Merge remote-tracking branch 'pku/aishell'

docs/source/recipes/aishell.rst

@@ -0,0 +1,10 @@
+Aishell
+=======
+
+We provide the following models for the Aishell dataset:
+
+.. toctree::
+   :maxdepth: 2
+
+   aishell/conformer_ctc
+   aishell/tdnn_lstm_ctc

docs/source/recipes/aishell/conformer_ctc.rst

@@ -0,0 +1,573 @@
+Confromer CTC
+=============
+
+This tutorial shows you how to run a conformer ctc model
+with the `Aishell <https://www.openslr.org/33>`_ dataset.
+
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+In this tutorial, you will learn:
+
+  - (1) How to prepare data for training and decoding
+  - (2) How to start the training, either with a single GPU or multiple GPUs
+  - (3) How to do decoding after training, with 1best and attention decoder rescoring
+  - (4) How to use a pre-trained model, provided by us
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+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/aishell/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 `Aishell <https://www.openslr.org/33>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/aishell`` 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.
+
+.. HINT::
+
+  A 3-gram language model will be downloaded from huggingface, we assume you have
+  intalled and initialized ``git-lfs``. If not, you could install ``git-lfs`` by
+
+  .. code-block:: bash
+
+    $ sudo apt-get install git-lfs
+    $ git-lfs install
+
+  If you don't have the ``sudo`` permission, you could download the
+  `git-lfs binary <https://github.com/git-lfs/git-lfs/releases>`_ here, then add it to you ``PATH``.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+
+Training
+--------
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./conformer_ctc/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-29.pt``
+    in the folder ``./conformer_ctc/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./conformer_ctc/train.py --start-epoch 10`` loads the
+    checkpoint ``./conformer_ctc/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/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./conformer_ctc/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/aishell/ASR
+          $ ./conformer_ctc/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/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./conformer_ctc/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. For instance, if
+    your are using V100 NVIDIA GPU, we recommend you to set it to ``200``.
+
+    .. 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
+`conformer_ctc/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/aishell/ASR/conformer_ctc/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./conformer_ctc/train.py`` directly.
+
+
+.. CAUTION::
+
+  The training set is perturbed by speed with two factors: 0.9 and 1.1.
+  Each epoch actually processes ``3x150 == 450`` hours of data.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``conformer_ctc/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-0.pt``, ``epoch-1.pt``, ...
+
+    These are checkpoint files, containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./conformer_ctc/train.py --start-epoch 11
+
+  - ``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 conformer_ctc/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "Conformer CTC training for Aishell 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/qvNrx6JIQAaN5Ly3uQotrg/
+
+        [2021-09-12T16:41:16] Started scanning logdir.
+        [2021-09-12T16:42:17] Total uploaded: 125346 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a URL in the above output, click it and you will see
+    the following screenshot:
+
+      .. figure:: images/aishell-conformer-ctc-tensorboard-log.jpg
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/qvNrx6JIQAaN5Ly3uQotrg/
+
+         TensorBoard screenshot.
+
+  - ``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 examples
+~~~~~~~~~~~~~~
+
+The following shows typical use cases:
+
+**Case 1**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/train.py --max-duration 200
+
+It uses ``--max-duration`` of 200 to avoid OOM.
+
+
+**Case 2**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ export CUDA_VISIBLE_DEVICES="0,3"
+  $ ./conformer_ctc/train.py --world-size 2
+
+It uses GPU 0 and GPU 3 for DDP training.
+
+**Case 3**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/train.py --num-epochs 10 --start-epoch 3
+
+It loads checkpoint ``./conformer_ctc/exp/epoch-2.pt`` and starts
+training from epoch 3. Also, it trains for 10 epochs.
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/decode.py --help
+
+shows the options for decoding.
+
+The commonly used options are:
+
+  - ``--method``
+
+    This specifies the decoding method.
+
+    The following command uses attention decoder for rescoring:
+
+    .. code-block::
+
+      $ cd egs/aishell/ASR
+      $ ./conformer_ctc/decode.py --method attention-decoder --max-duration 30 --lattice-score-scale 0.5
+
+  - ``--lattice-score-scale``
+
+    It is used to scale down lattice scores so that there are more unique
+    paths for rescoring.
+
+  - ``--max-duration``
+
+    It has the same meaning as the one during training. A larger
+    value may cause OOM.
+
+Pre-trained Model
+-----------------
+
+We have uploaded a pre-trained model to
+`<https://huggingface.co/pkufool/icefall_asr_aishell_conformer_ctc>`_.
+
+We describe how to use the pre-trained model to transcribe a sound file or
+multiple sound files in the following.
+
+Install kaldifeat
+~~~~~~~~~~~~~~~~~
+
+`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used to
+extract features for a single sound file or multiple sound files
+at the same time.
+
+Please refer to `<https://github.com/csukuangfj/kaldifeat>`_ for installation.
+
+Download the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The following commands describe how to download the pre-trained model:
+
+.. code-block::
+
+  $ cd egs/aishell/ASR
+  $ mkdir tmp
+  $ cd tmp
+  $ git lfs install
+  $ git clone https://huggingface.co/pkufool/icefall_asr_aishell_conformer_ctc
+
+.. CAUTION::
+
+  You have to use ``git lfs`` to download the pre-trained model.
+
+.. CAUTION::
+
+  In order to use this pre-trained model, your k2 version has to be v1.7 or later.
+
+After downloading, you will have the following files:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ tree tmp
+
+.. code-block:: bash
+
+  tmp/
+  `-- icefall_asr_aishell_conformer_ctc
+      |-- README.md
+      |-- data
+      |   `-- lang_char
+      |       |-- HLG.pt
+      |       |-- tokens.txt
+      |       `-- words.txt
+      |-- exp
+      |   `-- pretrained.pt
+      `-- test_waves
+          |-- BAC009S0764W0121.wav
+          |-- BAC009S0764W0122.wav
+          |-- BAC009S0764W0123.wav
+          `-- trans.txt
+
+  5 directories, 9 files
+
+**File descriptions**:
+
+  - ``data/lang_char/HLG.pt``
+
+      It is the decoding graph.
+
+  - ``data/lang_char/tokens.txt``
+
+      It contains tokens and their IDs.
+      Provided only for convenience so that you can look up the SOS/EOS ID easily.
+
+  - ``data/lang_char/words.txt``
+
+      It contains words and their IDs.
+
+  - ``exp/pretrained.pt``
+
+      It contains pre-trained model parameters, obtained by averaging
+      checkpoints from ``epoch-18.pt`` to ``epoch-40.pt``.
+      Note: We have removed optimizer ``state_dict`` to reduce file size.
+
+  - ``test_waves/*.wav``
+
+      It contains some test sound files from Aishell ``test`` dataset.
+
+  - ``test_waves/trans.txt``
+
+      It contains the reference transcripts for the sound files in `test_waves/`.
+
+The information of the test sound files is listed below:
+
+.. code-block:: bash
+
+  $ soxi tmp/icefall_asr_aishell_conformer_ctc/test_wavs/*.wav
+
+  Input File     : 'tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.20 = 67263 samples ~ 315.295 CDDA sectors
+  File Size      : 135k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.12 = 65840 samples ~ 308.625 CDDA sectors
+  File Size      : 132k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.00 = 64000 samples ~ 300 CDDA sectors
+  File Size      : 128k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+  Total Duration of 3 files: 00:00:12.32
+
+Usage
+~~~~~
+
+.. code-block::
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/pretrained.py --help
+
+displays the help information.
+
+It supports two decoding methods:
+
+  - HLG decoding
+  - HLG + attention decoder rescoring
+
+HLG decoding
+^^^^^^^^^^^^
+
+HLG decoding uses the best path of the decoding lattice as the decoding result.
+
+The command to run HLG decoding is:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_aishell_conformer_ctc/exp/pretrained.pt \
+    --words-file ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/words.txt \
+    --HLG ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt \
+    --method 1best \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_wavs/BAC009S0764W0121.wav \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_wavs/BAC009S0764W0122.wav \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_wavs/BAC009S0764W0123.wav 
+
+The output is given below:
+
+.. code-block::
+
+  2021-09-13 10:46:59,842 INFO [pretrained.py:219] device: cuda:0
+  2021-09-13 10:46:59,842 INFO [pretrained.py:221] Creating model
+  2021-09-13 10:47:54,682 INFO [pretrained.py:238] Loading HLG from ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt
+  2021-09-13 10:48:46,111 INFO [pretrained.py:245] Constructing Fbank computer
+  2021-09-13 10:48:46,113 INFO [pretrained.py:255] Reading sound files: ['./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav']
+  2021-09-13 10:48:46,368 INFO [pretrained.py:262] Decoding started
+  2021-09-13 10:48:46,847 INFO [pretrained.py:291] Use HLG decoding
+  2021-09-13 10:48:47,176 INFO [pretrained.py:322]
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav:
+  甚至 出现 交易 几乎 停止 的 情况
+
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav:
+  一二 线 城市 虽然 也 处于 调整 中
+
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav:
+  但 因为 聚集 了 过多 公共 资源
+
+
+  2021-09-13 10:48:47,177 INFO [pretrained.py:324] Decoding Done
+
+HLG decoding + attention decoder rescoring
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+It extracts n paths from the lattice, recores the extracted paths with
+an attention decoder. The path with the highest score is the decoding result.
+
+The command to run HLG decoding + attention decoder rescoring is:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./conformer_ctc/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_aishell_conformer_ctc/exp/pretrained.pt \
+    --words-file ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/words.txt \
+    --HLG ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt \
+    --method attention-decoder \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_wavs/BAC009S0764W0121.wav \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_wavs/BAC009S0764W0122.wav \
+    ./tmp/icefall_asr_aishell_conformer_ctc/test_wavs/BAC009S0764W0123.wav 
+
+The output is below:
+
+.. code-block::
+
+  2021-09-13 11:02:15,852 INFO [pretrained.py:219] device: cuda:0
+  2021-09-13 11:02:15,852 INFO [pretrained.py:221] Creating model
+  2021-09-13 11:02:22,292 INFO [pretrained.py:238] Loading HLG from ./tmp/icefall_asr_aishell_conformer_ctc/data/lang_char/HLG.pt
+  2021-09-13 11:02:27,060 INFO [pretrained.py:245] Constructing Fbank computer
+  2021-09-13 11:02:27,062 INFO [pretrained.py:255] Reading sound files: ['./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav', './tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav']
+  2021-09-13 11:02:27,129 INFO [pretrained.py:261] Decoding started
+  2021-09-13 11:02:27,241 INFO [pretrained.py:295] Use HLG + attention decoder rescoring
+  2021-09-13 11:02:27,823 INFO [pretrained.py:318]
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0121.wav:
+  甚至 出现 交易 几乎 停止 的 情况
+
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0122.wav:
+  一二 线 城市 虽然 也 处于 调整 中
+
+  ./tmp/icefall_asr_aishell_conformer_ctc/test_waves/BAC009S0764W0123.wav:
+  但 因为 聚集 了 过多 公共 资源
+
+
+  2021-09-13 11:02:27,823 INFO [pretrained.py:320] Decoding Done
+
+Colab notebook
+--------------
+
+We do provide a colab notebook for this recipe showing how to use a pre-trained model.
+
+|aishell asr conformer ctc colab notebook|
+
+.. |aishell asr conformer ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/1WnG17io5HEZ0Gn_cnh_VzK5QYOoiiklC
+
+.. HINT::
+
+  Due to limited memory provided by Colab, you have to upgrade to Colab Pro to
+  run ``HLG decoding + attention decoder rescoring``.
+  Otherwise, you can only run ``HLG decoding`` with Colab.
+
+**Congratulations!** You have finished the aishell ASR recipe with
+conformer CTC models in ``icefall``.

docs/source/recipes/aishell/tdnn_lstm_ctc.rst

@@ -0,0 +1,504 @@
+TDNN-LSTM CTC
+=============
+
+This tutorial shows you how to run a tdnn-lstm ctc model
+with the `Aishell <https://www.openslr.org/33>`_ dataset.
+
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+In this tutorial, you will learn:
+
+  - (1) How to prepare data for training and decoding
+  - (2) How to start the training, either with a single GPU or multiple GPUs
+  - (3) How to do decoding after training.
+  - (4) How to use a pre-trained model, provided by us
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+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/aishell/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 `Aishell <https://www.openslr.org/33>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/aishell`` 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.
+
+.. HINT::
+
+  A 3-gram language model will be downloaded from huggingface, we assume you have
+  intalled and initialized ``git-lfs``. If not, you could install ``git-lfs`` by
+
+  .. code-block:: bash
+
+    $ sudo apt-get install git-lfs
+    $ git-lfs install
+
+  If you don't have the ``sudo`` permission, you could download the
+  `git-lfs binary <https://github.com/git-lfs/git-lfs/releases>`_ here, then add it to you ``PATH``.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+
+Training
+--------
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./tdnn_lstm_ctc/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./tdnn_lstm_ctc/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-29.pt``
+    in the folder ``./tdnn_lstm_ctc/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./tdnn_lstm_ctc/train.py --start-epoch 10`` loads the
+    checkpoint ``./tdnn_lstm_ctc/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/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./tdnn_lstm_ctc/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/aishell/ASR
+          $ ./tdnn_lstm_ctc/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/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./tdnn_lstm_ctc/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. For instance, if
+    your are using V100 NVIDIA GPU, we recommend you to set it to ``2000``.
+
+    .. 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
+`tdnn_lstm_ctc/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/aishell/ASR/tdnn_lstm_ctc/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./tdnn_lstm_ctc/train.py`` directly.
+
+
+.. CAUTION::
+
+  The training set is perturbed by speed with two factors: 0.9 and 1.1.
+  Each epoch actually processes ``3x150 == 450`` hours of data.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``tdnn_lstm_ctc/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-0.pt``, ``epoch-1.pt``, ...
+
+    These are checkpoint files, containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./tdnn_lstm_ctc/train.py --start-epoch 11
+
+  - ``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 tdnn_lstm_ctc/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "TDNN-LSTM CTC training for Aishell 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/LJI9MWUORLOw3jkdhxwk8A/
+
+        [2021-09-13T11:59:23] Started scanning logdir.
+        [2021-09-13T11:59:24] Total uploaded: 4454 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a URL in the above output, click it and you will see
+    the following screenshot:
+
+      .. figure:: images/aishell-tdnn-lstm-ctc-tensorboard-log.jpg
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/LJI9MWUORLOw3jkdhxwk8A/
+
+         TensorBoard screenshot.
+
+  - ``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 examples
+~~~~~~~~~~~~~~
+
+The following shows typical use cases:
+
+**Case 1**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ export CUDA_VISIBLE_DEVICES="0,3"
+  $ ./tdnn_lstm_ctc/train.py --world-size 2
+
+It uses GPU 0 and GPU 3 for DDP training.
+
+**Case 2**
+^^^^^^^^^^
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./tdnn_lstm_ctc/train.py --num-epochs 10 --start-epoch 3
+
+It loads checkpoint ``./tdnn_lstm_ctc/exp/epoch-2.pt`` and starts
+training from epoch 3. Also, it trains for 10 epochs.
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./tdnn_lstm_ctc/decode.py --help
+
+shows the options for decoding.
+
+The commonly used options are:
+
+  - ``--method``
+
+    This specifies the decoding method.
+
+    The following command uses attention decoder for rescoring:
+
+    .. code-block::
+
+      $ cd egs/aishell/ASR
+      $ ./tdnn_lstm_ctc/decode.py --method 1best --max-duration 100
+
+  - ``--max-duration``
+
+    It has the same meaning as the one during training. A larger
+    value may cause OOM.
+
+Pre-trained Model
+-----------------
+
+We have uploaded a pre-trained model to
+`<https://huggingface.co/pkufool/icefall_asr_aishell_tdnn_lstm_ctc>`_.
+
+We describe how to use the pre-trained model to transcribe a sound file or
+multiple sound files in the following.
+
+Install kaldifeat
+~~~~~~~~~~~~~~~~~
+
+`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used to
+extract features for a single sound file or multiple sound files
+at the same time.
+
+Please refer to `<https://github.com/csukuangfj/kaldifeat>`_ for installation.
+
+Download the pre-trained model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The following commands describe how to download the pre-trained model:
+
+.. code-block::
+
+  $ cd egs/aishell/ASR
+  $ mkdir tmp
+  $ cd tmp
+  $ git lfs install
+  $ git clone https://huggingface.co/pkufool/icefall_asr_aishell_tdnn_lstm_ctc
+
+.. CAUTION::
+
+  You have to use ``git lfs`` to download the pre-trained model.
+
+.. CAUTION::
+
+  In order to use this pre-trained model, your k2 version has to be v1.7 or later.
+
+After downloading, you will have the following files:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ tree tmp
+
+.. code-block:: bash
+
+  tmp/
+  `-- icefall_asr_aishell_tdnn_lstm_ctc
+      |-- README.md
+      |-- data
+      |   `-- lang_phone
+      |       |-- HLG.pt
+      |       |-- tokens.txt
+      |       `-- words.txt
+      |-- exp
+      |   `-- pretrained.pt
+      `-- test_waves
+          |-- BAC009S0764W0121.wav
+          |-- BAC009S0764W0122.wav
+          |-- BAC009S0764W0123.wav
+          `-- trans.txt
+
+  5 directories, 9 files
+
+**File descriptions**:
+
+  - ``data/lang_phone/HLG.pt``
+
+      It is the decoding graph.
+
+  - ``data/lang_phone/tokens.txt``
+
+      It contains tokens and their IDs.
+      Provided only for convenience so that you can look up the SOS/EOS ID easily.
+
+  - ``data/lang_phone/words.txt``
+
+      It contains words and their IDs.
+
+  - ``exp/pretrained.pt``
+
+      It contains pre-trained model parameters, obtained by averaging
+      checkpoints from ``epoch-18.pt`` to ``epoch-40.pt``.
+      Note: We have removed optimizer ``state_dict`` to reduce file size.
+
+  - ``test_waves/*.wav``
+
+      It contains some test sound files from Aishell ``test`` dataset.
+
+  - ``test_waves/trans.txt``
+
+      It contains the reference transcripts for the sound files in `test_waves/`.
+
+The information of the test sound files is listed below:
+
+.. code-block:: bash
+
+  $ soxi tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_wavs/*.wav
+
+  Input File     : 'tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0121.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.20 = 67263 samples ~ 315.295 CDDA sectors
+  File Size      : 135k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0122.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.12 = 65840 samples ~ 308.625 CDDA sectors
+  File Size      : 132k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+
+  Input File     : 'tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0123.wav'
+  Channels       : 1
+  Sample Rate    : 16000
+  Precision      : 16-bit
+  Duration       : 00:00:04.00 = 64000 samples ~ 300 CDDA sectors
+  File Size      : 128k
+  Bit Rate       : 256k
+  Sample Encoding: 16-bit Signed Integer PCM
+
+  Total Duration of 3 files: 00:00:12.32
+
+Usage
+~~~~~
+
+.. code-block::
+
+  $ cd egs/aishell/ASR
+  $ ./tdnn_lstm_ctc/pretrained.py --help
+
+displays the help information.
+
+
+HLG decoding
+^^^^^^^^^^^^
+
+HLG decoding uses the best path of the decoding lattice as the decoding result.
+
+The command to run HLG decoding is:
+
+.. code-block:: bash
+
+  $ cd egs/aishell/ASR
+  $ ./tdnn_lstm_ctc/pretrained.py \
+    --checkpoint ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/exp/pretrained.pt \
+    --words-file ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/data/lang_phone/words.txt \
+    --HLG ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/data/lang_phone/HLG.pt \
+    --method 1best \
+    ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_wavs/BAC009S0764W0121.wav \
+    ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_wavs/BAC009S0764W0122.wav \
+    ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_wavs/BAC009S0764W0123.wav
+
+The output is given below:
+
+.. code-block::
+
+  2021-09-13 15:00:55,858 INFO [pretrained.py:140] device: cuda:0
+  2021-09-13 15:00:55,858 INFO [pretrained.py:142] Creating model
+  2021-09-13 15:01:05,389 INFO [pretrained.py:154] Loading HLG from ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/data/lang_phone/HLG.pt
+  2021-09-13 15:01:06,531 INFO [pretrained.py:161] Constructing Fbank computer
+  2021-09-13 15:01:06,536 INFO [pretrained.py:171] Reading sound files: ['./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0121.wav', './tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0122.wav', './tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0123.wav']
+  2021-09-13 15:01:06,539 INFO [pretrained.py:177] Decoding started
+  2021-09-13 15:01:06,917 INFO [pretrained.py:207] Use HLG decoding
+  2021-09-13 15:01:07,129 INFO [pretrained.py:220]
+  ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0121.wav:
+  甚至 出现 交易 几乎 停滞 的 情况
+
+  ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0122.wav:
+  一二 线 城市 虽然 也 处于 调整 中
+
+  ./tmp/icefall_asr_aishell_tdnn_lstm_ctc/test_waves/BAC009S0764W0123.wav:
+  但 因为 聚集 了 过多 公共 资源
+
+
+  2021-09-13 15:01:07,129 INFO [pretrained.py:222] Decoding Done
+
+
+Colab notebook
+--------------
+
+We do provide a colab notebook for this recipe showing how to use a pre-trained model.
+
+|aishell asr conformer ctc colab notebook|
+
+.. |aishell asr conformer ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
+   :target: https://colab.research.google.com/drive/1qULaGvXq7PCu_P61oubfz9b53JzY4H3z
+
+**Congratulations!** You have finished the aishell ASR recipe with
+TDNN-LSTM CTC models in ``icefall``.

docs/source/recipes/index.rst

@@ -15,3 +15,5 @@ We may add recipes for other tasks as well in the future.
    yesno
 
    librispeech
+
+   aishell

egs/aishell/ASR/README.md

@@ -0,0 +1,3 @@
+
+Please refer to <https://icefall.readthedocs.io/en/latest/recipes/aishell.html>
+for how to run models in this recipe.

egs/aishell/ASR/RESULTS.md

@@ -0,0 +1,56 @@
+## Results
+
+### Aishell training results (Conformer-CTC)
+#### 2021-09-13
+(Wei Kang): Result of https://github.com/k2-fsa/icefall/pull/30
+
+Pretrained model is available at https://huggingface.co/pkufool/icefall_asr_aishell_conformer_ctc
+
+The best decoding results (CER) are listed below, we got this results by averaging models from epoch 23 to 40, and using `attention-decoder` decoder with num_paths equals to 100.
+
+||test|
+|--|--|
+|CER| 4.74% |
+
+To get more unique paths, we scaled the lattice.scores with 0.5 (see https://github.com/k2-fsa/icefall/pull/10#discussion_r690951662 for more details), we searched the lm_score_scale and attention_score_scale for best results, the scales that produced the CER above are also listed below.
+
+||lm_scale|attention_scale|
+|--|--|--|
+|test|0.3|0.9|
+
+You can use the following commands to reproduce our results:
+
+```bash
+git clone https://github.com/k2-fsa/icefall
+cd icefall
+
+cd egs/aishell/ASR
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1"
+python conformer_ctc/train.py --bucketing-sampler False \
+                              --concatenate-cuts False \
+                              --max-duration 200 \
+                              --world-size 2
+
+python conformer_ctc/decode.py --lattice-score-scale 0.5 \
+                               --epoch 40 \
+                               --avg 18 \
+                               --method attention-decoder \
+                               --max-duration 50 \
+                               --num-paths 100
+```
+
+### Aishell training results (Tdnn-Lstm)
+#### 2021-09-13
+
+(Wei Kang): Result of phone based Tdnn-Lstm model, https://github.com/k2-fsa/icefall/pull/30
+
+Pretrained model is available at https://huggingface.co/pkufool/icefall_asr_aishell_conformer_ctc_lstm_ctc
+
+The best decoding results (CER) are listed below, we got this results by averaging models from epoch 19 to 8, and using `1best` decoding method.
+
+||test|
+|--|--|
+|CER| 10.16% |
+

egs/aishell/ASR/conformer_ctc/README.md

@@ -0,0 +1,4 @@
+
+Please visit
+<https://icefall.readthedocs.io/en/latest/recipes/aishell/conformer_ctc.html>
+for how to run this recipe.

egs/aishell/ASR/conformer_ctc/asr_datamodule.py

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

egs/aishell/ASR/conformer_ctc/conformer.py

@@ -0,0 +1,919 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from torch import Tensor, nn
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        num_classes (int): Number of output classes
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        num_decoder_layers (int): number of decoder layers
+        dropout (float): dropout rate
+        cnn_module_kernel (int): Kernel size of convolution module
+        normalize_before (bool): whether to use layer_norm before the first block.
+        vgg_frontend (bool): whether to use vgg frontend.
+        use_feat_batchnorm(bool): whether to use batch-normalize the input.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: bool = False,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            normalize_before=normalize_before,
+            vgg_frontend=vgg_frontend,
+            use_feat_batchnorm=use_feat_batchnorm,
+        )
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            cnn_module_kernel,
+            normalize_before,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+        self.normalize_before = normalize_before
+        if self.normalize_before:
+            self.after_norm = nn.LayerNorm(d_model)
+        else:
+            # Note: TorchScript detects that self.after_norm could be used inside forward()
+            #       and throws an error without this change.
+            self.after_norm = identity
+
+    def run_encoder(
+        self, x: Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        """
+        Args:
+          x:
+            The model input. Its shape is [N, T, C].
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        if mask is not None:
+            mask = mask.to(x.device)
+        x = self.encoder(x, pos_emb, src_key_padding_mask=mask)  # (T, B, F)
+
+        if self.normalize_before:
+            x = self.after_norm(x)
+
+        return x, mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+        normalize_before: whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        cnn_module_kernel: int = 31,
+        normalize_before: bool = True,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+        self.self_attn = RelPositionMultiheadAttention(
+            d_model, nhead, dropout=0.0
+        )
+
+        self.feed_forward = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            nn.Linear(d_model, dim_feedforward),
+            Swish(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_feedforward, d_model),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_ff_macaron = nn.LayerNorm(
+            d_model
+        )  # for the macaron style FNN module
+        self.norm_ff = nn.LayerNorm(d_model)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(d_model)  # for the MHA module
+
+        self.ff_scale = 0.5
+
+        self.norm_conv = nn.LayerNorm(d_model)  # for the CNN module
+        self.norm_final = nn.LayerNorm(
+            d_model
+        )  # for the final output of the block
+
+        self.dropout = nn.Dropout(dropout)
+
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+
+        # macaron style feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff_macaron(src)
+        src = residual + self.ff_scale * self.dropout(
+            self.feed_forward_macaron(src)
+        )
+        if not self.normalize_before:
+            src = self.norm_ff_macaron(src)
+
+        # multi-headed self-attention module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_mha(src)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout(src_att)
+        if not self.normalize_before:
+            src = self.norm_mha(src)
+
+        # convolution module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_conv(src)
+        src = residual + self.dropout(self.conv_module(src))
+        if not self.normalize_before:
+            src = self.norm_conv(src)
+
+        # feed forward module
+        residual = src
+        if self.normalize_before:
+            src = self.norm_ff(src)
+        src = residual + self.ff_scale * self.dropout(self.feed_forward(src))
+        if not self.normalize_before:
+            src = self.norm_ff(src)
+
+        if self.normalize_before:
+            src = self.norm_final(src)
+
+        return src
+
+
+class ConformerEncoder(nn.TransformerEncoder):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(
+        self, encoder_layer: nn.Module, num_layers: int, norm: nn.Module = None
+    ) -> None:
+        super(ConformerEncoder, self).__init__(
+            encoder_layer=encoder_layer, num_layers=num_layers, norm=norm
+        )
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(
+        self, d_model: int, dropout_rate: float, max_len: int = 5000
+    ) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(
+                    x.device
+                ):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vecotr and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True)
+
+        # linear transformation for positional encoding.
+        self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+
+        self._reset_parameters()
+
+
+    def _reset_parameters(self) -> None:
+        nn.init.xavier_uniform_(self.in_proj.weight)
+        nn.init.constant_(self.in_proj.bias, 0.0)
+        nn.init.constant_(self.out_proj.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.pos_bias_u)
+        nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.weight,
+            self.in_proj.bias,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(
+                query, in_proj_weight, in_proj_bias
+            ).chunk(3, dim=-1)
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError(
+                        "The size of the 2D attn_mask is not correct."
+                    )
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError(
+                        "The size of the 3D attn_mask is not correct."
+                    )
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(
+                        attn_mask.dim()
+                    )
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if (
+            key_padding_mask is not None
+            and key_padding_mask.dtype == torch.uint8
+        ):
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = q.contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self.pos_bias_u).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self.pos_bias_v).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(
+            q_with_bias_u, k
+        )  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = (
+            matrix_ac + matrix_bd
+        ) * scaling  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(
+            bsz * num_heads, tgt_len, -1
+        )
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1)
+            .contiguous()
+            .view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(
+            attn_output, out_proj_weight, out_proj_bias
+        )
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(
+        self, channels: int, kernel_size: int, bias: bool = True
+    ) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+        self.norm = nn.BatchNorm1d(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = Swish()
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        x = self.depthwise_conv(x)
+        x = self.activation(self.norm(x))
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+def identity(x):
+    return x

egs/aishell/ASR/conformer_ctc/decode.py

@@ -0,0 +1,515 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo,
+#                                            Fangjun Kuang,
+#                                            Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from conformer import Conformer
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=49,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        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.
+            - (3) attention-decoder. Extract n paths from the lattice,
+              the path with the highest score is the decoding result.
+            - (4) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    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, attention-decoder, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--lattice-score-scale",
+        type=float,
+        default=0.5,
+        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, attention-decoder, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        conformer_ctc/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "exp_dir": Path("conformer_ctc/exp"),
+            "lang_dir": Path("data/lang_char"),
+            "lm_dir": Path("data/lm"),
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            "nhead": 4,
+            "attention_dim": 512,
+            "num_encoder_layers": 12,
+            "num_decoder_layers": 6,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            # parameters for decoder
+            "search_beam": 20,
+            "output_beam": 7,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+) -> Dict[str, List[List[int]]]:
+    """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 decoding method is 1best, the key is the string `no_rescore`.
+               If attention rescoring is used, the key is the string
+               `ngram_lm_scale_xxx_attention_scale_xxx`, where `xxx` is the
+               value of `lm_scale` and `attention_scale`. An example key is
+               `ngram_lm_scale_0.7_attention_scale_0.5`
+        - 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.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "attention-decoder", it uses attention rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = HLG.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is [N, T, C]
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # 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,
+        HLG=HLG,
+        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 params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is slightly worse than that of rescored lattices.
+        return nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            scale=params.lattice_score_scale,
+        )
+
+    if params.method in ["1best", "nbest"]:
+        if params.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,
+                scale=params.lattice_score_scale,
+            )
+            key = f"no_rescore-scale-{params.lattice_score_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method == "attention-decoder"
+
+    best_path_dict = rescore_with_attention_decoder(
+        lattice=lattice,
+        num_paths=params.num_paths,
+        model=model,
+        memory=memory,
+        memory_key_padding_mask=memory_key_padding_mask,
+        sos_id=sos_id,
+        eos_id=eos_id,
+        scale=params.lattice_score_scale,
+    )
+    ans = dict()
+    for lm_scale_str, best_path in best_path_dict.items():
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        ans[lm_scale_str] = hyps
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+) -> Dict[str, List[Tuple[List[int], List[int]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if the decoding method is
+      1best or it may be "ngram_lm_scale_0.7_attention_scale_0.5" if attention
+      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.
+    """
+    results = []
+
+    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"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            batch=batch,
+            word_table=word_table,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            for hyp_words, ref_text in zip(hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        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[List[int], List[int]]]],
+):
+    if params.method == "attention-decoder":
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            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.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((list("".join(res[0])), list("".join(res[1]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            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.exp_dir / f"cer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER 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()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    HLG = k2.Fsa.from_dict(
+        torch.load(f"{params.lang_dir}/HLG.pt", map_location="cpu")
+    )
+    HLG = HLG.to(device)
+    assert HLG.requires_grad is False
+
+    if not hasattr(HLG, "lm_scores"):
+        HLG.lm_scores = HLG.scores.clone()
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    if 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 start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save(
+            {"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt"
+        )
+        return
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    aishell = AishellAsrDataModule(args)
+    # 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"]
+    for test_set, test_dl in zip(test_sets, aishell.test_dataloaders()):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            word_table=lexicon.word_table,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(
+            params=params, test_set_name=test_set, results_dict=results_dict
+        )
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()

egs/aishell/ASR/conformer_ctc/pretrained.py

@@ -0,0 +1,329 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from conformer import Conformer
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+)
+from icefall.utils import AttributeDict, 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(
+        "--words-file",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument(
+        "--HLG", type=str, required=True, help="Path to HLG.pt."
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Possible values are:
+        (1) 1best - Use the best path as decoding output. Only
+            the transformer encoder output is used for decoding.
+            We call it HLG decoding.
+        (2) attention-decoder - Extract n paths from the rescored
+            lattice and use the transformer attention decoder for
+            rescoring.
+            We call it HLG decoding + n-gram LM rescoring + attention
+            decoder rescoring.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the size of n-best list.""",
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.3,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the scale for n-gram LM scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--attention-decoder-scale",
+        type=float,
+        default=0.9,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the scale for attention decoder scores.
+        (Note: You need to tune it on a dataset.)
+        """,
+    )
+
+    parser.add_argument(
+        "--lattice-score-scale",
+        type=float,
+        default=0.5,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies the scale for lattice.scores when
+        extracting n-best lists. A smaller value results in
+        more unique number of paths with the risk of missing
+        the best path.
+        """,
+    )
+
+    parser.add_argument(
+        "--sos-id",
+        type=float,
+        default=1,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies ID for the SOS token.
+        """,
+    )
+
+    parser.add_argument(
+        "--eos-id",
+        type=float,
+        default=1,
+        help="""
+        Used only when method is attention-decoder.
+        It specifies ID for the EOS token.
+        """,
+    )
+
+    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 get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "sample_rate": 16000,
+            "num_classes": 4336,
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            "nhead": 4,
+            "attention_dim": 512,
+            "num_decoder_layers": 6,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            # parameters for deocding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+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}. "
+            f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+    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 = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=params.num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"])
+    model.to(device)
+    model.eval()
+
+    logging.info(f"Loading HLG from {params.HLG}")
+    HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+    HLG = HLG.to(device)
+    if not hasattr(HLG, "lm_scores"):
+        # For whole-lattice-rescoring and attention-decoder
+        HLG.lm_scores = HLG.scores.clone()
+
+    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)
+
+    features = pad_sequence(
+        features, batch_first=True, padding_value=math.log(1e-10)
+    )
+
+    # Note: We don't use key padding mask for attention during decoding
+    with torch.no_grad():
+        nnet_output, memory, memory_key_padding_mask = model(features)
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        HLG=HLG,
+        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 params.method == "1best":
+        logging.info("Use HLG decoding")
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+    elif params.method == "attention-decoder":
+        logging.info("Use HLG + attention decoder rescoring")
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=params.sos_id,
+            eos_id=params.eos_id,
+            scale=params.lattice_score_scale,
+            ngram_lm_scale=params.ngram_lm_scale,
+            attention_scale=params.attention_decoder_scale,
+        )
+        best_path = next(iter(best_path_dict.values()))
+
+    hyps = get_texts(best_path)
+    word_sym_table = k2.SymbolTable.from_file(params.words_file)
+    hyps = [[word_sym_table[i] for i in ids] for ids 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/aishell/ASR/conformer_ctc/subsampling.py

@@ -0,0 +1,161 @@
+# Copyright    2021  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.
+
+
+import torch
+import torch.nn as nn
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape [N, T, idim] to an output
+    with shape [N, T', odim], where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """
+        Args:
+          idim:
+            Input dim. The input shape is [N, T, idim].
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is [N, ((T-1)//2 - 1)//2, odim]
+        """
+        assert idim >= 7
+        super().__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels=1, out_channels=odim, kernel_size=3, stride=2
+            ),
+            nn.ReLU(),
+            nn.Conv2d(
+                in_channels=odim, out_channels=odim, kernel_size=3, stride=2
+            ),
+            nn.ReLU(),
+        )
+        self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is [N, T, idim].
+
+        Returns:
+          Return a tensor of shape [N, ((T-1)//2 - 1)//2, odim]
+        """
+        # On entry, x is [N, T, idim]
+        x = x.unsqueeze(1)  # [N, T, idim] -> [N, 1, T, idim] i.e., [N, C, H, W]
+        x = self.conv(x)
+        # Now x is of shape [N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2]
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape [N, ((T-1)//2 - 1))//2, odim]
+        return x
+
+
+class VggSubsampling(nn.Module):
+    """Trying to follow the setup described in the following paper:
+    https://arxiv.org/pdf/1910.09799.pdf
+
+    This paper is not 100% explicit so I am guessing to some extent,
+    and trying to compare with other VGG implementations.
+
+    Convert an input of shape [N, T, idim] to an output
+    with shape [N, T', odim], where
+    T' = ((T-1)//2 - 1)//2, which approximates T' = T//4
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """Construct a VggSubsampling object.
+
+        This uses 2 VGG blocks with 2 Conv2d layers each,
+        subsampling its input by a factor of 4 in the time dimensions.
+
+        Args:
+          idim:
+            Input dim. The input shape is [N, T, idim].
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is [N, ((T-1)//2 - 1)//2, odim]
+        """
+        super().__init__()
+
+        cur_channels = 1
+        layers = []
+        block_dims = [32, 64]
+
+        # The decision to use padding=1 for the 1st convolution, then padding=0
+        # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by
+        # a back-compatibility concern so that the number of frames at the
+        # output would be equal to:
+        #  (((T-1)//2)-1)//2.
+        # We can consider changing this by using padding=1 on the
+        # 2nd convolution, so the num-frames at the output would be T//4.
+        for block_dim in block_dims:
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=cur_channels,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=1,
+                    stride=1,
+                )
+            )
+            layers.append(torch.nn.ReLU())
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=block_dim,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=0,
+                    stride=1,
+                )
+            )
+            layers.append(
+                torch.nn.MaxPool2d(
+                    kernel_size=2, stride=2, padding=0, ceil_mode=True
+                )
+            )
+            cur_channels = block_dim
+
+        self.layers = nn.Sequential(*layers)
+
+        self.out = nn.Linear(
+            block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is [N, T, idim].
+
+        Returns:
+          Return a tensor of shape [N, ((T-1)//2 - 1)//2, odim]
+        """
+        x = x.unsqueeze(1)
+        x = self.layers(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        return x

egs/aishell/ASR/conformer_ctc/test_subsampling.py

@@ -0,0 +1,49 @@
+#!/usr/bin/env python3
+# Copyright    2021  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.
+
+
+from subsampling import Conv2dSubsampling
+from subsampling import VggSubsampling
+import torch
+
+
+def test_conv2d_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = Conv2dSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
+
+
+def test_vgg_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = VggSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim

egs/aishell/ASR/conformer_ctc/test_transformer.py

@@ -0,0 +1,105 @@
+#!/usr/bin/env python3
+# Copyright    2021  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.
+
+
+import torch
+from transformer import (
+    Transformer,
+    encoder_padding_mask,
+    generate_square_subsequent_mask,
+    decoder_padding_mask,
+    add_sos,
+    add_eos,
+)
+
+from torch.nn.utils.rnn import pad_sequence
+
+
+def test_encoder_padding_mask():
+    supervisions = {
+        "sequence_idx": torch.tensor([0, 1, 2]),
+        "start_frame": torch.tensor([0, 0, 0]),
+        "num_frames": torch.tensor([18, 7, 13]),
+    }
+
+    max_len = ((18 - 1) // 2 - 1) // 2
+    mask = encoder_padding_mask(max_len, supervisions)
+    expected_mask = torch.tensor(
+        [
+            [False, False, False],  # ((18 - 1)//2 - 1)//2 = 3,
+            [False, True, True],  # ((7 - 1)//2 - 1)//2 = 1,
+            [False, False, True],  # ((13 - 1)//2 - 1)//2 = 2,
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_transformer():
+    num_features = 40
+    num_classes = 87
+    model = Transformer(num_features=num_features, num_classes=num_classes)
+
+    N = 31
+
+    for T in range(7, 30):
+        x = torch.rand(N, T, num_features)
+        y, _, _ = model(x)
+        assert y.shape == (N, (((T - 1) // 2) - 1) // 2, num_classes)
+
+
+def test_generate_square_subsequent_mask():
+    s = 5
+    mask = generate_square_subsequent_mask(s)
+    inf = float("inf")
+    expected_mask = torch.tensor(
+        [
+            [0.0, -inf, -inf, -inf, -inf],
+            [0.0, 0.0, -inf, -inf, -inf],
+            [0.0, 0.0, 0.0, -inf, -inf],
+            [0.0, 0.0, 0.0, 0.0, -inf],
+            [0.0, 0.0, 0.0, 0.0, 0.0],
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_decoder_padding_mask():
+    x = [torch.tensor([1, 2]), torch.tensor([3]), torch.tensor([2, 5, 8])]
+    y = pad_sequence(x, batch_first=True, padding_value=-1)
+    mask = decoder_padding_mask(y, ignore_id=-1)
+    expected_mask = torch.tensor(
+        [
+            [False, False, True],
+            [False, True, True],
+            [False, False, False],
+        ]
+    )
+    assert torch.all(torch.eq(mask, expected_mask))
+
+
+def test_add_sos():
+    x = [[1, 2], [3], [2, 5, 8]]
+    y = add_sos(x, sos_id=0)
+    expected_y = [[0, 1, 2], [0, 3], [0, 2, 5, 8]]
+    assert y == expected_y
+
+
+def test_add_eos():
+    x = [[1, 2], [3], [2, 5, 8]]
+    y = add_eos(x, eos_id=0)
+    expected_y = [[1, 2, 0], [3, 0], [2, 5, 8, 0]]
+    assert y == expected_y

egs/aishell/ASR/conformer_ctc/train.py

@@ -0,0 +1,746 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional
+
+import k2
+import torch
+import torch.distributed as dist
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from conformer import Conformer
+from lhotse.utils import fix_random_seed
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=50,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    is saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    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_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - att_rate: The proportion of label smoothing loss, final loss will be
+                    (1 - att_rate) * ctc_loss + att_rate * label_smoothing_loss
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - attention_dim: Attention dimension.
+
+        - nhead: Number of heads in multi-head attention.
+                 Must satisfy attention_dim // nhead == 0.
+
+        - num_encoder_layers: Number of attention encoder layers.
+
+        - num_decoder_layers: Number of attention decoder layers.
+
+        - use_feat_batchnorm: Whether to do normalization in the input layer.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - lr_factor: The lr_factor for the optimizer.
+
+        - warm_step: The warm_step for the optimizer.
+    """
+    params = AttributeDict(
+        {
+            "exp_dir": Path("conformer_ctc/exp"),
+            "lang_dir": Path("data/lang_char"),
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 10,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for k2.ctc_loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            "att_rate": 0.7,
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "feature_dim": 80,
+            "attention_dim": 512,
+            "nhead": 4,
+            "num_encoder_layers": 12,
+            "num_decoder_layers": 6,
+            "use_feat_batchnorm": True,
+            # parameters for Noam
+            "weight_decay": 1e-5,
+            "lr_factor": 5.0,
+            "warm_step": 36000,
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    is_training: bool,
+):
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is [N, T, C]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is [N, T, C]
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    token_ids = graph_compiler.texts_to_ids(texts)
+
+    decoding_graph = graph_compiler.compile(token_ids)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    if params.att_rate != 0.0:
+        with torch.set_grad_enabled(is_training):
+            if hasattr(model, "module"):
+                att_loss = model.module.decoder_forward(
+                    encoder_memory,
+                    memory_mask,
+                    token_ids=token_ids,
+                    sos_id=graph_compiler.sos_id,
+                    eos_id=graph_compiler.eos_id,
+                )
+            else:
+                att_loss = model.decoder_forward(
+                    encoder_memory,
+                    memory_mask,
+                    token_ids=token_ids,
+                    sos_id=graph_compiler.sos_id,
+                    eos_id=graph_compiler.eos_id,
+                )
+        loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+    else:
+        loss = ctc_loss
+        att_loss = torch.tensor([0])
+
+    # train_frames and valid_frames are used for printing.
+    if is_training:
+        params.train_frames = supervision_segments[:, 2].sum().item()
+    else:
+        params.valid_frames = supervision_segments[:, 2].sum().item()
+
+    assert loss.requires_grad == is_training
+
+    return loss, ctc_loss.detach(), att_loss.detach()
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> None:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = 0.0
+    tot_ctc_loss = 0.0
+    tot_att_loss = 0.0
+    tot_frames = 0.0
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, ctc_loss, att_loss = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        assert ctc_loss.requires_grad is False
+        assert att_loss.requires_grad is False
+
+        loss_cpu = loss.detach().cpu().item()
+        tot_loss += loss_cpu
+
+        tot_ctc_loss += ctc_loss.detach().cpu().item()
+        tot_att_loss += att_loss.detach().cpu().item()
+
+        tot_frames += params.valid_frames
+
+    if world_size > 1:
+        s = torch.tensor(
+            [tot_loss, tot_ctc_loss, tot_att_loss, tot_frames],
+            device=loss.device,
+        )
+        dist.all_reduce(s, op=dist.ReduceOp.SUM)
+        s = s.cpu().tolist()
+        tot_loss = s[0]
+        tot_ctc_loss = s[1]
+        tot_att_loss = s[2]
+        tot_frames = s[3]
+
+    params.valid_loss = tot_loss / tot_frames
+    params.valid_ctc_loss = tot_ctc_loss / tot_frames
+    params.valid_att_loss = tot_att_loss / tot_frames
+
+    if params.valid_loss < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = params.valid_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = 0.0  # sum of losses over all batches
+    tot_ctc_loss = 0.0
+    tot_att_loss = 0.0
+
+    tot_frames = 0.0  # sum of frames over all batches
+    params.tot_loss = 0.0
+    params.tot_frames = 0.0
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, ctc_loss, att_loss = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        loss_cpu = loss.detach().cpu().item()
+        ctc_loss_cpu = ctc_loss.detach().cpu().item()
+        att_loss_cpu = att_loss.detach().cpu().item()
+
+        tot_frames += params.train_frames
+        tot_loss += loss_cpu
+        tot_ctc_loss += ctc_loss_cpu
+        tot_att_loss += att_loss_cpu
+
+        params.tot_frames += params.train_frames
+        params.tot_loss += loss_cpu
+
+        tot_avg_loss = tot_loss / tot_frames
+        tot_avg_ctc_loss = tot_ctc_loss / tot_frames
+        tot_avg_att_loss = tot_att_loss / tot_frames
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, batch {batch_idx}, "
+                f"batch avg ctc loss {ctc_loss_cpu/params.train_frames:.4f}, "
+                f"batch avg att loss {att_loss_cpu/params.train_frames:.4f}, "
+                f"batch avg loss {loss_cpu/params.train_frames:.4f}, "
+                f"total avg ctc loss: {tot_avg_ctc_loss:.4f}, "
+                f"total avg att loss: {tot_avg_att_loss:.4f}, "
+                f"total avg loss: {tot_avg_loss:.4f}, "
+                f"batch size: {batch_size}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/current_ctc_loss",
+                    ctc_loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/current_att_loss",
+                    att_loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/current_loss",
+                    loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/tot_avg_ctc_loss",
+                    tot_avg_ctc_loss,
+                    params.batch_idx_train,
+                )
+
+                tb_writer.add_scalar(
+                    "train/tot_avg_att_loss",
+                    tot_avg_att_loss,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/tot_avg_loss",
+                    tot_avg_loss,
+                    params.batch_idx_train,
+                )
+        if batch_idx > 0 and batch_idx % params.reset_interval == 0:
+            tot_loss = 0.0  # sum of losses over all batches
+            tot_ctc_loss = 0.0
+            tot_att_loss = 0.0
+
+            tot_frames = 0.0  # sum of frames over all batches
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"valid ctc loss {params.valid_ctc_loss:.4f},"
+                f"valid att loss {params.valid_att_loss:.4f},"
+                f"valid loss {params.valid_loss:.4f},"
+                f" best valid loss: {params.best_valid_loss:.4f} "
+                f"best valid epoch: {params.best_valid_epoch}"
+            )
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/valid_ctc_loss",
+                    params.valid_ctc_loss,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/valid_att_loss",
+                    params.valid_att_loss,
+                    params.batch_idx_train,
+                )
+                tb_writer.add_scalar(
+                    "train/valid_loss",
+                    params.valid_loss,
+                    params.batch_idx_train,
+                )
+
+    params.train_loss = params.tot_loss / params.tot_frames
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(42)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    aishell = AishellAsrDataModule(args)
+    train_dl = aishell.train_dataloaders()
+    valid_dl = aishell.valid_dataloaders()
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar(
+                "train/learning_rate", cur_lr, params.batch_idx_train
+            )
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()

egs/aishell/ASR/conformer_ctc/transformer.py

@@ -0,0 +1,998 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from subsampling import Conv2dSubsampling, VggSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+        use_feat_batchnorm: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+          use_feat_batchnorm:
+            True to use batchnorm for the input layer.
+        """
+        super().__init__()
+        self.use_feat_batchnorm = use_feat_batchnorm
+        if use_feat_batchnorm:
+            self.feat_batchnorm = nn.BatchNorm1d(num_features)
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape [N, T, num_classes]
+        # to the shape [N, T//subsampling_factor, d_model].
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, num_classes)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = nn.Embedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                normalize_before=normalize_before,
+            )
+
+            if normalize_before:
+                decoder_norm = nn.LayerNorm(d_model)
+            else:
+                decoder_norm = None
+
+            self.decoder = nn.TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+                norm=decoder_norm,
+            )
+
+            self.decoder_output_layer = torch.nn.Linear(
+                d_model, self.decoder_num_class
+            )
+
+            self.decoder_criterion = LabelSmoothingLoss(self.decoder_num_class)
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self, x: torch.Tensor, supervision: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is [N, T, C].
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is [N, T, C]
+            - Encoder output with shape [T, N, C]. It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is [N, T].
+              It is None if `supervision` is None.
+        """
+        if self.use_feat_batchnorm:
+            x = x.permute(0, 2, 1)  # [N, T, C] -> [N, C, T]
+            x = self.feat_batchnorm(x)
+            x = x.permute(0, 2, 1)  # [N, C, T] -> [N, T, C]
+        encoder_memory, memory_key_padding_mask = self.run_encoder(
+            x, supervision
+        )
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self, x: torch.Tensor, supervisions: Optional[Supervisions] = None
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is [N, T, C].
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape [T, N, C]
+            - encoder padding mask, with shape [N, T].
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is [T, N, C]
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is [N, T, C]
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape [T, N, C]
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=eos_id)
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=-1)
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(
+            device
+        )
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape [T, N, C]
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=eos_id)
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=-1)
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(
+            device
+        )
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerDecoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+        tgt2 = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout1(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm2(tgt)
+        tgt2 = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = residual + self.dropout2(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm2(tgt)
+
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = residual + self.dropout3(tgt2)
+        if not self.normalize_before:
+            tgt = self.norm3(tgt)
+        return tgt
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError(
+        "activation should be relu/gelu, not {}".format(activation)
+    )
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        self.pe = None
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is [1, T1, d_model]. The shape of the input x
+        is [N, T, d_model]. If T > T1, then we change the shape of self.pe
+        to [N, T, d_model]. Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape [N, T, C].
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                if self.pe.dtype != x.dtype or self.pe.device != x.device:
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape [1, T, d_model], where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is [N, T, C]
+
+        Returns:
+          Return a tensor of shape [N, T, C]
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+class LabelSmoothingLoss(nn.Module):
+    """
+    Label-smoothing loss. KL-divergence between
+    q_{smoothed ground truth prob.}(w)
+    and p_{prob. computed by model}(w) is minimized.
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/label_smoothing_loss.py  # noqa
+
+    Args:
+        size: the number of class
+        padding_idx: padding_idx: ignored class id
+        smoothing: smoothing rate (0.0 means the conventional CE)
+        normalize_length: normalize loss by sequence length if True
+        criterion: loss function to be smoothed
+    """
+
+    def __init__(
+        self,
+        size: int,
+        padding_idx: int = -1,
+        smoothing: float = 0.1,
+        normalize_length: bool = False,
+        criterion: nn.Module = nn.KLDivLoss(reduction="none"),
+    ) -> None:
+        """Construct an LabelSmoothingLoss object."""
+        super(LabelSmoothingLoss, self).__init__()
+        self.criterion = criterion
+        self.padding_idx = padding_idx
+        assert 0.0 < smoothing <= 1.0
+        self.confidence = 1.0 - smoothing
+        self.smoothing = smoothing
+        self.size = size
+        self.true_dist = None
+        self.normalize_length = normalize_length
+
+    def forward(self, x: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """
+        Compute loss between x and target.
+
+        Args:
+          x:
+            prediction of dimension
+            (batch_size, input_length, number_of_classes).
+          target:
+            target masked with self.padding_id of
+            dimension (batch_size, input_length).
+
+        Returns:
+          A scalar tensor containing the loss without normalization.
+        """
+        assert x.size(2) == self.size
+        #  batch_size = x.size(0)
+        x = x.view(-1, self.size)
+        target = target.view(-1)
+        with torch.no_grad():
+            true_dist = x.clone()
+            true_dist.fill_(self.smoothing / (self.size - 1))
+            ignore = target == self.padding_idx  # (B,)
+            total = len(target) - ignore.sum().item()
+            target = target.masked_fill(ignore, 0)  # avoid -1 index
+            true_dist.scatter_(1, target.unsqueeze(1), self.confidence)
+        kl = self.criterion(torch.log_softmax(x, dim=1), true_dist)
+        #  denom = total if self.normalize_length else batch_size
+        denom = total if self.normalize_length else 1
+        return kl.masked_fill(ignore.unsqueeze(1), 0).sum() / denom
+
+
+def encoder_padding_mask(
+    max_len: int, supervisions: Optional[Supervisions] = None
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [
+        0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)
+    ]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(
+    ys_pad: torch.Tensor, ignore_id: int = -1
+) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    ans = []
+    for utt in token_ids:
+        ans.append([sos_id] + utt)
+    return ans
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    ans = []
+    for utt in token_ids:
+        ans.append(utt + [eos_id])
+    return ans

egs/aishell/ASR/local/compile_hlg.py

@@ -0,0 +1,156 @@
+#!/usr/bin/env python3
+# Copyright    2021  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.
+
+
+"""
+This script takes as input lang_dir and generates HLG from
+
+    - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt
+    - L, the lexicon, built from lang_dir/L_disambig.pt
+
+        Caution: We use a lexicon that contains disambiguation symbols
+
+    - G, the LM, built from data/lm/G_3_gram.fst.txt
+
+The generated HLG is saved in $lang_dir/HLG.pt
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compile_HLG(lang_dir: str) -> k2.Fsa:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang_phone or data/lang_bpe_5000.
+
+    Return:
+      An FSA representing HLG.
+    """
+    lexicon = Lexicon(lang_dir)
+    max_token_id = max(lexicon.tokens)
+    logging.info(f"Building ctc_topo. max_token_id: {max_token_id}")
+    H = k2.ctc_topo(max_token_id)
+    L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt"))
+
+    if Path("data/lm/G_3_gram.pt").is_file():
+        logging.info("Loading pre-compiled G_3_gram")
+        d = torch.load("data/lm/G_3_gram.pt")
+        G = k2.Fsa.from_dict(d)
+    else:
+        logging.info("Loading G_3_gram.fst.txt")
+        with open("data/lm/G_3_gram.fst.txt") as f:
+            G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+            torch.save(G.as_dict(), "data/lm/G_3_gram.pt")
+
+    first_token_disambig_id = lexicon.token_table["#0"]
+    first_word_disambig_id = lexicon.word_table["#0"]
+
+    L = k2.arc_sort(L)
+    G = k2.arc_sort(G)
+
+    logging.info("Intersecting L and G")
+    LG = k2.compose(L, G)
+    logging.info(f"LG shape: {LG.shape}")
+
+    logging.info("Connecting LG")
+    LG = k2.connect(LG)
+    logging.info(f"LG shape after k2.connect: {LG.shape}")
+
+    logging.info(type(LG.aux_labels))
+    logging.info("Determinizing LG")
+
+    LG = k2.determinize(LG)
+    logging.info(type(LG.aux_labels))
+
+    logging.info("Connecting LG after k2.determinize")
+    LG = k2.connect(LG)
+
+    logging.info("Removing disambiguation symbols on LG")
+
+    LG.labels[LG.labels >= first_token_disambig_id] = 0
+
+    assert isinstance(LG.aux_labels, k2.RaggedTensor)
+    LG.aux_labels.data[LG.aux_labels.data >= first_word_disambig_id] = 0
+
+    LG = k2.remove_epsilon(LG)
+    logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}")
+
+    LG = k2.connect(LG)
+    LG.aux_labels = LG.aux_labels.remove_values_eq(0)
+
+    logging.info("Arc sorting LG")
+    LG = k2.arc_sort(LG)
+
+    logging.info("Composing H and LG")
+    # CAUTION: The name of the inner_labels is fixed
+    # to `tokens`. If you want to change it, please
+    # also change other places in icefall that are using
+    # it.
+    HLG = k2.compose(H, LG, inner_labels="tokens")
+
+    logging.info("Connecting LG")
+    HLG = k2.connect(HLG)
+
+    logging.info("Arc sorting LG")
+    HLG = k2.arc_sort(HLG)
+    logging.info(f"HLG.shape: {HLG.shape}")
+
+    return HLG
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    if (lang_dir / "HLG.pt").is_file():
+        logging.info(f"{lang_dir}/HLG.pt already exists - skipping")
+        return
+
+    logging.info(f"Processing {lang_dir}")
+
+    HLG = compile_HLG(lang_dir)
+    logging.info(f"Saving HLG.pt to {lang_dir}")
+    torch.save(HLG.as_dict(), f"{lang_dir}/HLG.pt")
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/aishell/ASR/local/compute_fbank_aishell.py

@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+# Copyright    2021  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.
+
+
+"""
+This file computes fbank features of the aishell dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomHdf5Writer
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_aishell(num_mel_bins: int = 80):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank40")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "train",
+        "dev",
+        "test",
+    )
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts, output_dir=src_dir
+    )
+    assert manifests is not None
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"cuts_{partition}.json.gz").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition:
+                cut_set = (
+                    cut_set
+                    + cut_set.perturb_speed(0.9)
+                    + cut_set.perturb_speed(1.1)
+                )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomHdf5Writer,
+            )
+            cut_set.to_json(output_dir / f"cuts_{partition}.json.gz")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_aishell(num_mel_bins=args.num_mel_bins)
+

egs/aishell/ASR/local/compute_fbank_musan.py

@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+# Copyright    2021  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.
+
+
+"""
+This file computes fbank features of the musan dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomHdf5Writer, combine
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_musan(num_mel_bins: int = 80):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank40")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "music",
+        "speech",
+        "noise",
+    )
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts, output_dir=src_dir
+    )
+    assert manifests is not None
+
+    musan_cuts_path = output_dir / "cuts_musan.json.gz"
+
+    if musan_cuts_path.is_file():
+        logging.info(f"{musan_cuts_path} already exists - skipping")
+        return
+
+    logging.info("Extracting features for Musan")
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        # create chunks of Musan with duration 5 - 10 seconds
+        musan_cuts = (
+            CutSet.from_manifests(
+                recordings=combine(
+                    part["recordings"] for part in manifests.values()
+                )
+            )
+            .cut_into_windows(10.0)
+            .filter(lambda c: c.duration > 5)
+            .compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/feats_musan",
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomHdf5Writer,
+            )
+        )
+        musan_cuts.to_json(musan_cuts_path)
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    compute_fbank_musan(num_mel_bins=args.num_mel_bins)
+

egs/aishell/ASR/local/prepare_char.py

@@ -0,0 +1,247 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang)
+#
+# 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 takes as input `lang_dir`, which should contain::
+
+    - lang_dir/text,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+import re
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] if i in token2id else token2id['<unk>'] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool:
+    """Check if all the given tokens are in token symbol table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that contains all the valid tokens.
+      tokens:
+        A list of tokens.
+    Returns:
+      Return True if there is any token not in the token_sym_table,
+      otherwise False.
+    """
+    for tok in tokens:
+        if not tok in token_sym_table:
+            return True
+    return False
+
+
+def generate_lexicon(
+    token_sym_table: Dict[str, int], words: List[str]
+) -> Lexicon:
+    """Generate a lexicon from a word list and token_sym_table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that mapping token to token ids.
+      words:
+        A list of strings representing words.
+    Returns:
+      Return a dict whose keys are words and values are the corresponding
+          tokens.
+    """
+    lexicon = []
+    for word in words:
+        chars = list(word.strip(" \t"))
+        if contain_oov(token_sym_table, chars):
+            continue
+        lexicon.append((word, chars))
+
+    # The OOV word is <UNK>
+    lexicon.append(("<UNK>", ["<unk>"]))
+    return lexicon
+
+
+def generate_tokens(text_file: str) -> Dict[str, int]:
+    """Generate tokens from the given text file.
+
+    Args:
+      text_file:
+        A file that contains text lines to generate tokens.
+    Returns:
+      Return a dict whose keys are tokens and values are token ids ranged
+      from 0 to len(keys) - 1.
+    """
+    tokens: Dict[str, int] = dict()
+    tokens['<blk>'] = 0
+    tokens['<sos/eos>'] = 1
+    tokens['<unk>'] = 2
+    whitespace = re.compile(r"([ \t\r\n]+)")
+    with open(text_file, "r", encoding="utf-8") as f:
+        for line in f:
+            line = re.sub(whitespace, "", line)
+            chars = list(line)
+            for char in chars:
+                if not char in tokens:
+                    tokens[char] = len(tokens)
+    return tokens 
+
+
+def main():
+    lang_dir = Path("data/lang_char")
+    text_file = lang_dir / "text"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    token_sym_table = generate_tokens(text_file)
+
+    lexicon = generate_lexicon(token_sym_table, words)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()

egs/aishell/ASR/local/prepare_lang.py

@@ -0,0 +1,381 @@
+#!/usr/bin/env python3
+# Copyright    2021  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.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def main():
+    out_dir = Path("data/lang_phone")
+    lexicon_filename = out_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(out_dir / "tokens.txt", token2id)
+    write_mapping(out_dir / "words.txt", word2id)
+    write_lexicon(out_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), out_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), out_dir / "L_disambig.pt")
+
+    if False:
+        # Just for debugging, will remove it
+        L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt")
+        L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt")
+        L_disambig.labels_sym = L.labels_sym
+        L_disambig.aux_labels_sym = L.aux_labels_sym
+        L.draw(out_dir / "L.png", title="L")
+        L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig")
+
+
+if __name__ == "__main__":
+    main()

egs/aishell/ASR/local/test_prepare_lang.py

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+# Copyright    2021  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.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import os
+import tempfile
+
+import k2
+from prepare_lang import (
+    add_disambig_symbols,
+    generate_id_map,
+    get_phones,
+    get_words,
+    lexicon_to_fst,
+    read_lexicon,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def generate_lexicon_file() -> str:
+    fd, filename = tempfile.mkstemp()
+    os.close(fd)
+    s = """
+    !SIL SIL
+    <SPOKEN_NOISE> SPN
+    <UNK> SPN
+    f f
+    a a
+    foo f o o
+    bar b a r
+    bark b a r k
+    food f o o d
+    food2 f o o d
+    fo  f o
+    """.strip()
+    with open(filename, "w") as f:
+        f.write(s)
+    return filename
+
+
+def test_read_lexicon(filename: str):
+    lexicon = read_lexicon(filename)
+    phones = get_phones(lexicon)
+    words = get_words(lexicon)
+    print(lexicon)
+    print(phones)
+    print(words)
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+    print(lexicon_disambig)
+    print("max disambig:", f"#{max_disambig}")
+
+    phones = ["<eps>", "SIL", "SPN"] + phones
+    for i in range(max_disambig + 1):
+        phones.append(f"#{i}")
+    words = ["<eps>"] + words
+
+    phone2id = generate_id_map(phones)
+    word2id = generate_id_map(words)
+
+    print(phone2id)
+    print(word2id)
+
+    write_mapping("phones.txt", phone2id)
+    write_mapping("words.txt", word2id)
+
+    write_lexicon("a.txt", lexicon)
+    write_lexicon("a_disambig.txt", lexicon_disambig)
+
+    fsa = lexicon_to_fst(lexicon, phone2id=phone2id, word2id=word2id)
+    fsa.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa.draw("L.pdf", title="L")
+
+    fsa_disambig = lexicon_to_fst(
+        lexicon_disambig, phone2id=phone2id, word2id=word2id
+    )
+    fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa_disambig.draw("L_disambig.pdf", title="L_disambig")
+
+
+def main():
+    filename = generate_lexicon_file()
+    test_read_lexicon(filename)
+    os.remove(filename)
+
+
+if __name__ == "__main__":
+    main()

egs/aishell/ASR/prepare.sh

@@ -0,0 +1,163 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=10
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/aishell
+#      You can find data_aishell, resource_aishell inside it.
+#      You can download them from https://www.openslr.org/33
+#
+#  - $dl_dir/lm
+#      This directory contains the language model downloaded from
+#        https://huggingface.co/pkufool/aishell_lm
+#
+#        - 3-gram.unpruned.apra
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "stage -1: Download LM"
+  # We assume that you have installed the git-lfs, if not, you could install it
+  # using: `sudo apt-get install git-lfs && git-lfs install`
+  [ ! -e $dl_dir/lm ] && mkdir -p $dl_dir/lm
+  git clone https://huggingface.co/pkufool/aishell_lm $dl_dir/lm
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/aishell,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/aishell $dl_dir/aishell
+  #
+  # The directory structure is
+  # aishell/
+  # |-- data_aishell
+  # |   |-- transcript
+  # |   `-- wav
+  # `-- resource_aishell
+  #     |-- lexicon.txt
+  #     `-- speaker.info
+
+  if [ ! -d $dl_dir/aishell/wav ]; then
+    lhotse download aishell $dl_dir
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/musan
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare aishell manifest"
+  # We assume that you have downloaded the aishell corpus
+  # to $dl_dir/aishell
+  mkdir -p data/manifests
+  lhotse prepare aishell -j $nj $dl_dir/aishell data/manifests
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  mkdir -p data/manifests
+  lhotse prepare musan $dl_dir/musan data/manifests
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for aishell"
+  mkdir -p data/fbank
+  ./local/compute_fbank_aishell.py
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  mkdir -p data/fbank
+  ./local/compute_fbank_musan.py
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare phone based lang"
+  mkdir -p data/lang_phone
+
+  (echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
+    cat - $dl_dir/aishell/resource_aishell/lexicon.txt |
+    sort | uniq > data/lang_phone/lexicon.txt
+
+  if [ ! -f data/lang_phone/L_disambig.pt ]; then
+    ./local/prepare_lang.py
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare char based lang"
+  mkdir -p data/lang_char
+  # We reuse words.txt from phone based lexicon
+  # so that the two can share G.pt later.
+  cp data/lang_phone/words.txt data/lang_char
+
+  cat $dl_dir/aishell/data_aishell/transcript/aishell_transcript_v0.8.txt |
+  cut -d " " -f 2- | sed -e 's/[ \t\r\n]*//g' > data/lang_char/text
+  
+  if [ ! -f data/lang_char/L_disambig.pt ]; then
+    ./local/prepare_char.py
+  fi
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Prepare G"
+  # We assume you have install kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  mkdir -p data/lm
+  if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="data/lang_phone/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      $dl_dir/lm/3-gram.unpruned.arpa > data/lm/G_3_gram.fst.txt
+  fi
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Compile HLG"
+  ./local/compile_hlg.py --lang-dir data/lang_phone
+  ./local/compile_hlg.py --lang-dir data/lang_char
+fi
+

egs/aishell/ASR/shared

@@ -0,0 +1 @@
+../../../icefall/shared/

egs/aishell/ASR/tdnn_lstm_ctc/README.md

@@ -0,0 +1,4 @@
+
+Please visit
+<https://icefall.readthedocs.io/en/latest/recipes/aishell/tdnn_lstm_ctc.html>
+for how to run this recipe.

egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py

@@ -0,0 +1,335 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import List, Union
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest
+from lhotse.dataset import (
+    BucketingSampler,
+    CutConcatenate,
+    CutMix,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SingleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.dataset.datamodule import DataModule
+from icefall.utils import str2bool
+
+
+class AishellAsrDataModule(DataModule):
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    """
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        super().add_arguments(parser)
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--feature-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the BucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+    def train_dataloaders(self) -> DataLoader:
+        logging.info("About to get train cuts")
+        cuts_train = self.train_cuts()
+
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(self.args.feature_dir / "cuts_musan.json.gz")
+
+        logging.info("About to create train dataset")
+        transforms = [CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20))]
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = [
+            SpecAugment(
+                num_frame_masks=2,
+                features_mask_size=27,
+                num_feature_masks=2,
+                frames_mask_size=100,
+            )
+        ]
+
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(
+                    Fbank(FbankConfig(num_mel_bins=80))
+                ),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using BucketingSampler.")
+            train_sampler = BucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                bucket_method="equal_duration",
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SingleCutSampler.")
+            train_sampler = SingleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self) -> DataLoader:
+        logging.info("About to get dev cuts")
+        cuts_valid = self.valid_cuts()
+
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(
+                    Fbank(FbankConfig(num_mel_bins=80))
+                ),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = SingleCutSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self) -> Union[DataLoader, List[DataLoader]]:
+        cuts = self.test_cuts()
+        is_list = isinstance(cuts, list)
+        test_loaders = []
+        if not is_list:
+            cuts = [cuts]
+
+        for cuts_test in cuts:
+            logging.debug("About to create test dataset")
+            test = K2SpeechRecognitionDataset(
+                input_strategy=OnTheFlyFeatures(
+                    Fbank(FbankConfig(num_mel_bins=80))
+                )
+                if self.args.on_the_fly_feats
+                else PrecomputedFeatures(),
+                return_cuts=self.args.return_cuts,
+            )
+            sampler = SingleCutSampler(
+                cuts_test, max_duration=self.args.max_duration
+            )
+            logging.debug("About to create test dataloader")
+            test_dl = DataLoader(
+                test, batch_size=None, sampler=sampler, num_workers=1
+            )
+            test_loaders.append(test_dl)
+
+        if is_list:
+            return test_loaders
+        else:
+            return test_loaders[0]
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        cuts_train = load_manifest(
+            self.args.feature_dir / "cuts_train.json.gz"
+        )
+        return cuts_train
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        cuts_valid = load_manifest(
+            self.args.feature_dir / "cuts_dev.json.gz"
+        )
+        return cuts_valid
+
+    @lru_cache()
+    def test_cuts(self) -> List[CutSet]:
+        test_sets = ["test"]
+        cuts = []
+        for test_set in test_sets:
+            logging.debug("About to get test cuts")
+            cuts.append(
+                load_manifest(
+                    self.args.feature_dir / f"cuts_{test_set}.json.gz"
+                )
+            )
+        return cuts

egs/aishell/ASR/tdnn_lstm_ctc/decode.py

@@ -0,0 +1,399 @@
+#!/usr/bin/env python3
+# Copyright      2021  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.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from model import TdnnLstm
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=19,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        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.
+        """,
+    )
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=30,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is nbest.
+        """,
+    )
+    parser.add_argument(
+        "--export",
+        type=str2bool,
+        default=False,
+        help="""When enabled, the averaged model is saved to
+        tdnn/exp/pretrained.pt. Note: only model.state_dict() is saved.
+        pretrained.pt contains a dict {"model": model.state_dict()},
+        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
+        """,
+    )
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn_lstm_ctc/exp/"),
+            "lang_dir": Path("data/lang_phone"),
+            "lm_dir": Path("data/lm"),
+            # parameters for tdnn_lstm_ctc
+            "subsampling_factor": 3,
+            "feature_dim": 80,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 7,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    batch: dict,
+    lexicon: Lexicon,
+) -> Dict[str, List[List[int]]]:
+    """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 the decoding method is 1best, the key is the string
+               `no_rescore`. If the decoding method is nbest, the key is the
+               string `no_rescore-xxx`, xxx is the num_paths.
+
+        - 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.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      lexicon:
+        It contains word symbol table.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = HLG.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is [N, T, C]
+
+    feature = feature.permute(0, 2, 1)  # now feature is [N, C, T]
+
+    nnet_output = model(feature)
+    # nnet_output is [N, T, C]
+
+    supervisions = batch["supervisions"]
+
+    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,
+        HLG=HLG,
+        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,
+    )
+
+    assert params.method in ["1best", "nbest"]
+    if params.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,
+        )
+        key = f"no_rescore-{params.num_paths}"
+    hyps = get_texts(best_path)
+    hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
+    return {key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: k2.Fsa,
+    lexicon: Lexicon,
+) -> Dict[str, List[Tuple[List[int], List[int]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph.
+      lexicon:
+        It contains word symbol table.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if decoding method is 1best,
+      or it may be "no-rescoer-100" if decoding method is nbest.
+      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.
+    """
+    results = []
+
+    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"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            batch=batch,
+            lexicon=lexicon,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for hyp_words, ref_text in zip(hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            results[lm_scale].extend(this_batch)
+
+        num_cuts += len(batch["supervisions"]["text"])
+
+        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[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        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.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        # We compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((list("".join(res[0])), list("".join(res[1]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results_char)
+            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.exp_dir / f"cer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER 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()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    HLG = k2.Fsa.from_dict(
+        torch.load(f"{params.lang_dir}/HLG.pt", map_location="cpu")
+    )
+    HLG = HLG.to(device)
+    assert HLG.requires_grad is False
+
+    if not hasattr(HLG, "lm_scores"):
+        HLG.lm_scores = HLG.scores.clone()
+
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+    if 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 start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    if params.export:
+        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
+        torch.save(
+            {"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt"
+        )
+
+    model.to(device)
+    model.eval()
+
+    aishell = AishellAsrDataModule(args)
+    # 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"]
+    for test_set, test_dl in zip(test_sets, aishell.test_dataloaders()):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            lexicon=lexicon,
+        )
+
+        save_results(
+            params=params, test_set_name=test_set, results_dict=results_dict
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/aishell/ASR/tdnn_lstm_ctc/model.py

@@ -0,0 +1,103 @@
+# Copyright      2021  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.
+
+
+import torch
+import torch.nn as nn
+
+
+class TdnnLstm(nn.Module):
+    def __init__(
+        self, num_features: int, num_classes: int, subsampling_factor: int = 3
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            It reduces the number of output frames by this factor.
+        """
+        super().__init__()
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        self.tdnn = nn.Sequential(
+            nn.Conv1d(
+                in_channels=num_features,
+                out_channels=500,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=500, affine=False),
+            nn.Conv1d(
+                in_channels=500,
+                out_channels=500,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=500, affine=False),
+            nn.Conv1d(
+                in_channels=500,
+                out_channels=500,
+                kernel_size=3,
+                stride=self.subsampling_factor,  # stride: subsampling_factor!
+                padding=1,
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm1d(num_features=500, affine=False),
+        )
+        self.lstms = nn.ModuleList(
+            [
+                nn.LSTM(input_size=500, hidden_size=500, num_layers=1)
+                for _ in range(5)
+            ]
+        )
+        self.lstm_bnorms = nn.ModuleList(
+            [nn.BatchNorm1d(num_features=500, affine=False) for _ in range(5)]
+        )
+        self.dropout = nn.Dropout(0.2)
+        self.linear = nn.Linear(in_features=500, out_features=self.num_classes)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            Its shape is [N, C, T]
+
+        Returns:
+          The output tensor has shape [N, T, C]
+        """
+        x = self.tdnn(x)
+        x = x.permute(2, 0, 1)  # (N, C, T) -> (T, N, C) -> how LSTM expects it
+        for lstm, bnorm in zip(self.lstms, self.lstm_bnorms):
+            x_new, _ = lstm(x)
+            x_new = bnorm(x_new.permute(1, 2, 0)).permute(
+                2, 0, 1
+            )  # (T, N, C) -> (N, C, T) -> (T, N, C)
+            x_new = self.dropout(x_new)
+            x = x_new + x  # skip connections
+        x = x.transpose(
+            1, 0
+        )  # (T, N, C) -> (N, T, C) -> linear expects "features" in the last dim
+        x = self.linear(x)
+        x = nn.functional.log_softmax(x, dim=-1)
+        return x

egs/aishell/ASR/tdnn_lstm_ctc/pretrained.py

@@ -0,0 +1,231 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from model import TdnnLstm
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.decode import (
+    get_lattice,
+    one_best_decoding,
+)
+from icefall.utils import AttributeDict, 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(
+        "--words-file",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument(
+        "--HLG", type=str, required=True, help="Path to HLG.pt."
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="1best",
+        help="""Decoding method.
+        Use the best path as decoding output. Only the transformer encoder
+        output is used for decoding. We call it HLG decoding.
+        """,
+    )
+
+    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 get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 80,
+            "subsampling_factor": 3,
+            "num_classes": 220,
+            "sample_rate": 16000,
+            "search_beam": 20,
+            "output_beam": 7,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+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}. "
+            f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+    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 = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=params.num_classes,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"])
+    model.to(device)
+    model.eval()
+
+    logging.info(f"Loading HLG from {params.HLG}")
+    HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
+    HLG = HLG.to(device)
+    if not hasattr(HLG, "lm_scores"):
+        # For whole-lattice-rescoring and attention-decoder
+        HLG.lm_scores = HLG.scores.clone()
+
+    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)
+
+    features = pad_sequence(
+        features, batch_first=True, padding_value=math.log(1e-10)
+    )
+    features = features.permute(0, 2, 1)  # now features is [N, C, T]
+
+    with torch.no_grad():
+        nnet_output = model(features)
+        # nnet_output is [N, T, C]
+
+    batch_size = nnet_output.shape[0]
+    supervision_segments = torch.tensor(
+        [[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
+        dtype=torch.int32,
+    )
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        HLG=HLG,
+        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,
+    )
+
+    assert(params.method == "1best")
+    logging.info("Use HLG decoding")
+    best_path = one_best_decoding(
+        lattice=lattice, use_double_scores=params.use_double_scores
+    )
+
+    hyps = get_texts(best_path)
+    word_sym_table = k2.SymbolTable.from_file(params.words_file)
+    hyps = [[word_sym_table[i] for i in ids] for ids 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/aishell/ASR/tdnn_lstm_ctc/train.py

@@ -0,0 +1,616 @@
+#!/usr/bin/env python3
+# Copyright      2021  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.
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional
+
+import k2
+import torch
+import torch.distributed as dist
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from asr_datamodule import AishellAsrDataModule
+from lhotse.utils import fix_random_seed
+from model import TdnnLstm
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.optim.lr_scheduler import StepLR
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=20,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        tdnn_lstm_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    is saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    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"
+
+        - lr: It specifies the initial learning rate
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - 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.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval` is 0
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+    """
+    params = AttributeDict(
+        {
+            "exp_dir": Path("tdnn_lstm_ctc/exp_lr1e-4"),
+            "lang_dir": Path("data/lang_phone"),
+            "lr": 1e-4,
+            "feature_dim": 80,
+            "weight_decay": 5e-4,
+            "subsampling_factor": 3,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 10,
+            "reset_interval": 200,
+            "valid_interval": 1000,
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: torch.optim.lr_scheduler._LRScheduler,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: CtcTrainingGraphCompiler,
+    is_training: bool,
+):
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of TdnnLstm in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is [N, T, C]
+    feature = feature.permute(0, 2, 1)  # now feature is [N, C, T]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output = model(feature)
+        # nnet_output is [N, T, C]
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervisions = batch["supervisions"]
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+    decoding_graph = graph_compiler.compile(texts)
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    assert loss.requires_grad == is_training
+
+    # train_frames and valid_frames are used for printing.
+    if is_training:
+        params.train_frames = supervision_segments[:, 2].sum().item()
+    else:
+        params.valid_frames = supervision_segments[:, 2].sum().item()
+
+    return loss
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> None:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = 0.0
+    tot_frames = 0.0
+    for batch_idx, batch in enumerate(valid_dl):
+        loss = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+
+        loss_cpu = loss.detach().cpu().item()
+        tot_loss += loss_cpu
+        tot_frames += params.valid_frames
+
+    if world_size > 1:
+        s = torch.tensor([tot_loss, tot_frames], device=loss.device)
+        dist.all_reduce(s, op=dist.ReduceOp.SUM)
+        s = s.cpu().tolist()
+        tot_loss = s[0]
+        tot_frames = s[1]
+
+    params.valid_loss = tot_loss / tot_frames
+
+    if params.valid_loss < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = params.valid_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = 0.0  # reset after params.reset_interval of batches
+    tot_frames = 0.0  # reset after params.reset_interval of batches
+
+    params.tot_loss = 0.0
+    params.tot_frames = 0.0
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        loss_cpu = loss.detach().cpu().item()
+
+        tot_frames += params.train_frames
+        tot_loss += loss_cpu
+        tot_avg_loss = tot_loss / tot_frames
+
+        params.tot_frames += params.train_frames
+        params.tot_loss += loss_cpu
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, batch {batch_idx}, "
+                f"batch avg loss {loss_cpu/params.train_frames:.4f}, "
+                f"total avg loss: {tot_avg_loss:.4f}, "
+                f"batch size: {batch_size}"
+            )
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/current_loss",
+                    loss_cpu / params.train_frames,
+                    params.batch_idx_train,
+                )
+
+                tb_writer.add_scalar(
+                    "train/tot_avg_loss",
+                    tot_avg_loss,
+                    params.batch_idx_train,
+                )
+
+        if batch_idx > 0 and batch_idx % params.reset_interval == 0:
+            tot_loss = 0
+            tot_frames = 0
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(
+                f"Epoch {params.cur_epoch}, valid loss {params.valid_loss:.4f},"
+                f" best valid loss: {params.best_valid_loss:.4f} "
+                f"best valid epoch: {params.best_valid_epoch}"
+            )
+
+    params.train_loss = params.tot_loss / params.tot_frames
+
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(42)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_phone_id = max(lexicon.tokens)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = CtcTrainingGraphCompiler(lexicon=lexicon, device=device)
+
+    model = TdnnLstm(
+        num_features=params.feature_dim,
+        num_classes=max_phone_id + 1,  # +1 for the blank symbol
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = optim.AdamW(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+    scheduler = StepLR(optimizer, step_size=8, gamma=0.1)
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    aishell = AishellAsrDataModule(args)
+    train_dl = aishell.train_dataloaders()
+    valid_dl = aishell.valid_dataloaders()
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        train_dl.sampler.set_epoch(epoch)
+
+        if epoch > params.start_epoch:
+            logging.info(f"epoch {epoch}, lr: {scheduler.get_last_lr()[0]}")
+
+        if tb_writer is not None:
+            tb_writer.add_scalar(
+                "train/lr",
+                scheduler.get_last_lr()[0],
+                params.batch_idx_train,
+            )
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        scheduler.step()
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+if __name__ == "__main__":
+    main()

icefall/char_graph_compiler.py

@@ -0,0 +1,98 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+from typing import List
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+class CharCtcTrainingGraphCompiler(object):
+    def __init__(
+        self,
+        lexicon: Lexicon,
+        device: torch.device,
+        sos_token: str = "<sos/eos>",
+        eos_token: str = "<sos/eos>",
+        oov: str = "<unk>",
+    ):
+        """
+        Args:
+          lexicon:
+            It is built from `data/lang/lexicon.txt`.
+          device:
+            The device to use for operations compiling transcripts to FSAs.
+          oov:
+            Out of vocabulary token. When a word(token) in the transcript
+            does not exist in the token list, it is replaced with `oov`.
+        """
+
+        assert oov in lexicon.token_table
+
+        self.oov_id = lexicon.token_table[oov]
+        self.token_table = lexicon.token_table
+
+        self.device = device
+
+        self.sos_id = self.token_table[sos_token]
+        self.eos_id = self.token_table[eos_token]
+
+
+    def texts_to_ids(self, texts: List[str]) -> List[List[int]]:
+        """Convert a list of texts to a list-of-list of token IDs.
+
+        Args:
+          texts:
+            It is a list of strings.
+            An example containing two strings is given below:
+
+                ['你好中国', '北京欢迎您']
+        Returns:
+          Return a list-of-list of token IDs.
+        """
+        ids: List[List[int]] = []
+        whitespace = re.compile(r"([ \t])")
+        for text in texts:
+            text = re.sub(whitespace, "", text)
+            sub_ids = [self.token_table[txt] if txt in self.token_table \
+                else self.oov_id for txt in text]
+            ids.append(sub_ids)
+        return ids
+
+
+    def compile(
+        self,
+        token_ids: List[List[int]],
+        modified: bool = False,
+    ) -> k2.Fsa:
+        """Build a ctc graph from a list-of-list token IDs.
+
+        Args:
+          piece_ids:
+            It is a list-of-list integer IDs.
+         modified:
+           See :func:`k2.ctc_graph` for its meaning.
+        Return:
+          Return an FsaVec, which is the result of composing a
+          CTC topology with linear FSAs constructed from the given
+          piece IDs.
+        """
+        return k2.ctc_graph(token_ids, modified=modified, device=self.device)
+

icefall/decode.py

@@ -903,3 +903,4 @@ def rescore_with_attention_decoder(
             key = f"ngram_lm_scale_{n_scale}_attention_scale_{a_scale}"
             ans[key] = best_path
     return ans
+

icefall/utils.py

@@ -104,7 +104,6 @@ def setup_logger(
     """
     now = datetime.now()
     date_time = now.strftime("%Y-%m-%d-%H-%M-%S")
-
     if dist.is_available() and dist.is_initialized():
         world_size = dist.get_world_size()
         rank = dist.get_rank()