
+    for lm_scale in lm_scale_list:
+        tot_scores = hp_scores.values / lm_scale + lm_scores
+        tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+        max_indexes = tot_scores.argmax()
+        best_path = k2.index_fsa(nbest.fsa, max_indexes)
+        key = f"lm_scale_{lm_scale}"
+        ans[key] = best_path
+    return ans
+
+
 def rescore_with_whole_lattice(
     lattice: k2.Fsa,
     G_with_epsilon_loops: k2.Fsa,

icefall/mmi.py

@@ -112,8 +112,12 @@ def _compute_mmi_loss_exact_non_optimized(
     num_graphs, den_graphs = graph_compiler.compile(texts, replicate_den=True)
 
     # TODO: pass output_beam as function argument
-    num_lats = k2.intersect_dense(num_graphs, dense_fsa_vec, output_beam=beam_size)
+    num_lats = k2.intersect_dense(
-    den_lats = k2.intersect_dense(den_graphs, dense_fsa_vec, output_beam=beam_size)
+        num_graphs, dense_fsa_vec, output_beam=beam_size, max_arcs=2147483600
+    )
+    den_lats = k2.intersect_dense(
+        den_graphs, dense_fsa_vec, output_beam=beam_size, max_arcs=2147483600
+    )
 
     num_tot_scores = num_lats.get_tot_scores(log_semiring=True, use_double_scores=True)
 
@@ -144,7 +148,7 @@ def _compute_mmi_loss_pruned(
     """
     num_graphs, den_graphs = graph_compiler.compile(texts, replicate_den=False)
 
-    num_lats = k2.intersect_dense(num_graphs, dense_fsa_vec, output_beam=10.0)
+    num_lats = k2.intersect_dense(num_graphs, dense_fsa_vec, output_beam=8.0)
 
     # the values for search_beam/output_beam/min_active_states/max_active_states
     # are not tuned. You may want to tune them.