mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-08-13 12:02:21 +00:00
Merge remote-tracking branch 'upstream/master' into conformer_lm
This commit is contained in:
Daniel Povey
commit
0cfa8c80e0
10
.github/workflows/run-yesno-recipe.yml
vendored
10
.github/workflows/run-yesno-recipe.yml
vendored
@ -21,11 +21,11 @@ on:
|
||||
branches:
|
||||
- master
|
||||
pull_request:
|
||||
branches:
|
||||
- master
|
||||
types: [labeled]
|
||||
|
||||
jobs:
|
||||
run-yesno-recipe:
|
||||
if: github.event.label.name == 'ready' || github.event_name == 'push'
|
||||
runs-on: ${{ matrix.os }}
|
||||
strategy:
|
||||
matrix:
|
||||
@ -33,6 +33,8 @@ jobs:
|
||||
# TODO: enable macOS for CPU testing
|
||||
os: [ubuntu-18.04]
|
||||
python-version: [3.8]
|
||||
torch: ["1.8.1"]
|
||||
k2-version: ["1.9.dev20210919"]
|
||||
fail-fast: false
|
||||
|
||||
steps:
|
||||
@ -54,10 +56,8 @@ jobs:
|
||||
|
||||
- name: Install Python dependencies
|
||||
run: |
|
||||
python3 -m pip install --upgrade pip black flake8
|
||||
python3 -m pip install -U pip
|
||||
python3 -m pip install k2==1.4.dev20210822+cpu.torch1.7.1 -f https://k2-fsa.org/nightly/
|
||||
python3 -m pip install torchaudio==0.7.2
|
||||
pip install k2==${{ matrix.k2-version }}+cpu.torch${{ matrix.torch }} -f https://k2-fsa.org/nightly/
|
||||
python3 -m pip install git+https://github.com/lhotse-speech/lhotse
|
||||
|
||||
# We are in ./icefall and there is a file: requirements.txt in it
|
||||
|
||||
2
.github/workflows/style_check.yml
vendored
2
.github/workflows/style_check.yml
vendored
@ -45,7 +45,7 @@ jobs:
|
||||
|
||||
- name: Install Python dependencies
|
||||
run: |
|
||||
python3 -m pip install --upgrade pip black flake8
|
||||
python3 -m pip install --upgrade pip black==21.6b0 flake8==3.9.2
|
||||
|
||||
- name: Run flake8
|
||||
shell: bash
|
||||
|
||||
21
.github/workflows/test.yml
vendored
21
.github/workflows/test.yml
vendored
@ -21,18 +21,19 @@ on:
|
||||
branches:
|
||||
- master
|
||||
pull_request:
|
||||
branches:
|
||||
- master
|
||||
types: [labeled]
|
||||
|
||||
jobs:
|
||||
test:
|
||||
if: github.event.label.name == 'ready' || github.event_name == 'push'
|
||||
runs-on: ${{ matrix.os }}
|
||||
strategy:
|
||||
matrix:
|
||||
os: [ubuntu-18.04, macos-10.15]
|
||||
python-version: [3.6, 3.7, 3.8, 3.9]
|
||||
torch: ["1.8.1"]
|
||||
k2-version: ["1.4.dev20210822"]
|
||||
k2-version: ["1.9.dev20210919"]
|
||||
|
||||
fail-fast: false
|
||||
|
||||
steps:
|
||||
@ -52,6 +53,20 @@ jobs:
|
||||
# icefall requirements
|
||||
pip install -r requirements.txt
|
||||
|
||||
- name: Install graphviz
|
||||
if: startsWith(matrix.os, 'ubuntu')
|
||||
shell: bash
|
||||
run: |
|
||||
python3 -m pip install -qq graphviz
|
||||
sudo apt-get -qq install graphviz
|
||||
|
||||
- name: Install graphviz
|
||||
if: startsWith(matrix.os, 'macos')
|
||||
shell: bash
|
||||
run: |
|
||||
python3 -m pip install -qq graphviz
|
||||
brew install -q graphviz
|
||||
|
||||
- name: Run tests
|
||||
if: startsWith(matrix.os, 'ubuntu')
|
||||
run: |
|
||||
|
||||
2
.gitignore
vendored
2
.gitignore
vendored
@ -4,4 +4,4 @@ path.sh
|
||||
exp
|
||||
exp*/
|
||||
*.pt
|
||||
download/
|
||||
download
|
||||
|
||||
107
README.md
107
README.md
@ -1,80 +1,61 @@
|
||||
|
||||
# Table of Contents
|
||||
|
||||
- [Installation](#installation)
|
||||
* [Install k2](#install-k2)
|
||||
* [Install lhotse](#install-lhotse)
|
||||
* [Install icefall](#install-icefall)
|
||||
- [Run recipes](#run-recipes)
|
||||
<div align="center">
|
||||
<img src="https://raw.githubusercontent.com/k2-fsa/icefall/master/docs/source/_static/logo.png" width=168>
|
||||
</div>
|
||||
|
||||
## Installation
|
||||
|
||||
`icefall` depends on [k2][k2] for FSA operations and [lhotse][lhotse] for
|
||||
data preparations. To use `icefall`, you have to install its dependencies first.
|
||||
The following subsections describe how to setup the environment.
|
||||
|
||||
CAUTION: There are various ways to setup the environment. What we describe
|
||||
here is just one alternative.
|
||||
|
||||
### Install k2
|
||||
|
||||
Please refer to [k2's installation documentation][k2-install] to install k2.
|
||||
If you have any issues about installing k2, please open an issue at
|
||||
<https://github.com/k2-fsa/k2/issues>.
|
||||
|
||||
### Install lhotse
|
||||
|
||||
Please refer to [lhotse's installation documentation][lhotse-install] to install
|
||||
lhotse.
|
||||
|
||||
### Install icefall
|
||||
|
||||
`icefall` is a set of Python scripts. What you need to do is just to set
|
||||
the environment variable `PYTHONPATH`:
|
||||
|
||||
```bash
|
||||
cd $HOME/open-source
|
||||
git clone https://github.com/k2-fsa/icefall
|
||||
cd icefall
|
||||
pip install -r requirements.txt
|
||||
export PYTHONPATH=$HOME/open-source/icefall:$PYTHONPATHON
|
||||
```
|
||||
|
||||
To verify `icefall` was installed successfully, you can run:
|
||||
|
||||
```bash
|
||||
python3 -c "import icefall; print(icefall.__file__)"
|
||||
```
|
||||
|
||||
It should print the path to `icefall`.
|
||||
Please refer to <https://icefall.readthedocs.io/en/latest/installation/index.html>
|
||||
for installation.
|
||||
|
||||
## Recipes
|
||||
|
||||
At present, two recipes are provided:
|
||||
Please refer to <https://icefall.readthedocs.io/en/latest/recipes/index.html>
|
||||
for more information.
|
||||
|
||||
- [LibriSpeech][LibriSpeech]
|
||||
- [yesno][yesno] [](https://colab.research.google.com/drive/1tIjjzaJc3IvGyKiMCDWO-TSnBgkcuN3B?usp=sharing)
|
||||
We provide two recipes at present:
|
||||
|
||||
### Yesno
|
||||
- [yesno][yesno]
|
||||
- [LibriSpeech][librispeech]
|
||||
|
||||
For the yesno recipe, training with 50 epochs takes less than 2 minutes using **CPU**.
|
||||
### yesno
|
||||
|
||||
The WER is
|
||||
This is the simplest ASR recipe in `icefall` and can be run on CPU.
|
||||
Training takes less than 30 seconds and gives you the following WER:
|
||||
|
||||
```
|
||||
[test_set] %WER 0.42% [1 / 240, 0 ins, 1 del, 0 sub ]
|
||||
```
|
||||
We do provide a Colab notebook for this recipe.
|
||||
|
||||
## Use Pre-trained models
|
||||
|
||||
See [egs/librispeech/ASR/conformer_ctc/README.md](egs/librispeech/ASR/conformer_ctc/README.md)
|
||||
for how to use pre-trained models.
|
||||
[](https://colab.research.google.com/drive/1huyupXAcHsUrKaWfI83iMEJ6J0Nh0213?usp=sharing)
|
||||
[](https://colab.research.google.com/drive/1tIjjzaJc3IvGyKiMCDWO-TSnBgkcuN3B?usp=sharing)
|
||||
|
||||
|
||||
[yesno]: egs/yesno/ASR/README.md
|
||||
[LibriSpeech]: egs/librispeech/ASR/README.md
|
||||
[k2-install]: https://k2.readthedocs.io/en/latest/installation/index.html#
|
||||
[k2]: https://github.com/k2-fsa/k2
|
||||
[lhotse]: https://github.com/lhotse-speech/lhotse
|
||||
[lhotse-install]: https://lhotse.readthedocs.io/en/latest/getting-started.html#installation
|
||||
### LibriSpeech
|
||||
|
||||
We provide two models for this recipe: [conformer CTC model][LibriSpeech_conformer_ctc]
|
||||
and [TDNN LSTM CTC model][LibriSpeech_tdnn_lstm_ctc].
|
||||
|
||||
#### Conformer CTC Model
|
||||
|
||||
The best WER we currently have is:
|
||||
|
||||
||test-clean|test-other|
|
||||
|--|--|--|
|
||||
|WER| 2.57% | 5.94% |
|
||||
|
||||
We provide a Colab notebook to run a pre-trained conformer CTC model: [](https://colab.research.google.com/drive/1huyupXAcHsUrKaWfI83iMEJ6J0Nh0213?usp=sharing)
|
||||
|
||||
#### TDNN LSTM CTC Model
|
||||
|
||||
The WER for this model is:
|
||||
|
||||
||test-clean|test-other|
|
||||
|--|--|--|
|
||||
|WER| 6.59% | 17.69% |
|
||||
|
||||
We provide a Colab notebook to run a pre-trained TDNN LSTM CTC model: [](https://colab.research.google.com/drive/1kNmDXNMwREi0rZGAOIAOJo93REBuOTcd?usp=sharing)
|
||||
|
||||
[LibriSpeech_tdnn_lstm_ctc]: egs/librispeech/ASR/tdnn_lstm_ctc
|
||||
[LibriSpeech_conformer_ctc]: egs/librispeech/ASR/conformer_ctc
|
||||
[yesno]: egs/yesno/ASR
|
||||
[librispeech]: egs/librispeech/ASR
|
||||
|
||||
1
docs/.gitignore
vendored
Normal file
1
docs/.gitignore
vendored
Normal file
@ -0,0 +1 @@
|
||||
build/
|
||||
20
docs/Makefile
Normal file
20
docs/Makefile
Normal file
@ -0,0 +1,20 @@
|
||||
# Minimal makefile for Sphinx documentation
|
||||
#
|
||||
|
||||
# You can set these variables from the command line, and also
|
||||
# from the environment for the first two.
|
||||
SPHINXOPTS ?=
|
||||
SPHINXBUILD ?= sphinx-build
|
||||
SOURCEDIR = source
|
||||
BUILDDIR = build
|
||||
|
||||
# Put it first so that "make" without argument is like "make help".
|
||||
help:
|
||||
@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
|
||||
|
||||
.PHONY: help Makefile
|
||||
|
||||
# Catch-all target: route all unknown targets to Sphinx using the new
|
||||
# "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS).
|
||||
%: Makefile
|
||||
@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
|
||||
35
docs/make.bat
Normal file
35
docs/make.bat
Normal file
@ -0,0 +1,35 @@
|
||||
@ECHO OFF
|
||||
|
||||
pushd %~dp0
|
||||
|
||||
REM Command file for Sphinx documentation
|
||||
|
||||
if "%SPHINXBUILD%" == "" (
|
||||
set SPHINXBUILD=sphinx-build
|
||||
)
|
||||
set SOURCEDIR=source
|
||||
set BUILDDIR=build
|
||||
|
||||
if "%1" == "" goto help
|
||||
|
||||
%SPHINXBUILD% >NUL 2>NUL
|
||||
if errorlevel 9009 (
|
||||
echo.
|
||||
echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
|
||||
echo.installed, then set the SPHINXBUILD environment variable to point
|
||||
echo.to the full path of the 'sphinx-build' executable. Alternatively you
|
||||
echo.may add the Sphinx directory to PATH.
|
||||
echo.
|
||||
echo.If you don't have Sphinx installed, grab it from
|
||||
echo.http://sphinx-doc.org/
|
||||
exit /b 1
|
||||
)
|
||||
|
||||
%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
|
||||
goto end
|
||||
|
||||
:help
|
||||
%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
|
||||
|
||||
:end
|
||||
popd
|
||||
2
docs/requirements.txt
Normal file
2
docs/requirements.txt
Normal file
@ -0,0 +1,2 @@
|
||||
sphinx_rtd_theme
|
||||
sphinx
|
||||
BIN
docs/source/_static/logo.png
Normal file
BIN
docs/source/_static/logo.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 666 KiB |
76
docs/source/conf.py
Normal file
76
docs/source/conf.py
Normal file
@ -0,0 +1,76 @@
|
||||
# Configuration file for the Sphinx documentation builder.
|
||||
#
|
||||
# This file only contains a selection of the most common options. For a full
|
||||
# list see the documentation:
|
||||
# https://www.sphinx-doc.org/en/master/usage/configuration.html
|
||||
|
||||
# -- Path setup --------------------------------------------------------------
|
||||
|
||||
# If extensions (or modules to document with autodoc) are in another directory,
|
||||
# add these directories to sys.path here. If the directory is relative to the
|
||||
# documentation root, use os.path.abspath to make it absolute, like shown here.
|
||||
#
|
||||
# import os
|
||||
# import sys
|
||||
# sys.path.insert(0, os.path.abspath('.'))
|
||||
|
||||
import sphinx_rtd_theme
|
||||
|
||||
# -- Project information -----------------------------------------------------
|
||||
|
||||
project = "icefall"
|
||||
copyright = "2021, icefall development team"
|
||||
author = "icefall development team"
|
||||
|
||||
# The full version, including alpha/beta/rc tags
|
||||
release = "0.1"
|
||||
|
||||
|
||||
# -- General configuration ---------------------------------------------------
|
||||
|
||||
# Add any Sphinx extension module names here, as strings. They can be
|
||||
# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
|
||||
# ones.
|
||||
extensions = [
|
||||
"sphinx_rtd_theme",
|
||||
]
|
||||
|
||||
# Add any paths that contain templates here, relative to this directory.
|
||||
templates_path = ["_templates"]
|
||||
|
||||
# List of patterns, relative to source directory, that match files and
|
||||
# directories to ignore when looking for source files.
|
||||
# This pattern also affects html_static_path and html_extra_path.
|
||||
exclude_patterns = []
|
||||
|
||||
source_suffix = {
|
||||
".rst": "restructuredtext",
|
||||
}
|
||||
master_doc = "index"
|
||||
|
||||
|
||||
# -- Options for HTML output -------------------------------------------------
|
||||
|
||||
# The theme to use for HTML and HTML Help pages. See the documentation for
|
||||
# a list of builtin themes.
|
||||
#
|
||||
html_theme = "sphinx_rtd_theme"
|
||||
html_theme_path = [sphinx_rtd_theme.get_html_theme_path()]
|
||||
html_show_sourcelink = True
|
||||
|
||||
# Add any paths that contain custom static files (such as style sheets) here,
|
||||
# relative to this directory. They are copied after the builtin static files,
|
||||
# so a file named "default.css" will overwrite the builtin "default.css".
|
||||
html_static_path = ["_static", "installation/images"]
|
||||
|
||||
pygments_style = "sphinx"
|
||||
|
||||
numfig = True
|
||||
|
||||
html_context = {
|
||||
"display_github": True,
|
||||
"github_user": "k2-fsa",
|
||||
"github_repo": "icefall",
|
||||
"github_version": "master",
|
||||
"conf_py_path": "/icefall/docs/source/",
|
||||
}
|
||||
67
docs/source/contributing/code-style.rst
Normal file
67
docs/source/contributing/code-style.rst
Normal file
@ -0,0 +1,67 @@
|
||||
.. _follow the code style:
|
||||
|
||||
Follow the code style
|
||||
=====================
|
||||
|
||||
We use the following tools to make the code style to be as consistent as possible:
|
||||
|
||||
- `black <https://github.com/psf/black>`_, to format the code
|
||||
- `flake8 <https://github.com/PyCQA/flake8>`_, to check the style and quality of the code
|
||||
- `isort <https://github.com/PyCQA/isort>`_, to sort ``imports``
|
||||
|
||||
The following versions of the above tools are used:
|
||||
|
||||
- ``black == 12.6b0``
|
||||
- ``flake8 == 3.9.2``
|
||||
- ``isort == 5.9.2``
|
||||
|
||||
After running the following commands:
|
||||
|
||||
.. code-block::
|
||||
|
||||
$ git clone https://github.com/k2-fsa/icefall
|
||||
$ cd icefall
|
||||
$ pip install pre-commit
|
||||
$ pre-commit install
|
||||
|
||||
it will run the following checks whenever you run ``git commit``, **automatically**:
|
||||
|
||||
.. figure:: images/pre-commit-check.png
|
||||
:width: 600
|
||||
:align: center
|
||||
|
||||
pre-commit hooks invoked by ``git commit`` (Failed).
|
||||
|
||||
If any of the above checks failed, your ``git commit`` was not successful.
|
||||
Please fix any issues reported by the check tools.
|
||||
|
||||
.. HINT::
|
||||
|
||||
Some of the check tools, i.e., ``black`` and ``isort`` will modify
|
||||
the files to be commited **in-place**. So please run ``git status``
|
||||
after failure to see which file has been modified by the tools
|
||||
before you make any further changes.
|
||||
|
||||
After fixing all the failures, run ``git commit`` again and
|
||||
it should succeed this time:
|
||||
|
||||
.. figure:: images/pre-commit-check-success.png
|
||||
:width: 600
|
||||
:align: center
|
||||
|
||||
pre-commit hooks invoked by ``git commit`` (Succeeded).
|
||||
|
||||
If you want to check the style of your code before ``git commit``, you
|
||||
can do the following:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd icefall
|
||||
$ pip install black==21.6b0 flake8==3.9.2 isort==5.9.2
|
||||
$ black --check your_changed_file.py
|
||||
$ black your_changed_file.py # modify it in-place
|
||||
$
|
||||
$ flake8 your_changed_file.py
|
||||
$
|
||||
$ isort --check your_changed_file.py # modify it in-place
|
||||
$ isort your_changed_file.py
|
||||
45
docs/source/contributing/doc.rst
Normal file
45
docs/source/contributing/doc.rst
Normal file
@ -0,0 +1,45 @@
|
||||
Contributing to Documentation
|
||||
=============================
|
||||
|
||||
We use `sphinx <https://www.sphinx-doc.org/en/master/>`_
|
||||
for documentation.
|
||||
|
||||
Before writing documentation, you have to prepare the environment:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd docs
|
||||
$ pip install -r requirements.txt
|
||||
|
||||
After setting up the environment, you are ready to write documentation.
|
||||
Please refer to `reStructuredText Primer <https://www.sphinx-doc.org/en/master/usage/restructuredtext/basics.html>`_
|
||||
if you are not familiar with ``reStructuredText``.
|
||||
|
||||
After writing some documentation, you can build the documentation **locally**
|
||||
to preview what it looks like if it is published:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd docs
|
||||
$ make html
|
||||
|
||||
The generated documentation is in ``docs/build/html`` and can be viewed
|
||||
with the following commands:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd docs/build/html
|
||||
$ python3 -m http.server
|
||||
|
||||
It will print::
|
||||
|
||||
Serving HTTP on 0.0.0.0 port 8000 (http://0.0.0.0:8000/) ...
|
||||
|
||||
Open your browser, go to `<http://0.0.0.0:8000/>`_, and you will see
|
||||
the following:
|
||||
|
||||
.. figure:: images/doc-contrib.png
|
||||
:width: 600
|
||||
:align: center
|
||||
|
||||
View generated documentation locally with ``python3 -m http.server``.
|
||||
156
docs/source/contributing/how-to-create-a-recipe.rst
Normal file
156
docs/source/contributing/how-to-create-a-recipe.rst
Normal file
@ -0,0 +1,156 @@
|
||||
How to create a recipe
|
||||
======================
|
||||
|
||||
.. HINT::
|
||||
|
||||
Please read :ref:`follow the code style` to adjust your code sytle.
|
||||
|
||||
.. CAUTION::
|
||||
|
||||
``icefall`` is designed to be as Pythonic as possible. Please use
|
||||
Python in your recipe if possible.
|
||||
|
||||
Data Preparation
|
||||
----------------
|
||||
|
||||
We recommend you to prepare your training/test/validate dataset
|
||||
with `lhotse <https://github.com/lhotse-speech/lhotse>`_.
|
||||
|
||||
Please refer to `<https://lhotse.readthedocs.io/en/latest/index.html>`_
|
||||
for how to create a recipe in ``lhotse``.
|
||||
|
||||
.. HINT::
|
||||
|
||||
The ``yesno`` recipe in ``lhotse`` is a very good example.
|
||||
|
||||
Please refer to `<https://github.com/lhotse-speech/lhotse/pull/380>`_,
|
||||
which shows how to add a new recipe to ``lhotse``.
|
||||
|
||||
Suppose you would like to add a recipe for a dataset named ``foo``.
|
||||
You can do the following:
|
||||
|
||||
.. code-block::
|
||||
|
||||
$ cd egs
|
||||
$ mkdir -p foo/ASR
|
||||
$ cd foo/ASR
|
||||
$ touch prepare.sh
|
||||
$ chmod +x prepare.sh
|
||||
|
||||
If your dataset is very simple, please follow
|
||||
`egs/yesno/ASR/prepare.sh <https://github.com/k2-fsa/icefall/blob/master/egs/yesno/ASR/prepare.sh>`_
|
||||
to write your own ``prepare.sh``.
|
||||
Otherwise, please refer to
|
||||
`egs/librispeech/ASR/prepare.sh <https://github.com/k2-fsa/icefall/blob/master/egs/yesno/ASR/prepare.sh>`_
|
||||
to prepare your data.
|
||||
|
||||
|
||||
Training
|
||||
--------
|
||||
|
||||
Assume you have a fancy model, called ``bar`` for the ``foo`` recipe, you can
|
||||
organize your files in the following way:
|
||||
|
||||
.. code-block::
|
||||
|
||||
$ cd egs/foo/ASR
|
||||
$ mkdir bar
|
||||
$ cd bar
|
||||
$ touch README.md model.py train.py decode.py asr_datamodule.py pretrained.py
|
||||
|
||||
For instance , the ``yesno`` recipe has a ``tdnn`` model and its directory structure
|
||||
looks like the following:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
egs/yesno/ASR/tdnn/
|
||||
|-- README.md
|
||||
|-- asr_datamodule.py
|
||||
|-- decode.py
|
||||
|-- model.py
|
||||
|-- pretrained.py
|
||||
`-- train.py
|
||||
|
||||
**File description**:
|
||||
|
||||
- ``README.md``
|
||||
|
||||
It contains information of this recipe, e.g., how to run it, what the WER is, etc.
|
||||
|
||||
- ``asr_datamodule.py``
|
||||
|
||||
It provides code to create PyTorch dataloaders with train/test/validation dataset.
|
||||
|
||||
- ``decode.py``
|
||||
|
||||
It takes as inputs the checkpoints saved during the training stage to decode the test
|
||||
dataset(s).
|
||||
|
||||
- ``model.py``
|
||||
|
||||
It contains the definition of your fancy neural network model.
|
||||
|
||||
- ``pretrained.py``
|
||||
|
||||
We can use this script to do inference with a pre-trained model.
|
||||
|
||||
- ``train.py``
|
||||
|
||||
It contains training code.
|
||||
|
||||
|
||||
.. HINT::
|
||||
|
||||
Please take a look at
|
||||
|
||||
- `egs/yesno/tdnn <https://github.com/k2-fsa/icefall/tree/master/egs/yesno/ASR/tdnn>`_
|
||||
- `egs/librispeech/tdnn_lstm_ctc <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/tdnn_lstm_ctc>`_
|
||||
- `egs/librispeech/conformer_ctc <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/conformer_ctc>`_
|
||||
|
||||
to get a feel what the resulting files look like.
|
||||
|
||||
.. NOTE::
|
||||
|
||||
Every model in a recipe is kept to be as self-contained as possible.
|
||||
We tolerate duplicate code among different recipes.
|
||||
|
||||
|
||||
The training stage should be invocable by:
|
||||
|
||||
.. code-block::
|
||||
|
||||
$ cd egs/foo/ASR
|
||||
$ ./bar/train.py
|
||||
$ ./bar/train.py --help
|
||||
|
||||
|
||||
Decoding
|
||||
--------
|
||||
|
||||
Please refer to
|
||||
|
||||
- `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/conformer_ctc/decode.py>`_
|
||||
|
||||
If your model is transformer/conformer based.
|
||||
|
||||
- `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/tdnn_lstm_ctc/decode.py>`_
|
||||
|
||||
If your model is TDNN/LSTM based, i.e., there is no attention decoder.
|
||||
|
||||
- `<https://github.com/k2-fsa/icefall/blob/master/egs/yesno/ASR/tdnn/decode.py>`_
|
||||
|
||||
If there is no LM rescoring.
|
||||
|
||||
The decoding stage should be invocable by:
|
||||
|
||||
.. code-block::
|
||||
|
||||
$ cd egs/foo/ASR
|
||||
$ ./bar/decode.py
|
||||
$ ./bar/decode.py --help
|
||||
|
||||
Pre-trained model
|
||||
-----------------
|
||||
|
||||
Please demonstrate how to use your model for inference in ``egs/foo/ASR/bar/pretrained.py``.
|
||||
If possible, please consider creating a Colab notebook to show that.
|
||||
BIN
docs/source/contributing/images/doc-contrib.png
Normal file
BIN
docs/source/contributing/images/doc-contrib.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 198 KiB |
BIN
docs/source/contributing/images/pre-commit-check-success.png
Normal file
BIN
docs/source/contributing/images/pre-commit-check-success.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 153 KiB |
BIN
docs/source/contributing/images/pre-commit-check.png
Normal file
BIN
docs/source/contributing/images/pre-commit-check.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 214 KiB |
22
docs/source/contributing/index.rst
Normal file
22
docs/source/contributing/index.rst
Normal file
@ -0,0 +1,22 @@
|
||||
Contributing
|
||||
============
|
||||
|
||||
Contributions to ``icefall`` are very welcomed.
|
||||
There are many possible ways to make contributions and
|
||||
two of them are:
|
||||
|
||||
- To write documentation
|
||||
- To write code
|
||||
|
||||
- (1) To follow the code style in the repository
|
||||
- (2) To write a new recipe
|
||||
|
||||
In this page, we describe how to contribute documentation
|
||||
and code to ``icefall``.
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 2
|
||||
|
||||
doc
|
||||
code-style
|
||||
how-to-create-a-recipe
|
||||
25
docs/source/index.rst
Normal file
25
docs/source/index.rst
Normal file
@ -0,0 +1,25 @@
|
||||
.. icefall documentation master file, created by
|
||||
sphinx-quickstart on Mon Aug 23 16:07:39 2021.
|
||||
You can adapt this file completely to your liking, but it should at least
|
||||
contain the root `toctree` directive.
|
||||
|
||||
Icefall
|
||||
=======
|
||||
|
||||
.. image:: _static/logo.png
|
||||
:alt: icefall logo
|
||||
:width: 168px
|
||||
:align: center
|
||||
:target: https://github.com/k2-fsa/icefall
|
||||
|
||||
|
||||
Documentation for `icefall <https://github.com/k2-fsa/icefall>`_, containing
|
||||
speech recognition recipes using `k2 <https://github.com/k2-fsa/k2>`_.
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 2
|
||||
:caption: Contents:
|
||||
|
||||
installation/index
|
||||
recipes/index
|
||||
contributing/index
|
||||
@ -0,0 +1 @@
|
||||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="122" height="20" role="img" aria-label="device: CPU | CUDA"><title>device: CPU | CUDA</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="122" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="45" height="20" fill="#555"/><rect x="45" width="77" height="20" fill="#fe7d37"/><rect width="122" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><text aria-hidden="true" x="235" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="350">device</text><text x="235" y="140" transform="scale(.1)" fill="#fff" textLength="350">device</text><text aria-hidden="true" x="825" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="670">CPU | CUDA</text><text x="825" y="140" transform="scale(.1)" fill="#fff" textLength="670">CPU | CUDA</text></g></svg>
|
||||
|
After Width: | Height: | Size: 1.1 KiB |
1
docs/source/installation/images/k2-v1.9-blueviolet.svg
Normal file
1
docs/source/installation/images/k2-v1.9-blueviolet.svg
Normal file
@ -0,0 +1 @@
|
||||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="58" height="20" role="img" aria-label="k2: v1.9"><title>k2: v1.9</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="58" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="23" height="20" fill="#555"/><rect x="23" width="35" height="20" fill="blueviolet"/><rect width="58" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><text aria-hidden="true" x="125" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="130">k2</text><text x="125" y="140" transform="scale(.1)" fill="#fff" textLength="130">k2</text><text aria-hidden="true" x="395" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="250">v1.9</text><text x="395" y="140" transform="scale(.1)" fill="#fff" textLength="250">v1.9</text></g></svg>
|
||||
|
After Width: | Height: | Size: 1.1 KiB |
@ -0,0 +1 @@
|
||||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="114" height="20" role="img" aria-label="os: Linux | macOS"><title>os: Linux | macOS</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="114" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="23" height="20" fill="#555"/><rect x="23" width="91" height="20" fill="#ff69b4"/><rect width="114" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><text aria-hidden="true" x="125" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="130">os</text><text x="125" y="140" transform="scale(.1)" fill="#fff" textLength="130">os</text><text aria-hidden="true" x="675" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="810">Linux | macOS</text><text x="675" y="140" transform="scale(.1)" fill="#fff" textLength="810">Linux | macOS</text></g></svg>
|
||||
|
After Width: | Height: | Size: 1.1 KiB |
@ -0,0 +1 @@
|
||||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="170" height="20" role="img" aria-label="python: 3.6 | 3.7 | 3.8 | 3.9"><title>python: 3.6 | 3.7 | 3.8 | 3.9</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="170" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="49" height="20" fill="#555"/><rect x="49" width="121" height="20" fill="#007ec6"/><rect width="170" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><text aria-hidden="true" x="255" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="390">python</text><text x="255" y="140" transform="scale(.1)" fill="#fff" textLength="390">python</text><text aria-hidden="true" x="1085" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="1110">3.6 | 3.7 | 3.8 | 3.9</text><text x="1085" y="140" transform="scale(.1)" fill="#fff" textLength="1110">3.6 | 3.7 | 3.8 | 3.9</text></g></svg>
|
||||
|
After Width: | Height: | Size: 1.2 KiB |
@ -0,0 +1 @@
|
||||
<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="286" height="20" role="img" aria-label="torch: 1.6.0 | 1.7.0 | 1.7.1 | 1.8.0 | 1.8.1 | 1.9.0"><title>torch: 1.6.0 | 1.7.0 | 1.7.1 | 1.8.0 | 1.8.1 | 1.9.0</title><linearGradient id="s" x2="0" y2="100%"><stop offset="0" stop-color="#bbb" stop-opacity=".1"/><stop offset="1" stop-opacity=".1"/></linearGradient><clipPath id="r"><rect width="286" height="20" rx="3" fill="#fff"/></clipPath><g clip-path="url(#r)"><rect width="39" height="20" fill="#555"/><rect x="39" width="247" height="20" fill="#97ca00"/><rect width="286" height="20" fill="url(#s)"/></g><g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" text-rendering="geometricPrecision" font-size="110"><text aria-hidden="true" x="205" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="290">torch</text><text x="205" y="140" transform="scale(.1)" fill="#fff" textLength="290">torch</text><text aria-hidden="true" x="1615" y="150" fill="#010101" fill-opacity=".3" transform="scale(.1)" textLength="2370">1.6.0 | 1.7.0 | 1.7.1 | 1.8.0 | 1.8.1 | 1.9.0</text><text x="1615" y="140" transform="scale(.1)" fill="#fff" textLength="2370">1.6.0 | 1.7.0 | 1.7.1 | 1.8.0 | 1.8.1 | 1.9.0</text></g></svg>
|
||||
|
After Width: | Height: | Size: 1.3 KiB |
466
docs/source/installation/index.rst
Normal file
466
docs/source/installation/index.rst
Normal file
@ -0,0 +1,466 @@
|
||||
.. _install icefall:
|
||||
|
||||
Installation
|
||||
============
|
||||
|
||||
- |os|
|
||||
- |device|
|
||||
- |python_versions|
|
||||
- |torch_versions|
|
||||
- |k2_versions|
|
||||
|
||||
.. |os| image:: ./images/os-Linux_macOS-ff69b4.svg
|
||||
:alt: Supported operating systems
|
||||
|
||||
.. |device| image:: ./images/device-CPU_CUDA-orange.svg
|
||||
:alt: Supported devices
|
||||
|
||||
.. |python_versions| image:: ./images/python-3.6_3.7_3.8_3.9-blue.svg
|
||||
:alt: Supported python versions
|
||||
|
||||
.. |torch_versions| image:: ./images/torch-1.6.0_1.7.0_1.7.1_1.8.0_1.8.1_1.9.0-green.svg
|
||||
:alt: Supported PyTorch versions
|
||||
|
||||
.. |k2_versions| image:: ./images/k2-v1.9-blueviolet.svg
|
||||
:alt: Supported k2 versions
|
||||
|
||||
``icefall`` depends on `k2 <https://github.com/k2-fsa/k2>`_ and
|
||||
`lhotse <https://github.com/lhotse-speech/lhotse>`_.
|
||||
|
||||
We recommend you to install ``k2`` first, as ``k2`` is bound to
|
||||
a specific version of PyTorch after compilation. Install ``k2`` also
|
||||
installs its dependency PyTorch, which can be reused by ``lhotse``.
|
||||
|
||||
|
||||
(1) Install k2
|
||||
--------------
|
||||
|
||||
Please refer to `<https://k2-fsa.github.io/k2/installation/index.html>`_
|
||||
to install ``k2``.
|
||||
|
||||
.. CAUTION::
|
||||
|
||||
You need to install ``k2`` with a version at least **v1.9**.
|
||||
|
||||
.. HINT::
|
||||
|
||||
If you have already installed PyTorch and don't want to replace it,
|
||||
please install a version of ``k2`` that is compiled against the version
|
||||
of PyTorch you are using.
|
||||
|
||||
(2) Install lhotse
|
||||
------------------
|
||||
|
||||
Please refer to `<https://lhotse.readthedocs.io/en/latest/getting-started.html#installation>`_
|
||||
to install ``lhotse``.
|
||||
|
||||
.. HINT::
|
||||
|
||||
Install ``lhotse`` also installs its dependency `torchaudio <https://github.com/pytorch/audio>`_.
|
||||
|
||||
.. CAUTION::
|
||||
|
||||
If you have installed ``torchaudio``, please consider uninstalling it before
|
||||
installing ``lhotse``. Otherwise, it may update your already installed PyTorch.
|
||||
|
||||
(3) Download icefall
|
||||
--------------------
|
||||
|
||||
``icefall`` is a collection of Python scripts, so you don't need to install it
|
||||
and we don't provide a ``setup.py`` to install it.
|
||||
|
||||
What you need is to download it and set the environment variable ``PYTHONPATH``
|
||||
to point to it.
|
||||
|
||||
Assume you want to place ``icefall`` in the folder ``/tmp``. The
|
||||
following commands show you how to setup ``icefall``:
|
||||
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
cd /tmp
|
||||
git clone https://github.com/k2-fsa/icefall
|
||||
cd icefall
|
||||
pip install -r requirements.txt
|
||||
export PYTHONPATH=/tmp/icefall:$PYTHONPATH
|
||||
|
||||
.. HINT::
|
||||
|
||||
You can put several versions of ``icefall`` in the same virtual environment.
|
||||
To switch among different versions of ``icefall``, just set ``PYTHONPATH``
|
||||
to point to the version you want.
|
||||
|
||||
|
||||
Installation example
|
||||
--------------------
|
||||
|
||||
The following shows an example about setting up the environment.
|
||||
|
||||
|
||||
(1) Create a virtual environment
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ virtualenv -p python3.8 test-icefall
|
||||
|
||||
created virtual environment CPython3.8.6.final.0-64 in 1540ms
|
||||
creator CPython3Posix(dest=/ceph-fj/fangjun/test-icefall, clear=False, no_vcs_ignore=False, global=False)
|
||||
seeder FromAppData(download=False, pip=bundle, setuptools=bundle, wheel=bundle, via=copy, app_data_dir=/root/fangjun/.local/share/v
|
||||
irtualenv)
|
||||
added seed packages: pip==21.1.3, setuptools==57.4.0, wheel==0.36.2
|
||||
activators BashActivator,CShellActivator,FishActivator,PowerShellActivator,PythonActivator,XonshActivator
|
||||
|
||||
|
||||
(2) Activate your virtual environment
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ source test-icefall/bin/activate
|
||||
|
||||
(3) Install k2
|
||||
~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ pip install k2==1.4.dev20210822+cpu.torch1.9.0 -f https://k2-fsa.org/nightly/index.html
|
||||
|
||||
Looking in links: https://k2-fsa.org/nightly/index.html
|
||||
Collecting k2==1.4.dev20210822+cpu.torch1.9.0
|
||||
Downloading https://k2-fsa.org/nightly/whl/k2-1.4.dev20210822%2Bcpu.torch1.9.0-cp38-cp38-linux_x86_64.whl (1.6 MB)
|
||||
|________________________________| 1.6 MB 185 kB/s
|
||||
Collecting graphviz
|
||||
Downloading graphviz-0.17-py3-none-any.whl (18 kB)
|
||||
Collecting torch==1.9.0
|
||||
Using cached torch-1.9.0-cp38-cp38-manylinux1_x86_64.whl (831.4 MB)
|
||||
Collecting typing-extensions
|
||||
Using cached typing_extensions-3.10.0.0-py3-none-any.whl (26 kB)
|
||||
Installing collected packages: typing-extensions, torch, graphviz, k2
|
||||
Successfully installed graphviz-0.17 k2-1.4.dev20210822+cpu.torch1.9.0 torch-1.9.0 typing-extensions-3.10.0.0
|
||||
|
||||
.. WARNING::
|
||||
|
||||
We choose to install a CPU version of k2 for testing. You would probably want to install
|
||||
a CUDA version of k2.
|
||||
|
||||
|
||||
(4) Install lhotse
|
||||
~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block::
|
||||
|
||||
$ pip install git+https://github.com/lhotse-speech/lhotse
|
||||
|
||||
Collecting git+https://github.com/lhotse-speech/lhotse
|
||||
Cloning https://github.com/lhotse-speech/lhotse to /tmp/pip-req-build-7b1b76ge
|
||||
Running command git clone -q https://github.com/lhotse-speech/lhotse /tmp/pip-req-build-7b1b76ge
|
||||
Collecting audioread>=2.1.9
|
||||
Using cached audioread-2.1.9-py3-none-any.whl
|
||||
Collecting SoundFile>=0.10
|
||||
Using cached SoundFile-0.10.3.post1-py2.py3-none-any.whl (21 kB)
|
||||
Collecting click>=7.1.1
|
||||
Using cached click-8.0.1-py3-none-any.whl (97 kB)
|
||||
Collecting cytoolz>=0.10.1
|
||||
Using cached cytoolz-0.11.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (1.9 MB)
|
||||
Collecting dataclasses
|
||||
Using cached dataclasses-0.6-py3-none-any.whl (14 kB)
|
||||
Collecting h5py>=2.10.0
|
||||
Downloading h5py-3.4.0-cp38-cp38-manylinux_2_12_x86_64.manylinux2010_x86_64.whl (4.5 MB)
|
||||
|________________________________| 4.5 MB 684 kB/s
|
||||
Collecting intervaltree>=3.1.0
|
||||
Using cached intervaltree-3.1.0-py2.py3-none-any.whl
|
||||
Collecting lilcom>=1.1.0
|
||||
Using cached lilcom-1.1.1-cp38-cp38-linux_x86_64.whl
|
||||
Collecting numpy>=1.18.1
|
||||
Using cached numpy-1.21.2-cp38-cp38-manylinux_2_12_x86_64.manylinux2010_x86_64.whl (15.8 MB)
|
||||
Collecting packaging
|
||||
Using cached packaging-21.0-py3-none-any.whl (40 kB)
|
||||
Collecting pyyaml>=5.3.1
|
||||
Using cached PyYAML-5.4.1-cp38-cp38-manylinux1_x86_64.whl (662 kB)
|
||||
Collecting tqdm
|
||||
Downloading tqdm-4.62.1-py2.py3-none-any.whl (76 kB)
|
||||
|________________________________| 76 kB 2.7 MB/s
|
||||
Collecting torchaudio==0.9.0
|
||||
Downloading torchaudio-0.9.0-cp38-cp38-manylinux1_x86_64.whl (1.9 MB)
|
||||
|________________________________| 1.9 MB 73.1 MB/s
|
||||
Requirement already satisfied: torch==1.9.0 in ./test-icefall/lib/python3.8/site-packages (from torchaudio==0.9.0->lhotse===0.8.0.dev
|
||||
-2a1410b-clean) (1.9.0)
|
||||
Requirement already satisfied: typing-extensions in ./test-icefall/lib/python3.8/site-packages (from torch==1.9.0->torchaudio==0.9.0-
|
||||
>lhotse===0.8.0.dev-2a1410b-clean) (3.10.0.0)
|
||||
Collecting toolz>=0.8.0
|
||||
Using cached toolz-0.11.1-py3-none-any.whl (55 kB)
|
||||
Collecting sortedcontainers<3.0,>=2.0
|
||||
Using cached sortedcontainers-2.4.0-py2.py3-none-any.whl (29 kB)
|
||||
Collecting cffi>=1.0
|
||||
Using cached cffi-1.14.6-cp38-cp38-manylinux1_x86_64.whl (411 kB)
|
||||
Collecting pycparser
|
||||
Using cached pycparser-2.20-py2.py3-none-any.whl (112 kB)
|
||||
Collecting pyparsing>=2.0.2
|
||||
Using cached pyparsing-2.4.7-py2.py3-none-any.whl (67 kB)
|
||||
Building wheels for collected packages: lhotse
|
||||
Building wheel for lhotse (setup.py) ... done
|
||||
Created wheel for lhotse: filename=lhotse-0.8.0.dev_2a1410b_clean-py3-none-any.whl size=342242 sha256=f683444afa4dc0881133206b4646a
|
||||
9d0f774224cc84000f55d0a67f6e4a37997
|
||||
Stored in directory: /tmp/pip-ephem-wheel-cache-ftu0qysz/wheels/7f/7a/8e/a0bf241336e2e3cb573e1e21e5600952d49f5162454f2e612f
|
||||
WARNING: Built wheel for lhotse is invalid: Metadata 1.2 mandates PEP 440 version, but '0.8.0.dev-2a1410b-clean' is not
|
||||
Failed to build lhotse
|
||||
Installing collected packages: pycparser, toolz, sortedcontainers, pyparsing, numpy, cffi, tqdm, torchaudio, SoundFile, pyyaml, packa
|
||||
ging, lilcom, intervaltree, h5py, dataclasses, cytoolz, click, audioread, lhotse
|
||||
Running setup.py install for lhotse ... done
|
||||
DEPRECATION: lhotse was installed using the legacy 'setup.py install' method, because a wheel could not be built for it. A possible
|
||||
replacement is to fix the wheel build issue reported above. You can find discussion regarding this at https://github.com/pypa/pip/is
|
||||
sues/8368.
|
||||
Successfully installed SoundFile-0.10.3.post1 audioread-2.1.9 cffi-1.14.6 click-8.0.1 cytoolz-0.11.0 dataclasses-0.6 h5py-3.4.0 inter
|
||||
valtree-3.1.0 lhotse-0.8.0.dev-2a1410b-clean lilcom-1.1.1 numpy-1.21.2 packaging-21.0 pycparser-2.20 pyparsing-2.4.7 pyyaml-5.4.1 sor
|
||||
tedcontainers-2.4.0 toolz-0.11.1 torchaudio-0.9.0 tqdm-4.62.1
|
||||
|
||||
(5) Download icefall
|
||||
~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block::
|
||||
|
||||
$ cd /tmp
|
||||
$ git clone https://github.com/k2-fsa/icefall
|
||||
|
||||
Cloning into 'icefall'...
|
||||
remote: Enumerating objects: 500, done.
|
||||
remote: Counting objects: 100% (500/500), done.
|
||||
remote: Compressing objects: 100% (308/308), done.
|
||||
remote: Total 500 (delta 263), reused 307 (delta 102), pack-reused 0
|
||||
Receiving objects: 100% (500/500), 172.49 KiB | 385.00 KiB/s, done.
|
||||
Resolving deltas: 100% (263/263), done.
|
||||
|
||||
$ cd icefall
|
||||
$ pip install -r requirements.txt
|
||||
|
||||
Collecting kaldilm
|
||||
Downloading kaldilm-1.8.tar.gz (48 kB)
|
||||
|________________________________| 48 kB 574 kB/s
|
||||
Collecting kaldialign
|
||||
Using cached kaldialign-0.2-cp38-cp38-linux_x86_64.whl
|
||||
Collecting sentencepiece>=0.1.96
|
||||
Using cached sentencepiece-0.1.96-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (1.2 MB)
|
||||
Collecting tensorboard
|
||||
Using cached tensorboard-2.6.0-py3-none-any.whl (5.6 MB)
|
||||
Requirement already satisfied: setuptools>=41.0.0 in /ceph-fj/fangjun/test-icefall/lib/python3.8/site-packages (from tensorboard->-r
|
||||
requirements.txt (line 4)) (57.4.0)
|
||||
Collecting absl-py>=0.4
|
||||
Using cached absl_py-0.13.0-py3-none-any.whl (132 kB)
|
||||
Collecting google-auth-oauthlib<0.5,>=0.4.1
|
||||
Using cached google_auth_oauthlib-0.4.5-py2.py3-none-any.whl (18 kB)
|
||||
Collecting grpcio>=1.24.3
|
||||
Using cached grpcio-1.39.0-cp38-cp38-manylinux2014_x86_64.whl (4.3 MB)
|
||||
Requirement already satisfied: wheel>=0.26 in /ceph-fj/fangjun/test-icefall/lib/python3.8/site-packages (from tensorboard->-r require
|
||||
ments.txt (line 4)) (0.36.2)
|
||||
Requirement already satisfied: numpy>=1.12.0 in /ceph-fj/fangjun/test-icefall/lib/python3.8/site-packages (from tensorboard->-r requi
|
||||
rements.txt (line 4)) (1.21.2)
|
||||
Collecting protobuf>=3.6.0
|
||||
Using cached protobuf-3.17.3-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.whl (1.0 MB)
|
||||
Collecting werkzeug>=0.11.15
|
||||
Using cached Werkzeug-2.0.1-py3-none-any.whl (288 kB)
|
||||
Collecting tensorboard-data-server<0.7.0,>=0.6.0
|
||||
Using cached tensorboard_data_server-0.6.1-py3-none-manylinux2010_x86_64.whl (4.9 MB)
|
||||
Collecting google-auth<2,>=1.6.3
|
||||
Downloading google_auth-1.35.0-py2.py3-none-any.whl (152 kB)
|
||||
|________________________________| 152 kB 1.4 MB/s
|
||||
Collecting requests<3,>=2.21.0
|
||||
Using cached requests-2.26.0-py2.py3-none-any.whl (62 kB)
|
||||
Collecting tensorboard-plugin-wit>=1.6.0
|
||||
Using cached tensorboard_plugin_wit-1.8.0-py3-none-any.whl (781 kB)
|
||||
Collecting markdown>=2.6.8
|
||||
Using cached Markdown-3.3.4-py3-none-any.whl (97 kB)
|
||||
Collecting six
|
||||
Using cached six-1.16.0-py2.py3-none-any.whl (11 kB)
|
||||
Collecting cachetools<5.0,>=2.0.0
|
||||
Using cached cachetools-4.2.2-py3-none-any.whl (11 kB)
|
||||
Collecting rsa<5,>=3.1.4
|
||||
Using cached rsa-4.7.2-py3-none-any.whl (34 kB)
|
||||
Collecting pyasn1-modules>=0.2.1
|
||||
Using cached pyasn1_modules-0.2.8-py2.py3-none-any.whl (155 kB)
|
||||
Collecting requests-oauthlib>=0.7.0
|
||||
Using cached requests_oauthlib-1.3.0-py2.py3-none-any.whl (23 kB)
|
||||
Collecting pyasn1<0.5.0,>=0.4.6
|
||||
Using cached pyasn1-0.4.8-py2.py3-none-any.whl (77 kB)
|
||||
Collecting urllib3<1.27,>=1.21.1
|
||||
Using cached urllib3-1.26.6-py2.py3-none-any.whl (138 kB)
|
||||
Collecting certifi>=2017.4.17
|
||||
Using cached certifi-2021.5.30-py2.py3-none-any.whl (145 kB)
|
||||
Collecting charset-normalizer~=2.0.0
|
||||
Using cached charset_normalizer-2.0.4-py3-none-any.whl (36 kB)
|
||||
Collecting idna<4,>=2.5
|
||||
Using cached idna-3.2-py3-none-any.whl (59 kB)
|
||||
Collecting oauthlib>=3.0.0
|
||||
Using cached oauthlib-3.1.1-py2.py3-none-any.whl (146 kB)
|
||||
Building wheels for collected packages: kaldilm
|
||||
Building wheel for kaldilm (setup.py) ... done
|
||||
Created wheel for kaldilm: filename=kaldilm-1.8-cp38-cp38-linux_x86_64.whl size=897233 sha256=eccb906cafcd45bf9a7e1a1718e4534254bfb
|
||||
f4c0d0cbc66eee6c88d68a63862
|
||||
Stored in directory: /root/fangjun/.cache/pip/wheels/85/7d/63/f2dd586369b8797cb36d213bf3a84a789eeb92db93d2e723c9
|
||||
Successfully built kaldilm
|
||||
Installing collected packages: urllib3, pyasn1, idna, charset-normalizer, certifi, six, rsa, requests, pyasn1-modules, oauthlib, cach
|
||||
etools, requests-oauthlib, google-auth, werkzeug, tensorboard-plugin-wit, tensorboard-data-server, protobuf, markdown, grpcio, google
|
||||
-auth-oauthlib, absl-py, tensorboard, sentencepiece, kaldilm, kaldialign
|
||||
Successfully installed absl-py-0.13.0 cachetools-4.2.2 certifi-2021.5.30 charset-normalizer-2.0.4 google-auth-1.35.0 google-auth-oaut
|
||||
hlib-0.4.5 grpcio-1.39.0 idna-3.2 kaldialign-0.2 kaldilm-1.8 markdown-3.3.4 oauthlib-3.1.1 protobuf-3.17.3 pyasn1-0.4.8 pyasn1-module
|
||||
s-0.2.8 requests-2.26.0 requests-oauthlib-1.3.0 rsa-4.7.2 sentencepiece-0.1.96 six-1.16.0 tensorboard-2.6.0 tensorboard-data-server-0
|
||||
.6.1 tensorboard-plugin-wit-1.8.0 urllib3-1.26.6 werkzeug-2.0.1
|
||||
|
||||
|
||||
Test Your Installation
|
||||
----------------------
|
||||
|
||||
To test that your installation is successful, let us run
|
||||
the `yesno recipe <https://github.com/k2-fsa/icefall/tree/master/egs/yesno/ASR>`_
|
||||
on CPU.
|
||||
|
||||
Data preparation
|
||||
~~~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ export PYTHONPATH=/tmp/icefall:$PYTHONPATH
|
||||
$ cd /tmp/icefall
|
||||
$ cd egs/yesno/ASR
|
||||
$ ./prepare.sh
|
||||
|
||||
The log of running ``./prepare.sh`` is:
|
||||
|
||||
.. code-block::
|
||||
|
||||
2021-08-23 19:27:26 (prepare.sh:24:main) dl_dir: /tmp/icefall/egs/yesno/ASR/download
|
||||
2021-08-23 19:27:26 (prepare.sh:27:main) stage 0: Download data
|
||||
Downloading waves_yesno.tar.gz: 4.49MB [00:03, 1.39MB/s]
|
||||
2021-08-23 19:27:30 (prepare.sh:36:main) Stage 1: Prepare yesno manifest
|
||||
2021-08-23 19:27:31 (prepare.sh:42:main) Stage 2: Compute fbank for yesno
|
||||
2021-08-23 19:27:32,803 INFO [compute_fbank_yesno.py:52] Processing train
|
||||
Extracting and storing features: 100%|_______________________________________________________________| 90/90 [00:01<00:00, 80.57it/s]
|
||||
2021-08-23 19:27:34,085 INFO [compute_fbank_yesno.py:52] Processing test
|
||||
Extracting and storing features: 100%|______________________________________________________________| 30/30 [00:00<00:00, 248.21it/s]
|
||||
2021-08-23 19:27:34 (prepare.sh:48:main) Stage 3: Prepare lang
|
||||
2021-08-23 19:27:35 (prepare.sh:63:main) Stage 4: Prepare G
|
||||
/tmp/pip-install-fcordre9/kaldilm_6899d26f2d684ad48f21025950cd2866/kaldilm/csrc/arpa_file_parser.cc:void kaldilm::ArpaFileParser::Rea
|
||||
d(std::istream&):79
|
||||
[I] Reading \data\ section.
|
||||
/tmp/pip-install-fcordre9/kaldilm_6899d26f2d684ad48f21025950cd2866/kaldilm/csrc/arpa_file_parser.cc:void kaldilm::ArpaFileParser::Rea
|
||||
d(std::istream&):140
|
||||
[I] Reading \1-grams: section.
|
||||
2021-08-23 19:27:35 (prepare.sh:89:main) Stage 5: Compile HLG
|
||||
2021-08-23 19:27:35,928 INFO [compile_hlg.py:120] Processing data/lang_phone
|
||||
2021-08-23 19:27:35,929 INFO [lexicon.py:116] Converting L.pt to Linv.pt
|
||||
2021-08-23 19:27:35,931 INFO [compile_hlg.py:48] Building ctc_topo. max_token_id: 3
|
||||
2021-08-23 19:27:35,932 INFO [compile_hlg.py:52] Loading G.fst.txt
|
||||
2021-08-23 19:27:35,932 INFO [compile_hlg.py:62] Intersecting L and G
|
||||
2021-08-23 19:27:35,933 INFO [compile_hlg.py:64] LG shape: (4, None)
|
||||
2021-08-23 19:27:35,933 INFO [compile_hlg.py:66] Connecting LG
|
||||
2021-08-23 19:27:35,933 INFO [compile_hlg.py:68] LG shape after k2.connect: (4, None)
|
||||
2021-08-23 19:27:35,933 INFO [compile_hlg.py:70] <class 'torch.Tensor'>
|
||||
2021-08-23 19:27:35,933 INFO [compile_hlg.py:71] Determinizing LG
|
||||
2021-08-23 19:27:35,934 INFO [compile_hlg.py:74] <class '_k2.RaggedInt'>
|
||||
2021-08-23 19:27:35,934 INFO [compile_hlg.py:76] Connecting LG after k2.determinize
|
||||
2021-08-23 19:27:35,934 INFO [compile_hlg.py:79] Removing disambiguation symbols on LG
|
||||
2021-08-23 19:27:35,934 INFO [compile_hlg.py:87] LG shape after k2.remove_epsilon: (6, None)
|
||||
2021-08-23 19:27:35,935 INFO [compile_hlg.py:92] Arc sorting LG
|
||||
2021-08-23 19:27:35,935 INFO [compile_hlg.py:95] Composing H and LG
|
||||
2021-08-23 19:27:35,935 INFO [compile_hlg.py:102] Connecting LG
|
||||
2021-08-23 19:27:35,935 INFO [compile_hlg.py:105] Arc sorting LG
|
||||
2021-08-23 19:27:35,936 INFO [compile_hlg.py:107] HLG.shape: (8, None)
|
||||
2021-08-23 19:27:35,936 INFO [compile_hlg.py:123] Saving HLG.pt to data/lang_phone
|
||||
|
||||
|
||||
Training
|
||||
~~~~~~~~
|
||||
|
||||
Now let us run the training part:
|
||||
|
||||
.. code-block::
|
||||
|
||||
$ export CUDA_VISIBLE_DEVICES=""
|
||||
$ ./tdnn/train.py
|
||||
|
||||
.. CAUTION::
|
||||
|
||||
We use ``export CUDA_VISIBLE_DEVICES=""`` so that ``icefall`` uses CPU
|
||||
even if there are GPUs available.
|
||||
|
||||
The training log is given below:
|
||||
|
||||
.. code-block::
|
||||
|
||||
2021-08-23 19:30:31,072 INFO [train.py:465] Training started
|
||||
2021-08-23 19:30:31,072 INFO [train.py:466] {'exp_dir': PosixPath('tdnn/exp'), 'lang_dir': PosixPath('data/lang_phone'), 'lr': 0.01,
|
||||
'feature_dim': 23, 'weight_decay': 1e-06, 'start_epoch': 0, 'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, '
|
||||
best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 10, 'valid_interval': 10, 'beam_size': 10, 'reduction': 'sum', 'use_doub
|
||||
le_scores': True, 'world_size': 1, 'master_port': 12354, 'tensorboard': True, 'num_epochs': 15, 'feature_dir': PosixPath('data/fbank'
|
||||
), 'max_duration': 30.0, 'bucketing_sampler': False, 'num_buckets': 10, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0
|
||||
, 'on_the_fly_feats': False, 'shuffle': True, 'return_cuts': True, 'num_workers': 2}
|
||||
2021-08-23 19:30:31,074 INFO [lexicon.py:113] Loading pre-compiled data/lang_phone/Linv.pt
|
||||
2021-08-23 19:30:31,098 INFO [asr_datamodule.py:146] About to get train cuts
|
||||
2021-08-23 19:30:31,098 INFO [asr_datamodule.py:240] About to get train cuts
|
||||
2021-08-23 19:30:31,102 INFO [asr_datamodule.py:149] About to create train dataset
|
||||
2021-08-23 19:30:31,102 INFO [asr_datamodule.py:200] Using SingleCutSampler.
|
||||
2021-08-23 19:30:31,102 INFO [asr_datamodule.py:206] About to create train dataloader
|
||||
2021-08-23 19:30:31,102 INFO [asr_datamodule.py:219] About to get test cuts
|
||||
2021-08-23 19:30:31,102 INFO [asr_datamodule.py:246] About to get test cuts
|
||||
2021-08-23 19:30:31,357 INFO [train.py:416] Epoch 0, batch 0, batch avg loss 1.0789, total avg loss: 1.0789, batch size: 4
|
||||
2021-08-23 19:30:31,848 INFO [train.py:416] Epoch 0, batch 10, batch avg loss 0.5356, total avg loss: 0.7556, batch size: 4
|
||||
2021-08-23 19:30:32,301 INFO [train.py:432] Epoch 0, valid loss 0.9972, best valid loss: 0.9972 best valid epoch: 0
|
||||
2021-08-23 19:30:32,805 INFO [train.py:416] Epoch 0, batch 20, batch avg loss 0.2436, total avg loss: 0.5717, batch size: 3
|
||||
2021-08-23 19:30:33,109 INFO [train.py:432] Epoch 0, valid loss 0.4167, best valid loss: 0.4167 best valid epoch: 0
|
||||
2021-08-23 19:30:33,121 INFO [checkpoint.py:62] Saving checkpoint to tdnn/exp/epoch-0.pt
|
||||
2021-08-23 19:30:33,325 INFO [train.py:416] Epoch 1, batch 0, batch avg loss 0.2214, total avg loss: 0.2214, batch size: 5
|
||||
2021-08-23 19:30:33,798 INFO [train.py:416] Epoch 1, batch 10, batch avg loss 0.0781, total avg loss: 0.1343, batch size: 5
|
||||
2021-08-23 19:30:34,065 INFO [train.py:432] Epoch 1, valid loss 0.0859, best valid loss: 0.0859 best valid epoch: 1
|
||||
2021-08-23 19:30:34,556 INFO [train.py:416] Epoch 1, batch 20, batch avg loss 0.0421, total avg loss: 0.0975, batch size: 3
|
||||
2021-08-23 19:30:34,810 INFO [train.py:432] Epoch 1, valid loss 0.0431, best valid loss: 0.0431 best valid epoch: 1
|
||||
2021-08-23 19:30:34,824 INFO [checkpoint.py:62] Saving checkpoint to tdnn/exp/epoch-1.pt
|
||||
|
||||
... ...
|
||||
|
||||
2021-08-23 19:30:49,657 INFO [train.py:416] Epoch 13, batch 0, batch avg loss 0.0109, total avg loss: 0.0109, batch size: 5
|
||||
2021-08-23 19:30:49,984 INFO [train.py:416] Epoch 13, batch 10, batch avg loss 0.0093, total avg loss: 0.0096, batch size: 4
|
||||
2021-08-23 19:30:50,239 INFO [train.py:432] Epoch 13, valid loss 0.0104, best valid loss: 0.0101 best valid epoch: 12
|
||||
2021-08-23 19:30:50,569 INFO [train.py:416] Epoch 13, batch 20, batch avg loss 0.0092, total avg loss: 0.0096, batch size: 2
|
||||
2021-08-23 19:30:50,819 INFO [train.py:432] Epoch 13, valid loss 0.0101, best valid loss: 0.0101 best valid epoch: 13
|
||||
2021-08-23 19:30:50,835 INFO [checkpoint.py:62] Saving checkpoint to tdnn/exp/epoch-13.pt
|
||||
2021-08-23 19:30:51,024 INFO [train.py:416] Epoch 14, batch 0, batch avg loss 0.0105, total avg loss: 0.0105, batch size: 5
|
||||
2021-08-23 19:30:51,317 INFO [train.py:416] Epoch 14, batch 10, batch avg loss 0.0099, total avg loss: 0.0097, batch size: 4
|
||||
2021-08-23 19:30:51,552 INFO [train.py:432] Epoch 14, valid loss 0.0108, best valid loss: 0.0101 best valid epoch: 13
|
||||
2021-08-23 19:30:51,869 INFO [train.py:416] Epoch 14, batch 20, batch avg loss 0.0096, total avg loss: 0.0097, batch size: 5
|
||||
2021-08-23 19:30:52,107 INFO [train.py:432] Epoch 14, valid loss 0.0102, best valid loss: 0.0101 best valid epoch: 13
|
||||
2021-08-23 19:30:52,126 INFO [checkpoint.py:62] Saving checkpoint to tdnn/exp/epoch-14.pt
|
||||
2021-08-23 19:30:52,128 INFO [train.py:537] Done!
|
||||
|
||||
Decoding
|
||||
~~~~~~~~
|
||||
|
||||
Let us use the trained model to decode the test set:
|
||||
|
||||
.. code-block::
|
||||
|
||||
$ ./tdnn/decode.py
|
||||
|
||||
The decoding log is:
|
||||
|
||||
.. code-block::
|
||||
|
||||
2021-08-23 19:35:30,192 INFO [decode.py:249] Decoding started
|
||||
2021-08-23 19:35:30,192 INFO [decode.py:250] {'exp_dir': PosixPath('tdnn/exp'), 'lang_dir': PosixPath('data/lang_phone'), 'lm_dir': PosixPath('data/lm'), 'feature_dim': 23, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'epoch': 14, 'avg': 2, 'feature_dir': PosixPath('data/fbank'), 'max_duration': 30.0, 'bucketing_sampler': False, 'num_buckets': 10, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': True, 'return_cuts': True, 'num_workers': 2}
|
||||
2021-08-23 19:35:30,193 INFO [lexicon.py:113] Loading pre-compiled data/lang_phone/Linv.pt
|
||||
2021-08-23 19:35:30,213 INFO [decode.py:259] device: cpu
|
||||
2021-08-23 19:35:30,217 INFO [decode.py:279] averaging ['tdnn/exp/epoch-13.pt', 'tdnn/exp/epoch-14.pt']
|
||||
/tmp/icefall/icefall/checkpoint.py:146: UserWarning: floor_divide is deprecated, and will be removed in a future version of pytorch.
|
||||
It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values.
|
||||
To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor'). (Triggered internally at /pytorch/aten/src/ATen/native/BinaryOps.cpp:450.)
|
||||
avg[k] //= n
|
||||
2021-08-23 19:35:30,220 INFO [asr_datamodule.py:219] About to get test cuts
|
||||
2021-08-23 19:35:30,220 INFO [asr_datamodule.py:246] About to get test cuts
|
||||
2021-08-23 19:35:30,409 INFO [decode.py:190] batch 0/8, cuts processed until now is 4
|
||||
2021-08-23 19:35:30,571 INFO [decode.py:228] The transcripts are stored in tdnn/exp/recogs-test_set.txt
|
||||
2021-08-23 19:35:30,572 INFO [utils.py:317] [test_set] %WER 0.42% [1 / 240, 0 ins, 1 del, 0 sub ]
|
||||
2021-08-23 19:35:30,573 INFO [decode.py:236] Wrote detailed error stats to tdnn/exp/errs-test_set.txt
|
||||
2021-08-23 19:35:30,573 INFO [decode.py:299] Done!
|
||||
|
||||
**Congratulations!** You have successfully setup the environment and have run the first recipe in ``icefall``.
|
||||
|
||||
Have fun with ``icefall``!
|
||||
BIN
docs/source/recipes/images/yesno-tdnn-tensorboard-log.png
Normal file
BIN
docs/source/recipes/images/yesno-tdnn-tensorboard-log.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 121 KiB |
17
docs/source/recipes/index.rst
Normal file
17
docs/source/recipes/index.rst
Normal file
@ -0,0 +1,17 @@
|
||||
Recipes
|
||||
=======
|
||||
|
||||
This page contains various recipes in ``icefall``.
|
||||
Currently, only speech recognition recipes are provided.
|
||||
|
||||
We may add recipes for other tasks as well in the future.
|
||||
|
||||
.. we put the yesno recipe as the first recipe since it is the simplest one.
|
||||
.. Other recipes are listed in a alphabetical order.
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 2
|
||||
|
||||
yesno
|
||||
|
||||
librispeech
|
||||
10
docs/source/recipes/librispeech.rst
Normal file
10
docs/source/recipes/librispeech.rst
Normal file
@ -0,0 +1,10 @@
|
||||
LibriSpeech
|
||||
===========
|
||||
|
||||
We provide the following models for the LibriSpeech dataset:
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 2
|
||||
|
||||
librispeech/tdnn_lstm_ctc
|
||||
librispeech/conformer_ctc
|
||||
631
docs/source/recipes/librispeech/conformer_ctc.rst
Normal file
631
docs/source/recipes/librispeech/conformer_ctc.rst
Normal file
@ -0,0 +1,631 @@
|
||||
Confromer CTC
|
||||
=============
|
||||
|
||||
This tutorial shows you how to run a conformer ctc model
|
||||
with the `LibriSpeech <https://www.openslr.org/12>`_ 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 n-gram LM rescoring and attention decoder rescoring
|
||||
- (4) How to use a pre-trained model, provided by us
|
||||
|
||||
Data preparation
|
||||
----------------
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/ASR
|
||||
$ ./prepare.sh
|
||||
|
||||
The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
|
||||
All you need to do is to run it.
|
||||
|
||||
The data preparation contains several stages, you can use the following two
|
||||
options:
|
||||
|
||||
- ``--stage``
|
||||
- ``--stop-stage``
|
||||
|
||||
to control which stage(s) should be run. By default, all stages are executed.
|
||||
|
||||
|
||||
For example,
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/ASR
|
||||
$ ./prepare.sh --stage 0 --stop-stage 0
|
||||
|
||||
means to run only stage 0.
|
||||
|
||||
To run stage 2 to stage 5, use:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ ./prepare.sh --stage 2 --stop-stage 5
|
||||
|
||||
.. HINT::
|
||||
|
||||
If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
|
||||
dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
|
||||
they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
|
||||
the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
|
||||
``./prepare.sh`` won't re-download them.
|
||||
|
||||
.. NOTE::
|
||||
|
||||
All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
|
||||
are saved in ``./data`` directory.
|
||||
|
||||
|
||||
Training
|
||||
--------
|
||||
|
||||
Configurable options
|
||||
~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/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:
|
||||
|
||||
- ``--full-libri``
|
||||
|
||||
If it's True, the training part uses all the training data, i.e.,
|
||||
960 hours. Otherwise, the training part uses only the subset
|
||||
``train-clean-100``, which has 100 hours of training data.
|
||||
|
||||
.. CAUTION::
|
||||
|
||||
The training set is perturbed by speed with two factors: 0.9 and 1.1.
|
||||
If ``--full-libri`` is True, each epoch actually processes
|
||||
``3x960 == 2880`` hours of data.
|
||||
|
||||
- ``--num-epochs``
|
||||
|
||||
It is the number of epochs to train. For instance,
|
||||
``./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/librispeech/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/librispeech/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/librispeech/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/librispeech/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.
|
||||
|
||||
|
||||
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 LibriSpeech with icefall"
|
||||
|
||||
It will print something like below:
|
||||
|
||||
.. code-block::
|
||||
|
||||
TensorFlow installation not found - running with reduced feature set.
|
||||
Upload started and will continue reading any new data as it's added to the logdir.
|
||||
|
||||
To stop uploading, press Ctrl-C.
|
||||
|
||||
New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/lzGnETjwRxC3yghNMd4kPw/
|
||||
|
||||
[2021-08-24T16:42:43] Started scanning logdir.
|
||||
Uploading 4540 scalars...
|
||||
|
||||
Note there is a URL in the above output, click it and you will see
|
||||
the following screenshot:
|
||||
|
||||
.. figure:: images/librispeech-conformer-ctc-tensorboard-log.png
|
||||
:width: 600
|
||||
:alt: TensorBoard screenshot
|
||||
:align: center
|
||||
:target: https://tensorboard.dev/experiment/lzGnETjwRxC3yghNMd4kPw/
|
||||
|
||||
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/librispeech/ASR
|
||||
$ ./conformer_ctc/train.py --max-duration 200 --full-libri 0
|
||||
|
||||
It uses ``--max-duration`` of 200 to avoid OOM. Also, it uses only
|
||||
a subset of the LibriSpeech data for training.
|
||||
|
||||
|
||||
**Case 2**
|
||||
^^^^^^^^^^
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/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/librispeech/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/librispeech/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/librispeech/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_librispeech_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/librispeech/ASR
|
||||
$ mkdir tmp
|
||||
$ cd tmp
|
||||
$ git lfs install
|
||||
$ git clone https://huggingface.co/pkufool/icefall_asr_librispeech_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/librispeech/ASR
|
||||
$ tree tmp
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
tmp
|
||||
`-- icefall_asr_librispeech_conformer_ctc
|
||||
|-- README.md
|
||||
|-- data
|
||||
| |-- lang_bpe
|
||||
| | |-- HLG.pt
|
||||
| | |-- bpe.model
|
||||
| | |-- tokens.txt
|
||||
| | `-- words.txt
|
||||
| `-- lm
|
||||
| `-- G_4_gram.pt
|
||||
|-- exp
|
||||
| `-- pretrained.pt
|
||||
`-- test_wavs
|
||||
|-- 1089-134686-0001.flac
|
||||
|-- 1221-135766-0001.flac
|
||||
|-- 1221-135766-0002.flac
|
||||
`-- trans.txt
|
||||
|
||||
6 directories, 11 files
|
||||
|
||||
**File descriptions**:
|
||||
|
||||
- ``data/lang_bpe/HLG.pt``
|
||||
|
||||
It is the decoding graph.
|
||||
|
||||
- ``data/lang_bpe/bpe.model``
|
||||
|
||||
It is a sentencepiece model. You can use it to reproduce our results.
|
||||
|
||||
- ``data/lang_bpe/tokens.txt``
|
||||
|
||||
It contains tokens and their IDs, generated from ``bpe.model``.
|
||||
Provided only for convenience so that you can look up the SOS/EOS ID easily.
|
||||
|
||||
- ``data/lang_bpe/words.txt``
|
||||
|
||||
It contains words and their IDs.
|
||||
|
||||
- ``data/lm/G_4_gram.pt``
|
||||
|
||||
It is a 4-gram LM, used for n-gram LM rescoring.
|
||||
|
||||
- ``exp/pretrained.pt``
|
||||
|
||||
It contains pre-trained model parameters, obtained by averaging
|
||||
checkpoints from ``epoch-15.pt`` to ``epoch-34.pt``.
|
||||
Note: We have removed optimizer ``state_dict`` to reduce file size.
|
||||
|
||||
- ``test_waves/*.flac``
|
||||
|
||||
It contains some test sound files from LibriSpeech ``test-clean`` 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_librispeech_conformer_ctc/test_wavs/*.flac
|
||||
|
||||
Input File : 'tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1089-134686-0001.flac'
|
||||
Channels : 1
|
||||
Sample Rate : 16000
|
||||
Precision : 16-bit
|
||||
Duration : 00:00:06.62 = 106000 samples ~ 496.875 CDDA sectors
|
||||
File Size : 116k
|
||||
Bit Rate : 140k
|
||||
Sample Encoding: 16-bit FLAC
|
||||
|
||||
Input File : 'tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0001.flac'
|
||||
Channels : 1
|
||||
Sample Rate : 16000
|
||||
Precision : 16-bit
|
||||
Duration : 00:00:16.71 = 267440 samples ~ 1253.62 CDDA sectors
|
||||
File Size : 343k
|
||||
Bit Rate : 164k
|
||||
Sample Encoding: 16-bit FLAC
|
||||
|
||||
Input File : 'tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0002.flac'
|
||||
Channels : 1
|
||||
Sample Rate : 16000
|
||||
Precision : 16-bit
|
||||
Duration : 00:00:04.83 = 77200 samples ~ 361.875 CDDA sectors
|
||||
File Size : 105k
|
||||
Bit Rate : 174k
|
||||
Sample Encoding: 16-bit FLAC
|
||||
|
||||
Total Duration of 3 files: 00:00:28.16
|
||||
|
||||
Usage
|
||||
~~~~~
|
||||
|
||||
.. code-block::
|
||||
|
||||
$ cd egs/librispeech/ASR
|
||||
$ ./conformer_ctc/pretrained.py --help
|
||||
|
||||
displays the help information.
|
||||
|
||||
It supports three decoding methods:
|
||||
|
||||
- HLG decoding
|
||||
- HLG + n-gram LM rescoring
|
||||
- HLG + n-gram LM rescoring + 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/librispeech/ASR
|
||||
$ ./conformer_ctc/pretrained.py \
|
||||
--checkpoint ./tmp/icefall_asr_librispeech_conformer_ctc/exp/pretrained.pt \
|
||||
--words-file ./tmp/icefall_asr_librispeech_conformer_ctc/data/lang_bpe/words.txt \
|
||||
--HLG ./tmp/icefall_asr_librispeech_conformer_ctc/data/lang_bpe/HLG.pt \
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1089-134686-0001.flac \
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0001.flac \
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0002.flac
|
||||
|
||||
The output is given below:
|
||||
|
||||
.. code-block::
|
||||
|
||||
2021-08-20 11:03:05,712 INFO [pretrained.py:217] device: cuda:0
|
||||
2021-08-20 11:03:05,712 INFO [pretrained.py:219] Creating model
|
||||
2021-08-20 11:03:11,345 INFO [pretrained.py:238] Loading HLG from ./tmp/icefall_asr_librispeech_conformer_ctc/data/lang_bpe/HLG.pt
|
||||
2021-08-20 11:03:18,442 INFO [pretrained.py:255] Constructing Fbank computer
|
||||
2021-08-20 11:03:18,444 INFO [pretrained.py:265] Reading sound files: ['./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1089-134686-0001.flac', './tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0001.flac', './tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0002.flac']
|
||||
2021-08-20 11:03:18,507 INFO [pretrained.py:271] Decoding started
|
||||
2021-08-20 11:03:18,795 INFO [pretrained.py:300] Use HLG decoding
|
||||
2021-08-20 11:03:19,149 INFO [pretrained.py:339]
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1089-134686-0001.flac:
|
||||
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
|
||||
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0001.flac:
|
||||
GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONOURED
|
||||
BOSOM TO CONNECT HER PARENT FOR EVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
|
||||
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0002.flac:
|
||||
YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
|
||||
|
||||
2021-08-20 11:03:19,149 INFO [pretrained.py:341] Decoding Done
|
||||
|
||||
HLG decoding + LM rescoring
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
It uses an n-gram LM to rescore the decoding lattice and the best
|
||||
path of the rescored lattice is the decoding result.
|
||||
|
||||
The command to run HLG decoding + LM rescoring is:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/ASR
|
||||
$ ./conformer_ctc/pretrained.py \
|
||||
--checkpoint ./tmp/icefall_asr_librispeech_conformer_ctc/exp/pretrained.pt \
|
||||
--words-file ./tmp/icefall_asr_librispeech_conformer_ctc/data/lang_bpe/words.txt \
|
||||
--HLG ./tmp/icefall_asr_librispeech_conformer_ctc/data/lang_bpe/HLG.pt \
|
||||
--method whole-lattice-rescoring \
|
||||
--G ./tmp/icefall_asr_librispeech_conformer_ctc/data/lm/G_4_gram.pt \
|
||||
--ngram-lm-scale 0.8 \
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1089-134686-0001.flac \
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0001.flac \
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0002.flac
|
||||
|
||||
Its output is:
|
||||
|
||||
.. code-block::
|
||||
|
||||
2021-08-20 11:12:17,565 INFO [pretrained.py:217] device: cuda:0
|
||||
2021-08-20 11:12:17,565 INFO [pretrained.py:219] Creating model
|
||||
2021-08-20 11:12:23,728 INFO [pretrained.py:238] Loading HLG from ./tmp/icefall_asr_librispeech_conformer_ctc/data/lang_bpe/HLG.pt
|
||||
2021-08-20 11:12:30,035 INFO [pretrained.py:246] Loading G from ./tmp/icefall_asr_librispeech_conformer_ctc/data/lm/G_4_gram.pt
|
||||
2021-08-20 11:13:10,779 INFO [pretrained.py:255] Constructing Fbank computer
|
||||
2021-08-20 11:13:10,787 INFO [pretrained.py:265] Reading sound files: ['./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1089-134686-0001.flac', './tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0001.flac', './tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0002.flac']
|
||||
2021-08-20 11:13:10,798 INFO [pretrained.py:271] Decoding started
|
||||
2021-08-20 11:13:11,085 INFO [pretrained.py:305] Use HLG decoding + LM rescoring
|
||||
2021-08-20 11:13:11,736 INFO [pretrained.py:339]
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1089-134686-0001.flac:
|
||||
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
|
||||
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0001.flac:
|
||||
GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONOURED
|
||||
BOSOM TO CONNECT HER PARENT FOR EVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
|
||||
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0002.flac:
|
||||
YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
|
||||
|
||||
2021-08-20 11:13:11,737 INFO [pretrained.py:341] Decoding Done
|
||||
|
||||
HLG decoding + LM rescoring + attention decoder rescoring
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
It uses an n-gram LM to rescore the decoding lattice, extracts
|
||||
n paths from the rescored 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 + LM rescoring + attention decoder rescoring is:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/ASR
|
||||
$ ./conformer_ctc/pretrained.py \
|
||||
--checkpoint ./tmp/icefall_asr_librispeech_conformer_ctc/exp/pretrained.pt \
|
||||
--words-file ./tmp/icefall_asr_librispeech_conformer_ctc/data/lang_bpe/words.txt \
|
||||
--HLG ./tmp/icefall_asr_librispeech_conformer_ctc/data/lang_bpe/HLG.pt \
|
||||
--method attention-decoder \
|
||||
--G ./tmp/icefall_asr_librispeech_conformer_ctc/data/lm/G_4_gram.pt \
|
||||
--ngram-lm-scale 1.3 \
|
||||
--attention-decoder-scale 1.2 \
|
||||
--lattice-score-scale 0.5 \
|
||||
--num-paths 100 \
|
||||
--sos-id 1 \
|
||||
--eos-id 1 \
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1089-134686-0001.flac \
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0001.flac \
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0002.flac
|
||||
|
||||
The output is below:
|
||||
|
||||
.. code-block::
|
||||
|
||||
2021-08-20 11:19:11,397 INFO [pretrained.py:217] device: cuda:0
|
||||
2021-08-20 11:19:11,397 INFO [pretrained.py:219] Creating model
|
||||
2021-08-20 11:19:17,354 INFO [pretrained.py:238] Loading HLG from ./tmp/icefall_asr_librispeech_conformer_ctc/data/lang_bpe/HLG.pt
|
||||
2021-08-20 11:19:24,615 INFO [pretrained.py:246] Loading G from ./tmp/icefall_asr_librispeech_conformer_ctc/data/lm/G_4_gram.pt
|
||||
2021-08-20 11:20:04,576 INFO [pretrained.py:255] Constructing Fbank computer
|
||||
2021-08-20 11:20:04,584 INFO [pretrained.py:265] Reading sound files: ['./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1089-134686-0001.flac', './tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0001.flac', './tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0002.flac']
|
||||
2021-08-20 11:20:04,595 INFO [pretrained.py:271] Decoding started
|
||||
2021-08-20 11:20:04,854 INFO [pretrained.py:313] Use HLG + LM rescoring + attention decoder rescoring
|
||||
2021-08-20 11:20:05,805 INFO [pretrained.py:339]
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1089-134686-0001.flac:
|
||||
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
|
||||
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0001.flac:
|
||||
GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONOURED
|
||||
BOSOM TO CONNECT HER PARENT FOR EVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
|
||||
|
||||
./tmp/icefall_asr_librispeech_conformer_ctc/test_wavs/1221-135766-0002.flac:
|
||||
YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
|
||||
|
||||
2021-08-20 11:20:05,805 INFO [pretrained.py:341] Decoding Done
|
||||
|
||||
Colab notebook
|
||||
--------------
|
||||
|
||||
We do provide a colab notebook for this recipe showing how to use a pre-trained model.
|
||||
|
||||
|librispeech asr conformer ctc colab notebook|
|
||||
|
||||
.. |librispeech asr conformer ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
|
||||
:target: https://colab.research.google.com/drive/1huyupXAcHsUrKaWfI83iMEJ6J0Nh0213?usp=sharing
|
||||
|
||||
.. HINT::
|
||||
|
||||
Due to limited memory provided by Colab, you have to upgrade to Colab Pro to
|
||||
run ``HLG decoding + LM rescoring`` and
|
||||
``HLG decoding + LM rescoring + attention decoder rescoring``.
|
||||
Otherwise, you can only run ``HLG decoding`` with Colab.
|
||||
|
||||
**Congratulations!** You have finished the librispeech ASR recipe with
|
||||
conformer CTC models in ``icefall``.
|
||||
Binary file not shown.
|
After Width: | Height: | Size: 422 KiB |
394
docs/source/recipes/librispeech/tdnn_lstm_ctc.rst
Normal file
394
docs/source/recipes/librispeech/tdnn_lstm_ctc.rst
Normal file
@ -0,0 +1,394 @@
|
||||
TDNN-LSTM-CTC
|
||||
=============
|
||||
|
||||
This tutorial shows you how to run a TDNN-LSTM-CTC model with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
|
||||
|
||||
|
||||
.. HINT::
|
||||
|
||||
We assume you have read the page :ref:`install icefall` and have setup
|
||||
the environment for ``icefall``.
|
||||
|
||||
|
||||
Data preparation
|
||||
----------------
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/ASR
|
||||
$ ./prepare.sh
|
||||
|
||||
The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
|
||||
All you need to do is to run it.
|
||||
|
||||
The data preparation contains several stages, you can use the following two
|
||||
options:
|
||||
|
||||
- ``--stage``
|
||||
- ``--stop-stage``
|
||||
|
||||
to control which stage(s) should be run. By default, all stages are executed.
|
||||
|
||||
|
||||
For example,
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/ASR
|
||||
$ ./prepare.sh --stage 0 --stop-stage 0
|
||||
|
||||
means to run only stage 0.
|
||||
|
||||
To run stage 2 to stage 5, use:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ ./prepare.sh --stage 2 --stop-stage 5
|
||||
|
||||
|
||||
Training
|
||||
--------
|
||||
|
||||
Now describing the training of TDNN-LSTM-CTC model, contained in
|
||||
the `tdnn_lstm_ctc <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/tdnn_lstm_ctc>`_
|
||||
folder.
|
||||
|
||||
The command to run the training part is:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/ASR
|
||||
$ export CUDA_VISIBLE_DEVICES="0,1,2,3"
|
||||
$ ./tdnn_lstm_ctc/train.py --world-size 4
|
||||
|
||||
By default, it will run ``20`` epochs. Training logs and checkpoints are saved
|
||||
in ``tdnn_lstm_ctc/exp``.
|
||||
|
||||
In ``tdnn_lstm_ctc/exp``, you will find the following files:
|
||||
|
||||
- ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-19.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 training for librispeech with icefall"
|
||||
|
||||
- ``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.
|
||||
|
||||
|
||||
To see available training options, you can use:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ ./tdnn_lstm_ctc/train.py --help
|
||||
|
||||
Other training options, e.g., learning rate, results dir, etc., are
|
||||
pre-configured in the function ``get_params()``
|
||||
in `tdnn_lstm_ctc/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/tdnn_lstm_ctc/train.py>`_.
|
||||
Normally, you don't need to change them. You can change them by modifying the code, if
|
||||
you want.
|
||||
|
||||
Decoding
|
||||
--------
|
||||
|
||||
The decoding part uses checkpoints saved by the training part, so you have
|
||||
to run the training part first.
|
||||
|
||||
The command for decoding is:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ export CUDA_VISIBLE_DEVICES="0"
|
||||
$ ./tdnn_lstm_ctc/decode.py
|
||||
|
||||
You will see the WER in the output log.
|
||||
|
||||
Decoded results are saved in ``tdnn_lstm_ctc/exp``.
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ ./tdnn_lstm_ctc/decode.py --help
|
||||
|
||||
shows you the available decoding options.
|
||||
|
||||
Some commonly used options are:
|
||||
|
||||
- ``--epoch``
|
||||
|
||||
You can select which checkpoint to be used for decoding.
|
||||
For instance, ``./tdnn_lstm_ctc/decode.py --epoch 10`` means to use
|
||||
``./tdnn_lstm_ctc/exp/epoch-10.pt`` for decoding.
|
||||
|
||||
- ``--avg``
|
||||
|
||||
It's related to model averaging. It specifies number of checkpoints
|
||||
to be averaged. The averaged model is used for decoding.
|
||||
For example, the following command:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ ./tdnn_lstm_ctc/decode.py --epoch 10 --avg 3
|
||||
|
||||
uses the average of ``epoch-8.pt``, ``epoch-9.pt`` and ``epoch-10.pt``
|
||||
for decoding.
|
||||
|
||||
- ``--export``
|
||||
|
||||
If it is ``True``, i.e., ``./tdnn_lstm_ctc/decode.py --export 1``, the code
|
||||
will save the averaged model to ``tdnn_lstm_ctc/exp/pretrained.pt``.
|
||||
See :ref:`tdnn_lstm_ctc use a pre-trained model` for how to use it.
|
||||
|
||||
|
||||
.. _tdnn_lstm_ctc use a pre-trained model:
|
||||
|
||||
Pre-trained Model
|
||||
-----------------
|
||||
|
||||
We have uploaded the pre-trained model to
|
||||
`<https://huggingface.co/pkufool/icefall_asr_librispeech_tdnn-lstm_ctc>`_.
|
||||
|
||||
The following shows you how to use the pre-trained model.
|
||||
|
||||
|
||||
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
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/ASR
|
||||
$ mkdir tmp
|
||||
$ cd tmp
|
||||
$ git lfs install
|
||||
$ git clone https://huggingface.co/pkufool/icefall_asr_librispeech_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/librispeech/ASR
|
||||
$ tree tmp
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
tmp/
|
||||
`-- icefall_asr_librispeech_tdnn-lstm_ctc
|
||||
|-- README.md
|
||||
|-- data
|
||||
| |-- lang_phone
|
||||
| | |-- HLG.pt
|
||||
| | |-- tokens.txt
|
||||
| | `-- words.txt
|
||||
| `-- lm
|
||||
| `-- G_4_gram.pt
|
||||
|-- exp
|
||||
| `-- pretrained.pt
|
||||
`-- test_wavs
|
||||
|-- 1089-134686-0001.flac
|
||||
|-- 1221-135766-0001.flac
|
||||
|-- 1221-135766-0002.flac
|
||||
`-- trans.txt
|
||||
|
||||
6 directories, 10 files
|
||||
|
||||
**File descriptions**:
|
||||
|
||||
- ``data/lang_phone/HLG.pt``
|
||||
|
||||
It is the decoding graph.
|
||||
|
||||
- ``data/lang_phone/tokens.txt``
|
||||
|
||||
It contains tokens and their IDs.
|
||||
|
||||
- ``data/lang_phone/words.txt``
|
||||
|
||||
It contains words and their IDs.
|
||||
|
||||
- ``data/lm/G_4_gram.pt``
|
||||
|
||||
It is a 4-gram LM, useful for LM rescoring.
|
||||
|
||||
- ``exp/pretrained.pt``
|
||||
|
||||
It contains pre-trained model parameters, obtained by averaging
|
||||
checkpoints from ``epoch-14.pt`` to ``epoch-19.pt``.
|
||||
Note: We have removed optimizer ``state_dict`` to reduce file size.
|
||||
|
||||
- ``test_waves/*.flac``
|
||||
|
||||
It contains some test sound files from LibriSpeech ``test-clean`` 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_librispeech_tdnn-lstm_ctc/test_wavs/*.flac
|
||||
|
||||
Input File : 'tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac'
|
||||
Channels : 1
|
||||
Sample Rate : 16000
|
||||
Precision : 16-bit
|
||||
Duration : 00:00:06.62 = 106000 samples ~ 496.875 CDDA sectors
|
||||
File Size : 116k
|
||||
Bit Rate : 140k
|
||||
Sample Encoding: 16-bit FLAC
|
||||
|
||||
|
||||
Input File : 'tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac'
|
||||
Channels : 1
|
||||
Sample Rate : 16000
|
||||
Precision : 16-bit
|
||||
Duration : 00:00:16.71 = 267440 samples ~ 1253.62 CDDA sectors
|
||||
File Size : 343k
|
||||
Bit Rate : 164k
|
||||
Sample Encoding: 16-bit FLAC
|
||||
|
||||
|
||||
Input File : 'tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac'
|
||||
Channels : 1
|
||||
Sample Rate : 16000
|
||||
Precision : 16-bit
|
||||
Duration : 00:00:04.83 = 77200 samples ~ 361.875 CDDA sectors
|
||||
File Size : 105k
|
||||
Bit Rate : 174k
|
||||
Sample Encoding: 16-bit FLAC
|
||||
|
||||
Total Duration of 3 files: 00:00:28.16
|
||||
|
||||
|
||||
Inference with a pre-trained model
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/librispeech/ASR
|
||||
$ ./tdnn_lstm_ctc/pretrained.py --help
|
||||
|
||||
shows the usage information of ``./tdnn_lstm_ctc/pretrained.py``.
|
||||
|
||||
To decode with ``1best`` method, we can use:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
./tdnn_lstm_ctc/pretrained.py \
|
||||
--checkpoint ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/exp/pretraind.pt \
|
||||
--words-file ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/words.txt \
|
||||
--HLG ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/HLG.pt \
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac \
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac \
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac
|
||||
|
||||
The output is:
|
||||
|
||||
.. code-block::
|
||||
|
||||
2021-08-24 16:57:13,315 INFO [pretrained.py:168] device: cuda:0
|
||||
2021-08-24 16:57:13,315 INFO [pretrained.py:170] Creating model
|
||||
2021-08-24 16:57:18,331 INFO [pretrained.py:182] Loading HLG from ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/HLG.pt
|
||||
2021-08-24 16:57:27,581 INFO [pretrained.py:199] Constructing Fbank computer
|
||||
2021-08-24 16:57:27,584 INFO [pretrained.py:209] Reading sound files: ['./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac', './tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac', './tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac']
|
||||
2021-08-24 16:57:27,599 INFO [pretrained.py:215] Decoding started
|
||||
2021-08-24 16:57:27,791 INFO [pretrained.py:245] Use HLG decoding
|
||||
2021-08-24 16:57:28,098 INFO [pretrained.py:266]
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac:
|
||||
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
|
||||
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac:
|
||||
GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
|
||||
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac:
|
||||
YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
|
||||
|
||||
|
||||
2021-08-24 16:57:28,099 INFO [pretrained.py:268] Decoding Done
|
||||
|
||||
|
||||
To decode with ``whole-lattice-rescoring`` methond, you can use
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
./tdnn_lstm_ctc/pretrained.py \
|
||||
--checkpoint ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/exp/pretraind.pt \
|
||||
--words-file ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/words.txt \
|
||||
--HLG ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/HLG.pt \
|
||||
--method whole-lattice-rescoring \
|
||||
--G ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lm/G_4_gram.pt \
|
||||
--ngram-lm-scale 0.8 \
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac \
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac \
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac
|
||||
|
||||
The decoding output is:
|
||||
|
||||
.. code-block::
|
||||
|
||||
2021-08-24 16:39:24,725 INFO [pretrained.py:168] device: cuda:0
|
||||
2021-08-24 16:39:24,725 INFO [pretrained.py:170] Creating model
|
||||
2021-08-24 16:39:29,403 INFO [pretrained.py:182] Loading HLG from ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lang_phone/HLG.pt
|
||||
2021-08-24 16:39:40,631 INFO [pretrained.py:190] Loading G from ./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/data/lm/G_4_gram.pt
|
||||
2021-08-24 16:39:53,098 INFO [pretrained.py:199] Constructing Fbank computer
|
||||
2021-08-24 16:39:53,107 INFO [pretrained.py:209] Reading sound files: ['./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac', './tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac', './tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac']
|
||||
2021-08-24 16:39:53,121 INFO [pretrained.py:215] Decoding started
|
||||
2021-08-24 16:39:53,443 INFO [pretrained.py:250] Use HLG decoding + LM rescoring
|
||||
2021-08-24 16:39:54,010 INFO [pretrained.py:266]
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1089-134686-0001.flac:
|
||||
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
|
||||
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0001.flac:
|
||||
GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONORED BOSOM TO CONNECT HER PARENT FOREVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
|
||||
|
||||
./tmp/icefall_asr_librispeech_tdnn-lstm_ctc/test_wavs/1221-135766-0002.flac:
|
||||
YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
|
||||
|
||||
|
||||
2021-08-24 16:39:54,010 INFO [pretrained.py:268] Decoding Done
|
||||
|
||||
|
||||
Colab notebook
|
||||
--------------
|
||||
|
||||
We provide a colab notebook for decoding with pre-trained model.
|
||||
|
||||
|librispeech tdnn_lstm_ctc colab notebook|
|
||||
|
||||
.. |librispeech tdnn_lstm_ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
|
||||
:target: https://colab.research.google.com/drive/1kNmDXNMwREi0rZGAOIAOJo93REBuOTcd
|
||||
|
||||
|
||||
**Congratulations!** You have finished the TDNN-LSTM-CTC recipe on librispeech in ``icefall``.
|
||||
445
docs/source/recipes/yesno.rst
Normal file
445
docs/source/recipes/yesno.rst
Normal file
@ -0,0 +1,445 @@
|
||||
yesno
|
||||
=====
|
||||
|
||||
This page shows you how to run the `yesno <https://www.openslr.org/1>`_ recipe. It contains:
|
||||
|
||||
- (1) Prepare data for training
|
||||
- (2) Train a TDNN model
|
||||
|
||||
- (a) View text format logs and visualize TensorBoard logs
|
||||
- (b) Select device type, i.e., CPU and GPU, for training
|
||||
- (c) Change training options
|
||||
- (d) Resume training from a checkpoint
|
||||
|
||||
- (3) Decode with a trained model
|
||||
|
||||
- (a) Select a checkpoint for decoding
|
||||
- (b) Model averaging
|
||||
|
||||
- (4) Colab notebook
|
||||
|
||||
- (a) It shows you step by step how to setup the environment, how to do training,
|
||||
and how to do decoding
|
||||
- (b) How to use a pre-trained model
|
||||
|
||||
- (5) Inference with a pre-trained model
|
||||
|
||||
- (a) Download a pre-trained model, provided by us
|
||||
- (b) Decode a single sound file with a pre-trained model
|
||||
- (c) Decode multiple sound files at the same time
|
||||
|
||||
It does **NOT** show you:
|
||||
|
||||
- (1) How to train with multiple GPUs
|
||||
|
||||
The ``yesno`` dataset is so small that CPU is more than enough
|
||||
for training as well as for decoding.
|
||||
|
||||
- (2) How to use LM rescoring for decoding
|
||||
|
||||
The dataset does not have an LM for rescoring.
|
||||
|
||||
.. HINT::
|
||||
|
||||
We assume you have read the page :ref:`install icefall` and have setup
|
||||
the environment for ``icefall``.
|
||||
|
||||
.. HINT::
|
||||
|
||||
You **don't** need a **GPU** to run this recipe. It can be run on a **CPU**.
|
||||
The training part takes less than 30 **seconds** on a CPU and you will get
|
||||
the following WER at the end::
|
||||
|
||||
[test_set] %WER 0.42% [1 / 240, 0 ins, 1 del, 0 sub ]
|
||||
|
||||
Data preparation
|
||||
----------------
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/yesno/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/yesno/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
|
||||
|
||||
|
||||
Training
|
||||
--------
|
||||
|
||||
We provide only a TDNN model, contained in
|
||||
the `tdnn <https://github.com/k2-fsa/icefall/tree/master/egs/yesno/ASR/tdnn>`_
|
||||
folder, for ``yesno``.
|
||||
|
||||
The command to run the training part is:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/yesno/ASR
|
||||
$ export CUDA_VISIBLE_DEVICES=""
|
||||
$ ./tdnn/train.py
|
||||
|
||||
By default, it will run ``15`` epochs. Training logs and checkpoints are saved
|
||||
in ``tdnn/exp``.
|
||||
|
||||
In ``tdnn/exp``, you will find the following files:
|
||||
|
||||
- ``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/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/exp/tensorboard
|
||||
$ tensorboard dev upload --logdir . --description "TDNN training for yesno 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/yKUbhb5wRmOSXYkId1z9eg/
|
||||
|
||||
[2021-08-23T23:49:41] Started scanning logdir.
|
||||
[2021-08-23T23:49:42] Total uploaded: 135 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/yesno-tdnn-tensorboard-log.png
|
||||
:width: 600
|
||||
:alt: TensorBoard screenshot
|
||||
:align: center
|
||||
:target: https://tensorboard.dev/experiment/yKUbhb5wRmOSXYkId1z9eg/
|
||||
|
||||
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.
|
||||
|
||||
|
||||
|
||||
.. NOTE::
|
||||
|
||||
By default, ``./tdnn/train.py`` uses GPU 0 for training if GPUs are available.
|
||||
If you have two GPUs, say, GPU 0 and GPU 1, and you want to use GPU 1 for
|
||||
training, you can run:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ export CUDA_VISIBLE_DEVICES="1"
|
||||
$ ./tdnn/train.py
|
||||
|
||||
Since the ``yesno`` dataset is very small, containing only 30 sound files
|
||||
for training, and the model in use is also very small, we use:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ export CUDA_VISIBLE_DEVICES=""
|
||||
|
||||
so that ``./tdnn/train.py`` uses CPU during training.
|
||||
|
||||
If you don't have GPUs, then you don't need to
|
||||
run ``export CUDA_VISIBLE_DEVICES=""``.
|
||||
|
||||
To see available training options, you can use:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ ./tdnn/train.py --help
|
||||
|
||||
Other training options, e.g., learning rate, results dir, etc., are
|
||||
pre-configured in the function ``get_params()``
|
||||
in `tdnn/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/yesno/ASR/tdnn/train.py>`_.
|
||||
Normally, you don't need to change them. You can change them by modifying the code, if
|
||||
you want.
|
||||
|
||||
Decoding
|
||||
--------
|
||||
|
||||
The decoding part uses checkpoints saved by the training part, so you have
|
||||
to run the training part first.
|
||||
|
||||
The command for decoding is:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ export CUDA_VISIBLE_DEVICES=""
|
||||
$ ./tdnn/decode.py
|
||||
|
||||
You will see the WER in the output log.
|
||||
|
||||
Decoded results are saved in ``tdnn/exp``.
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ ./tdnn/decode.py --help
|
||||
|
||||
shows you the available decoding options.
|
||||
|
||||
Some commonly used options are:
|
||||
|
||||
- ``--epoch``
|
||||
|
||||
You can select which checkpoint to be used for decoding.
|
||||
For instance, ``./tdnn/decode.py --epoch 10`` means to use
|
||||
``./tdnn/exp/epoch-10.pt`` for decoding.
|
||||
|
||||
- ``--avg``
|
||||
|
||||
It's related to model averaging. It specifies number of checkpoints
|
||||
to be averaged. The averaged model is used for decoding.
|
||||
For example, the following command:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ ./tdnn/decode.py --epoch 10 --avg 3
|
||||
|
||||
uses the average of ``epoch-8.pt``, ``epoch-9.pt`` and ``epoch-10.pt``
|
||||
for decoding.
|
||||
|
||||
- ``--export``
|
||||
|
||||
If it is ``True``, i.e., ``./tdnn/decode.py --export 1``, the code
|
||||
will save the averaged model to ``tdnn/exp/pretrained.pt``.
|
||||
See :ref:`yesno use a pre-trained model` for how to use it.
|
||||
|
||||
|
||||
.. _yesno use a pre-trained model:
|
||||
|
||||
Pre-trained Model
|
||||
-----------------
|
||||
|
||||
We have uploaded the pre-trained model to
|
||||
`<https://huggingface.co/csukuangfj/icefall_asr_yesno_tdnn>`_.
|
||||
|
||||
The following shows you how to use the pre-trained model.
|
||||
|
||||
Download the pre-trained model
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/yesno/ASR
|
||||
$ mkdir tmp
|
||||
$ cd tmp
|
||||
$ git lfs install
|
||||
$ git clone https://huggingface.co/csukuangfj/icefall_asr_yesno_tdnn
|
||||
|
||||
.. CAUTION::
|
||||
|
||||
You have to use ``git lfs`` to download the pre-trained model.
|
||||
|
||||
After downloading, you will have the following files:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/yesno/ASR
|
||||
$ tree tmp
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
tmp/
|
||||
`-- icefall_asr_yesno_tdnn
|
||||
|-- README.md
|
||||
|-- lang_phone
|
||||
| |-- HLG.pt
|
||||
| |-- L.pt
|
||||
| |-- L_disambig.pt
|
||||
| |-- Linv.pt
|
||||
| |-- lexicon.txt
|
||||
| |-- lexicon_disambig.txt
|
||||
| |-- tokens.txt
|
||||
| `-- words.txt
|
||||
|-- lm
|
||||
| |-- G.arpa
|
||||
| `-- G.fst.txt
|
||||
|-- pretrained.pt
|
||||
`-- test_waves
|
||||
|-- 0_0_0_1_0_0_0_1.wav
|
||||
|-- 0_0_1_0_0_0_1_0.wav
|
||||
|-- 0_0_1_0_0_1_1_1.wav
|
||||
|-- 0_0_1_0_1_0_0_1.wav
|
||||
|-- 0_0_1_1_0_0_0_1.wav
|
||||
|-- 0_0_1_1_0_1_1_0.wav
|
||||
|-- 0_0_1_1_1_0_0_0.wav
|
||||
|-- 0_0_1_1_1_1_0_0.wav
|
||||
|-- 0_1_0_0_0_1_0_0.wav
|
||||
|-- 0_1_0_0_1_0_1_0.wav
|
||||
|-- 0_1_0_1_0_0_0_0.wav
|
||||
|-- 0_1_0_1_1_1_0_0.wav
|
||||
|-- 0_1_1_0_0_1_1_1.wav
|
||||
|-- 0_1_1_1_0_0_1_0.wav
|
||||
|-- 0_1_1_1_1_0_1_0.wav
|
||||
|-- 1_0_0_0_0_0_0_0.wav
|
||||
|-- 1_0_0_0_0_0_1_1.wav
|
||||
|-- 1_0_0_1_0_1_1_1.wav
|
||||
|-- 1_0_1_1_0_1_1_1.wav
|
||||
|-- 1_0_1_1_1_1_0_1.wav
|
||||
|-- 1_1_0_0_0_1_1_1.wav
|
||||
|-- 1_1_0_0_1_0_1_1.wav
|
||||
|-- 1_1_0_1_0_1_0_0.wav
|
||||
|-- 1_1_0_1_1_0_0_1.wav
|
||||
|-- 1_1_0_1_1_1_1_0.wav
|
||||
|-- 1_1_1_0_0_1_0_1.wav
|
||||
|-- 1_1_1_0_1_0_1_0.wav
|
||||
|-- 1_1_1_1_0_0_1_0.wav
|
||||
|-- 1_1_1_1_1_0_0_0.wav
|
||||
`-- 1_1_1_1_1_1_1_1.wav
|
||||
|
||||
4 directories, 42 files
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ soxi tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav
|
||||
|
||||
Input File : 'tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav'
|
||||
Channels : 1
|
||||
Sample Rate : 8000
|
||||
Precision : 16-bit
|
||||
Duration : 00:00:06.76 = 54080 samples ~ 507 CDDA sectors
|
||||
File Size : 108k
|
||||
Bit Rate : 128k
|
||||
Sample Encoding: 16-bit Signed Integer PCM
|
||||
|
||||
- ``0_0_1_0_1_0_0_1.wav``
|
||||
|
||||
0 means No; 1 means Yes. No and Yes are not in English,
|
||||
but in `Hebrew <https://en.wikipedia.org/wiki/Hebrew_language>`_.
|
||||
So this file contains ``NO NO YES NO YES NO NO YES``.
|
||||
|
||||
Download kaldifeat
|
||||
~~~~~~~~~~~~~~~~~~
|
||||
|
||||
`kaldifeat <https://github.com/csukuangfj/kaldifeat>`_ is used for extracting
|
||||
features from a single or multiple sound files. Please refer to
|
||||
`<https://github.com/csukuangfj/kaldifeat>`_ to install ``kaldifeat`` first.
|
||||
|
||||
Inference with a pre-trained model
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
$ cd egs/yesno/ASR
|
||||
$ ./tdnn/pretrained.py --help
|
||||
|
||||
shows the usage information of ``./tdnn/pretrained.py``.
|
||||
|
||||
To decode a single file, we can use:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
./tdnn/pretrained.py \
|
||||
--checkpoint ./tmp/icefall_asr_yesno_tdnn/pretrained.pt \
|
||||
--words-file ./tmp/icefall_asr_yesno_tdnn/lang_phone/words.txt \
|
||||
--HLG ./tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt \
|
||||
./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav
|
||||
|
||||
The output is:
|
||||
|
||||
.. code-block::
|
||||
|
||||
2021-08-24 12:22:51,621 INFO [pretrained.py:119] {'feature_dim': 23, 'num_classes': 4, 'sample_rate': 8000, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'checkpoint': './tmp/icefall_asr_yesno_tdnn/pretrained.pt', 'words_file': './tmp/icefall_asr_yesno_tdnn/lang_phone/words.txt', 'HLG': './tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt', 'sound_files': ['./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav']}
|
||||
2021-08-24 12:22:51,645 INFO [pretrained.py:125] device: cpu
|
||||
2021-08-24 12:22:51,645 INFO [pretrained.py:127] Creating model
|
||||
2021-08-24 12:22:51,650 INFO [pretrained.py:139] Loading HLG from ./tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt
|
||||
2021-08-24 12:22:51,651 INFO [pretrained.py:143] Constructing Fbank computer
|
||||
2021-08-24 12:22:51,652 INFO [pretrained.py:153] Reading sound files: ['./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav']
|
||||
2021-08-24 12:22:51,684 INFO [pretrained.py:159] Decoding started
|
||||
2021-08-24 12:22:51,708 INFO [pretrained.py:198]
|
||||
./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav:
|
||||
NO NO YES NO YES NO NO YES
|
||||
|
||||
|
||||
2021-08-24 12:22:51,708 INFO [pretrained.py:200] Decoding Done
|
||||
|
||||
You can see that for the sound file ``0_0_1_0_1_0_0_1.wav``, the decoding result is
|
||||
``NO NO YES NO YES NO NO YES``.
|
||||
|
||||
To decode **multiple** files at the same time, you can use
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
./tdnn/pretrained.py \
|
||||
--checkpoint ./tmp/icefall_asr_yesno_tdnn/pretrained.pt \
|
||||
--words-file ./tmp/icefall_asr_yesno_tdnn/lang_phone/words.txt \
|
||||
--HLG ./tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt \
|
||||
./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav \
|
||||
./tmp/icefall_asr_yesno_tdnn/test_waves/1_0_1_1_0_1_1_1.wav
|
||||
|
||||
The decoding output is:
|
||||
|
||||
.. code-block::
|
||||
|
||||
2021-08-24 12:25:20,159 INFO [pretrained.py:119] {'feature_dim': 23, 'num_classes': 4, 'sample_rate': 8000, 'search_beam': 20, 'output_beam': 8, 'min_active_states': 30, 'max_active_states': 10000, 'use_double_scores': True, 'checkpoint': './tmp/icefall_asr_yesno_tdnn/pretrained.pt', 'words_file': './tmp/icefall_asr_yesno_tdnn/lang_phone/words.txt', 'HLG': './tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt', 'sound_files': ['./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav', './tmp/icefall_asr_yesno_tdnn/test_waves/1_0_1_1_0_1_1_1.wav']}
|
||||
2021-08-24 12:25:20,181 INFO [pretrained.py:125] device: cpu
|
||||
2021-08-24 12:25:20,181 INFO [pretrained.py:127] Creating model
|
||||
2021-08-24 12:25:20,185 INFO [pretrained.py:139] Loading HLG from ./tmp/icefall_asr_yesno_tdnn/lang_phone/HLG.pt
|
||||
2021-08-24 12:25:20,186 INFO [pretrained.py:143] Constructing Fbank computer
|
||||
2021-08-24 12:25:20,187 INFO [pretrained.py:153] Reading sound files: ['./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav',
|
||||
'./tmp/icefall_asr_yesno_tdnn/test_waves/1_0_1_1_0_1_1_1.wav']
|
||||
2021-08-24 12:25:20,213 INFO [pretrained.py:159] Decoding started
|
||||
2021-08-24 12:25:20,287 INFO [pretrained.py:198]
|
||||
./tmp/icefall_asr_yesno_tdnn/test_waves/0_0_1_0_1_0_0_1.wav:
|
||||
NO NO YES NO YES NO NO YES
|
||||
|
||||
./tmp/icefall_asr_yesno_tdnn/test_waves/1_0_1_1_0_1_1_1.wav:
|
||||
YES NO YES YES NO YES YES YES
|
||||
|
||||
2021-08-24 12:25:20,287 INFO [pretrained.py:200] Decoding Done
|
||||
|
||||
You can see again that it decodes correctly.
|
||||
|
||||
Colab notebook
|
||||
--------------
|
||||
|
||||
We do provide a colab notebook for this recipe.
|
||||
|
||||
|yesno colab notebook|
|
||||
|
||||
.. |yesno colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
|
||||
:target: https://colab.research.google.com/drive/1tIjjzaJc3IvGyKiMCDWO-TSnBgkcuN3B?usp=sharing
|
||||
|
||||
|
||||
**Congratulations!** You have finished the simplest speech recognition recipe in ``icefall``.
|
||||
@ -1,64 +1,3 @@
|
||||
|
||||
## Data preparation
|
||||
|
||||
If you want to use `./prepare.sh` to download everything for you,
|
||||
you can just run
|
||||
|
||||
```
|
||||
./prepare.sh
|
||||
```
|
||||
|
||||
If you have pre-downloaded the LibriSpeech dataset, please
|
||||
read `./prepare.sh` and modify it to point to the location
|
||||
of your dataset so that it won't re-download it. After modification,
|
||||
please run
|
||||
|
||||
```
|
||||
./prepare.sh
|
||||
```
|
||||
|
||||
The script `./prepare.sh` prepares features, lexicon, LMs, etc.
|
||||
All generated files are saved in the folder `./data`.
|
||||
|
||||
**HINT:** `./prepare.sh` supports options `--stage` and `--stop-stage`.
|
||||
|
||||
## TDNN-LSTM CTC training
|
||||
|
||||
The folder `tdnn_lstm_ctc` contains scripts for CTC training
|
||||
with TDNN-LSTM models.
|
||||
|
||||
Pre-configured parameters for training and decoding are set in the function
|
||||
`get_params()` within `tdnn_lstm_ctc/train.py`
|
||||
and `tdnn_lstm_ctc/decode.py`.
|
||||
|
||||
Parameters that can be passed from the command-line can be found by
|
||||
|
||||
```
|
||||
./tdnn_lstm_ctc/train.py --help
|
||||
./tdnn_lstm_ctc/decode.py --help
|
||||
```
|
||||
|
||||
If you have 4 GPUs on a machine and want to use GPU 0, 2, 3 for
|
||||
mutli-GPU training, you can run
|
||||
|
||||
```
|
||||
export CUDA_VISIBLE_DEVICES="0,2,3"
|
||||
./tdnn_lstm_ctc/train.py \
|
||||
--master-port 12345 \
|
||||
--world-size 3
|
||||
```
|
||||
|
||||
If you want to decode by averaging checkpoints `epoch-8.pt`,
|
||||
`epoch-9.pt` and `epoch-10.pt`, you can run
|
||||
|
||||
```
|
||||
./tdnn_lstm_ctc/decode.py \
|
||||
--epoch 10 \
|
||||
--avg 3
|
||||
```
|
||||
|
||||
## Conformer CTC training
|
||||
|
||||
The folder `conformer-ctc` contains scripts for CTC training
|
||||
with conformer models. The steps of running the training and
|
||||
decoding are similar to `tdnn_lstm_ctc`.
|
||||
Please refer to <https://icefall.readthedocs.io/en/latest/recipes/librispeech.html>
|
||||
for how to run models in this recipe.
|
||||
|
||||
@ -6,7 +6,7 @@
|
||||
|
||||
TensorBoard log is available at https://tensorboard.dev/experiment/GnRzq8WWQW62dK4bklXBTg/#scalars
|
||||
|
||||
Pretrained model is available at https://huggingface.co/pkufool/conformer_ctc
|
||||
Pretrained model is available at https://huggingface.co/pkufool/icefall_asr_librispeech_conformer_ctc
|
||||
|
||||
The best decoding results (WER) are listed below, we got this results by averaging models from epoch 15 to 34, and using `attention-decoder` decoder with num_paths equals to 100.
|
||||
|
||||
@ -21,3 +21,51 @@ To get more unique paths, we scaled the lattice.scores with 0.5 (see https://git
|
||||
|test-clean|1.3|1.2|
|
||||
|test-other|1.2|1.1|
|
||||
|
||||
You can use the following commands to reproduce our results:
|
||||
|
||||
```bash
|
||||
git clone https://github.com/k2-fsa/icefall
|
||||
cd icefall
|
||||
|
||||
# It was using ef233486, you may not need to switch to it
|
||||
# git checkout ef233486
|
||||
|
||||
cd egs/librispeech/ASR
|
||||
./prepare.sh
|
||||
|
||||
export CUDA_VISIBLE_DEVICES="0,1,2,3"
|
||||
python conformer_ctc/train.py --bucketing-sampler True \
|
||||
--concatenate-cuts False \
|
||||
--max-duration 200 \
|
||||
--full-libri True \
|
||||
--world-size 4
|
||||
|
||||
python conformer_ctc/decode.py --lattice-score-scale 0.5 \
|
||||
--epoch 34 \
|
||||
--avg 20 \
|
||||
--method attention-decoder \
|
||||
--max-duration 20 \
|
||||
--num-paths 100
|
||||
```
|
||||
|
||||
### LibriSpeech training results (Tdnn-Lstm)
|
||||
#### 2021-08-24
|
||||
|
||||
(Wei Kang): Result of phone based Tdnn-Lstm model.
|
||||
|
||||
Icefall version: https://github.com/k2-fsa/icefall/commit/caa0b9e9425af27e0c6211048acb55a76ed5d315
|
||||
|
||||
Pretrained model is available at https://huggingface.co/pkufool/icefall_asr_librispeech_tdnn-lstm_ctc
|
||||
|
||||
The best decoding results (WER) are listed below, we got this results by averaging models from epoch 19 to 14, and using `whole-lattice-rescoring` decoding method.
|
||||
|
||||
||test-clean|test-other|
|
||||
|--|--|--|
|
||||
|WER| 6.59% | 17.69% |
|
||||
|
||||
We searched the lm_score_scale for best results, the scales that produced the WER above are also listed below.
|
||||
|
||||
||lm_scale|
|
||||
|--|--|
|
||||
|test-clean|0.8|
|
||||
|test-other|0.9|
|
||||
|
||||
@ -1,351 +1,3 @@
|
||||
|
||||
# How to use a pre-trained model to transcribe a sound file or multiple sound files
|
||||
|
||||
(See the bottom of this document for the link to a colab notebook.)
|
||||
|
||||
You need to prepare 4 files:
|
||||
|
||||
- a model checkpoint file, e.g., epoch-20.pt
|
||||
- HLG.pt, the decoding graph
|
||||
- words.txt, the word symbol table
|
||||
- a sound file, whose sampling rate has to be 16 kHz.
|
||||
Supported formats are those supported by `torchaudio.load()`,
|
||||
e.g., wav and flac.
|
||||
|
||||
Also, you need to install `kaldifeat`. Please refer to
|
||||
<https://github.com/csukuangfj/kaldifeat> for installation.
|
||||
|
||||
```bash
|
||||
./conformer_ctc/pretrained.py --help
|
||||
```
|
||||
|
||||
displays the help information.
|
||||
|
||||
## HLG decoding
|
||||
|
||||
Once you have the above files ready and have `kaldifeat` installed,
|
||||
you can run:
|
||||
|
||||
```bash
|
||||
./conformer_ctc/pretrained.py \
|
||||
--checkpoint /path/to/your/checkpoint.pt \
|
||||
--words-file /path/to/words.txt \
|
||||
--HLG /path/to/HLG.pt \
|
||||
/path/to/your/sound.wav
|
||||
```
|
||||
|
||||
and you will see the transcribed result.
|
||||
|
||||
If you want to transcribe multiple files at the same time, you can use:
|
||||
|
||||
```bash
|
||||
./conformer_ctc/pretrained.py \
|
||||
--checkpoint /path/to/your/checkpoint.pt \
|
||||
--words-file /path/to/words.txt \
|
||||
--HLG /path/to/HLG.pt \
|
||||
/path/to/your/sound1.wav \
|
||||
/path/to/your/sound2.wav \
|
||||
/path/to/your/sound3.wav
|
||||
```
|
||||
|
||||
**Note**: This is the fastest decoding method.
|
||||
|
||||
## HLG decoding + LM rescoring
|
||||
|
||||
`./conformer_ctc/pretrained.py` also supports `whole lattice LM rescoring`
|
||||
and `attention decoder rescoring`.
|
||||
|
||||
To use whole lattice LM rescoring, you also need the following files:
|
||||
|
||||
- G.pt, e.g., `data/lm/G_4_gram.pt` if you have run `./prepare.sh`
|
||||
|
||||
The command to run decoding with LM rescoring is:
|
||||
|
||||
```bash
|
||||
./conformer_ctc/pretrained.py \
|
||||
--checkpoint /path/to/your/checkpoint.pt \
|
||||
--words-file /path/to/words.txt \
|
||||
--HLG /path/to/HLG.pt \
|
||||
--method whole-lattice-rescoring \
|
||||
--G data/lm/G_4_gram.pt \
|
||||
--ngram-lm-scale 0.8 \
|
||||
/path/to/your/sound1.wav \
|
||||
/path/to/your/sound2.wav \
|
||||
/path/to/your/sound3.wav
|
||||
```
|
||||
|
||||
## HLG Decoding + LM rescoring + attention decoder rescoring
|
||||
|
||||
To use attention decoder for rescoring, you need the following extra information:
|
||||
|
||||
- sos token ID
|
||||
- eos token ID
|
||||
|
||||
The command to run decoding with attention decoder rescoring is:
|
||||
|
||||
```bash
|
||||
./conformer_ctc/pretrained.py \
|
||||
--checkpoint /path/to/your/checkpoint.pt \
|
||||
--words-file /path/to/words.txt \
|
||||
--HLG /path/to/HLG.pt \
|
||||
--method attention-decoder \
|
||||
--G data/lm/G_4_gram.pt \
|
||||
--ngram-lm-scale 1.3 \
|
||||
--attention-decoder-scale 1.2 \
|
||||
--lattice-score-scale 0.5 \
|
||||
--num-paths 100 \
|
||||
--sos-id 1 \
|
||||
--eos-id 1 \
|
||||
/path/to/your/sound1.wav \
|
||||
/path/to/your/sound2.wav \
|
||||
/path/to/your/sound3.wav
|
||||
```
|
||||
|
||||
# Decoding with a pre-trained model in action
|
||||
|
||||
We have uploaded a pre-trained model to <https://huggingface.co/pkufool/conformer_ctc>
|
||||
|
||||
The following shows the steps about the usage of the provided pre-trained model.
|
||||
|
||||
### (1) Download the pre-trained model
|
||||
|
||||
```bash
|
||||
sudo apt-get install git-lfs
|
||||
cd /path/to/icefall/egs/librispeech/ASR
|
||||
git lfs install
|
||||
mkdir tmp
|
||||
cd tmp
|
||||
git clone https://huggingface.co/pkufool/conformer_ctc
|
||||
```
|
||||
|
||||
**CAUTION**: You have to install `git-lfst` to download the pre-trained model.
|
||||
|
||||
You will find the following files:
|
||||
|
||||
```
|
||||
tmp
|
||||
`-- conformer_ctc
|
||||
|-- README.md
|
||||
|-- data
|
||||
| |-- lang_bpe
|
||||
| | |-- HLG.pt
|
||||
| | |-- bpe.model
|
||||
| | |-- tokens.txt
|
||||
| | `-- words.txt
|
||||
| `-- lm
|
||||
| `-- G_4_gram.pt
|
||||
|-- exp
|
||||
| `-- pretraind.pt
|
||||
`-- test_wavs
|
||||
|-- 1089-134686-0001.flac
|
||||
|-- 1221-135766-0001.flac
|
||||
|-- 1221-135766-0002.flac
|
||||
`-- trans.txt
|
||||
|
||||
6 directories, 11 files
|
||||
```
|
||||
|
||||
**File descriptions**:
|
||||
|
||||
- `data/lang_bpe/HLG.pt`
|
||||
|
||||
It is the decoding graph.
|
||||
|
||||
- `data/lang_bpe/bpe.model`
|
||||
|
||||
It is a sentencepiece model. You can use it to reproduce our results.
|
||||
|
||||
- `data/lang_bpe/tokens.txt`
|
||||
|
||||
It contains tokens and their IDs, generated from `bpe.model`.
|
||||
Provided only for convienice so that you can look up the SOS/EOS ID easily.
|
||||
|
||||
- `data/lang_bpe/words.txt`
|
||||
|
||||
It contains words and their IDs.
|
||||
|
||||
- `data/lm/G_4_gram.pt`
|
||||
|
||||
It is a 4-gram LM, useful for LM rescoring.
|
||||
|
||||
- `exp/pretrained.pt`
|
||||
|
||||
It contains pre-trained model parameters, obtained by averaging
|
||||
checkpoints from `epoch-15.pt` to `epoch-34.pt`.
|
||||
Note: We have removed optimizer `state_dict` to reduce file size.
|
||||
|
||||
- `test_waves/*.flac`
|
||||
|
||||
It contains some test sound files from LibriSpeech `test-clean` 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:
|
||||
|
||||
```
|
||||
$ soxi tmp/conformer_ctc/test_wavs/*.flac
|
||||
|
||||
Input File : 'tmp/conformer_ctc/test_wavs/1089-134686-0001.flac'
|
||||
Channels : 1
|
||||
Sample Rate : 16000
|
||||
Precision : 16-bit
|
||||
Duration : 00:00:06.62 = 106000 samples ~ 496.875 CDDA sectors
|
||||
File Size : 116k
|
||||
Bit Rate : 140k
|
||||
Sample Encoding: 16-bit FLAC
|
||||
|
||||
Input File : 'tmp/conformer_ctc/test_wavs/1221-135766-0001.flac'
|
||||
Channels : 1
|
||||
Sample Rate : 16000
|
||||
Precision : 16-bit
|
||||
Duration : 00:00:16.71 = 267440 samples ~ 1253.62 CDDA sectors
|
||||
File Size : 343k
|
||||
Bit Rate : 164k
|
||||
Sample Encoding: 16-bit FLAC
|
||||
|
||||
Input File : 'tmp/conformer_ctc/test_wavs/1221-135766-0002.flac'
|
||||
Channels : 1
|
||||
Sample Rate : 16000
|
||||
Precision : 16-bit
|
||||
Duration : 00:00:04.83 = 77200 samples ~ 361.875 CDDA sectors
|
||||
File Size : 105k
|
||||
Bit Rate : 174k
|
||||
Sample Encoding: 16-bit FLAC
|
||||
|
||||
Total Duration of 3 files: 00:00:28.16
|
||||
```
|
||||
|
||||
### (2) Use HLG decoding
|
||||
|
||||
```bash
|
||||
cd /path/to/icefall/egs/librispeech/ASR
|
||||
|
||||
./conformer_ctc/pretrained.py \
|
||||
--checkpoint ./tmp/conformer_ctc/exp/pretraind.pt \
|
||||
--words-file ./tmp/conformer_ctc/data/lang_bpe/words.txt \
|
||||
--HLG ./tmp/conformer_ctc/data/lang_bpe/HLG.pt \
|
||||
./tmp/conformer_ctc/test_wavs/1089-134686-0001.flac \
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0001.flac \
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0002.flac
|
||||
```
|
||||
|
||||
The output is given below:
|
||||
|
||||
```
|
||||
2021-08-20 11:03:05,712 INFO [pretrained.py:217] device: cuda:0
|
||||
2021-08-20 11:03:05,712 INFO [pretrained.py:219] Creating model
|
||||
2021-08-20 11:03:11,345 INFO [pretrained.py:238] Loading HLG from ./tmp/conformer_ctc/data/lang_bpe/HLG.pt
|
||||
2021-08-20 11:03:18,442 INFO [pretrained.py:255] Constructing Fbank computer
|
||||
2021-08-20 11:03:18,444 INFO [pretrained.py:265] Reading sound files: ['./tmp/conformer_ctc/test_wavs/1089-134686-0001.flac', './tmp/conformer_ctc/test_wavs/1221-135766-0001.flac', './tmp/conformer_ctc/test_wavs/1221-135766-0002.flac']
|
||||
2021-08-20 11:03:18,507 INFO [pretrained.py:271] Decoding started
|
||||
2021-08-20 11:03:18,795 INFO [pretrained.py:300] Use HLG decoding
|
||||
2021-08-20 11:03:19,149 INFO [pretrained.py:339]
|
||||
./tmp/conformer_ctc/test_wavs/1089-134686-0001.flac:
|
||||
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
|
||||
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0001.flac:
|
||||
GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONOURED
|
||||
BOSOM TO CONNECT HER PARENT FOR EVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
|
||||
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0002.flac:
|
||||
YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
|
||||
|
||||
|
||||
2021-08-20 11:03:19,149 INFO [pretrained.py:341] Decoding Done
|
||||
```
|
||||
|
||||
### (3) Use HLG decoding + LM rescoring
|
||||
|
||||
```bash
|
||||
./conformer_ctc/pretrained.py \
|
||||
--checkpoint ./tmp/conformer_ctc/exp/pretraind.pt \
|
||||
--words-file ./tmp/conformer_ctc/data/lang_bpe/words.txt \
|
||||
--HLG ./tmp/conformer_ctc/data/lang_bpe/HLG.pt \
|
||||
--method whole-lattice-rescoring \
|
||||
--G ./tmp/conformer_ctc/data/lm/G_4_gram.pt \
|
||||
--ngram-lm-scale 0.8 \
|
||||
./tmp/conformer_ctc/test_wavs/1089-134686-0001.flac \
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0001.flac \
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0002.flac
|
||||
```
|
||||
|
||||
The output is:
|
||||
|
||||
```
|
||||
2021-08-20 11:12:17,565 INFO [pretrained.py:217] device: cuda:0
|
||||
2021-08-20 11:12:17,565 INFO [pretrained.py:219] Creating model
|
||||
2021-08-20 11:12:23,728 INFO [pretrained.py:238] Loading HLG from ./tmp/conformer_ctc/data/lang_bpe/HLG.pt
|
||||
2021-08-20 11:12:30,035 INFO [pretrained.py:246] Loading G from ./tmp/conformer_ctc/data/lm/G_4_gram.pt
|
||||
2021-08-20 11:13:10,779 INFO [pretrained.py:255] Constructing Fbank computer
|
||||
2021-08-20 11:13:10,787 INFO [pretrained.py:265] Reading sound files: ['./tmp/conformer_ctc/test_wavs/1089-134686-0001.flac', './tmp/conformer_ctc/test_wavs/1221-135766-0001.flac', './tmp/conformer_ctc/test_wavs/1221-135766-0002.flac']
|
||||
2021-08-20 11:13:10,798 INFO [pretrained.py:271] Decoding started
|
||||
2021-08-20 11:13:11,085 INFO [pretrained.py:305] Use HLG decoding + LM rescoring
|
||||
2021-08-20 11:13:11,736 INFO [pretrained.py:339]
|
||||
./tmp/conformer_ctc/test_wavs/1089-134686-0001.flac:
|
||||
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
|
||||
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0001.flac:
|
||||
GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONOURED
|
||||
BOSOM TO CONNECT HER PARENT FOR EVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
|
||||
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0002.flac:
|
||||
YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
|
||||
|
||||
|
||||
2021-08-20 11:13:11,737 INFO [pretrained.py:341] Decoding Done
|
||||
```
|
||||
|
||||
### (4) Use HLG decoding + LM rescoring + attention decoder rescoring
|
||||
|
||||
```bash
|
||||
./conformer_ctc/pretrained.py \
|
||||
--checkpoint ./tmp/conformer_ctc/exp/pretraind.pt \
|
||||
--words-file ./tmp/conformer_ctc/data/lang_bpe/words.txt \
|
||||
--HLG ./tmp/conformer_ctc/data/lang_bpe/HLG.pt \
|
||||
--method attention-decoder \
|
||||
--G ./tmp/conformer_ctc/data/lm/G_4_gram.pt \
|
||||
--ngram-lm-scale 1.3 \
|
||||
--attention-decoder-scale 1.2 \
|
||||
--lattice-score-scale 0.5 \
|
||||
--num-paths 100 \
|
||||
--sos-id 1 \
|
||||
--eos-id 1 \
|
||||
./tmp/conformer_ctc/test_wavs/1089-134686-0001.flac \
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0001.flac \
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0002.flac
|
||||
```
|
||||
|
||||
The output is:
|
||||
|
||||
```
|
||||
2021-08-20 11:19:11,397 INFO [pretrained.py:217] device: cuda:0
|
||||
2021-08-20 11:19:11,397 INFO [pretrained.py:219] Creating model
|
||||
2021-08-20 11:19:17,354 INFO [pretrained.py:238] Loading HLG from ./tmp/conformer_ctc/data/lang_bpe/HLG.pt
|
||||
2021-08-20 11:19:24,615 INFO [pretrained.py:246] Loading G from ./tmp/conformer_ctc/data/lm/G_4_gram.pt
|
||||
2021-08-20 11:20:04,576 INFO [pretrained.py:255] Constructing Fbank computer
|
||||
2021-08-20 11:20:04,584 INFO [pretrained.py:265] Reading sound files: ['./tmp/conformer_ctc/test_wavs/1089-134686-0001.flac', './tmp/conformer_ctc/test_wavs/1221-135766-0001.flac', './tmp/conformer_ctc/test_wavs/1221-135766-0002.flac']
|
||||
2021-08-20 11:20:04,595 INFO [pretrained.py:271] Decoding started
|
||||
2021-08-20 11:20:04,854 INFO [pretrained.py:313] Use HLG + LM rescoring + attention decoder rescoring
|
||||
2021-08-20 11:20:05,805 INFO [pretrained.py:339]
|
||||
./tmp/conformer_ctc/test_wavs/1089-134686-0001.flac:
|
||||
AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS
|
||||
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0001.flac:
|
||||
GOD AS A DIRECT CONSEQUENCE OF THE SIN WHICH MAN THUS PUNISHED HAD GIVEN HER A LOVELY CHILD WHOSE PLACE WAS ON THAT SAME DISHONOURED
|
||||
BOSOM TO CONNECT HER PARENT FOR EVER WITH THE RACE AND DESCENT OF MORTALS AND TO BE FINALLY A BLESSED SOUL IN HEAVEN
|
||||
|
||||
./tmp/conformer_ctc/test_wavs/1221-135766-0002.flac:
|
||||
YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
|
||||
|
||||
|
||||
2021-08-20 11:20:05,805 INFO [pretrained.py:341] Decoding Done
|
||||
```
|
||||
|
||||
**NOTE**: We provide a colab notebook for demonstration.
|
||||
[](https://colab.research.google.com/drive/1huyupXAcHsUrKaWfI83iMEJ6J0Nh0213?usp=sharing)
|
||||
|
||||
Due to limited memory provided by Colab, you have to upgrade to Colab Pro to
|
||||
run `HLG decoding + LM rescoring` and `HLG decoding + LM rescoring + attention decoder rescoring`.
|
||||
Otherwise, you can only run `HLG decoding` with Colab.
|
||||
Please visit
|
||||
<https://icefall.readthedocs.io/en/latest/recipes/librispeech/conformer_ctc.html>
|
||||
for how to run this recipe.
|
||||
|
||||
@ -56,8 +56,6 @@ class Conformer(Transformer):
|
||||
cnn_module_kernel: int = 31,
|
||||
normalize_before: bool = True,
|
||||
vgg_frontend: bool = False,
|
||||
is_espnet_structure: bool = False,
|
||||
mmi_loss: bool = True,
|
||||
use_feat_batchnorm: bool = False,
|
||||
) -> None:
|
||||
super(Conformer, self).__init__(
|
||||
@ -72,7 +70,6 @@ class Conformer(Transformer):
|
||||
dropout=dropout,
|
||||
normalize_before=normalize_before,
|
||||
vgg_frontend=vgg_frontend,
|
||||
mmi_loss=mmi_loss,
|
||||
use_feat_batchnorm=use_feat_batchnorm,
|
||||
)
|
||||
|
||||
@ -85,12 +82,10 @@ class Conformer(Transformer):
|
||||
dropout,
|
||||
cnn_module_kernel,
|
||||
normalize_before,
|
||||
is_espnet_structure,
|
||||
)
|
||||
self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
|
||||
self.normalize_before = normalize_before
|
||||
self.is_espnet_structure = is_espnet_structure
|
||||
if self.normalize_before and self.is_espnet_structure:
|
||||
if self.normalize_before:
|
||||
self.after_norm = nn.LayerNorm(d_model)
|
||||
else:
|
||||
# Note: TorchScript detects that self.after_norm could be used inside forward()
|
||||
@ -103,7 +98,7 @@ class Conformer(Transformer):
|
||||
"""
|
||||
Args:
|
||||
x:
|
||||
The model input. Its shape is [N, T, C].
|
||||
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
|
||||
@ -125,7 +120,7 @@ class Conformer(Transformer):
|
||||
mask = mask.to(x.device)
|
||||
x = self.encoder(x, pos_emb, src_key_padding_mask=mask) # (T, B, F)
|
||||
|
||||
if self.normalize_before and self.is_espnet_structure:
|
||||
if self.normalize_before:
|
||||
x = self.after_norm(x)
|
||||
|
||||
return x, mask
|
||||
@ -159,11 +154,10 @@ class ConformerEncoderLayer(nn.Module):
|
||||
dropout: float = 0.1,
|
||||
cnn_module_kernel: int = 31,
|
||||
normalize_before: bool = True,
|
||||
is_espnet_structure: bool = False,
|
||||
) -> None:
|
||||
super(ConformerEncoderLayer, self).__init__()
|
||||
self.self_attn = RelPositionMultiheadAttention(
|
||||
d_model, nhead, dropout=0.0, is_espnet_structure=is_espnet_structure
|
||||
d_model, nhead, dropout=0.0
|
||||
)
|
||||
|
||||
self.feed_forward = nn.Sequential(
|
||||
@ -436,7 +430,6 @@ class RelPositionMultiheadAttention(nn.Module):
|
||||
embed_dim: int,
|
||||
num_heads: int,
|
||||
dropout: float = 0.0,
|
||||
is_espnet_structure: bool = False,
|
||||
) -> None:
|
||||
super(RelPositionMultiheadAttention, self).__init__()
|
||||
self.embed_dim = embed_dim
|
||||
@ -459,8 +452,6 @@ class RelPositionMultiheadAttention(nn.Module):
|
||||
|
||||
self._reset_parameters()
|
||||
|
||||
self.is_espnet_structure = is_espnet_structure
|
||||
|
||||
def _reset_parameters(self) -> None:
|
||||
nn.init.xavier_uniform_(self.in_proj.weight)
|
||||
nn.init.constant_(self.in_proj.bias, 0.0)
|
||||
@ -690,9 +681,6 @@ class RelPositionMultiheadAttention(nn.Module):
|
||||
_b = _b[_start:]
|
||||
v = nn.functional.linear(value, _w, _b)
|
||||
|
||||
if not self.is_espnet_structure:
|
||||
q = q * scaling
|
||||
|
||||
if attn_mask is not None:
|
||||
assert (
|
||||
attn_mask.dtype == torch.float32
|
||||
@ -785,14 +773,9 @@ class RelPositionMultiheadAttention(nn.Module):
|
||||
) # (batch, head, time1, 2*time1-1)
|
||||
matrix_bd = self.rel_shift(matrix_bd)
|
||||
|
||||
if not self.is_espnet_structure:
|
||||
attn_output_weights = (
|
||||
matrix_ac + matrix_bd
|
||||
) # (batch, head, time1, time2)
|
||||
else:
|
||||
attn_output_weights = (
|
||||
matrix_ac + matrix_bd
|
||||
) * scaling # (batch, head, time1, time2)
|
||||
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
|
||||
|
||||
@ -45,6 +45,7 @@ from icefall.utils import (
|
||||
get_texts,
|
||||
setup_logger,
|
||||
store_transcripts,
|
||||
str2bool,
|
||||
write_error_stats,
|
||||
)
|
||||
|
||||
@ -57,28 +58,74 @@ def get_parser():
|
||||
parser.add_argument(
|
||||
"--epoch",
|
||||
type=int,
|
||||
default=9,
|
||||
default=34,
|
||||
help="It specifies the checkpoint to use for decoding."
|
||||
"Note: Epoch counts from 0.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--avg",
|
||||
type=int,
|
||||
default=1,
|
||||
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) nbest-rescoring. Extract n paths from the decoding lattice,
|
||||
rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
|
||||
the highest score is the decoding result.
|
||||
- (4) whole-lattice-rescoring. Rescore the decoding lattice with an
|
||||
n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
|
||||
is the decoding result.
|
||||
- (5) attention-decoder. Extract n paths from the LM rescored
|
||||
lattice, the path with the highest score is the decoding result.
|
||||
- (6) 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, nbest-rescoring, attention-decoder, and nbest-oracle
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--lattice-score-scale",
|
||||
type=float,
|
||||
default=1.0,
|
||||
help="The scale to be applied to `lattice.scores`."
|
||||
"It's needed if you use any kinds of n-best based rescoring. "
|
||||
"Currently, it is used when the decoding method is: nbest, "
|
||||
"nbest-rescoring, attention-decoder, and nbest-oracle. "
|
||||
"A smaller value results in more unique paths.",
|
||||
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, nbest-rescoring, 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
|
||||
@ -90,35 +137,20 @@ def get_params() -> AttributeDict:
|
||||
"exp_dir": Path("conformer_ctc/exp"),
|
||||
"lang_dir": Path("data/lang_bpe"),
|
||||
"lm_dir": Path("data/lm"),
|
||||
# parameters for conformer
|
||||
"subsampling_factor": 4,
|
||||
"vgg_frontend": False,
|
||||
"use_feat_batchnorm": True,
|
||||
"feature_dim": 80,
|
||||
"nhead": 8,
|
||||
"attention_dim": 512,
|
||||
"subsampling_factor": 4,
|
||||
"num_decoder_layers": 6,
|
||||
"vgg_frontend": False,
|
||||
"is_espnet_structure": True,
|
||||
"mmi_loss": False,
|
||||
"use_feat_batchnorm": True,
|
||||
# parameters for decoding
|
||||
"search_beam": 20,
|
||||
"output_beam": 8,
|
||||
"min_active_states": 30,
|
||||
"max_active_states": 10000,
|
||||
"use_double_scores": True,
|
||||
# Possible values for method:
|
||||
# - 1best
|
||||
# - nbest
|
||||
# - nbest-rescoring
|
||||
# - whole-lattice-rescoring
|
||||
# - attention-decoder
|
||||
# - nbest-oracle
|
||||
# "method": "nbest",
|
||||
# "method": "nbest-rescoring",
|
||||
# "method": "whole-lattice-rescoring",
|
||||
"method": "attention-decoder",
|
||||
# "method": "nbest-oracle",
|
||||
# num_paths is used when method is "nbest", "nbest-rescoring",
|
||||
# attention-decoder, and nbest-oracle
|
||||
"num_paths": 100,
|
||||
}
|
||||
)
|
||||
return params
|
||||
@ -129,7 +161,7 @@ def decode_one_batch(
|
||||
model: nn.Module,
|
||||
HLG: k2.Fsa,
|
||||
batch: dict,
|
||||
lexicon: Lexicon,
|
||||
word_table: k2.SymbolTable,
|
||||
sos_id: int,
|
||||
eos_id: int,
|
||||
G: Optional[k2.Fsa] = None,
|
||||
@ -163,8 +195,8 @@ def decode_one_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.
|
||||
word_table:
|
||||
The word symbol table.
|
||||
sos_id:
|
||||
The token ID of the SOS.
|
||||
eos_id:
|
||||
@ -181,12 +213,12 @@ def decode_one_batch(
|
||||
feature = batch["inputs"]
|
||||
assert feature.ndim == 3
|
||||
feature = feature.to(device)
|
||||
# at entry, feature is [N, T, C]
|
||||
# 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]
|
||||
# nnet_output is (N, T, C)
|
||||
|
||||
supervision_segments = torch.stack(
|
||||
(
|
||||
@ -212,14 +244,19 @@ def decode_one_batch(
|
||||
# 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(
|
||||
# is only slightly worse than that of rescored lattices.
|
||||
best_path = nbest_oracle(
|
||||
lattice=lattice,
|
||||
num_paths=params.num_paths,
|
||||
ref_texts=supervisions["text"],
|
||||
lexicon=lexicon,
|
||||
scale=params.lattice_score_scale,
|
||||
word_table=word_table,
|
||||
lattice_score_scale=params.lattice_score_scale,
|
||||
oov="<UNK>",
|
||||
)
|
||||
hyps = get_texts(best_path)
|
||||
hyps = [[word_table[i] for i in ids] for ids in hyps]
|
||||
key = f"oracle_{params.num_paths}_lattice_score_scale_{params.lattice_score_scale}" # noqa
|
||||
return {key: hyps}
|
||||
|
||||
if params.method in ["1best", "nbest"]:
|
||||
if params.method == "1best":
|
||||
@ -232,12 +269,12 @@ def decode_one_batch(
|
||||
lattice=lattice,
|
||||
num_paths=params.num_paths,
|
||||
use_double_scores=params.use_double_scores,
|
||||
scale=params.lattice_score_scale,
|
||||
lattice_score_scale=params.lattice_score_scale,
|
||||
)
|
||||
key = f"no_rescore-scale-{params.lattice_score_scale}-{params.num_paths}" # noqa
|
||||
|
||||
hyps = get_texts(best_path)
|
||||
hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
|
||||
hyps = [[word_table[i] for i in ids] for ids in hyps]
|
||||
return {key: hyps}
|
||||
|
||||
assert params.method in [
|
||||
@ -246,7 +283,8 @@ def decode_one_batch(
|
||||
"attention-decoder",
|
||||
]
|
||||
|
||||
lm_scale_list = [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
|
||||
lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
|
||||
lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
|
||||
lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
|
||||
|
||||
if params.method == "nbest-rescoring":
|
||||
@ -255,17 +293,23 @@ def decode_one_batch(
|
||||
G=G,
|
||||
num_paths=params.num_paths,
|
||||
lm_scale_list=lm_scale_list,
|
||||
scale=params.lattice_score_scale,
|
||||
lattice_score_scale=params.lattice_score_scale,
|
||||
)
|
||||
elif params.method == "whole-lattice-rescoring":
|
||||
best_path_dict = rescore_with_whole_lattice(
|
||||
lattice=lattice, G_with_epsilon_loops=G, lm_scale_list=lm_scale_list
|
||||
lattice=lattice,
|
||||
G_with_epsilon_loops=G,
|
||||
lm_scale_list=lm_scale_list,
|
||||
)
|
||||
elif params.method == "attention-decoder":
|
||||
# lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
|
||||
rescored_lattice = rescore_with_whole_lattice(
|
||||
lattice=lattice, G_with_epsilon_loops=G, lm_scale_list=None
|
||||
lattice=lattice,
|
||||
G_with_epsilon_loops=G,
|
||||
lm_scale_list=None,
|
||||
)
|
||||
# TODO: pass `lattice` instead of `rescored_lattice` to
|
||||
# `rescore_with_attention_decoder`
|
||||
|
||||
best_path_dict = rescore_with_attention_decoder(
|
||||
lattice=rescored_lattice,
|
||||
@ -275,16 +319,20 @@ def decode_one_batch(
|
||||
memory_key_padding_mask=memory_key_padding_mask,
|
||||
sos_id=sos_id,
|
||||
eos_id=eos_id,
|
||||
scale=params.lattice_score_scale,
|
||||
lattice_score_scale=params.lattice_score_scale,
|
||||
)
|
||||
else:
|
||||
assert False, f"Unsupported decoding method: {params.method}"
|
||||
|
||||
ans = dict()
|
||||
for lm_scale_str, best_path in best_path_dict.items():
|
||||
hyps = get_texts(best_path)
|
||||
hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
|
||||
ans[lm_scale_str] = hyps
|
||||
if best_path_dict is not None:
|
||||
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
|
||||
else:
|
||||
for lm_scale in lm_scale_list:
|
||||
ans[lm_scale_str] = [[] * lattice.shape[0]]
|
||||
return ans
|
||||
|
||||
|
||||
@ -293,7 +341,7 @@ def decode_dataset(
|
||||
params: AttributeDict,
|
||||
model: nn.Module,
|
||||
HLG: k2.Fsa,
|
||||
lexicon: Lexicon,
|
||||
word_table: k2.SymbolTable,
|
||||
sos_id: int,
|
||||
eos_id: int,
|
||||
G: Optional[k2.Fsa] = None,
|
||||
@ -309,8 +357,8 @@ def decode_dataset(
|
||||
The neural model.
|
||||
HLG:
|
||||
The decoding graph.
|
||||
lexicon:
|
||||
It contains word symbol table.
|
||||
word_table:
|
||||
It is the word symbol table.
|
||||
sos_id:
|
||||
The token ID for SOS.
|
||||
eos_id:
|
||||
@ -344,7 +392,7 @@ def decode_dataset(
|
||||
model=model,
|
||||
HLG=HLG,
|
||||
batch=batch,
|
||||
lexicon=lexicon,
|
||||
word_table=word_table,
|
||||
G=G,
|
||||
sos_id=sos_id,
|
||||
eos_id=eos_id,
|
||||
@ -505,8 +553,6 @@ def main():
|
||||
subsampling_factor=params.subsampling_factor,
|
||||
num_decoder_layers=params.num_decoder_layers,
|
||||
vgg_frontend=params.vgg_frontend,
|
||||
is_espnet_structure=params.is_espnet_structure,
|
||||
mmi_loss=params.mmi_loss,
|
||||
use_feat_batchnorm=params.use_feat_batchnorm,
|
||||
)
|
||||
|
||||
@ -521,6 +567,13 @@ def main():
|
||||
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()])
|
||||
@ -540,7 +593,7 @@ def main():
|
||||
params=params,
|
||||
model=model,
|
||||
HLG=HLG,
|
||||
lexicon=lexicon,
|
||||
word_table=lexicon.word_table,
|
||||
G=G,
|
||||
sos_id=sos_id,
|
||||
eos_id=eos_id,
|
||||
|
||||
@ -173,17 +173,17 @@ def get_parser():
|
||||
def get_params() -> AttributeDict:
|
||||
params = AttributeDict(
|
||||
{
|
||||
"sample_rate": 16000,
|
||||
# parameters for conformer
|
||||
"subsampling_factor": 4,
|
||||
"vgg_frontend": False,
|
||||
"use_feat_batchnorm": True,
|
||||
"feature_dim": 80,
|
||||
"nhead": 8,
|
||||
"num_classes": 5000,
|
||||
"sample_rate": 16000,
|
||||
"attention_dim": 512,
|
||||
"subsampling_factor": 4,
|
||||
"num_decoder_layers": 6,
|
||||
"vgg_frontend": False,
|
||||
"is_espnet_structure": True,
|
||||
"mmi_loss": False,
|
||||
"use_feat_batchnorm": True,
|
||||
# parameters for decoding
|
||||
"search_beam": 20,
|
||||
"output_beam": 8,
|
||||
"min_active_states": 30,
|
||||
@ -241,8 +241,6 @@ def main():
|
||||
subsampling_factor=params.subsampling_factor,
|
||||
num_decoder_layers=params.num_decoder_layers,
|
||||
vgg_frontend=params.vgg_frontend,
|
||||
is_espnet_structure=params.is_espnet_structure,
|
||||
mmi_loss=params.mmi_loss,
|
||||
use_feat_batchnorm=params.use_feat_batchnorm,
|
||||
)
|
||||
|
||||
@ -338,7 +336,7 @@ def main():
|
||||
memory_key_padding_mask=memory_key_padding_mask,
|
||||
sos_id=params.sos_id,
|
||||
eos_id=params.eos_id,
|
||||
scale=params.lattice_score_scale,
|
||||
lattice_score_scale=params.lattice_score_scale,
|
||||
ngram_lm_scale=params.ngram_lm_scale,
|
||||
attention_scale=params.attention_decoder_scale,
|
||||
)
|
||||
|
||||
@ -22,8 +22,8 @@ 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
|
||||
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
|
||||
@ -34,10 +34,10 @@ class Conv2dSubsampling(nn.Module):
|
||||
"""
|
||||
Args:
|
||||
idim:
|
||||
Input dim. The input shape is [N, T, 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]
|
||||
Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
|
||||
"""
|
||||
assert idim >= 7
|
||||
super().__init__()
|
||||
@ -58,18 +58,18 @@ class Conv2dSubsampling(nn.Module):
|
||||
|
||||
Args:
|
||||
x:
|
||||
Its shape is [N, T, idim].
|
||||
Its shape is (N, T, idim).
|
||||
|
||||
Returns:
|
||||
Return a tensor of shape [N, ((T-1)//2 - 1)//2, odim]
|
||||
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]
|
||||
# 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]
|
||||
# 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]
|
||||
# Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
|
||||
return x
|
||||
|
||||
|
||||
@ -80,8 +80,8 @@ class VggSubsampling(nn.Module):
|
||||
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
|
||||
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
|
||||
"""
|
||||
|
||||
@ -93,10 +93,10 @@ class VggSubsampling(nn.Module):
|
||||
|
||||
Args:
|
||||
idim:
|
||||
Input dim. The input shape is [N, T, 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]
|
||||
Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
|
||||
"""
|
||||
super().__init__()
|
||||
|
||||
@ -149,10 +149,10 @@ class VggSubsampling(nn.Module):
|
||||
|
||||
Args:
|
||||
x:
|
||||
Its shape is [N, T, idim].
|
||||
Its shape is (N, T, idim).
|
||||
|
||||
Returns:
|
||||
Return a tensor of shape [N, ((T-1)//2 - 1)//2, odim]
|
||||
Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
|
||||
"""
|
||||
x = x.unsqueeze(1)
|
||||
x = self.layers(x)
|
||||
|
||||
@ -16,9 +16,8 @@
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
from subsampling import Conv2dSubsampling
|
||||
from subsampling import VggSubsampling
|
||||
import torch
|
||||
from subsampling import Conv2dSubsampling, VggSubsampling
|
||||
|
||||
|
||||
def test_conv2d_subsampling():
|
||||
|
||||
@ -17,17 +17,16 @@
|
||||
|
||||
|
||||
import torch
|
||||
from torch.nn.utils.rnn import pad_sequence
|
||||
from transformer import (
|
||||
Transformer,
|
||||
add_eos,
|
||||
add_sos,
|
||||
decoder_padding_mask,
|
||||
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 = {
|
||||
|
||||
@ -1,5 +1,6 @@
|
||||
#!/usr/bin/env python3
|
||||
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang)
|
||||
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang,
|
||||
# Wei Kang)
|
||||
#
|
||||
# See ../../../../LICENSE for clarification regarding multiple authors
|
||||
#
|
||||
@ -74,6 +75,23 @@ def get_parser():
|
||||
help="Should various information be logged in tensorboard.",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--num-epochs",
|
||||
type=int,
|
||||
default=35,
|
||||
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
|
||||
|
||||
|
||||
@ -94,20 +112,6 @@ def get_params() -> AttributeDict:
|
||||
- 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.
|
||||
|
||||
- start_epoch: If it is not zero, load checkpoint `start_epoch-1`
|
||||
and continue training from that checkpoint.
|
||||
|
||||
- num_epochs: Number of epochs to train.
|
||||
|
||||
- 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.
|
||||
@ -126,25 +130,40 @@ def get_params() -> AttributeDict:
|
||||
|
||||
- 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
|
||||
|
||||
- reset_interval: Reset statistics if batch_idx % reset_interval is 0
|
||||
- feature_dim: The model input dim. It has to match the one used
|
||||
in computing features.
|
||||
|
||||
- subsampling_factor: The subsampling factor for the model.
|
||||
|
||||
- use_feat_batchnorm: Whether to do batch normalization for the
|
||||
input features.
|
||||
|
||||
- attention_dim: Hidden dim for multi-head attention model.
|
||||
|
||||
- head: Number of heads of multi-head attention model.
|
||||
|
||||
- num_decoder_layers: Number of decoder layer of transformer decoder.
|
||||
|
||||
- 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
|
||||
|
||||
- weight_decay: The weight_decay for the optimizer.
|
||||
|
||||
- lr_factor: The lr_factor for Noam optimizer.
|
||||
|
||||
- warm_step: The warm_step for Noam optimizer.
|
||||
"""
|
||||
params = AttributeDict(
|
||||
{
|
||||
"exp_dir": Path("conformer_ctc/exp"),
|
||||
"lang_dir": Path("data/lang_bpe"),
|
||||
"feature_dim": 80,
|
||||
"weight_decay": 1e-6,
|
||||
"subsampling_factor": 4,
|
||||
"start_epoch": 0,
|
||||
"num_epochs": 20,
|
||||
"best_train_loss": float("inf"),
|
||||
"best_valid_loss": float("inf"),
|
||||
"best_train_epoch": -1,
|
||||
@ -153,17 +172,20 @@ def get_params() -> AttributeDict:
|
||||
"log_interval": 10,
|
||||
"reset_interval": 200,
|
||||
"valid_interval": 3000,
|
||||
"beam_size": 10,
|
||||
"reduction": "sum",
|
||||
"use_double_scores": True,
|
||||
"accum_grad": 1,
|
||||
"att_rate": 0.7,
|
||||
# parameters for conformer
|
||||
"feature_dim": 80,
|
||||
"subsampling_factor": 4,
|
||||
"use_feat_batchnorm": True,
|
||||
"attention_dim": 512,
|
||||
"nhead": 8,
|
||||
"num_decoder_layers": 6,
|
||||
"is_espnet_structure": True,
|
||||
"mmi_loss": False,
|
||||
"use_feat_batchnorm": True,
|
||||
# parameters for loss
|
||||
"beam_size": 10,
|
||||
"reduction": "sum",
|
||||
"use_double_scores": True,
|
||||
"att_rate": 0.7,
|
||||
# parameters for Noam
|
||||
"weight_decay": 1e-6,
|
||||
"lr_factor": 5.0,
|
||||
"warm_step": 80000,
|
||||
}
|
||||
@ -288,14 +310,14 @@ def compute_loss(
|
||||
"""
|
||||
device = graph_compiler.device
|
||||
feature = batch["inputs"]
|
||||
# at entry, feature is [N, T, C]
|
||||
# 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]
|
||||
# nnet_output is (N, T, C)
|
||||
|
||||
# NOTE: We need `encode_supervisions` to sort sequences with
|
||||
# different duration in decreasing order, required by
|
||||
@ -636,8 +658,6 @@ def run(rank, world_size, args):
|
||||
subsampling_factor=params.subsampling_factor,
|
||||
num_decoder_layers=params.num_decoder_layers,
|
||||
vgg_frontend=False,
|
||||
is_espnet_structure=params.is_espnet_structure,
|
||||
mmi_loss=params.mmi_loss,
|
||||
use_feat_batchnorm=params.use_feat_batchnorm,
|
||||
)
|
||||
|
||||
|
||||
@ -41,7 +41,6 @@ class Transformer(nn.Module):
|
||||
dropout: float = 0.1,
|
||||
normalize_before: bool = True,
|
||||
vgg_frontend: bool = False,
|
||||
mmi_loss: bool = True,
|
||||
use_feat_batchnorm: bool = False,
|
||||
) -> None:
|
||||
"""
|
||||
@ -70,7 +69,6 @@ class Transformer(nn.Module):
|
||||
If True, use pre-layer norm; False to use post-layer norm.
|
||||
vgg_frontend:
|
||||
True to use vgg style frontend for subsampling.
|
||||
mmi_loss:
|
||||
use_feat_batchnorm:
|
||||
True to use batchnorm for the input layer.
|
||||
"""
|
||||
@ -85,8 +83,8 @@ class Transformer(nn.Module):
|
||||
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].
|
||||
# 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
|
||||
@ -122,14 +120,9 @@ class Transformer(nn.Module):
|
||||
)
|
||||
|
||||
if num_decoder_layers > 0:
|
||||
if mmi_loss:
|
||||
self.decoder_num_class = (
|
||||
self.num_classes + 1
|
||||
) # +1 for the sos/eos symbol
|
||||
else:
|
||||
self.decoder_num_class = (
|
||||
self.num_classes
|
||||
) # bpe model already has sos/eos symbol
|
||||
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
|
||||
@ -169,7 +162,7 @@ class Transformer(nn.Module):
|
||||
"""
|
||||
Args:
|
||||
x:
|
||||
The input tensor. Its shape is [N, T, C].
|
||||
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
|
||||
@ -178,17 +171,17 @@ class Transformer(nn.Module):
|
||||
|
||||
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
|
||||
- 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].
|
||||
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 = 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]
|
||||
x = x.permute(0, 2, 1) # (N, C, T) -> (N, T, C)
|
||||
encoder_memory, memory_key_padding_mask = self.run_encoder(
|
||||
x, supervision
|
||||
)
|
||||
@ -202,7 +195,7 @@ class Transformer(nn.Module):
|
||||
|
||||
Args:
|
||||
x:
|
||||
The model input. Its shape is [N, T, C].
|
||||
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
|
||||
@ -213,8 +206,8 @@ class Transformer(nn.Module):
|
||||
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 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.
|
||||
"""
|
||||
@ -232,11 +225,11 @@ class Transformer(nn.Module):
|
||||
Args:
|
||||
x:
|
||||
The output tensor from the transformer encoder.
|
||||
Its shape is [T, N, C]
|
||||
Its shape is (T, N, C)
|
||||
|
||||
Returns:
|
||||
Return a tensor that can be used for CTC decoding.
|
||||
Its shape is [N, T, C]
|
||||
Its shape is (N, T, C)
|
||||
"""
|
||||
x = self.encoder_output_layer(x)
|
||||
x = x.permute(1, 0, 2) # (T, N, C) ->(N, T, C)
|
||||
@ -254,7 +247,7 @@ class Transformer(nn.Module):
|
||||
"""
|
||||
Args:
|
||||
memory:
|
||||
It's the output of the encoder with shape [T, N, C]
|
||||
It's the output of the encoder with shape (T, N, C)
|
||||
memory_key_padding_mask:
|
||||
The padding mask from the encoder.
|
||||
token_ids:
|
||||
@ -319,7 +312,7 @@ class Transformer(nn.Module):
|
||||
"""
|
||||
Args:
|
||||
memory:
|
||||
It's the output of the encoder with shape [T, N, C]
|
||||
It's the output of the encoder with shape (T, N, C)
|
||||
memory_key_padding_mask:
|
||||
The padding mask from the encoder.
|
||||
token_ids:
|
||||
@ -661,13 +654,13 @@ class PositionalEncoding(nn.Module):
|
||||
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.
|
||||
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].
|
||||
It is a tensor of shape (N, T, C).
|
||||
Returns:
|
||||
Return None.
|
||||
"""
|
||||
@ -685,7 +678,7 @@ class PositionalEncoding(nn.Module):
|
||||
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)
|
||||
# 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:
|
||||
@ -694,10 +687,10 @@ class PositionalEncoding(nn.Module):
|
||||
|
||||
Args:
|
||||
x:
|
||||
Its shape is [N, T, C]
|
||||
Its shape is (N, T, C)
|
||||
|
||||
Returns:
|
||||
Return a tensor of shape [N, T, C]
|
||||
Return a tensor of shape (N, T, C)
|
||||
"""
|
||||
self.extend_pe(x)
|
||||
x = x * self.xscale + self.pe[:, : x.size(1), :]
|
||||
|
||||
@ -102,14 +102,14 @@ def compile_HLG(lang_dir: str) -> k2.Fsa:
|
||||
|
||||
LG.labels[LG.labels >= first_token_disambig_id] = 0
|
||||
|
||||
assert isinstance(LG.aux_labels, k2.RaggedInt)
|
||||
LG.aux_labels.values()[LG.aux_labels.values() >= first_word_disambig_id] = 0
|
||||
assert isinstance(LG.aux_labels, k2.RaggedTensor)
|
||||
LG.aux_labels.values[LG.aux_labels.values >= 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 = k2.ragged.remove_values_eq(LG.aux_labels, 0)
|
||||
LG.aux_labels = LG.aux_labels.remove_values_eq(0)
|
||||
|
||||
logging.info("Arc sorting LG")
|
||||
LG = k2.arc_sort(LG)
|
||||
|
||||
@ -1,2 +1,4 @@
|
||||
|
||||
Will add results later.
|
||||
Please visit
|
||||
<https://icefall.readthedocs.io/en/latest/recipes/librispeech/tdnn_lstm_ctc.html>
|
||||
for how to run this recipe.
|
||||
|
||||
@ -82,14 +82,14 @@ class LibriSpeechAsrDataModule(DataModule):
|
||||
group.add_argument(
|
||||
"--max-duration",
|
||||
type=int,
|
||||
default=500.0,
|
||||
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=False,
|
||||
default=True,
|
||||
help="When enabled, the batches will come from buckets of "
|
||||
"similar duration (saves padding frames).",
|
||||
)
|
||||
|
||||
@ -42,6 +42,7 @@ from icefall.utils import (
|
||||
get_texts,
|
||||
setup_logger,
|
||||
store_transcripts,
|
||||
str2bool,
|
||||
write_error_stats,
|
||||
)
|
||||
|
||||
@ -54,7 +55,7 @@ def get_parser():
|
||||
parser.add_argument(
|
||||
"--epoch",
|
||||
type=int,
|
||||
default=9,
|
||||
default=19,
|
||||
help="It specifies the checkpoint to use for decoding."
|
||||
"Note: Epoch counts from 0.",
|
||||
)
|
||||
@ -66,6 +67,57 @@ def get_parser():
|
||||
"consecutive checkpoints before the checkpoint specified by "
|
||||
"'--epoch'. ",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--method",
|
||||
type=str,
|
||||
default="whole-lattice-rescoring",
|
||||
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) nbest-rescoring. Extract n paths from the decoding lattice,
|
||||
rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
|
||||
the highest score is the decoding result.
|
||||
- (4) whole-lattice-rescoring. Rescore the decoding lattice with an
|
||||
n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
|
||||
is the decoding result.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--num-paths",
|
||||
type=int,
|
||||
default=100,
|
||||
help="""Number of paths for n-best based decoding method.
|
||||
Used only when "method" is one of the following values:
|
||||
nbest, nbest-rescoring
|
||||
""",
|
||||
)
|
||||
|
||||
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, nbest-rescoring
|
||||
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
|
||||
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
|
||||
|
||||
|
||||
@ -82,14 +134,6 @@ def get_params() -> AttributeDict:
|
||||
"min_active_states": 30,
|
||||
"max_active_states": 10000,
|
||||
"use_double_scores": True,
|
||||
# Possible values for method:
|
||||
# - 1best
|
||||
# - nbest
|
||||
# - nbest-rescoring
|
||||
# - whole-lattice-rescoring
|
||||
"method": "1best",
|
||||
# num_paths is used when method is "nbest" and "nbest-rescoring"
|
||||
"num_paths": 30,
|
||||
}
|
||||
)
|
||||
return params
|
||||
@ -146,12 +190,12 @@ def decode_one_batch(
|
||||
feature = batch["inputs"]
|
||||
assert feature.ndim == 3
|
||||
feature = feature.to(device)
|
||||
# at entry, feature is [N, T, C]
|
||||
# at entry, feature is (N, T, C)
|
||||
|
||||
feature = feature.permute(0, 2, 1) # now feature is [N, C, T]
|
||||
feature = feature.permute(0, 2, 1) # now feature is (N, C, T)
|
||||
|
||||
nnet_output = model(feature)
|
||||
# nnet_output is [N, T, C]
|
||||
# nnet_output is (N, T, C)
|
||||
|
||||
supervisions = batch["supervisions"]
|
||||
|
||||
@ -185,6 +229,7 @@ def decode_one_batch(
|
||||
lattice=lattice,
|
||||
num_paths=params.num_paths,
|
||||
use_double_scores=params.use_double_scores,
|
||||
lattice_score_scale=params.lattice_score_scale,
|
||||
)
|
||||
key = f"no_rescore-{params.num_paths}"
|
||||
hyps = get_texts(best_path)
|
||||
@ -193,7 +238,8 @@ def decode_one_batch(
|
||||
|
||||
assert params.method in ["nbest-rescoring", "whole-lattice-rescoring"]
|
||||
|
||||
lm_scale_list = [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
|
||||
lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
|
||||
lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
|
||||
lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
|
||||
|
||||
if params.method == "nbest-rescoring":
|
||||
@ -202,10 +248,13 @@ def decode_one_batch(
|
||||
G=G,
|
||||
num_paths=params.num_paths,
|
||||
lm_scale_list=lm_scale_list,
|
||||
lattice_score_scale=params.lattice_score_scale,
|
||||
)
|
||||
else:
|
||||
best_path_dict = rescore_with_whole_lattice(
|
||||
lattice=lattice, G_with_epsilon_loops=G, lm_scale_list=lm_scale_list
|
||||
lattice=lattice,
|
||||
G_with_epsilon_loops=G,
|
||||
lm_scale_list=lm_scale_list,
|
||||
)
|
||||
|
||||
ans = dict()
|
||||
@ -408,6 +457,13 @@ def main():
|
||||
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()
|
||||
|
||||
|
||||
277
egs/librispeech/ASR/tdnn_lstm_ctc/pretrained.py
Executable file
277
egs/librispeech/ASR/tdnn_lstm_ctc/pretrained.py
Executable file
@ -0,0 +1,277 @@
|
||||
#!/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,
|
||||
rescore_with_whole_lattice,
|
||||
)
|
||||
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) whole-lattice-rescoring - Use an LM to rescore the
|
||||
decoding lattice and then use 1best to decode the
|
||||
rescored lattice.
|
||||
We call it HLG decoding + n-gram LM rescoring.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--G",
|
||||
type=str,
|
||||
help="""An LM for rescoring.
|
||||
Used only when method is
|
||||
whole-lattice-rescoring.
|
||||
It's usually a 4-gram LM.
|
||||
""",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--ngram-lm-scale",
|
||||
type=float,
|
||||
default=0.8,
|
||||
help="""
|
||||
Used only when method is whole-lattice-rescoring.
|
||||
It specifies the scale for n-gram LM scores.
|
||||
(Note: You need to tune it on a dataset.)
|
||||
""",
|
||||
)
|
||||
|
||||
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": 72,
|
||||
"sample_rate": 16000,
|
||||
"search_beam": 20,
|
||||
"output_beam": 5,
|
||||
"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()
|
||||
|
||||
if params.method == "whole-lattice-rescoring":
|
||||
logging.info(f"Loading G from {params.G}")
|
||||
G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
|
||||
# Add epsilon self-loops to G as we will compose
|
||||
# it with the whole lattice later
|
||||
G = G.to(device)
|
||||
G = k2.add_epsilon_self_loops(G)
|
||||
G = k2.arc_sort(G)
|
||||
G.lm_scores = G.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,
|
||||
)
|
||||
|
||||
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 == "whole-lattice-rescoring":
|
||||
logging.info("Use HLG decoding + LM rescoring")
|
||||
best_path_dict = rescore_with_whole_lattice(
|
||||
lattice=lattice,
|
||||
G_with_epsilon_loops=G,
|
||||
lm_scale_list=[params.ngram_lm_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()
|
||||
@ -75,6 +75,23 @@ def get_parser():
|
||||
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
|
||||
|
||||
|
||||
@ -104,11 +121,6 @@ def get_params() -> AttributeDict:
|
||||
|
||||
- subsampling_factor: The subsampling factor for the model.
|
||||
|
||||
- start_epoch: If it is not zero, load checkpoint `start_epoch-1`
|
||||
and continue training from that checkpoint.
|
||||
|
||||
- num_epochs: Number of epochs to train.
|
||||
|
||||
- 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.
|
||||
@ -127,6 +139,8 @@ def get_params() -> AttributeDict:
|
||||
|
||||
- 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
|
||||
@ -143,14 +157,13 @@ def get_params() -> AttributeDict:
|
||||
"feature_dim": 80,
|
||||
"weight_decay": 5e-4,
|
||||
"subsampling_factor": 3,
|
||||
"start_epoch": 0,
|
||||
"num_epochs": 10,
|
||||
"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",
|
||||
@ -277,14 +290,14 @@ def compute_loss(
|
||||
"""
|
||||
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]
|
||||
# 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]
|
||||
# nnet_output is (N, T, C)
|
||||
|
||||
# NOTE: We need `encode_supervisions` to sort sequences with
|
||||
# different duration in decreasing order, required by
|
||||
@ -398,8 +411,12 @@ def train_one_epoch(
|
||||
"""
|
||||
model.train()
|
||||
|
||||
tot_loss = 0.0 # sum of losses over all batches
|
||||
tot_frames = 0.0 # sum of frames over all batches
|
||||
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"])
|
||||
@ -426,6 +443,9 @@ def train_one_epoch(
|
||||
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}, "
|
||||
@ -433,6 +453,22 @@ def train_one_epoch(
|
||||
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(
|
||||
@ -449,7 +485,7 @@ def train_one_epoch(
|
||||
f"best valid epoch: {params.best_valid_epoch}"
|
||||
)
|
||||
|
||||
params.train_loss = tot_loss / tot_frames
|
||||
params.train_loss = params.tot_loss / params.tot_frames
|
||||
|
||||
if params.train_loss < params.best_train_loss:
|
||||
params.best_train_epoch = params.cur_epoch
|
||||
|
||||
@ -1,15 +1,14 @@
|
||||
## Yesno recipe
|
||||
|
||||
You can run the recipe with **CPU**.
|
||||
This is the simplest ASR recipe in `icefall`.
|
||||
|
||||
|
||||
[](https://colab.research.google.com/drive/1tIjjzaJc3IvGyKiMCDWO-TSnBgkcuN3B?usp=sharing)
|
||||
|
||||
The above Colab notebook finishes the training using **CPU**
|
||||
within two minutes (50 epochs in total).
|
||||
|
||||
The WER is
|
||||
It can be run on CPU and takes less than 30 seconds to
|
||||
get the following WER:
|
||||
|
||||
```
|
||||
[test_set] %WER 0.42% [1 / 240, 0 ins, 1 del, 0 sub ]
|
||||
```
|
||||
|
||||
Please refer to
|
||||
<https://icefall.readthedocs.io/en/latest/recipes/yesno.html>
|
||||
for detailed instructions.
|
||||
|
||||
@ -80,14 +80,14 @@ def compile_HLG(lang_dir: str) -> k2.Fsa:
|
||||
|
||||
LG.labels[LG.labels >= first_token_disambig_id] = 0
|
||||
|
||||
assert isinstance(LG.aux_labels, k2.RaggedInt)
|
||||
LG.aux_labels.values()[LG.aux_labels.values() >= first_word_disambig_id] = 0
|
||||
assert isinstance(LG.aux_labels, k2.RaggedTensor)
|
||||
LG.aux_labels.values[LG.aux_labels.values >= 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 = k2.ragged.remove_values_eq(LG.aux_labels, 0)
|
||||
LG.aux_labels = LG.aux_labels.remove_values_eq(0)
|
||||
|
||||
logging.info("Arc sorting LG")
|
||||
LG = k2.arc_sort(LG)
|
||||
|
||||
8
egs/yesno/ASR/tdnn/README.md
Normal file
8
egs/yesno/ASR/tdnn/README.md
Normal file
@ -0,0 +1,8 @@
|
||||
|
||||
## How to run this recipe
|
||||
|
||||
You can find detailed instructions by visiting
|
||||
<https://icefall.readthedocs.io/en/latest/recipes/yesno.html>
|
||||
|
||||
It describes how to run this recipe and how to use
|
||||
a pre-trained model with `./pretrained.py`.
|
||||
@ -20,6 +20,7 @@ from icefall.utils import (
|
||||
get_texts,
|
||||
setup_logger,
|
||||
store_transcripts,
|
||||
str2bool,
|
||||
write_error_stats,
|
||||
)
|
||||
|
||||
@ -44,6 +45,17 @@ def get_parser():
|
||||
"consecutive checkpoints before the checkpoint specified by "
|
||||
"'--epoch'. ",
|
||||
)
|
||||
|
||||
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
|
||||
|
||||
|
||||
@ -99,10 +111,10 @@ def decode_one_batch(
|
||||
feature = batch["inputs"]
|
||||
assert feature.ndim == 3
|
||||
feature = feature.to(device)
|
||||
# at entry, feature is [N, T, C]
|
||||
# at entry, feature is (N, T, C)
|
||||
|
||||
nnet_output = model(feature)
|
||||
# nnet_output is [N, T, C]
|
||||
# nnet_output is (N, T, C)
|
||||
|
||||
batch_size = nnet_output.shape[0]
|
||||
supervision_segments = torch.tensor(
|
||||
@ -279,6 +291,13 @@ def main():
|
||||
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()
|
||||
|
||||
|
||||
209
egs/yesno/ASR/tdnn/pretrained.py
Executable file
209
egs/yesno/ASR/tdnn/pretrained.py
Executable file
@ -0,0 +1,209 @@
|
||||
#!/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
|
||||
import math
|
||||
from typing import List
|
||||
|
||||
import k2
|
||||
import kaldifeat
|
||||
import torch
|
||||
import torchaudio
|
||||
from model import Tdnn
|
||||
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(
|
||||
"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": 23,
|
||||
"num_classes": 4, # [<blk>, N, SIL, Y]
|
||||
"sample_rate": 8000,
|
||||
"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 = Tdnn(
|
||||
num_features=params.feature_dim,
|
||||
num_classes=params.num_classes,
|
||||
)
|
||||
|
||||
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)
|
||||
|
||||
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 = 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,
|
||||
)
|
||||
|
||||
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()
|
||||
@ -60,6 +60,16 @@ def get_parser():
|
||||
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/exp/epoch-{start_epoch-1}.pt
|
||||
""",
|
||||
)
|
||||
|
||||
return parser
|
||||
|
||||
|
||||
@ -92,8 +102,6 @@ def get_params() -> AttributeDict:
|
||||
- start_epoch: If it is not zero, load checkpoint `start_epoch-1`
|
||||
and continue training from that checkpoint.
|
||||
|
||||
- num_epochs: Number of epochs to train.
|
||||
|
||||
- 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.
|
||||
@ -260,13 +268,13 @@ def compute_loss(
|
||||
"""
|
||||
device = graph_compiler.device
|
||||
feature = batch["inputs"]
|
||||
# at entry, feature is [N, T, C]
|
||||
# at entry, feature is (N, T, C)
|
||||
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]
|
||||
# nnet_output is (N, T, C)
|
||||
|
||||
# NOTE: We need `encode_supervisions` to sort sequences with
|
||||
# different duration in decreasing order, required by
|
||||
@ -420,6 +428,19 @@ def train_one_epoch(
|
||||
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.valid_interval == 0:
|
||||
compute_validation_loss(
|
||||
params=params,
|
||||
@ -434,6 +455,12 @@ def train_one_epoch(
|
||||
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_loss",
|
||||
params.valid_loss,
|
||||
params.batch_idx_train,
|
||||
)
|
||||
|
||||
params.train_loss = tot_loss / tot_frames
|
||||
|
||||
|
||||
1045
icefall/decode.py
1045
icefall/decode.py
File diff suppressed because it is too large
Load Diff
@ -106,7 +106,7 @@ class CtcTrainingGraphCompiler(object):
|
||||
word_ids_list = []
|
||||
for text in texts:
|
||||
word_ids = []
|
||||
for word in text.split(" "):
|
||||
for word in text.split():
|
||||
if word in self.word_table:
|
||||
word_ids.append(self.word_table[word])
|
||||
else:
|
||||
|
||||
@ -157,7 +157,7 @@ class BpeLexicon(Lexicon):
|
||||
lang_dir / "lexicon.txt"
|
||||
)
|
||||
|
||||
def convert_lexicon_to_ragged(self, filename: str) -> k2.RaggedInt:
|
||||
def convert_lexicon_to_ragged(self, filename: str) -> k2.RaggedTensor:
|
||||
"""Read a BPE lexicon from file and convert it to a
|
||||
k2 ragged tensor.
|
||||
|
||||
@ -200,19 +200,18 @@ class BpeLexicon(Lexicon):
|
||||
)
|
||||
values = torch.tensor(token_ids, dtype=torch.int32)
|
||||
|
||||
return k2.RaggedInt(shape, values)
|
||||
return k2.RaggedTensor(shape, values)
|
||||
|
||||
def words_to_piece_ids(self, words: List[str]) -> k2.RaggedInt:
|
||||
def words_to_piece_ids(self, words: List[str]) -> k2.RaggedTensor:
|
||||
"""Convert a list of words to a ragged tensor contained
|
||||
word piece IDs.
|
||||
"""
|
||||
word_ids = [self.word_table[w] for w in words]
|
||||
word_ids = torch.tensor(word_ids, dtype=torch.int32)
|
||||
|
||||
ragged, _ = k2.ragged.index(
|
||||
self.ragged_lexicon,
|
||||
ragged, _ = self.ragged_lexicon.index(
|
||||
indexes=word_ids,
|
||||
need_value_indexes=False,
|
||||
axis=0,
|
||||
need_value_indexes=False,
|
||||
)
|
||||
return ragged
|
||||
|
||||
@ -26,7 +26,6 @@ from pathlib import Path
|
||||
from typing import Dict, Iterable, List, TextIO, Tuple, Union
|
||||
|
||||
import k2
|
||||
import k2.ragged as k2r
|
||||
import kaldialign
|
||||
import torch
|
||||
import torch.distributed as dist
|
||||
@ -147,12 +146,20 @@ def get_env_info():
|
||||
}
|
||||
|
||||
|
||||
# See
|
||||
# https://stackoverflow.com/questions/4984647/accessing-dict-keys-like-an-attribute # noqa
|
||||
class AttributeDict(dict):
|
||||
__slots__ = ()
|
||||
__getattr__ = dict.__getitem__
|
||||
__setattr__ = dict.__setitem__
|
||||
def __getattr__(self, key):
|
||||
if key in self:
|
||||
return self[key]
|
||||
raise AttributeError(f"No such attribute '{key}'")
|
||||
|
||||
def __setattr__(self, key, value):
|
||||
self[key] = value
|
||||
|
||||
def __delattr__(self, key):
|
||||
if key in self:
|
||||
del self[key]
|
||||
return
|
||||
raise AttributeError(f"No such attribute '{key}'")
|
||||
|
||||
|
||||
def encode_supervisions(
|
||||
@ -187,7 +194,9 @@ def encode_supervisions(
|
||||
return supervision_segments, texts
|
||||
|
||||
|
||||
def get_texts(best_paths: k2.Fsa) -> List[List[int]]:
|
||||
def get_texts(
|
||||
best_paths: k2.Fsa, return_ragged: bool = False
|
||||
) -> Union[List[List[int]], k2.RaggedTensor]:
|
||||
"""Extract the texts (as word IDs) from the best-path FSAs.
|
||||
Args:
|
||||
best_paths:
|
||||
@ -195,30 +204,35 @@ def get_texts(best_paths: k2.Fsa) -> List[List[int]]:
|
||||
containing multiple FSAs, which is expected to be the result
|
||||
of k2.shortest_path (otherwise the returned values won't
|
||||
be meaningful).
|
||||
return_ragged:
|
||||
True to return a ragged tensor with two axes [utt][word_id].
|
||||
False to return a list-of-list word IDs.
|
||||
Returns:
|
||||
Returns a list of lists of int, containing the label sequences we
|
||||
decoded.
|
||||
"""
|
||||
if isinstance(best_paths.aux_labels, k2.RaggedInt):
|
||||
if isinstance(best_paths.aux_labels, k2.RaggedTensor):
|
||||
# remove 0's and -1's.
|
||||
aux_labels = k2r.remove_values_leq(best_paths.aux_labels, 0)
|
||||
aux_shape = k2r.compose_ragged_shapes(
|
||||
best_paths.arcs.shape(), aux_labels.shape()
|
||||
)
|
||||
aux_labels = best_paths.aux_labels.remove_values_leq(0)
|
||||
# TODO: change arcs.shape() to arcs.shape
|
||||
aux_shape = best_paths.arcs.shape().compose(aux_labels.shape)
|
||||
|
||||
# remove the states and arcs axes.
|
||||
aux_shape = k2r.remove_axis(aux_shape, 1)
|
||||
aux_shape = k2r.remove_axis(aux_shape, 1)
|
||||
aux_labels = k2.RaggedInt(aux_shape, aux_labels.values())
|
||||
aux_shape = aux_shape.remove_axis(1)
|
||||
aux_shape = aux_shape.remove_axis(1)
|
||||
aux_labels = k2.RaggedTensor(aux_shape, aux_labels.values)
|
||||
else:
|
||||
# remove axis corresponding to states.
|
||||
aux_shape = k2r.remove_axis(best_paths.arcs.shape(), 1)
|
||||
aux_labels = k2.RaggedInt(aux_shape, best_paths.aux_labels)
|
||||
aux_shape = best_paths.arcs.shape().remove_axis(1)
|
||||
aux_labels = k2.RaggedTensor(aux_shape, best_paths.aux_labels)
|
||||
# remove 0's and -1's.
|
||||
aux_labels = k2r.remove_values_leq(aux_labels, 0)
|
||||
aux_labels = aux_labels.remove_values_leq(0)
|
||||
|
||||
assert aux_labels.num_axes() == 2
|
||||
return k2r.to_list(aux_labels)
|
||||
assert aux_labels.num_axes == 2
|
||||
if return_ragged:
|
||||
return aux_labels
|
||||
else:
|
||||
return aux_labels.tolist()
|
||||
|
||||
|
||||
def store_transcripts(
|
||||
|
||||
@ -16,9 +16,10 @@
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
from pathlib import Path
|
||||
|
||||
from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
|
||||
from icefall.lexicon import BpeLexicon
|
||||
from pathlib import Path
|
||||
|
||||
|
||||
def test():
|
||||
|
||||
62
test/test_decode.py
Normal file
62
test/test_decode.py
Normal file
@ -0,0 +1,62 @@
|
||||
#!/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.
|
||||
|
||||
"""
|
||||
You can run this file in one of the two ways:
|
||||
|
||||
(1) cd icefall; pytest test/test_decode.py
|
||||
(2) cd icefall; ./test/test_decode.py
|
||||
"""
|
||||
|
||||
import k2
|
||||
from icefall.decode import Nbest
|
||||
|
||||
|
||||
def test_nbest_from_lattice():
|
||||
s = """
|
||||
0 1 1 10 0.1
|
||||
0 1 5 10 0.11
|
||||
0 1 2 20 0.2
|
||||
1 2 3 30 0.3
|
||||
1 2 4 40 0.4
|
||||
2 3 -1 -1 0.5
|
||||
3
|
||||
"""
|
||||
lattice = k2.Fsa.from_str(s, acceptor=False)
|
||||
lattice = k2.Fsa.from_fsas([lattice, lattice])
|
||||
|
||||
nbest = Nbest.from_lattice(
|
||||
lattice=lattice,
|
||||
num_paths=10,
|
||||
use_double_scores=True,
|
||||
lattice_score_scale=0.5,
|
||||
)
|
||||
# each lattice has only 4 distinct paths that have different word sequences:
|
||||
# 10->30
|
||||
# 10->40
|
||||
# 20->30
|
||||
# 20->40
|
||||
#
|
||||
# So there should be only 4 paths for each lattice in the Nbest object
|
||||
assert nbest.fsa.shape[0] == 4 * 2
|
||||
assert nbest.shape.row_splits(1).tolist() == [0, 4, 8]
|
||||
|
||||
nbest2 = nbest.intersect(lattice)
|
||||
tot_scores = nbest2.tot_scores()
|
||||
argmax = tot_scores.argmax()
|
||||
best_path = k2.index_fsa(nbest2.fsa, argmax)
|
||||
print(best_path[0])
|
||||
@ -60,7 +60,7 @@ def test_get_texts_ragged():
|
||||
4
|
||||
"""
|
||||
)
|
||||
fsa1.aux_labels = k2.RaggedInt("[ [1 3 0 2] [] [4 0 1] [-1]]")
|
||||
fsa1.aux_labels = k2.RaggedTensor("[ [1 3 0 2] [] [4 0 1] [-1]]")
|
||||
|
||||
fsa2 = k2.Fsa.from_str(
|
||||
"""
|
||||
@ -70,7 +70,7 @@ def test_get_texts_ragged():
|
||||
3
|
||||
"""
|
||||
)
|
||||
fsa2.aux_labels = k2.RaggedInt("[[3 0 5 0 8] [0 9 7 0] [-1]]")
|
||||
fsa2.aux_labels = k2.RaggedTensor("[[3 0 5 0 8] [0 9 7 0] [-1]]")
|
||||
fsas = k2.Fsa.from_fsas([fsa1, fsa2])
|
||||
texts = get_texts(fsas)
|
||||
assert texts == [[1, 3, 2, 4, 1], [3, 5, 8, 9, 7]]
|
||||
@ -108,3 +108,14 @@ def test_attribute_dict():
|
||||
assert s["b"] == 20
|
||||
s.c = 100
|
||||
assert s["c"] == 100
|
||||
assert hasattr(s, "a")
|
||||
assert hasattr(s, "b")
|
||||
assert getattr(s, "a") == 10
|
||||
del s.a
|
||||
assert hasattr(s, "a") is False
|
||||
setattr(s, "c", 100)
|
||||
s.c = 100
|
||||
try:
|
||||
del s.a
|
||||
except AttributeError as ex:
|
||||
print(f"Caught exception: {ex}")
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user