Spaces:
Running
on
A10G
Running
on
A10G
Stereo demo update (#60)
Browse files- updated demo (a16e65ea0e58cfa993cbd11f06f2ded8eafb559c)
Co-authored-by: Alexandre Défossez <adefossez@users.noreply.huggingface.co>
This view is limited to 50 files because it contains too many changes.
See raw diff
- .github/actions/audiocraft_build/action.yml +2 -0
- .github/workflows/audiocraft_docs.yml +3 -3
- .github/workflows/audiocraft_tests.yml +6 -1
- .gitignore +8 -1
- CHANGELOG.md +31 -1
- CONTRIBUTING.md +2 -2
- LICENSE_weights +399 -157
- MANIFEST.in +7 -0
- Makefile +23 -4
- README.md +43 -83
- assets/a_duck_quacking_as_birds_chirp_and_a_pigeon_cooing.mp3 +0 -0
- assets/sirens_and_a_humming_engine_approach_and_pass.mp3 +0 -0
- audiocraft/__init__.py +17 -1
- audiocraft/adversarial/__init__.py +22 -0
- audiocraft/adversarial/discriminators/__init__.py +10 -0
- audiocraft/adversarial/discriminators/base.py +34 -0
- audiocraft/adversarial/discriminators/mpd.py +106 -0
- audiocraft/adversarial/discriminators/msd.py +126 -0
- audiocraft/adversarial/discriminators/msstftd.py +134 -0
- audiocraft/adversarial/losses.py +228 -0
- audiocraft/data/__init__.py +3 -1
- audiocraft/data/audio.py +37 -21
- audiocraft/data/audio_dataset.py +93 -31
- audiocraft/data/audio_utils.py +12 -10
- audiocraft/data/info_audio_dataset.py +110 -0
- audiocraft/data/music_dataset.py +270 -0
- audiocraft/data/sound_dataset.py +330 -0
- audiocraft/data/zip.py +8 -6
- audiocraft/environment.py +176 -0
- audiocraft/grids/__init__.py +6 -0
- audiocraft/grids/_base_explorers.py +80 -0
- audiocraft/grids/audiogen/__init__.py +6 -0
- audiocraft/grids/audiogen/audiogen_base_16khz.py +23 -0
- audiocraft/grids/audiogen/audiogen_pretrained_16khz_eval.py +68 -0
- audiocraft/grids/compression/__init__.py +6 -0
- audiocraft/grids/compression/_explorers.py +55 -0
- audiocraft/grids/compression/debug.py +31 -0
- audiocraft/grids/compression/encodec_audiogen_16khz.py +29 -0
- audiocraft/grids/compression/encodec_base_24khz.py +28 -0
- audiocraft/grids/compression/encodec_musicgen_32khz.py +34 -0
- audiocraft/grids/diffusion/4_bands_base_32khz.py +27 -0
- audiocraft/grids/diffusion/__init__.py +6 -0
- audiocraft/grids/diffusion/_explorers.py +66 -0
- audiocraft/grids/musicgen/__init__.py +6 -0
- audiocraft/grids/musicgen/_explorers.py +93 -0
- audiocraft/grids/musicgen/musicgen_base_32khz.py +43 -0
- audiocraft/grids/musicgen/musicgen_base_cached_32khz.py +67 -0
- audiocraft/grids/musicgen/musicgen_clapemb_32khz.py +32 -0
- audiocraft/grids/musicgen/musicgen_melody_32khz.py +65 -0
- audiocraft/grids/musicgen/musicgen_pretrained_32khz_eval.py +99 -0
.github/actions/audiocraft_build/action.yml
CHANGED
@@ -21,6 +21,8 @@ runs:
|
|
21 |
python3 -m venv env
|
22 |
. env/bin/activate
|
23 |
python -m pip install --upgrade pip
|
|
|
|
|
24 |
pip install -e '.[dev]'
|
25 |
- name: System Dependencies
|
26 |
shell: bash
|
|
|
21 |
python3 -m venv env
|
22 |
. env/bin/activate
|
23 |
python -m pip install --upgrade pip
|
24 |
+
pip install torch torchvision torchaudio
|
25 |
+
pip install xformers
|
26 |
pip install -e '.[dev]'
|
27 |
- name: System Dependencies
|
28 |
shell: bash
|
.github/workflows/audiocraft_docs.yml
CHANGED
@@ -23,9 +23,9 @@ jobs:
|
|
23 |
- name: Make docs
|
24 |
run: |
|
25 |
. env/bin/activate
|
26 |
-
make
|
27 |
-
git add -f
|
28 |
-
git commit -m
|
29 |
|
30 |
- name: Push branch
|
31 |
run: |
|
|
|
23 |
- name: Make docs
|
24 |
run: |
|
25 |
. env/bin/activate
|
26 |
+
make api_docs
|
27 |
+
git add -f api_docs
|
28 |
+
git commit -m api_docs
|
29 |
|
30 |
- name: Push branch
|
31 |
run: |
|
.github/workflows/audiocraft_tests.yml
CHANGED
@@ -12,6 +12,11 @@ jobs:
|
|
12 |
steps:
|
13 |
- uses: actions/checkout@v2
|
14 |
- uses: ./.github/actions/audiocraft_build
|
15 |
-
-
|
|
|
16 |
. env/bin/activate
|
17 |
make tests
|
|
|
|
|
|
|
|
|
|
12 |
steps:
|
13 |
- uses: actions/checkout@v2
|
14 |
- uses: ./.github/actions/audiocraft_build
|
15 |
+
- name: Run unit tests
|
16 |
+
run: |
|
17 |
. env/bin/activate
|
18 |
make tests
|
19 |
+
- name: Run integration tests
|
20 |
+
run: |
|
21 |
+
. env/bin/activate
|
22 |
+
make tests_integ
|
.gitignore
CHANGED
@@ -35,7 +35,7 @@ wheels/
|
|
35 |
.coverage
|
36 |
|
37 |
# docs
|
38 |
-
/
|
39 |
|
40 |
# dotenv
|
41 |
.env
|
@@ -46,6 +46,13 @@ wheels/
|
|
46 |
venv/
|
47 |
ENV/
|
48 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
# personal notebooks & scripts
|
50 |
*/local_scripts
|
51 |
*/notes
|
|
|
35 |
.coverage
|
36 |
|
37 |
# docs
|
38 |
+
/api_docs
|
39 |
|
40 |
# dotenv
|
41 |
.env
|
|
|
46 |
venv/
|
47 |
ENV/
|
48 |
|
49 |
+
# egs with manifest files
|
50 |
+
egs/*
|
51 |
+
!egs/example
|
52 |
+
# local datasets
|
53 |
+
dataset/*
|
54 |
+
!dataset/example
|
55 |
+
|
56 |
# personal notebooks & scripts
|
57 |
*/local_scripts
|
58 |
*/notes
|
CHANGELOG.md
CHANGED
@@ -4,7 +4,37 @@ All notable changes to this project will be documented in this file.
|
|
4 |
|
5 |
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
|
6 |
|
7 |
-
## [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
Improved demo, fixed top p (thanks @jnordberg).
|
10 |
|
|
|
4 |
|
5 |
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
|
6 |
|
7 |
+
## [1.2.0a] - TBD
|
8 |
+
|
9 |
+
Adding stereo models.
|
10 |
+
|
11 |
+
|
12 |
+
## [1.1.0] - 2023-11-06
|
13 |
+
|
14 |
+
Not using torchaudio anymore when writing audio files, relying instead directly on the commandline ffmpeg. Also not using it anymore for reading audio files, for similar reasons.
|
15 |
+
|
16 |
+
Fixed DAC support with non default number of codebooks.
|
17 |
+
|
18 |
+
Fixed bug when `two_step_cfg` was overriden when calling `generate()`.
|
19 |
+
|
20 |
+
Fixed samples being always prompted with audio, rather than having both prompted and unprompted.
|
21 |
+
|
22 |
+
**Backward incompatible change:** A `torch.no_grad` around the computation of the conditioning made its way in the public release.
|
23 |
+
The released models were trained without this. Those impact linear layers applied to the output of the T5 or melody conditioners.
|
24 |
+
We removed it, so you might need to retrain models.
|
25 |
+
|
26 |
+
**Backward incompatible change:** Fixing wrong sample rate in CLAP (WARNING if you trained model with CLAP before).
|
27 |
+
|
28 |
+
**Backward incompatible change:** Renamed VALLEPattern to CoarseFirstPattern, as it was wrongly named. Probably no one
|
29 |
+
retrained a model with this pattern, so hopefully this won't impact you!
|
30 |
+
|
31 |
+
|
32 |
+
## [1.0.0] - 2023-09-07
|
33 |
+
|
34 |
+
Major revision, added training code for EnCodec, AudioGen, MusicGen, and MultiBandDiffusion.
|
35 |
+
Added pretrained model for AudioGen and MultiBandDiffusion.
|
36 |
+
|
37 |
+
## [0.0.2] - 2023-08-01
|
38 |
|
39 |
Improved demo, fixed top p (thanks @jnordberg).
|
40 |
|
CONTRIBUTING.md
CHANGED
@@ -1,11 +1,11 @@
|
|
1 |
-
# Contributing to
|
2 |
|
3 |
We want to make contributing to this project as easy and transparent as
|
4 |
possible.
|
5 |
|
6 |
## Pull Requests
|
7 |
|
8 |
-
|
9 |
Therefore, we do not plan on accepting many pull requests for new features.
|
10 |
We certainly welcome them for bug fixes.
|
11 |
|
|
|
1 |
+
# Contributing to AudioCraft
|
2 |
|
3 |
We want to make contributing to this project as easy and transparent as
|
4 |
possible.
|
5 |
|
6 |
## Pull Requests
|
7 |
|
8 |
+
AudioCraft is the implementation of a research paper.
|
9 |
Therefore, we do not plan on accepting many pull requests for new features.
|
10 |
We certainly welcome them for bug fixes.
|
11 |
|
LICENSE_weights
CHANGED
@@ -1,157 +1,399 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Attribution-NonCommercial 4.0 International
|
2 |
+
|
3 |
+
=======================================================================
|
4 |
+
|
5 |
+
Creative Commons Corporation ("Creative Commons") is not a law firm and
|
6 |
+
does not provide legal services or legal advice. Distribution of
|
7 |
+
Creative Commons public licenses does not create a lawyer-client or
|
8 |
+
other relationship. Creative Commons makes its licenses and related
|
9 |
+
information available on an "as-is" basis. Creative Commons gives no
|
10 |
+
warranties regarding its licenses, any material licensed under their
|
11 |
+
terms and conditions, or any related information. Creative Commons
|
12 |
+
disclaims all liability for damages resulting from their use to the
|
13 |
+
fullest extent possible.
|
14 |
+
|
15 |
+
Using Creative Commons Public Licenses
|
16 |
+
|
17 |
+
Creative Commons public licenses provide a standard set of terms and
|
18 |
+
conditions that creators and other rights holders may use to share
|
19 |
+
original works of authorship and other material subject to copyright
|
20 |
+
and certain other rights specified in the public license below. The
|
21 |
+
following considerations are for informational purposes only, are not
|
22 |
+
exhaustive, and do not form part of our licenses.
|
23 |
+
|
24 |
+
Considerations for licensors: Our public licenses are
|
25 |
+
intended for use by those authorized to give the public
|
26 |
+
permission to use material in ways otherwise restricted by
|
27 |
+
copyright and certain other rights. Our licenses are
|
28 |
+
irrevocable. Licensors should read and understand the terms
|
29 |
+
and conditions of the license they choose before applying it.
|
30 |
+
Licensors should also secure all rights necessary before
|
31 |
+
applying our licenses so that the public can reuse the
|
32 |
+
material as expected. Licensors should clearly mark any
|
33 |
+
material not subject to the license. This includes other CC-
|
34 |
+
licensed material, or material used under an exception or
|
35 |
+
limitation to copyright. More considerations for licensors:
|
36 |
+
wiki.creativecommons.org/Considerations_for_licensors
|
37 |
+
|
38 |
+
Considerations for the public: By using one of our public
|
39 |
+
licenses, a licensor grants the public permission to use the
|
40 |
+
licensed material under specified terms and conditions. If
|
41 |
+
the licensor's permission is not necessary for any reason--for
|
42 |
+
example, because of any applicable exception or limitation to
|
43 |
+
copyright--then that use is not regulated by the license. Our
|
44 |
+
licenses grant only permissions under copyright and certain
|
45 |
+
other rights that a licensor has authority to grant. Use of
|
46 |
+
the licensed material may still be restricted for other
|
47 |
+
reasons, including because others have copyright or other
|
48 |
+
rights in the material. A licensor may make special requests,
|
49 |
+
such as asking that all changes be marked or described.
|
50 |
+
Although not required by our licenses, you are encouraged to
|
51 |
+
respect those requests where reasonable. More_considerations
|
52 |
+
for the public:
|
53 |
+
wiki.creativecommons.org/Considerations_for_licensees
|
54 |
+
|
55 |
+
=======================================================================
|
56 |
+
|
57 |
+
Creative Commons Attribution-NonCommercial 4.0 International Public
|
58 |
+
License
|
59 |
+
|
60 |
+
By exercising the Licensed Rights (defined below), You accept and agree
|
61 |
+
to be bound by the terms and conditions of this Creative Commons
|
62 |
+
Attribution-NonCommercial 4.0 International Public License ("Public
|
63 |
+
License"). To the extent this Public License may be interpreted as a
|
64 |
+
contract, You are granted the Licensed Rights in consideration of Your
|
65 |
+
acceptance of these terms and conditions, and the Licensor grants You
|
66 |
+
such rights in consideration of benefits the Licensor receives from
|
67 |
+
making the Licensed Material available under these terms and
|
68 |
+
conditions.
|
69 |
+
|
70 |
+
Section 1 -- Definitions.
|
71 |
+
|
72 |
+
a. Adapted Material means material subject to Copyright and Similar
|
73 |
+
Rights that is derived from or based upon the Licensed Material
|
74 |
+
and in which the Licensed Material is translated, altered,
|
75 |
+
arranged, transformed, or otherwise modified in a manner requiring
|
76 |
+
permission under the Copyright and Similar Rights held by the
|
77 |
+
Licensor. For purposes of this Public License, where the Licensed
|
78 |
+
Material is a musical work, performance, or sound recording,
|
79 |
+
Adapted Material is always produced where the Licensed Material is
|
80 |
+
synched in timed relation with a moving image.
|
81 |
+
|
82 |
+
b. Adapter's License means the license You apply to Your Copyright
|
83 |
+
and Similar Rights in Your contributions to Adapted Material in
|
84 |
+
accordance with the terms and conditions of this Public License.
|
85 |
+
|
86 |
+
c. Copyright and Similar Rights means copyright and/or similar rights
|
87 |
+
closely related to copyright including, without limitation,
|
88 |
+
performance, broadcast, sound recording, and Sui Generis Database
|
89 |
+
Rights, without regard to how the rights are labeled or
|
90 |
+
categorized. For purposes of this Public License, the rights
|
91 |
+
specified in Section 2(b)(1)-(2) are not Copyright and Similar
|
92 |
+
Rights.
|
93 |
+
d. Effective Technological Measures means those measures that, in the
|
94 |
+
absence of proper authority, may not be circumvented under laws
|
95 |
+
fulfilling obligations under Article 11 of the WIPO Copyright
|
96 |
+
Treaty adopted on December 20, 1996, and/or similar international
|
97 |
+
agreements.
|
98 |
+
|
99 |
+
e. Exceptions and Limitations means fair use, fair dealing, and/or
|
100 |
+
any other exception or limitation to Copyright and Similar Rights
|
101 |
+
that applies to Your use of the Licensed Material.
|
102 |
+
|
103 |
+
f. Licensed Material means the artistic or literary work, database,
|
104 |
+
or other material to which the Licensor applied this Public
|
105 |
+
License.
|
106 |
+
|
107 |
+
g. Licensed Rights means the rights granted to You subject to the
|
108 |
+
terms and conditions of this Public License, which are limited to
|
109 |
+
all Copyright and Similar Rights that apply to Your use of the
|
110 |
+
Licensed Material and that the Licensor has authority to license.
|
111 |
+
|
112 |
+
h. Licensor means the individual(s) or entity(ies) granting rights
|
113 |
+
under this Public License.
|
114 |
+
|
115 |
+
i. NonCommercial means not primarily intended for or directed towards
|
116 |
+
commercial advantage or monetary compensation. For purposes of
|
117 |
+
this Public License, the exchange of the Licensed Material for
|
118 |
+
other material subject to Copyright and Similar Rights by digital
|
119 |
+
file-sharing or similar means is NonCommercial provided there is
|
120 |
+
no payment of monetary compensation in connection with the
|
121 |
+
exchange.
|
122 |
+
|
123 |
+
j. Share means to provide material to the public by any means or
|
124 |
+
process that requires permission under the Licensed Rights, such
|
125 |
+
as reproduction, public display, public performance, distribution,
|
126 |
+
dissemination, communication, or importation, and to make material
|
127 |
+
available to the public including in ways that members of the
|
128 |
+
public may access the material from a place and at a time
|
129 |
+
individually chosen by them.
|
130 |
+
|
131 |
+
k. Sui Generis Database Rights means rights other than copyright
|
132 |
+
resulting from Directive 96/9/EC of the European Parliament and of
|
133 |
+
the Council of 11 March 1996 on the legal protection of databases,
|
134 |
+
as amended and/or succeeded, as well as other essentially
|
135 |
+
equivalent rights anywhere in the world.
|
136 |
+
|
137 |
+
l. You means the individual or entity exercising the Licensed Rights
|
138 |
+
under this Public License. Your has a corresponding meaning.
|
139 |
+
|
140 |
+
Section 2 -- Scope.
|
141 |
+
|
142 |
+
a. License grant.
|
143 |
+
|
144 |
+
1. Subject to the terms and conditions of this Public License,
|
145 |
+
the Licensor hereby grants You a worldwide, royalty-free,
|
146 |
+
non-sublicensable, non-exclusive, irrevocable license to
|
147 |
+
exercise the Licensed Rights in the Licensed Material to:
|
148 |
+
|
149 |
+
a. reproduce and Share the Licensed Material, in whole or
|
150 |
+
in part, for NonCommercial purposes only; and
|
151 |
+
|
152 |
+
b. produce, reproduce, and Share Adapted Material for
|
153 |
+
NonCommercial purposes only.
|
154 |
+
|
155 |
+
2. Exceptions and Limitations. For the avoidance of doubt, where
|
156 |
+
Exceptions and Limitations apply to Your use, this Public
|
157 |
+
License does not apply, and You do not need to comply with
|
158 |
+
its terms and conditions.
|
159 |
+
|
160 |
+
3. Term. The term of this Public License is specified in Section
|
161 |
+
6(a).
|
162 |
+
|
163 |
+
4. Media and formats; technical modifications allowed. The
|
164 |
+
Licensor authorizes You to exercise the Licensed Rights in
|
165 |
+
all media and formats whether now known or hereafter created,
|
166 |
+
and to make technical modifications necessary to do so. The
|
167 |
+
Licensor waives and/or agrees not to assert any right or
|
168 |
+
authority to forbid You from making technical modifications
|
169 |
+
necessary to exercise the Licensed Rights, including
|
170 |
+
technical modifications necessary to circumvent Effective
|
171 |
+
Technological Measures. For purposes of this Public License,
|
172 |
+
simply making modifications authorized by this Section 2(a)
|
173 |
+
(4) never produces Adapted Material.
|
174 |
+
|
175 |
+
5. Downstream recipients.
|
176 |
+
|
177 |
+
a. Offer from the Licensor -- Licensed Material. Every
|
178 |
+
recipient of the Licensed Material automatically
|
179 |
+
receives an offer from the Licensor to exercise the
|
180 |
+
Licensed Rights under the terms and conditions of this
|
181 |
+
Public License.
|
182 |
+
|
183 |
+
b. No downstream restrictions. You may not offer or impose
|
184 |
+
any additional or different terms or conditions on, or
|
185 |
+
apply any Effective Technological Measures to, the
|
186 |
+
Licensed Material if doing so restricts exercise of the
|
187 |
+
Licensed Rights by any recipient of the Licensed
|
188 |
+
Material.
|
189 |
+
|
190 |
+
6. No endorsement. Nothing in this Public License constitutes or
|
191 |
+
may be construed as permission to assert or imply that You
|
192 |
+
are, or that Your use of the Licensed Material is, connected
|
193 |
+
with, or sponsored, endorsed, or granted official status by,
|
194 |
+
the Licensor or others designated to receive attribution as
|
195 |
+
provided in Section 3(a)(1)(A)(i).
|
196 |
+
|
197 |
+
b. Other rights.
|
198 |
+
|
199 |
+
1. Moral rights, such as the right of integrity, are not
|
200 |
+
licensed under this Public License, nor are publicity,
|
201 |
+
privacy, and/or other similar personality rights; however, to
|
202 |
+
the extent possible, the Licensor waives and/or agrees not to
|
203 |
+
assert any such rights held by the Licensor to the limited
|
204 |
+
extent necessary to allow You to exercise the Licensed
|
205 |
+
Rights, but not otherwise.
|
206 |
+
|
207 |
+
2. Patent and trademark rights are not licensed under this
|
208 |
+
Public License.
|
209 |
+
|
210 |
+
3. To the extent possible, the Licensor waives any right to
|
211 |
+
collect royalties from You for the exercise of the Licensed
|
212 |
+
Rights, whether directly or through a collecting society
|
213 |
+
under any voluntary or waivable statutory or compulsory
|
214 |
+
licensing scheme. In all other cases the Licensor expressly
|
215 |
+
reserves any right to collect such royalties, including when
|
216 |
+
the Licensed Material is used other than for NonCommercial
|
217 |
+
purposes.
|
218 |
+
|
219 |
+
Section 3 -- License Conditions.
|
220 |
+
|
221 |
+
Your exercise of the Licensed Rights is expressly made subject to the
|
222 |
+
following conditions.
|
223 |
+
|
224 |
+
a. Attribution.
|
225 |
+
|
226 |
+
1. If You Share the Licensed Material (including in modified
|
227 |
+
form), You must:
|
228 |
+
|
229 |
+
a. retain the following if it is supplied by the Licensor
|
230 |
+
with the Licensed Material:
|
231 |
+
|
232 |
+
i. identification of the creator(s) of the Licensed
|
233 |
+
Material and any others designated to receive
|
234 |
+
attribution, in any reasonable manner requested by
|
235 |
+
the Licensor (including by pseudonym if
|
236 |
+
designated);
|
237 |
+
|
238 |
+
ii. a copyright notice;
|
239 |
+
|
240 |
+
iii. a notice that refers to this Public License;
|
241 |
+
|
242 |
+
iv. a notice that refers to the disclaimer of
|
243 |
+
warranties;
|
244 |
+
|
245 |
+
v. a URI or hyperlink to the Licensed Material to the
|
246 |
+
extent reasonably practicable;
|
247 |
+
|
248 |
+
b. indicate if You modified the Licensed Material and
|
249 |
+
retain an indication of any previous modifications; and
|
250 |
+
|
251 |
+
c. indicate the Licensed Material is licensed under this
|
252 |
+
Public License, and include the text of, or the URI or
|
253 |
+
hyperlink to, this Public License.
|
254 |
+
|
255 |
+
2. You may satisfy the conditions in Section 3(a)(1) in any
|
256 |
+
reasonable manner based on the medium, means, and context in
|
257 |
+
which You Share the Licensed Material. For example, it may be
|
258 |
+
reasonable to satisfy the conditions by providing a URI or
|
259 |
+
hyperlink to a resource that includes the required
|
260 |
+
information.
|
261 |
+
|
262 |
+
3. If requested by the Licensor, You must remove any of the
|
263 |
+
information required by Section 3(a)(1)(A) to the extent
|
264 |
+
reasonably practicable.
|
265 |
+
|
266 |
+
4. If You Share Adapted Material You produce, the Adapter's
|
267 |
+
License You apply must not prevent recipients of the Adapted
|
268 |
+
Material from complying with this Public License.
|
269 |
+
|
270 |
+
Section 4 -- Sui Generis Database Rights.
|
271 |
+
|
272 |
+
Where the Licensed Rights include Sui Generis Database Rights that
|
273 |
+
apply to Your use of the Licensed Material:
|
274 |
+
|
275 |
+
a. for the avoidance of doubt, Section 2(a)(1) grants You the right
|
276 |
+
to extract, reuse, reproduce, and Share all or a substantial
|
277 |
+
portion of the contents of the database for NonCommercial purposes
|
278 |
+
only;
|
279 |
+
|
280 |
+
b. if You include all or a substantial portion of the database
|
281 |
+
contents in a database in which You have Sui Generis Database
|
282 |
+
Rights, then the database in which You have Sui Generis Database
|
283 |
+
Rights (but not its individual contents) is Adapted Material; and
|
284 |
+
|
285 |
+
c. You must comply with the conditions in Section 3(a) if You Share
|
286 |
+
all or a substantial portion of the contents of the database.
|
287 |
+
|
288 |
+
For the avoidance of doubt, this Section 4 supplements and does not
|
289 |
+
replace Your obligations under this Public License where the Licensed
|
290 |
+
Rights include other Copyright and Similar Rights.
|
291 |
+
|
292 |
+
Section 5 -- Disclaimer of Warranties and Limitation of Liability.
|
293 |
+
|
294 |
+
a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
|
295 |
+
EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
|
296 |
+
AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
|
297 |
+
ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
|
298 |
+
IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
|
299 |
+
WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
|
300 |
+
PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
|
301 |
+
ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
|
302 |
+
KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
|
303 |
+
ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
|
304 |
+
|
305 |
+
b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
|
306 |
+
TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
|
307 |
+
NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
|
308 |
+
INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
|
309 |
+
COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
|
310 |
+
USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
|
311 |
+
ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
|
312 |
+
DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
|
313 |
+
IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
|
314 |
+
|
315 |
+
c. The disclaimer of warranties and limitation of liability provided
|
316 |
+
above shall be interpreted in a manner that, to the extent
|
317 |
+
possible, most closely approximates an absolute disclaimer and
|
318 |
+
waiver of all liability.
|
319 |
+
|
320 |
+
Section 6 -- Term and Termination.
|
321 |
+
|
322 |
+
a. This Public License applies for the term of the Copyright and
|
323 |
+
Similar Rights licensed here. However, if You fail to comply with
|
324 |
+
this Public License, then Your rights under this Public License
|
325 |
+
terminate automatically.
|
326 |
+
|
327 |
+
b. Where Your right to use the Licensed Material has terminated under
|
328 |
+
Section 6(a), it reinstates:
|
329 |
+
|
330 |
+
1. automatically as of the date the violation is cured, provided
|
331 |
+
it is cured within 30 days of Your discovery of the
|
332 |
+
violation; or
|
333 |
+
|
334 |
+
2. upon express reinstatement by the Licensor.
|
335 |
+
|
336 |
+
For the avoidance of doubt, this Section 6(b) does not affect any
|
337 |
+
right the Licensor may have to seek remedies for Your violations
|
338 |
+
of this Public License.
|
339 |
+
|
340 |
+
c. For the avoidance of doubt, the Licensor may also offer the
|
341 |
+
Licensed Material under separate terms or conditions or stop
|
342 |
+
distributing the Licensed Material at any time; however, doing so
|
343 |
+
will not terminate this Public License.
|
344 |
+
|
345 |
+
d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
|
346 |
+
License.
|
347 |
+
|
348 |
+
Section 7 -- Other Terms and Conditions.
|
349 |
+
|
350 |
+
a. The Licensor shall not be bound by any additional or different
|
351 |
+
terms or conditions communicated by You unless expressly agreed.
|
352 |
+
|
353 |
+
b. Any arrangements, understandings, or agreements regarding the
|
354 |
+
Licensed Material not stated herein are separate from and
|
355 |
+
independent of the terms and conditions of this Public License.
|
356 |
+
|
357 |
+
Section 8 -- Interpretation.
|
358 |
+
|
359 |
+
a. For the avoidance of doubt, this Public License does not, and
|
360 |
+
shall not be interpreted to, reduce, limit, restrict, or impose
|
361 |
+
conditions on any use of the Licensed Material that could lawfully
|
362 |
+
be made without permission under this Public License.
|
363 |
+
|
364 |
+
b. To the extent possible, if any provision of this Public License is
|
365 |
+
deemed unenforceable, it shall be automatically reformed to the
|
366 |
+
minimum extent necessary to make it enforceable. If the provision
|
367 |
+
cannot be reformed, it shall be severed from this Public License
|
368 |
+
without affecting the enforceability of the remaining terms and
|
369 |
+
conditions.
|
370 |
+
|
371 |
+
c. No term or condition of this Public License will be waived and no
|
372 |
+
failure to comply consented to unless expressly agreed to by the
|
373 |
+
Licensor.
|
374 |
+
|
375 |
+
d. Nothing in this Public License constitutes or may be interpreted
|
376 |
+
as a limitation upon, or waiver of, any privileges and immunities
|
377 |
+
that apply to the Licensor or You, including from the legal
|
378 |
+
processes of any jurisdiction or authority.
|
379 |
+
|
380 |
+
=======================================================================
|
381 |
+
|
382 |
+
Creative Commons is not a party to its public
|
383 |
+
licenses. Notwithstanding, Creative Commons may elect to apply one of
|
384 |
+
its public licenses to material it publishes and in those instances
|
385 |
+
will be considered the “Licensor.” The text of the Creative Commons
|
386 |
+
public licenses is dedicated to the public domain under the CC0 Public
|
387 |
+
Domain Dedication. Except for the limited purpose of indicating that
|
388 |
+
material is shared under a Creative Commons public license or as
|
389 |
+
otherwise permitted by the Creative Commons policies published at
|
390 |
+
creativecommons.org/policies, Creative Commons does not authorize the
|
391 |
+
use of the trademark "Creative Commons" or any other trademark or logo
|
392 |
+
of Creative Commons without its prior written consent including,
|
393 |
+
without limitation, in connection with any unauthorized modifications
|
394 |
+
to any of its public licenses or any other arrangements,
|
395 |
+
understandings, or agreements concerning use of licensed material. For
|
396 |
+
the avoidance of doubt, this paragraph does not form part of the
|
397 |
+
public licenses.
|
398 |
+
|
399 |
+
Creative Commons may be contacted at creativecommons.org.
|
MANIFEST.in
CHANGED
@@ -6,3 +6,10 @@ include *.ini
|
|
6 |
include requirements.txt
|
7 |
include audiocraft/py.typed
|
8 |
include assets/*.mp3
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
include requirements.txt
|
7 |
include audiocraft/py.typed
|
8 |
include assets/*.mp3
|
9 |
+
include datasets/*.mp3
|
10 |
+
recursive-include config *.yaml
|
11 |
+
recursive-include demos *.py
|
12 |
+
recursive-include demos *.ipynb
|
13 |
+
recursive-include scripts *.py
|
14 |
+
recursive-include model_cards *.md
|
15 |
+
recursive-include docs *.md
|
Makefile
CHANGED
@@ -1,3 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
default: linter tests
|
2 |
|
3 |
install:
|
@@ -10,12 +22,19 @@ linter:
|
|
10 |
|
11 |
tests:
|
12 |
coverage run -m pytest tests
|
13 |
-
coverage report
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
-
|
16 |
-
pdoc3 --html -o
|
17 |
|
18 |
dist:
|
19 |
python setup.py sdist
|
20 |
|
21 |
-
.PHONY: linter tests
|
|
|
1 |
+
INTEG=AUDIOCRAFT_DORA_DIR="/tmp/magma_$(USER)" python3 -m dora -v run --clear device=cpu dataset.num_workers=0 optim.epochs=1 \
|
2 |
+
dataset.train.num_samples=10 dataset.valid.num_samples=10 \
|
3 |
+
dataset.evaluate.num_samples=10 dataset.generate.num_samples=2 sample_rate=16000 \
|
4 |
+
logging.level=DEBUG
|
5 |
+
INTEG_COMPRESSION = $(INTEG) solver=compression/debug rvq.n_q=2 rvq.bins=48 checkpoint.save_last=true # SIG is 5091833e
|
6 |
+
INTEG_MUSICGEN = $(INTEG) solver=musicgen/debug dset=audio/example compression_model_checkpoint=//sig/5091833e \
|
7 |
+
transformer_lm.n_q=2 transformer_lm.card=48 transformer_lm.dim=16 checkpoint.save_last=false # Using compression model from 5091833e
|
8 |
+
INTEG_AUDIOGEN = $(INTEG) solver=audiogen/debug dset=audio/example compression_model_checkpoint=//sig/5091833e \
|
9 |
+
transformer_lm.n_q=2 transformer_lm.card=48 transformer_lm.dim=16 checkpoint.save_last=false # Using compression model from 5091833e
|
10 |
+
INTEG_MBD = $(INTEG) solver=diffusion/debug dset=audio/example \
|
11 |
+
checkpoint.save_last=false # Using compression model from 616d7b3c
|
12 |
+
|
13 |
default: linter tests
|
14 |
|
15 |
install:
|
|
|
22 |
|
23 |
tests:
|
24 |
coverage run -m pytest tests
|
25 |
+
coverage report
|
26 |
+
|
27 |
+
tests_integ:
|
28 |
+
$(INTEG_COMPRESSION)
|
29 |
+
$(INTEG_MBD)
|
30 |
+
$(INTEG_MUSICGEN)
|
31 |
+
$(INTEG_AUDIOGEN)
|
32 |
+
|
33 |
|
34 |
+
api_docs:
|
35 |
+
pdoc3 --html -o api_docs -f audiocraft
|
36 |
|
37 |
dist:
|
38 |
python setup.py sdist
|
39 |
|
40 |
+
.PHONY: linter tests api_docs dist
|
README.md
CHANGED
@@ -5,7 +5,7 @@ tags:
|
|
5 |
- "music generation"
|
6 |
- "language models"
|
7 |
- "LLMs"
|
8 |
-
app_file: "
|
9 |
emoji: 🎵
|
10 |
colorFrom: gray
|
11 |
colorTo: blue
|
@@ -14,33 +14,17 @@ sdk_version: 3.34.0
|
|
14 |
pinned: true
|
15 |
license: "cc-by-nc-4.0"
|
16 |
---
|
17 |
-
#
|
18 |
![docs badge](https://github.com/facebookresearch/audiocraft/workflows/audiocraft_docs/badge.svg)
|
19 |
![linter badge](https://github.com/facebookresearch/audiocraft/workflows/audiocraft_linter/badge.svg)
|
20 |
![tests badge](https://github.com/facebookresearch/audiocraft/workflows/audiocraft_tests/badge.svg)
|
21 |
|
22 |
-
|
|
|
23 |
|
24 |
-
## MusicGen
|
25 |
-
|
26 |
-
Audiocraft provides the code and models for MusicGen, [a simple and controllable model for music generation][arxiv]. MusicGen is a single stage auto-regressive
|
27 |
-
Transformer model trained over a 32kHz <a href="https://github.com/facebookresearch/encodec">EnCodec tokenizer</a> with 4 codebooks sampled at 50 Hz. Unlike existing methods like [MusicLM](https://arxiv.org/abs/2301.11325), MusicGen doesn't require a self-supervised semantic representation, and it generates
|
28 |
-
all 4 codebooks in one pass. By introducing a small delay between the codebooks, we show we can predict
|
29 |
-
them in parallel, thus having only 50 auto-regressive steps per second of audio.
|
30 |
-
Check out our [sample page][musicgen_samples] or test the available demo!
|
31 |
-
|
32 |
-
<a target="_blank" href="https://colab.research.google.com/drive/1-Xe9NCdIs2sCUbiSmwHXozK6AAhMm7_i?usp=sharing">
|
33 |
-
<img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
|
34 |
-
</a>
|
35 |
-
<a target="_blank" href="https://huggingface.co/spaces/facebook/MusicGen">
|
36 |
-
<img src="https://huggingface.co/datasets/huggingface/badges/raw/main/open-in-hf-spaces-sm.svg" alt="Open in HugginFace"/>
|
37 |
-
</a>
|
38 |
-
<br>
|
39 |
-
|
40 |
-
We use 20K hours of licensed music to train MusicGen. Specifically, we rely on an internal dataset of 10K high-quality music tracks, and on the ShutterStock and Pond5 music data.
|
41 |
|
42 |
## Installation
|
43 |
-
|
44 |
|
45 |
```shell
|
46 |
# Best to make sure you have torch installed first, in particular before installing xformers.
|
@@ -49,92 +33,68 @@ pip install 'torch>=2.0'
|
|
49 |
# Then proceed to one of the following
|
50 |
pip install -U audiocraft # stable release
|
51 |
pip install -U git+https://git@github.com/facebookresearch/audiocraft#egg=audiocraft # bleeding edge
|
52 |
-
pip install -e . # or if you cloned the repo locally
|
53 |
```
|
54 |
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
4. You can play with MusicGen by running the jupyter notebook at [`demo.ipynb`](./demo.ipynb) locally (if you have a GPU).
|
61 |
-
5. Finally, checkout [@camenduru Colab page](https://github.com/camenduru/MusicGen-colab) which is regularly
|
62 |
-
updated with contributions from @camenduru and the community.
|
63 |
-
|
64 |
-
## API
|
65 |
-
|
66 |
-
We provide a simple API and 4 pre-trained models. The pre trained models are:
|
67 |
-
- `small`: 300M model, text to music only - [🤗 Hub](https://huggingface.co/facebook/musicgen-small)
|
68 |
-
- `medium`: 1.5B model, text to music only - [🤗 Hub](https://huggingface.co/facebook/musicgen-medium)
|
69 |
-
- `melody`: 1.5B model, text to music and text+melody to music - [🤗 Hub](https://huggingface.co/facebook/musicgen-melody)
|
70 |
-
- `large`: 3.3B model, text to music only - [🤗 Hub](https://huggingface.co/facebook/musicgen-large)
|
71 |
-
|
72 |
-
We observe the best trade-off between quality and compute with the `medium` or `melody` model.
|
73 |
-
In order to use MusicGen locally **you must have a GPU**. We recommend 16GB of memory, but smaller
|
74 |
-
GPUs will be able to generate short sequences, or longer sequences with the `small` model.
|
75 |
-
|
76 |
-
**Note**: Please make sure to have [ffmpeg](https://ffmpeg.org/download.html) installed when using newer version of `torchaudio`.
|
77 |
-
You can install it with:
|
78 |
-
```
|
79 |
-
apt-get install ffmpeg
|
80 |
```
|
81 |
|
82 |
-
|
83 |
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
|
|
88 |
|
89 |
-
|
90 |
-
model.set_generation_params(duration=8) # generate 8 seconds.
|
91 |
-
wav = model.generate_unconditional(4) # generates 4 unconditional audio samples
|
92 |
-
descriptions = ['happy rock', 'energetic EDM', 'sad jazz']
|
93 |
-
wav = model.generate(descriptions) # generates 3 samples.
|
94 |
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
|
99 |
-
|
100 |
-
|
101 |
-
audio_write(f'{idx}', one_wav.cpu(), model.sample_rate, strategy="loudness", loudness_compressor=True)
|
102 |
-
```
|
103 |
|
104 |
|
105 |
-
##
|
106 |
|
107 |
-
|
108 |
|
109 |
-
## FAQ
|
110 |
|
111 |
-
|
112 |
|
113 |
-
|
114 |
|
|
|
115 |
|
116 |
-
####
|
117 |
|
118 |
-
|
|
|
|
|
119 |
|
120 |
-
#### I need help for running the demo on Colab
|
121 |
|
122 |
-
|
|
|
|
|
123 |
|
124 |
|
125 |
## Citation
|
|
|
|
|
126 |
```
|
127 |
@article{copet2023simple,
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
}
|
133 |
```
|
134 |
|
135 |
-
|
136 |
-
|
137 |
-
* The weights in this repository are released under the CC-BY-NC 4.0 license as found in the [LICENSE_weights file](LICENSE_weights).
|
138 |
-
|
139 |
-
[arxiv]: https://arxiv.org/abs/2306.05284
|
140 |
-
[musicgen_samples]: https://ai.honu.io/papers/musicgen/
|
|
|
5 |
- "music generation"
|
6 |
- "language models"
|
7 |
- "LLMs"
|
8 |
+
app_file: "demos/musicgen_app.py"
|
9 |
emoji: 🎵
|
10 |
colorFrom: gray
|
11 |
colorTo: blue
|
|
|
14 |
pinned: true
|
15 |
license: "cc-by-nc-4.0"
|
16 |
---
|
17 |
+
# AudioCraft
|
18 |
![docs badge](https://github.com/facebookresearch/audiocraft/workflows/audiocraft_docs/badge.svg)
|
19 |
![linter badge](https://github.com/facebookresearch/audiocraft/workflows/audiocraft_linter/badge.svg)
|
20 |
![tests badge](https://github.com/facebookresearch/audiocraft/workflows/audiocraft_tests/badge.svg)
|
21 |
|
22 |
+
AudioCraft is a PyTorch library for deep learning research on audio generation. AudioCraft contains inference and training code
|
23 |
+
for two state-of-the-art AI generative models producing high-quality audio: AudioGen and MusicGen.
|
24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
|
26 |
## Installation
|
27 |
+
AudioCraft requires Python 3.9, PyTorch 2.0.0. To install AudioCraft, you can run the following:
|
28 |
|
29 |
```shell
|
30 |
# Best to make sure you have torch installed first, in particular before installing xformers.
|
|
|
33 |
# Then proceed to one of the following
|
34 |
pip install -U audiocraft # stable release
|
35 |
pip install -U git+https://git@github.com/facebookresearch/audiocraft#egg=audiocraft # bleeding edge
|
36 |
+
pip install -e . # or if you cloned the repo locally (mandatory if you want to train).
|
37 |
```
|
38 |
|
39 |
+
We also recommend having `ffmpeg` installed, either through your system or Anaconda:
|
40 |
+
```bash
|
41 |
+
sudo apt-get install ffmpeg
|
42 |
+
# Or if you are using Anaconda or Miniconda
|
43 |
+
conda install "ffmpeg<5" -c conda-forge
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
```
|
45 |
|
46 |
+
## Models
|
47 |
|
48 |
+
At the moment, AudioCraft contains the training code and inference code for:
|
49 |
+
* [MusicGen](./docs/MUSICGEN.md): A state-of-the-art controllable text-to-music model.
|
50 |
+
* [AudioGen](./docs/AUDIOGEN.md): A state-of-the-art text-to-sound model.
|
51 |
+
* [EnCodec](./docs/ENCODEC.md): A state-of-the-art high fidelity neural audio codec.
|
52 |
+
* [Multi Band Diffusion](./docs/MBD.md): An EnCodec compatible decoder using diffusion.
|
53 |
|
54 |
+
## Training code
|
|
|
|
|
|
|
|
|
55 |
|
56 |
+
AudioCraft contains PyTorch components for deep learning research in audio and training pipelines for the developed models.
|
57 |
+
For a general introduction of AudioCraft design principles and instructions to develop your own training pipeline, refer to
|
58 |
+
the [AudioCraft training documentation](./docs/TRAINING.md).
|
59 |
|
60 |
+
For reproducing existing work and using the developed training pipelines, refer to the instructions for each specific model
|
61 |
+
that provides pointers to configuration, example grids and model/task-specific information and FAQ.
|
|
|
|
|
62 |
|
63 |
|
64 |
+
## API documentation
|
65 |
|
66 |
+
We provide some [API documentation](https://facebookresearch.github.io/audiocraft/api_docs/audiocraft/index.html) for AudioCraft.
|
67 |
|
|
|
68 |
|
69 |
+
## FAQ
|
70 |
|
71 |
+
#### Is the training code available?
|
72 |
|
73 |
+
Yes! We provide the training code for [EnCodec](./docs/ENCODEC.md), [MusicGen](./docs/MUSICGEN.md) and [Multi Band Diffusion](./docs/MBD.md).
|
74 |
|
75 |
+
#### Where are the models stored?
|
76 |
|
77 |
+
Hugging Face stored the model in a specific location, which can be overriden by setting the `AUDIOCRAFT_CACHE_DIR` environment variable for the AudioCraft models.
|
78 |
+
In order to change the cache location of the other Hugging Face models, please check out the [Hugging Face Transformers documentation for the cache setup](https://huggingface.co/docs/transformers/installation#cache-setup).
|
79 |
+
Finally, if you use a model that relies on Demucs (e.g. `musicgen-melody`) and want to change the download location for Demucs, refer to the [Torch Hub documentation](https://pytorch.org/docs/stable/hub.html#where-are-my-downloaded-models-saved).
|
80 |
|
|
|
81 |
|
82 |
+
## License
|
83 |
+
* The code in this repository is released under the MIT license as found in the [LICENSE file](LICENSE).
|
84 |
+
* The models weights in this repository are released under the CC-BY-NC 4.0 license as found in the [LICENSE_weights file](LICENSE_weights).
|
85 |
|
86 |
|
87 |
## Citation
|
88 |
+
|
89 |
+
For the general framework of AudioCraft, please cite the following.
|
90 |
```
|
91 |
@article{copet2023simple,
|
92 |
+
title={Simple and Controllable Music Generation},
|
93 |
+
author={Jade Copet and Felix Kreuk and Itai Gat and Tal Remez and David Kant and Gabriel Synnaeve and Yossi Adi and Alexandre Défossez},
|
94 |
+
year={2023},
|
95 |
+
journal={arXiv preprint arXiv:2306.05284},
|
96 |
}
|
97 |
```
|
98 |
|
99 |
+
When referring to a specific model, please cite as mentioned in the model specific README, e.g
|
100 |
+
[./docs/MUSICGEN.md](./docs/MUSICGEN.md), [./docs/AUDIOGEN.md](./docs/AUDIOGEN.md), etc.
|
|
|
|
|
|
|
|
assets/a_duck_quacking_as_birds_chirp_and_a_pigeon_cooing.mp3
ADDED
Binary file (15.2 kB). View file
|
|
assets/sirens_and_a_humming_engine_approach_and_pass.mp3
ADDED
Binary file (15.2 kB). View file
|
|
audiocraft/__init__.py
CHANGED
@@ -3,8 +3,24 @@
|
|
3 |
#
|
4 |
# This source code is licensed under the license found in the
|
5 |
# LICENSE file in the root directory of this source tree.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
# flake8: noqa
|
8 |
from . import data, modules, models
|
9 |
|
10 |
-
__version__ = '
|
|
|
3 |
#
|
4 |
# This source code is licensed under the license found in the
|
5 |
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""
|
7 |
+
AudioCraft is a general framework for training audio generative models.
|
8 |
+
At the moment we provide the training code for:
|
9 |
+
|
10 |
+
- [MusicGen](https://arxiv.org/abs/2306.05284), a state-of-the-art
|
11 |
+
text-to-music and melody+text autoregressive generative model.
|
12 |
+
For the solver, see `audiocraft.solvers.musicgen.MusicGenSolver`, and for the model,
|
13 |
+
`audiocraft.models.musicgen.MusicGen`.
|
14 |
+
- [AudioGen](https://arxiv.org/abs/2209.15352), a state-of-the-art
|
15 |
+
text-to-general-audio generative model.
|
16 |
+
- [EnCodec](https://arxiv.org/abs/2210.13438), efficient and high fidelity
|
17 |
+
neural audio codec which provides an excellent tokenizer for autoregressive language models.
|
18 |
+
See `audiocraft.solvers.compression.CompressionSolver`, and `audiocraft.models.encodec.EncodecModel`.
|
19 |
+
- [MultiBandDiffusion](TODO), alternative diffusion-based decoder compatible with EnCodec that
|
20 |
+
improves the perceived quality and reduces the artifacts coming from adversarial decoders.
|
21 |
+
"""
|
22 |
|
23 |
# flake8: noqa
|
24 |
from . import data, modules, models
|
25 |
|
26 |
+
__version__ = '1.1.0'
|
audiocraft/adversarial/__init__.py
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""Adversarial losses and discriminator architectures."""
|
7 |
+
|
8 |
+
# flake8: noqa
|
9 |
+
from .discriminators import (
|
10 |
+
MultiPeriodDiscriminator,
|
11 |
+
MultiScaleDiscriminator,
|
12 |
+
MultiScaleSTFTDiscriminator
|
13 |
+
)
|
14 |
+
from .losses import (
|
15 |
+
AdversarialLoss,
|
16 |
+
AdvLossType,
|
17 |
+
get_adv_criterion,
|
18 |
+
get_fake_criterion,
|
19 |
+
get_real_criterion,
|
20 |
+
FeatLossType,
|
21 |
+
FeatureMatchingLoss
|
22 |
+
)
|
audiocraft/adversarial/discriminators/__init__.py
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
# flake8: noqa
|
8 |
+
from .mpd import MultiPeriodDiscriminator
|
9 |
+
from .msd import MultiScaleDiscriminator
|
10 |
+
from .msstftd import MultiScaleSTFTDiscriminator
|
audiocraft/adversarial/discriminators/base.py
ADDED
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from abc import ABC, abstractmethod
|
8 |
+
import typing as tp
|
9 |
+
|
10 |
+
import torch
|
11 |
+
import torch.nn as nn
|
12 |
+
|
13 |
+
|
14 |
+
FeatureMapType = tp.List[torch.Tensor]
|
15 |
+
LogitsType = torch.Tensor
|
16 |
+
MultiDiscriminatorOutputType = tp.Tuple[tp.List[LogitsType], tp.List[FeatureMapType]]
|
17 |
+
|
18 |
+
|
19 |
+
class MultiDiscriminator(ABC, nn.Module):
|
20 |
+
"""Base implementation for discriminators composed of sub-discriminators acting at different scales.
|
21 |
+
"""
|
22 |
+
def __init__(self):
|
23 |
+
super().__init__()
|
24 |
+
|
25 |
+
@abstractmethod
|
26 |
+
def forward(self, x: torch.Tensor) -> MultiDiscriminatorOutputType:
|
27 |
+
...
|
28 |
+
|
29 |
+
@property
|
30 |
+
@abstractmethod
|
31 |
+
def num_discriminators(self) -> int:
|
32 |
+
"""Number of discriminators.
|
33 |
+
"""
|
34 |
+
...
|
audiocraft/adversarial/discriminators/mpd.py
ADDED
@@ -0,0 +1,106 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import typing as tp
|
8 |
+
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
import torch.nn.functional as F
|
12 |
+
|
13 |
+
from ...modules import NormConv2d
|
14 |
+
from .base import MultiDiscriminator, MultiDiscriminatorOutputType
|
15 |
+
|
16 |
+
|
17 |
+
def get_padding(kernel_size: int, dilation: int = 1) -> int:
|
18 |
+
return int((kernel_size * dilation - dilation) / 2)
|
19 |
+
|
20 |
+
|
21 |
+
class PeriodDiscriminator(nn.Module):
|
22 |
+
"""Period sub-discriminator.
|
23 |
+
|
24 |
+
Args:
|
25 |
+
period (int): Period between samples of audio.
|
26 |
+
in_channels (int): Number of input channels.
|
27 |
+
out_channels (int): Number of output channels.
|
28 |
+
n_layers (int): Number of convolutional layers.
|
29 |
+
kernel_sizes (list of int): Kernel sizes for convolutions.
|
30 |
+
stride (int): Stride for convolutions.
|
31 |
+
filters (int): Initial number of filters in convolutions.
|
32 |
+
filters_scale (int): Multiplier of number of filters as we increase depth.
|
33 |
+
max_filters (int): Maximum number of filters.
|
34 |
+
norm (str): Normalization method.
|
35 |
+
activation (str): Activation function.
|
36 |
+
activation_params (dict): Parameters to provide to the activation function.
|
37 |
+
"""
|
38 |
+
def __init__(self, period: int, in_channels: int = 1, out_channels: int = 1,
|
39 |
+
n_layers: int = 5, kernel_sizes: tp.List[int] = [5, 3], stride: int = 3,
|
40 |
+
filters: int = 8, filters_scale: int = 4, max_filters: int = 1024,
|
41 |
+
norm: str = 'weight_norm', activation: str = 'LeakyReLU',
|
42 |
+
activation_params: dict = {'negative_slope': 0.2}):
|
43 |
+
super().__init__()
|
44 |
+
self.period = period
|
45 |
+
self.n_layers = n_layers
|
46 |
+
self.activation = getattr(torch.nn, activation)(**activation_params)
|
47 |
+
self.convs = nn.ModuleList()
|
48 |
+
in_chs = in_channels
|
49 |
+
for i in range(self.n_layers):
|
50 |
+
out_chs = min(filters * (filters_scale ** (i + 1)), max_filters)
|
51 |
+
eff_stride = 1 if i == self.n_layers - 1 else stride
|
52 |
+
self.convs.append(NormConv2d(in_chs, out_chs, kernel_size=(kernel_sizes[0], 1), stride=(eff_stride, 1),
|
53 |
+
padding=((kernel_sizes[0] - 1) // 2, 0), norm=norm))
|
54 |
+
in_chs = out_chs
|
55 |
+
self.conv_post = NormConv2d(in_chs, out_channels, kernel_size=(kernel_sizes[1], 1), stride=1,
|
56 |
+
padding=((kernel_sizes[1] - 1) // 2, 0), norm=norm)
|
57 |
+
|
58 |
+
def forward(self, x: torch.Tensor):
|
59 |
+
fmap = []
|
60 |
+
# 1d to 2d
|
61 |
+
b, c, t = x.shape
|
62 |
+
if t % self.period != 0: # pad first
|
63 |
+
n_pad = self.period - (t % self.period)
|
64 |
+
x = F.pad(x, (0, n_pad), 'reflect')
|
65 |
+
t = t + n_pad
|
66 |
+
x = x.view(b, c, t // self.period, self.period)
|
67 |
+
|
68 |
+
for conv in self.convs:
|
69 |
+
x = conv(x)
|
70 |
+
x = self.activation(x)
|
71 |
+
fmap.append(x)
|
72 |
+
x = self.conv_post(x)
|
73 |
+
fmap.append(x)
|
74 |
+
# x = torch.flatten(x, 1, -1)
|
75 |
+
|
76 |
+
return x, fmap
|
77 |
+
|
78 |
+
|
79 |
+
class MultiPeriodDiscriminator(MultiDiscriminator):
|
80 |
+
"""Multi-Period (MPD) Discriminator.
|
81 |
+
|
82 |
+
Args:
|
83 |
+
in_channels (int): Number of input channels.
|
84 |
+
out_channels (int): Number of output channels.
|
85 |
+
periods (Sequence[int]): Periods between samples of audio for the sub-discriminators.
|
86 |
+
**kwargs: Additional args for `PeriodDiscriminator`
|
87 |
+
"""
|
88 |
+
def __init__(self, in_channels: int = 1, out_channels: int = 1,
|
89 |
+
periods: tp.Sequence[int] = [2, 3, 5, 7, 11], **kwargs):
|
90 |
+
super().__init__()
|
91 |
+
self.discriminators = nn.ModuleList([
|
92 |
+
PeriodDiscriminator(p, in_channels, out_channels, **kwargs) for p in periods
|
93 |
+
])
|
94 |
+
|
95 |
+
@property
|
96 |
+
def num_discriminators(self):
|
97 |
+
return len(self.discriminators)
|
98 |
+
|
99 |
+
def forward(self, x: torch.Tensor) -> MultiDiscriminatorOutputType:
|
100 |
+
logits = []
|
101 |
+
fmaps = []
|
102 |
+
for disc in self.discriminators:
|
103 |
+
logit, fmap = disc(x)
|
104 |
+
logits.append(logit)
|
105 |
+
fmaps.append(fmap)
|
106 |
+
return logits, fmaps
|
audiocraft/adversarial/discriminators/msd.py
ADDED
@@ -0,0 +1,126 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import typing as tp
|
8 |
+
|
9 |
+
import numpy as np
|
10 |
+
import torch
|
11 |
+
import torch.nn as nn
|
12 |
+
|
13 |
+
from ...modules import NormConv1d
|
14 |
+
from .base import MultiDiscriminator, MultiDiscriminatorOutputType
|
15 |
+
|
16 |
+
|
17 |
+
class ScaleDiscriminator(nn.Module):
|
18 |
+
"""Waveform sub-discriminator.
|
19 |
+
|
20 |
+
Args:
|
21 |
+
in_channels (int): Number of input channels.
|
22 |
+
out_channels (int): Number of output channels.
|
23 |
+
kernel_sizes (Sequence[int]): Kernel sizes for first and last convolutions.
|
24 |
+
filters (int): Number of initial filters for convolutions.
|
25 |
+
max_filters (int): Maximum number of filters.
|
26 |
+
downsample_scales (Sequence[int]): Scale for downsampling implemented as strided convolutions.
|
27 |
+
inner_kernel_sizes (Sequence[int] or None): Kernel sizes for inner convolutions.
|
28 |
+
groups (Sequence[int] or None): Groups for inner convolutions.
|
29 |
+
strides (Sequence[int] or None): Strides for inner convolutions.
|
30 |
+
paddings (Sequence[int] or None): Paddings for inner convolutions.
|
31 |
+
norm (str): Normalization method.
|
32 |
+
activation (str): Activation function.
|
33 |
+
activation_params (dict): Parameters to provide to the activation function.
|
34 |
+
pad (str): Padding for initial convolution.
|
35 |
+
pad_params (dict): Parameters to provide to the padding module.
|
36 |
+
"""
|
37 |
+
def __init__(self, in_channels=1, out_channels=1, kernel_sizes: tp.Sequence[int] = [5, 3],
|
38 |
+
filters: int = 16, max_filters: int = 1024, downsample_scales: tp.Sequence[int] = [4, 4, 4, 4],
|
39 |
+
inner_kernel_sizes: tp.Optional[tp.Sequence[int]] = None, groups: tp.Optional[tp.Sequence[int]] = None,
|
40 |
+
strides: tp.Optional[tp.Sequence[int]] = None, paddings: tp.Optional[tp.Sequence[int]] = None,
|
41 |
+
norm: str = 'weight_norm', activation: str = 'LeakyReLU',
|
42 |
+
activation_params: dict = {'negative_slope': 0.2}, pad: str = 'ReflectionPad1d',
|
43 |
+
pad_params: dict = {}):
|
44 |
+
super().__init__()
|
45 |
+
assert len(kernel_sizes) == 2
|
46 |
+
assert kernel_sizes[0] % 2 == 1
|
47 |
+
assert kernel_sizes[1] % 2 == 1
|
48 |
+
assert (inner_kernel_sizes is None or len(inner_kernel_sizes) == len(downsample_scales))
|
49 |
+
assert (groups is None or len(groups) == len(downsample_scales))
|
50 |
+
assert (strides is None or len(strides) == len(downsample_scales))
|
51 |
+
assert (paddings is None or len(paddings) == len(downsample_scales))
|
52 |
+
self.activation = getattr(torch.nn, activation)(**activation_params)
|
53 |
+
self.convs = nn.ModuleList()
|
54 |
+
self.convs.append(
|
55 |
+
nn.Sequential(
|
56 |
+
getattr(torch.nn, pad)((np.prod(kernel_sizes) - 1) // 2, **pad_params),
|
57 |
+
NormConv1d(in_channels, filters, kernel_size=np.prod(kernel_sizes), stride=1, norm=norm)
|
58 |
+
)
|
59 |
+
)
|
60 |
+
|
61 |
+
in_chs = filters
|
62 |
+
for i, downsample_scale in enumerate(downsample_scales):
|
63 |
+
out_chs = min(in_chs * downsample_scale, max_filters)
|
64 |
+
default_kernel_size = downsample_scale * 10 + 1
|
65 |
+
default_stride = downsample_scale
|
66 |
+
default_padding = (default_kernel_size - 1) // 2
|
67 |
+
default_groups = in_chs // 4
|
68 |
+
self.convs.append(
|
69 |
+
NormConv1d(in_chs, out_chs,
|
70 |
+
kernel_size=inner_kernel_sizes[i] if inner_kernel_sizes else default_kernel_size,
|
71 |
+
stride=strides[i] if strides else default_stride,
|
72 |
+
groups=groups[i] if groups else default_groups,
|
73 |
+
padding=paddings[i] if paddings else default_padding,
|
74 |
+
norm=norm))
|
75 |
+
in_chs = out_chs
|
76 |
+
|
77 |
+
out_chs = min(in_chs * 2, max_filters)
|
78 |
+
self.convs.append(NormConv1d(in_chs, out_chs, kernel_size=kernel_sizes[0], stride=1,
|
79 |
+
padding=(kernel_sizes[0] - 1) // 2, norm=norm))
|
80 |
+
self.conv_post = NormConv1d(out_chs, out_channels, kernel_size=kernel_sizes[1], stride=1,
|
81 |
+
padding=(kernel_sizes[1] - 1) // 2, norm=norm)
|
82 |
+
|
83 |
+
def forward(self, x: torch.Tensor):
|
84 |
+
fmap = []
|
85 |
+
for layer in self.convs:
|
86 |
+
x = layer(x)
|
87 |
+
x = self.activation(x)
|
88 |
+
fmap.append(x)
|
89 |
+
x = self.conv_post(x)
|
90 |
+
fmap.append(x)
|
91 |
+
# x = torch.flatten(x, 1, -1)
|
92 |
+
return x, fmap
|
93 |
+
|
94 |
+
|
95 |
+
class MultiScaleDiscriminator(MultiDiscriminator):
|
96 |
+
"""Multi-Scale (MSD) Discriminator,
|
97 |
+
|
98 |
+
Args:
|
99 |
+
in_channels (int): Number of input channels.
|
100 |
+
out_channels (int): Number of output channels.
|
101 |
+
downsample_factor (int): Downsampling factor between the different scales.
|
102 |
+
scale_norms (Sequence[str]): Normalization for each sub-discriminator.
|
103 |
+
**kwargs: Additional args for ScaleDiscriminator.
|
104 |
+
"""
|
105 |
+
def __init__(self, in_channels: int = 1, out_channels: int = 1, downsample_factor: int = 2,
|
106 |
+
scale_norms: tp.Sequence[str] = ['weight_norm', 'weight_norm', 'weight_norm'], **kwargs):
|
107 |
+
super().__init__()
|
108 |
+
self.discriminators = nn.ModuleList([
|
109 |
+
ScaleDiscriminator(in_channels, out_channels, norm=norm, **kwargs) for norm in scale_norms
|
110 |
+
])
|
111 |
+
self.downsample = nn.AvgPool1d(downsample_factor * 2, downsample_factor, padding=downsample_factor)
|
112 |
+
|
113 |
+
@property
|
114 |
+
def num_discriminators(self):
|
115 |
+
return len(self.discriminators)
|
116 |
+
|
117 |
+
def forward(self, x: torch.Tensor) -> MultiDiscriminatorOutputType:
|
118 |
+
logits = []
|
119 |
+
fmaps = []
|
120 |
+
for i, disc in enumerate(self.discriminators):
|
121 |
+
if i != 0:
|
122 |
+
self.downsample(x)
|
123 |
+
logit, fmap = disc(x)
|
124 |
+
logits.append(logit)
|
125 |
+
fmaps.append(fmap)
|
126 |
+
return logits, fmaps
|
audiocraft/adversarial/discriminators/msstftd.py
ADDED
@@ -0,0 +1,134 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import typing as tp
|
8 |
+
|
9 |
+
import torchaudio
|
10 |
+
import torch
|
11 |
+
from torch import nn
|
12 |
+
from einops import rearrange
|
13 |
+
|
14 |
+
from ...modules import NormConv2d
|
15 |
+
from .base import MultiDiscriminator, MultiDiscriminatorOutputType
|
16 |
+
|
17 |
+
|
18 |
+
def get_2d_padding(kernel_size: tp.Tuple[int, int], dilation: tp.Tuple[int, int] = (1, 1)):
|
19 |
+
return (((kernel_size[0] - 1) * dilation[0]) // 2, ((kernel_size[1] - 1) * dilation[1]) // 2)
|
20 |
+
|
21 |
+
|
22 |
+
class DiscriminatorSTFT(nn.Module):
|
23 |
+
"""STFT sub-discriminator.
|
24 |
+
|
25 |
+
Args:
|
26 |
+
filters (int): Number of filters in convolutions.
|
27 |
+
in_channels (int): Number of input channels.
|
28 |
+
out_channels (int): Number of output channels.
|
29 |
+
n_fft (int): Size of FFT for each scale.
|
30 |
+
hop_length (int): Length of hop between STFT windows for each scale.
|
31 |
+
kernel_size (tuple of int): Inner Conv2d kernel sizes.
|
32 |
+
stride (tuple of int): Inner Conv2d strides.
|
33 |
+
dilations (list of int): Inner Conv2d dilation on the time dimension.
|
34 |
+
win_length (int): Window size for each scale.
|
35 |
+
normalized (bool): Whether to normalize by magnitude after stft.
|
36 |
+
norm (str): Normalization method.
|
37 |
+
activation (str): Activation function.
|
38 |
+
activation_params (dict): Parameters to provide to the activation function.
|
39 |
+
growth (int): Growth factor for the filters.
|
40 |
+
"""
|
41 |
+
def __init__(self, filters: int, in_channels: int = 1, out_channels: int = 1,
|
42 |
+
n_fft: int = 1024, hop_length: int = 256, win_length: int = 1024, max_filters: int = 1024,
|
43 |
+
filters_scale: int = 1, kernel_size: tp.Tuple[int, int] = (3, 9), dilations: tp.List = [1, 2, 4],
|
44 |
+
stride: tp.Tuple[int, int] = (1, 2), normalized: bool = True, norm: str = 'weight_norm',
|
45 |
+
activation: str = 'LeakyReLU', activation_params: dict = {'negative_slope': 0.2}):
|
46 |
+
super().__init__()
|
47 |
+
assert len(kernel_size) == 2
|
48 |
+
assert len(stride) == 2
|
49 |
+
self.filters = filters
|
50 |
+
self.in_channels = in_channels
|
51 |
+
self.out_channels = out_channels
|
52 |
+
self.n_fft = n_fft
|
53 |
+
self.hop_length = hop_length
|
54 |
+
self.win_length = win_length
|
55 |
+
self.normalized = normalized
|
56 |
+
self.activation = getattr(torch.nn, activation)(**activation_params)
|
57 |
+
self.spec_transform = torchaudio.transforms.Spectrogram(
|
58 |
+
n_fft=self.n_fft, hop_length=self.hop_length, win_length=self.win_length, window_fn=torch.hann_window,
|
59 |
+
normalized=self.normalized, center=False, pad_mode=None, power=None)
|
60 |
+
spec_channels = 2 * self.in_channels
|
61 |
+
self.convs = nn.ModuleList()
|
62 |
+
self.convs.append(
|
63 |
+
NormConv2d(spec_channels, self.filters, kernel_size=kernel_size, padding=get_2d_padding(kernel_size))
|
64 |
+
)
|
65 |
+
in_chs = min(filters_scale * self.filters, max_filters)
|
66 |
+
for i, dilation in enumerate(dilations):
|
67 |
+
out_chs = min((filters_scale ** (i + 1)) * self.filters, max_filters)
|
68 |
+
self.convs.append(NormConv2d(in_chs, out_chs, kernel_size=kernel_size, stride=stride,
|
69 |
+
dilation=(dilation, 1), padding=get_2d_padding(kernel_size, (dilation, 1)),
|
70 |
+
norm=norm))
|
71 |
+
in_chs = out_chs
|
72 |
+
out_chs = min((filters_scale ** (len(dilations) + 1)) * self.filters, max_filters)
|
73 |
+
self.convs.append(NormConv2d(in_chs, out_chs, kernel_size=(kernel_size[0], kernel_size[0]),
|
74 |
+
padding=get_2d_padding((kernel_size[0], kernel_size[0])),
|
75 |
+
norm=norm))
|
76 |
+
self.conv_post = NormConv2d(out_chs, self.out_channels,
|
77 |
+
kernel_size=(kernel_size[0], kernel_size[0]),
|
78 |
+
padding=get_2d_padding((kernel_size[0], kernel_size[0])),
|
79 |
+
norm=norm)
|
80 |
+
|
81 |
+
def forward(self, x: torch.Tensor):
|
82 |
+
fmap = []
|
83 |
+
z = self.spec_transform(x) # [B, 2, Freq, Frames, 2]
|
84 |
+
z = torch.cat([z.real, z.imag], dim=1)
|
85 |
+
z = rearrange(z, 'b c w t -> b c t w')
|
86 |
+
for i, layer in enumerate(self.convs):
|
87 |
+
z = layer(z)
|
88 |
+
z = self.activation(z)
|
89 |
+
fmap.append(z)
|
90 |
+
z = self.conv_post(z)
|
91 |
+
return z, fmap
|
92 |
+
|
93 |
+
|
94 |
+
class MultiScaleSTFTDiscriminator(MultiDiscriminator):
|
95 |
+
"""Multi-Scale STFT (MS-STFT) discriminator.
|
96 |
+
|
97 |
+
Args:
|
98 |
+
filters (int): Number of filters in convolutions.
|
99 |
+
in_channels (int): Number of input channels.
|
100 |
+
out_channels (int): Number of output channels.
|
101 |
+
sep_channels (bool): Separate channels to distinct samples for stereo support.
|
102 |
+
n_ffts (Sequence[int]): Size of FFT for each scale.
|
103 |
+
hop_lengths (Sequence[int]): Length of hop between STFT windows for each scale.
|
104 |
+
win_lengths (Sequence[int]): Window size for each scale.
|
105 |
+
**kwargs: Additional args for STFTDiscriminator.
|
106 |
+
"""
|
107 |
+
def __init__(self, filters: int, in_channels: int = 1, out_channels: int = 1, sep_channels: bool = False,
|
108 |
+
n_ffts: tp.List[int] = [1024, 2048, 512], hop_lengths: tp.List[int] = [256, 512, 128],
|
109 |
+
win_lengths: tp.List[int] = [1024, 2048, 512], **kwargs):
|
110 |
+
super().__init__()
|
111 |
+
assert len(n_ffts) == len(hop_lengths) == len(win_lengths)
|
112 |
+
self.sep_channels = sep_channels
|
113 |
+
self.discriminators = nn.ModuleList([
|
114 |
+
DiscriminatorSTFT(filters, in_channels=in_channels, out_channels=out_channels,
|
115 |
+
n_fft=n_ffts[i], win_length=win_lengths[i], hop_length=hop_lengths[i], **kwargs)
|
116 |
+
for i in range(len(n_ffts))
|
117 |
+
])
|
118 |
+
|
119 |
+
@property
|
120 |
+
def num_discriminators(self):
|
121 |
+
return len(self.discriminators)
|
122 |
+
|
123 |
+
def _separate_channels(self, x: torch.Tensor) -> torch.Tensor:
|
124 |
+
B, C, T = x.shape
|
125 |
+
return x.view(-1, 1, T)
|
126 |
+
|
127 |
+
def forward(self, x: torch.Tensor) -> MultiDiscriminatorOutputType:
|
128 |
+
logits = []
|
129 |
+
fmaps = []
|
130 |
+
for disc in self.discriminators:
|
131 |
+
logit, fmap = disc(x)
|
132 |
+
logits.append(logit)
|
133 |
+
fmaps.append(fmap)
|
134 |
+
return logits, fmaps
|
audiocraft/adversarial/losses.py
ADDED
@@ -0,0 +1,228 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
"""
|
8 |
+
Utility module to handle adversarial losses without requiring to mess up the main training loop.
|
9 |
+
"""
|
10 |
+
|
11 |
+
import typing as tp
|
12 |
+
|
13 |
+
import flashy
|
14 |
+
import torch
|
15 |
+
import torch.nn as nn
|
16 |
+
import torch.nn.functional as F
|
17 |
+
|
18 |
+
|
19 |
+
ADVERSARIAL_LOSSES = ['mse', 'hinge', 'hinge2']
|
20 |
+
|
21 |
+
|
22 |
+
AdvLossType = tp.Union[nn.Module, tp.Callable[[torch.Tensor], torch.Tensor]]
|
23 |
+
FeatLossType = tp.Union[nn.Module, tp.Callable[[torch.Tensor, torch.Tensor], torch.Tensor]]
|
24 |
+
|
25 |
+
|
26 |
+
class AdversarialLoss(nn.Module):
|
27 |
+
"""Adversary training wrapper.
|
28 |
+
|
29 |
+
Args:
|
30 |
+
adversary (nn.Module): The adversary module will be used to estimate the logits given the fake and real samples.
|
31 |
+
We assume here the adversary output is ``Tuple[List[torch.Tensor], List[List[torch.Tensor]]]``
|
32 |
+
where the first item is a list of logits and the second item is a list of feature maps.
|
33 |
+
optimizer (torch.optim.Optimizer): Optimizer used for training the given module.
|
34 |
+
loss (AdvLossType): Loss function for generator training.
|
35 |
+
loss_real (AdvLossType): Loss function for adversarial training on logits from real samples.
|
36 |
+
loss_fake (AdvLossType): Loss function for adversarial training on logits from fake samples.
|
37 |
+
loss_feat (FeatLossType): Feature matching loss function for generator training.
|
38 |
+
normalize (bool): Whether to normalize by number of sub-discriminators.
|
39 |
+
|
40 |
+
Example of usage:
|
41 |
+
adv_loss = AdversarialLoss(adversaries, optimizer, loss, loss_real, loss_fake)
|
42 |
+
for real in loader:
|
43 |
+
noise = torch.randn(...)
|
44 |
+
fake = model(noise)
|
45 |
+
adv_loss.train_adv(fake, real)
|
46 |
+
loss, _ = adv_loss(fake, real)
|
47 |
+
loss.backward()
|
48 |
+
"""
|
49 |
+
def __init__(self,
|
50 |
+
adversary: nn.Module,
|
51 |
+
optimizer: torch.optim.Optimizer,
|
52 |
+
loss: AdvLossType,
|
53 |
+
loss_real: AdvLossType,
|
54 |
+
loss_fake: AdvLossType,
|
55 |
+
loss_feat: tp.Optional[FeatLossType] = None,
|
56 |
+
normalize: bool = True):
|
57 |
+
super().__init__()
|
58 |
+
self.adversary: nn.Module = adversary
|
59 |
+
flashy.distrib.broadcast_model(self.adversary)
|
60 |
+
self.optimizer = optimizer
|
61 |
+
self.loss = loss
|
62 |
+
self.loss_real = loss_real
|
63 |
+
self.loss_fake = loss_fake
|
64 |
+
self.loss_feat = loss_feat
|
65 |
+
self.normalize = normalize
|
66 |
+
|
67 |
+
def _save_to_state_dict(self, destination, prefix, keep_vars):
|
68 |
+
# Add the optimizer state dict inside our own.
|
69 |
+
super()._save_to_state_dict(destination, prefix, keep_vars)
|
70 |
+
destination[prefix + 'optimizer'] = self.optimizer.state_dict()
|
71 |
+
return destination
|
72 |
+
|
73 |
+
def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
|
74 |
+
# Load optimizer state.
|
75 |
+
self.optimizer.load_state_dict(state_dict.pop(prefix + 'optimizer'))
|
76 |
+
super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
|
77 |
+
|
78 |
+
def get_adversary_pred(self, x):
|
79 |
+
"""Run adversary model, validating expected output format."""
|
80 |
+
logits, fmaps = self.adversary(x)
|
81 |
+
assert isinstance(logits, list) and all([isinstance(t, torch.Tensor) for t in logits]), \
|
82 |
+
f'Expecting a list of tensors as logits but {type(logits)} found.'
|
83 |
+
assert isinstance(fmaps, list), f'Expecting a list of features maps but {type(fmaps)} found.'
|
84 |
+
for fmap in fmaps:
|
85 |
+
assert isinstance(fmap, list) and all([isinstance(f, torch.Tensor) for f in fmap]), \
|
86 |
+
f'Expecting a list of tensors as feature maps but {type(fmap)} found.'
|
87 |
+
return logits, fmaps
|
88 |
+
|
89 |
+
def train_adv(self, fake: torch.Tensor, real: torch.Tensor) -> torch.Tensor:
|
90 |
+
"""Train the adversary with the given fake and real example.
|
91 |
+
|
92 |
+
We assume the adversary output is the following format: Tuple[List[torch.Tensor], List[List[torch.Tensor]]].
|
93 |
+
The first item being the logits and second item being a list of feature maps for each sub-discriminator.
|
94 |
+
|
95 |
+
This will automatically synchronize gradients (with `flashy.distrib.eager_sync_model`)
|
96 |
+
and call the optimizer.
|
97 |
+
"""
|
98 |
+
loss = torch.tensor(0., device=fake.device)
|
99 |
+
all_logits_fake_is_fake, _ = self.get_adversary_pred(fake.detach())
|
100 |
+
all_logits_real_is_fake, _ = self.get_adversary_pred(real.detach())
|
101 |
+
n_sub_adversaries = len(all_logits_fake_is_fake)
|
102 |
+
for logit_fake_is_fake, logit_real_is_fake in zip(all_logits_fake_is_fake, all_logits_real_is_fake):
|
103 |
+
loss += self.loss_fake(logit_fake_is_fake) + self.loss_real(logit_real_is_fake)
|
104 |
+
|
105 |
+
if self.normalize:
|
106 |
+
loss /= n_sub_adversaries
|
107 |
+
|
108 |
+
self.optimizer.zero_grad()
|
109 |
+
with flashy.distrib.eager_sync_model(self.adversary):
|
110 |
+
loss.backward()
|
111 |
+
self.optimizer.step()
|
112 |
+
|
113 |
+
return loss
|
114 |
+
|
115 |
+
def forward(self, fake: torch.Tensor, real: torch.Tensor) -> tp.Tuple[torch.Tensor, torch.Tensor]:
|
116 |
+
"""Return the loss for the generator, i.e. trying to fool the adversary,
|
117 |
+
and feature matching loss if provided.
|
118 |
+
"""
|
119 |
+
adv = torch.tensor(0., device=fake.device)
|
120 |
+
feat = torch.tensor(0., device=fake.device)
|
121 |
+
with flashy.utils.readonly(self.adversary):
|
122 |
+
all_logits_fake_is_fake, all_fmap_fake = self.get_adversary_pred(fake)
|
123 |
+
all_logits_real_is_fake, all_fmap_real = self.get_adversary_pred(real)
|
124 |
+
n_sub_adversaries = len(all_logits_fake_is_fake)
|
125 |
+
for logit_fake_is_fake in all_logits_fake_is_fake:
|
126 |
+
adv += self.loss(logit_fake_is_fake)
|
127 |
+
if self.loss_feat:
|
128 |
+
for fmap_fake, fmap_real in zip(all_fmap_fake, all_fmap_real):
|
129 |
+
feat += self.loss_feat(fmap_fake, fmap_real)
|
130 |
+
|
131 |
+
if self.normalize:
|
132 |
+
adv /= n_sub_adversaries
|
133 |
+
feat /= n_sub_adversaries
|
134 |
+
|
135 |
+
return adv, feat
|
136 |
+
|
137 |
+
|
138 |
+
def get_adv_criterion(loss_type: str) -> tp.Callable:
|
139 |
+
assert loss_type in ADVERSARIAL_LOSSES
|
140 |
+
if loss_type == 'mse':
|
141 |
+
return mse_loss
|
142 |
+
elif loss_type == 'hinge':
|
143 |
+
return hinge_loss
|
144 |
+
elif loss_type == 'hinge2':
|
145 |
+
return hinge2_loss
|
146 |
+
raise ValueError('Unsupported loss')
|
147 |
+
|
148 |
+
|
149 |
+
def get_fake_criterion(loss_type: str) -> tp.Callable:
|
150 |
+
assert loss_type in ADVERSARIAL_LOSSES
|
151 |
+
if loss_type == 'mse':
|
152 |
+
return mse_fake_loss
|
153 |
+
elif loss_type in ['hinge', 'hinge2']:
|
154 |
+
return hinge_fake_loss
|
155 |
+
raise ValueError('Unsupported loss')
|
156 |
+
|
157 |
+
|
158 |
+
def get_real_criterion(loss_type: str) -> tp.Callable:
|
159 |
+
assert loss_type in ADVERSARIAL_LOSSES
|
160 |
+
if loss_type == 'mse':
|
161 |
+
return mse_real_loss
|
162 |
+
elif loss_type in ['hinge', 'hinge2']:
|
163 |
+
return hinge_real_loss
|
164 |
+
raise ValueError('Unsupported loss')
|
165 |
+
|
166 |
+
|
167 |
+
def mse_real_loss(x: torch.Tensor) -> torch.Tensor:
|
168 |
+
return F.mse_loss(x, torch.tensor(1., device=x.device).expand_as(x))
|
169 |
+
|
170 |
+
|
171 |
+
def mse_fake_loss(x: torch.Tensor) -> torch.Tensor:
|
172 |
+
return F.mse_loss(x, torch.tensor(0., device=x.device).expand_as(x))
|
173 |
+
|
174 |
+
|
175 |
+
def hinge_real_loss(x: torch.Tensor) -> torch.Tensor:
|
176 |
+
return -torch.mean(torch.min(x - 1, torch.tensor(0., device=x.device).expand_as(x)))
|
177 |
+
|
178 |
+
|
179 |
+
def hinge_fake_loss(x: torch.Tensor) -> torch.Tensor:
|
180 |
+
return -torch.mean(torch.min(-x - 1, torch.tensor(0., device=x.device).expand_as(x)))
|
181 |
+
|
182 |
+
|
183 |
+
def mse_loss(x: torch.Tensor) -> torch.Tensor:
|
184 |
+
if x.numel() == 0:
|
185 |
+
return torch.tensor([0.0], device=x.device)
|
186 |
+
return F.mse_loss(x, torch.tensor(1., device=x.device).expand_as(x))
|
187 |
+
|
188 |
+
|
189 |
+
def hinge_loss(x: torch.Tensor) -> torch.Tensor:
|
190 |
+
if x.numel() == 0:
|
191 |
+
return torch.tensor([0.0], device=x.device)
|
192 |
+
return -x.mean()
|
193 |
+
|
194 |
+
|
195 |
+
def hinge2_loss(x: torch.Tensor) -> torch.Tensor:
|
196 |
+
if x.numel() == 0:
|
197 |
+
return torch.tensor([0.0])
|
198 |
+
return -torch.mean(torch.min(x - 1, torch.tensor(0., device=x.device).expand_as(x)))
|
199 |
+
|
200 |
+
|
201 |
+
class FeatureMatchingLoss(nn.Module):
|
202 |
+
"""Feature matching loss for adversarial training.
|
203 |
+
|
204 |
+
Args:
|
205 |
+
loss (nn.Module): Loss to use for feature matching (default=torch.nn.L1).
|
206 |
+
normalize (bool): Whether to normalize the loss.
|
207 |
+
by number of feature maps.
|
208 |
+
"""
|
209 |
+
def __init__(self, loss: nn.Module = torch.nn.L1Loss(), normalize: bool = True):
|
210 |
+
super().__init__()
|
211 |
+
self.loss = loss
|
212 |
+
self.normalize = normalize
|
213 |
+
|
214 |
+
def forward(self, fmap_fake: tp.List[torch.Tensor], fmap_real: tp.List[torch.Tensor]) -> torch.Tensor:
|
215 |
+
assert len(fmap_fake) == len(fmap_real) and len(fmap_fake) > 0
|
216 |
+
feat_loss = torch.tensor(0., device=fmap_fake[0].device)
|
217 |
+
feat_scale = torch.tensor(0., device=fmap_fake[0].device)
|
218 |
+
n_fmaps = 0
|
219 |
+
for (feat_fake, feat_real) in zip(fmap_fake, fmap_real):
|
220 |
+
assert feat_fake.shape == feat_real.shape
|
221 |
+
n_fmaps += 1
|
222 |
+
feat_loss += self.loss(feat_fake, feat_real)
|
223 |
+
feat_scale += torch.mean(torch.abs(feat_real))
|
224 |
+
|
225 |
+
if self.normalize:
|
226 |
+
feat_loss /= n_fmaps
|
227 |
+
|
228 |
+
return feat_loss
|
audiocraft/data/__init__.py
CHANGED
@@ -3,6 +3,8 @@
|
|
3 |
#
|
4 |
# This source code is licensed under the license found in the
|
5 |
# LICENSE file in the root directory of this source tree.
|
|
|
|
|
6 |
|
7 |
# flake8: noqa
|
8 |
-
from . import audio, audio_dataset
|
|
|
3 |
#
|
4 |
# This source code is licensed under the license found in the
|
5 |
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""Audio loading and writing support. Datasets for raw audio
|
7 |
+
or also including some metadata."""
|
8 |
|
9 |
# flake8: noqa
|
10 |
+
from . import audio, audio_dataset, info_audio_dataset, music_dataset, sound_dataset
|
audiocraft/data/audio.py
CHANGED
@@ -18,11 +18,11 @@ import numpy as np
|
|
18 |
import soundfile
|
19 |
import torch
|
20 |
from torch.nn import functional as F
|
21 |
-
import torchaudio as ta
|
22 |
|
23 |
import av
|
|
|
24 |
|
25 |
-
from .audio_utils import f32_pcm,
|
26 |
|
27 |
|
28 |
_av_initialized = False
|
@@ -78,7 +78,7 @@ def _av_read(filepath: tp.Union[str, Path], seek_time: float = 0, duration: floa
|
|
78 |
seek_time (float): Time at which to start reading in the file.
|
79 |
duration (float): Duration to read from the file. If set to -1, the whole file is read.
|
80 |
Returns:
|
81 |
-
|
82 |
"""
|
83 |
_init_av()
|
84 |
with av.open(str(filepath)) as af:
|
@@ -123,7 +123,7 @@ def audio_read(filepath: tp.Union[str, Path], seek_time: float = 0.,
|
|
123 |
duration (float): Duration to read from the file. If set to -1, the whole file is read.
|
124 |
pad (bool): Pad output audio if not reaching expected duration.
|
125 |
Returns:
|
126 |
-
|
127 |
"""
|
128 |
fp = Path(filepath)
|
129 |
if fp.suffix in ['.flac', '.ogg']: # TODO: check if we can safely use av_read for .ogg
|
@@ -136,12 +136,6 @@ def audio_read(filepath: tp.Union[str, Path], seek_time: float = 0.,
|
|
136 |
wav = torch.from_numpy(wav).t().contiguous()
|
137 |
if len(wav.shape) == 1:
|
138 |
wav = torch.unsqueeze(wav, 0)
|
139 |
-
elif (
|
140 |
-
fp.suffix in ['.wav', '.mp3'] and fp.suffix[1:] in ta.utils.sox_utils.list_read_formats()
|
141 |
-
and duration <= 0 and seek_time == 0
|
142 |
-
):
|
143 |
-
# Torchaudio is faster if we load an entire file at once.
|
144 |
-
wav, sr = ta.load(fp)
|
145 |
else:
|
146 |
wav, sr = _av_read(filepath, seek_time, duration)
|
147 |
if pad and duration > 0:
|
@@ -150,10 +144,22 @@ def audio_read(filepath: tp.Union[str, Path], seek_time: float = 0.,
|
|
150 |
return wav, sr
|
151 |
|
152 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
153 |
def audio_write(stem_name: tp.Union[str, Path],
|
154 |
wav: torch.Tensor, sample_rate: int,
|
155 |
-
format: str = 'wav', mp3_rate: int = 320,
|
156 |
-
strategy: str = 'peak', peak_clip_headroom_db: float = 1,
|
157 |
rms_headroom_db: float = 18, loudness_headroom_db: float = 14,
|
158 |
loudness_compressor: bool = False,
|
159 |
log_clipping: bool = True, make_parent_dir: bool = True,
|
@@ -162,8 +168,11 @@ def audio_write(stem_name: tp.Union[str, Path],
|
|
162 |
|
163 |
Args:
|
164 |
stem_name (str or Path): Filename without extension which will be added automatically.
|
165 |
-
|
|
|
|
|
166 |
mp3_rate (int): kbps when using mp3s.
|
|
|
167 |
normalize (bool): if `True` (default), normalizes according to the prescribed
|
168 |
strategy (see after). If `False`, the strategy is only used in case clipping
|
169 |
would happen.
|
@@ -175,7 +184,7 @@ def audio_write(stem_name: tp.Union[str, Path],
|
|
175 |
than the `peak_clip` one to avoid further clipping.
|
176 |
loudness_headroom_db (float): Target loudness for loudness normalization.
|
177 |
loudness_compressor (bool): Uses tanh for soft clipping when strategy is 'loudness'.
|
178 |
-
when strategy is 'loudness'log_clipping (bool): If True, basic logging on stderr when clipping still
|
179 |
occurs despite strategy (only for 'rms').
|
180 |
make_parent_dir (bool): Make parent directory if it doesn't exist.
|
181 |
Returns:
|
@@ -188,16 +197,23 @@ def audio_write(stem_name: tp.Union[str, Path],
|
|
188 |
raise ValueError("Input wav should be at most 2 dimension.")
|
189 |
assert wav.isfinite().all()
|
190 |
wav = normalize_audio(wav, normalize, strategy, peak_clip_headroom_db,
|
191 |
-
rms_headroom_db, loudness_headroom_db,
|
192 |
-
sample_rate=sample_rate,
|
193 |
-
|
194 |
if format == 'mp3':
|
195 |
suffix = '.mp3'
|
196 |
-
|
197 |
elif format == 'wav':
|
198 |
-
wav = i16_pcm(wav)
|
199 |
suffix = '.wav'
|
200 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
201 |
else:
|
202 |
raise RuntimeError(f"Invalid format {format}. Only wav or mp3 are supported.")
|
203 |
if not add_suffix:
|
@@ -206,7 +222,7 @@ def audio_write(stem_name: tp.Union[str, Path],
|
|
206 |
if make_parent_dir:
|
207 |
path.parent.mkdir(exist_ok=True, parents=True)
|
208 |
try:
|
209 |
-
|
210 |
except Exception:
|
211 |
if path.exists():
|
212 |
# we do not want to leave half written files around.
|
|
|
18 |
import soundfile
|
19 |
import torch
|
20 |
from torch.nn import functional as F
|
|
|
21 |
|
22 |
import av
|
23 |
+
import subprocess as sp
|
24 |
|
25 |
+
from .audio_utils import f32_pcm, normalize_audio
|
26 |
|
27 |
|
28 |
_av_initialized = False
|
|
|
78 |
seek_time (float): Time at which to start reading in the file.
|
79 |
duration (float): Duration to read from the file. If set to -1, the whole file is read.
|
80 |
Returns:
|
81 |
+
tuple of torch.Tensor, int: Tuple containing audio data and sample rate
|
82 |
"""
|
83 |
_init_av()
|
84 |
with av.open(str(filepath)) as af:
|
|
|
123 |
duration (float): Duration to read from the file. If set to -1, the whole file is read.
|
124 |
pad (bool): Pad output audio if not reaching expected duration.
|
125 |
Returns:
|
126 |
+
tuple of torch.Tensor, int: Tuple containing audio data and sample rate.
|
127 |
"""
|
128 |
fp = Path(filepath)
|
129 |
if fp.suffix in ['.flac', '.ogg']: # TODO: check if we can safely use av_read for .ogg
|
|
|
136 |
wav = torch.from_numpy(wav).t().contiguous()
|
137 |
if len(wav.shape) == 1:
|
138 |
wav = torch.unsqueeze(wav, 0)
|
|
|
|
|
|
|
|
|
|
|
|
|
139 |
else:
|
140 |
wav, sr = _av_read(filepath, seek_time, duration)
|
141 |
if pad and duration > 0:
|
|
|
144 |
return wav, sr
|
145 |
|
146 |
|
147 |
+
def _piping_to_ffmpeg(out_path: tp.Union[str, Path], wav: torch.Tensor, sample_rate: int, flags: tp.List[str]):
|
148 |
+
# ffmpeg is always installed and torchaudio is a bit unstable lately, so let's bypass it entirely.
|
149 |
+
assert wav.dim() == 2, wav.shape
|
150 |
+
command = [
|
151 |
+
'ffmpeg',
|
152 |
+
'-loglevel', 'error',
|
153 |
+
'-y', '-f', 'f32le', '-ar', str(sample_rate), '-ac', str(wav.shape[0]),
|
154 |
+
'-i', '-'] + flags + [str(out_path)]
|
155 |
+
input_ = f32_pcm(wav).t().detach().cpu().numpy().tobytes()
|
156 |
+
sp.run(command, input=input_, check=True)
|
157 |
+
|
158 |
+
|
159 |
def audio_write(stem_name: tp.Union[str, Path],
|
160 |
wav: torch.Tensor, sample_rate: int,
|
161 |
+
format: str = 'wav', mp3_rate: int = 320, ogg_rate: tp.Optional[int] = None,
|
162 |
+
normalize: bool = True, strategy: str = 'peak', peak_clip_headroom_db: float = 1,
|
163 |
rms_headroom_db: float = 18, loudness_headroom_db: float = 14,
|
164 |
loudness_compressor: bool = False,
|
165 |
log_clipping: bool = True, make_parent_dir: bool = True,
|
|
|
168 |
|
169 |
Args:
|
170 |
stem_name (str or Path): Filename without extension which will be added automatically.
|
171 |
+
wav (torch.Tensor): Audio data to save.
|
172 |
+
sample_rate (int): Sample rate of audio data.
|
173 |
+
format (str): Either "wav", "mp3", "ogg", or "flac".
|
174 |
mp3_rate (int): kbps when using mp3s.
|
175 |
+
ogg_rate (int): kbps when using ogg/vorbis. If not provided, let ffmpeg decide for itself.
|
176 |
normalize (bool): if `True` (default), normalizes according to the prescribed
|
177 |
strategy (see after). If `False`, the strategy is only used in case clipping
|
178 |
would happen.
|
|
|
184 |
than the `peak_clip` one to avoid further clipping.
|
185 |
loudness_headroom_db (float): Target loudness for loudness normalization.
|
186 |
loudness_compressor (bool): Uses tanh for soft clipping when strategy is 'loudness'.
|
187 |
+
when strategy is 'loudness' log_clipping (bool): If True, basic logging on stderr when clipping still
|
188 |
occurs despite strategy (only for 'rms').
|
189 |
make_parent_dir (bool): Make parent directory if it doesn't exist.
|
190 |
Returns:
|
|
|
197 |
raise ValueError("Input wav should be at most 2 dimension.")
|
198 |
assert wav.isfinite().all()
|
199 |
wav = normalize_audio(wav, normalize, strategy, peak_clip_headroom_db,
|
200 |
+
rms_headroom_db, loudness_headroom_db, loudness_compressor,
|
201 |
+
log_clipping=log_clipping, sample_rate=sample_rate,
|
202 |
+
stem_name=str(stem_name))
|
203 |
if format == 'mp3':
|
204 |
suffix = '.mp3'
|
205 |
+
flags = ['-f', 'mp3', '-c:a', 'libmp3lame', '-b:a', f'{mp3_rate}k']
|
206 |
elif format == 'wav':
|
|
|
207 |
suffix = '.wav'
|
208 |
+
flags = ['-f', 'wav', '-c:a', 'pcm_s16le']
|
209 |
+
elif format == 'ogg':
|
210 |
+
suffix = '.ogg'
|
211 |
+
flags = ['-f', 'ogg', '-c:a', 'libvorbis']
|
212 |
+
if ogg_rate is not None:
|
213 |
+
flags += ['-b:a', f'{ogg_rate}k']
|
214 |
+
elif format == 'flac':
|
215 |
+
suffix = '.flac'
|
216 |
+
flags = ['-f', 'flac']
|
217 |
else:
|
218 |
raise RuntimeError(f"Invalid format {format}. Only wav or mp3 are supported.")
|
219 |
if not add_suffix:
|
|
|
222 |
if make_parent_dir:
|
223 |
path.parent.mkdir(exist_ok=True, parents=True)
|
224 |
try:
|
225 |
+
_piping_to_ffmpeg(path, wav, sample_rate, flags)
|
226 |
except Exception:
|
227 |
if path.exists():
|
228 |
# we do not want to leave half written files around.
|
audiocraft/data/audio_dataset.py
CHANGED
@@ -3,12 +3,16 @@
|
|
3 |
#
|
4 |
# This source code is licensed under the license found in the
|
5 |
# LICENSE file in the root directory of this source tree.
|
6 |
-
|
|
|
|
|
|
|
7 |
import argparse
|
8 |
import copy
|
9 |
from concurrent.futures import ThreadPoolExecutor, Future
|
10 |
from dataclasses import dataclass, fields
|
11 |
from contextlib import ExitStack
|
|
|
12 |
import gzip
|
13 |
import json
|
14 |
import logging
|
@@ -81,9 +85,12 @@ class AudioMeta(BaseInfo):
|
|
81 |
class SegmentInfo(BaseInfo):
|
82 |
meta: AudioMeta
|
83 |
seek_time: float
|
84 |
-
|
|
|
|
|
85 |
total_frames: int # total number of frames, padding included
|
86 |
-
sample_rate: int
|
|
|
87 |
|
88 |
|
89 |
DEFAULT_EXTS = ['.wav', '.mp3', '.flac', '.ogg', '.m4a']
|
@@ -114,8 +121,8 @@ def _resolve_audio_meta(m: AudioMeta, fast: bool = True) -> AudioMeta:
|
|
114 |
|
115 |
Args:
|
116 |
m (AudioMeta): Audio meta to resolve.
|
117 |
-
fast (bool): If True, uses a really fast check for determining if a file
|
118 |
-
Only valid on Linux/Mac.
|
119 |
Returns:
|
120 |
AudioMeta: Audio meta with resolved path.
|
121 |
"""
|
@@ -151,7 +158,7 @@ def find_audio_files(path: tp.Union[Path, str],
|
|
151 |
progress (bool): Whether to log progress on audio files collection.
|
152 |
workers (int): number of parallel workers, if 0, use only the current thread.
|
153 |
Returns:
|
154 |
-
|
155 |
"""
|
156 |
audio_files = []
|
157 |
futures: tp.List[Future] = []
|
@@ -203,7 +210,7 @@ def load_audio_meta(path: tp.Union[str, Path],
|
|
203 |
resolve (bool): Whether to resolve the path from AudioMeta (default=True).
|
204 |
fast (bool): activates some tricks to make things faster.
|
205 |
Returns:
|
206 |
-
|
207 |
"""
|
208 |
open_fn = gzip.open if str(path).lower().endswith('.gz') else open
|
209 |
with open_fn(path, 'rb') as fp: # type: ignore
|
@@ -250,9 +257,14 @@ class AudioDataset:
|
|
250 |
allows to return a tuple containing the torch Tensor and additional metadata on the segment and the
|
251 |
original audio meta.
|
252 |
|
|
|
|
|
|
|
|
|
|
|
253 |
Args:
|
254 |
-
meta (
|
255 |
-
segment_duration (float): Optional segment duration of audio to load.
|
256 |
If not specified, the dataset will load the full audio segment from the file.
|
257 |
shuffle (bool): Set to `True` to have the data reshuffled at every epoch.
|
258 |
sample_rate (int): Target sample rate of the loaded audio samples.
|
@@ -266,10 +278,19 @@ class AudioDataset:
|
|
266 |
is shorter than the desired segment.
|
267 |
max_read_retry (int): Maximum number of retries to sample an audio segment from the dataset.
|
268 |
return_info (bool): Whether to return the wav only or return wav along with segment info and metadata.
|
269 |
-
min_audio_duration (
|
270 |
audio shorter than this will be filtered out.
|
271 |
-
max_audio_duration (
|
272 |
audio longer than this will be filtered out.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
273 |
"""
|
274 |
def __init__(self,
|
275 |
meta: tp.List[AudioMeta],
|
@@ -285,16 +306,14 @@ class AudioDataset:
|
|
285 |
max_read_retry: int = 10,
|
286 |
return_info: bool = False,
|
287 |
min_audio_duration: tp.Optional[float] = None,
|
288 |
-
max_audio_duration: tp.Optional[float] = None
|
|
|
|
|
|
|
289 |
):
|
290 |
-
assert len(meta) > 0,
|
291 |
assert segment_duration is None or segment_duration > 0
|
292 |
assert segment_duration is None or min_segment_ratio >= 0
|
293 |
-
logging.debug(f'sample_on_duration: {sample_on_duration}')
|
294 |
-
logging.debug(f'sample_on_weight: {sample_on_weight}')
|
295 |
-
logging.debug(f'pad: {pad}')
|
296 |
-
logging.debug(f'min_segment_ratio: {min_segment_ratio}')
|
297 |
-
|
298 |
self.segment_duration = segment_duration
|
299 |
self.min_segment_ratio = min_segment_ratio
|
300 |
self.max_audio_duration = max_audio_duration
|
@@ -317,13 +336,25 @@ class AudioDataset:
|
|
317 |
self.sampling_probabilities = self._get_sampling_probabilities()
|
318 |
self.max_read_retry = max_read_retry
|
319 |
self.return_info = return_info
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
320 |
|
321 |
def __len__(self):
|
322 |
return self.num_samples
|
323 |
|
324 |
def _get_sampling_probabilities(self, normalized: bool = True):
|
325 |
-
"""Return the sampling probabilities for each file inside `self.meta`.
|
326 |
-
"""
|
327 |
scores: tp.List[float] = []
|
328 |
for file_meta in self.meta:
|
329 |
score = 1.
|
@@ -337,12 +368,32 @@ class AudioDataset:
|
|
337 |
probabilities /= probabilities.sum()
|
338 |
return probabilities
|
339 |
|
340 |
-
|
341 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
342 |
This is only called if `segment_duration` is not None.
|
343 |
|
344 |
You must use the provided random number generator `rng` for reproducibility.
|
|
|
345 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
346 |
if not self.sample_on_weight and not self.sample_on_duration:
|
347 |
file_index = int(torch.randint(len(self.sampling_probabilities), (1,), generator=rng).item())
|
348 |
else:
|
@@ -350,6 +401,15 @@ class AudioDataset:
|
|
350 |
|
351 |
return self.meta[file_index]
|
352 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
353 |
def __getitem__(self, index: int) -> tp.Union[torch.Tensor, tp.Tuple[torch.Tensor, SegmentInfo]]:
|
354 |
if self.segment_duration is None:
|
355 |
file_meta = self.meta[index]
|
@@ -357,18 +417,22 @@ class AudioDataset:
|
|
357 |
out = convert_audio(out, sr, self.sample_rate, self.channels)
|
358 |
n_frames = out.shape[-1]
|
359 |
segment_info = SegmentInfo(file_meta, seek_time=0., n_frames=n_frames, total_frames=n_frames,
|
360 |
-
sample_rate=self.sample_rate)
|
361 |
else:
|
362 |
rng = torch.Generator()
|
363 |
if self.shuffle:
|
364 |
-
# We use index, plus extra randomness
|
365 |
-
|
|
|
|
|
|
|
|
|
366 |
else:
|
367 |
# We only use index
|
368 |
rng.manual_seed(index)
|
369 |
|
370 |
for retry in range(self.max_read_retry):
|
371 |
-
file_meta = self.sample_file(rng)
|
372 |
# We add some variance in the file position even if audio file is smaller than segment
|
373 |
# without ending up with empty segments
|
374 |
max_seek = max(0, file_meta.duration - self.segment_duration * self.min_segment_ratio)
|
@@ -381,7 +445,7 @@ class AudioDataset:
|
|
381 |
if self.pad:
|
382 |
out = F.pad(out, (0, target_frames - n_frames))
|
383 |
segment_info = SegmentInfo(file_meta, seek_time, n_frames=n_frames, total_frames=target_frames,
|
384 |
-
sample_rate=self.sample_rate)
|
385 |
except Exception as exc:
|
386 |
logger.warning("Error opening file %s: %r", file_meta.path, exc)
|
387 |
if retry == self.max_read_retry - 1:
|
@@ -423,7 +487,7 @@ class AudioDataset:
|
|
423 |
if to_pad:
|
424 |
# Each wav could be of a different duration as they are not segmented.
|
425 |
for i in range(len(samples)):
|
426 |
-
# Determines the total
|
427 |
segment_infos[i].total_frames = max_len
|
428 |
wavs[i] = _pad_wav(wavs[i])
|
429 |
|
@@ -436,9 +500,7 @@ class AudioDataset:
|
|
436 |
return torch.stack(samples)
|
437 |
|
438 |
def _filter_duration(self, meta: tp.List[AudioMeta]) -> tp.List[AudioMeta]:
|
439 |
-
"""Filters out audio files with
|
440 |
-
Removes from meta files that have durations that will not allow to samples examples from them.
|
441 |
-
"""
|
442 |
orig_len = len(meta)
|
443 |
|
444 |
# Filter data that is too short.
|
|
|
3 |
#
|
4 |
# This source code is licensed under the license found in the
|
5 |
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""AudioDataset support. In order to handle a larger number of files
|
7 |
+
without having to scan again the folders, we precompute some metadata
|
8 |
+
(filename, sample rate, duration), and use that to efficiently sample audio segments.
|
9 |
+
"""
|
10 |
import argparse
|
11 |
import copy
|
12 |
from concurrent.futures import ThreadPoolExecutor, Future
|
13 |
from dataclasses import dataclass, fields
|
14 |
from contextlib import ExitStack
|
15 |
+
from functools import lru_cache
|
16 |
import gzip
|
17 |
import json
|
18 |
import logging
|
|
|
85 |
class SegmentInfo(BaseInfo):
|
86 |
meta: AudioMeta
|
87 |
seek_time: float
|
88 |
+
# The following values are given once the audio is processed, e.g.
|
89 |
+
# at the target sample rate and target number of channels.
|
90 |
+
n_frames: int # actual number of frames without padding
|
91 |
total_frames: int # total number of frames, padding included
|
92 |
+
sample_rate: int # actual sample rate
|
93 |
+
channels: int # number of audio channels.
|
94 |
|
95 |
|
96 |
DEFAULT_EXTS = ['.wav', '.mp3', '.flac', '.ogg', '.m4a']
|
|
|
121 |
|
122 |
Args:
|
123 |
m (AudioMeta): Audio meta to resolve.
|
124 |
+
fast (bool): If True, uses a really fast check for determining if a file
|
125 |
+
is already absolute or not. Only valid on Linux/Mac.
|
126 |
Returns:
|
127 |
AudioMeta: Audio meta with resolved path.
|
128 |
"""
|
|
|
158 |
progress (bool): Whether to log progress on audio files collection.
|
159 |
workers (int): number of parallel workers, if 0, use only the current thread.
|
160 |
Returns:
|
161 |
+
list of AudioMeta: List of audio file path and its metadata.
|
162 |
"""
|
163 |
audio_files = []
|
164 |
futures: tp.List[Future] = []
|
|
|
210 |
resolve (bool): Whether to resolve the path from AudioMeta (default=True).
|
211 |
fast (bool): activates some tricks to make things faster.
|
212 |
Returns:
|
213 |
+
list of AudioMeta: List of audio file path and its total duration.
|
214 |
"""
|
215 |
open_fn = gzip.open if str(path).lower().endswith('.gz') else open
|
216 |
with open_fn(path, 'rb') as fp: # type: ignore
|
|
|
257 |
allows to return a tuple containing the torch Tensor and additional metadata on the segment and the
|
258 |
original audio meta.
|
259 |
|
260 |
+
Note that you can call `start_epoch(epoch)` in order to get
|
261 |
+
a deterministic "randomization" for `shuffle=True`.
|
262 |
+
For a given epoch and dataset index, this will always return the same extract.
|
263 |
+
You can get back some diversity by setting the `shuffle_seed` param.
|
264 |
+
|
265 |
Args:
|
266 |
+
meta (list of AudioMeta): List of audio files metadata.
|
267 |
+
segment_duration (float, optional): Optional segment duration of audio to load.
|
268 |
If not specified, the dataset will load the full audio segment from the file.
|
269 |
shuffle (bool): Set to `True` to have the data reshuffled at every epoch.
|
270 |
sample_rate (int): Target sample rate of the loaded audio samples.
|
|
|
278 |
is shorter than the desired segment.
|
279 |
max_read_retry (int): Maximum number of retries to sample an audio segment from the dataset.
|
280 |
return_info (bool): Whether to return the wav only or return wav along with segment info and metadata.
|
281 |
+
min_audio_duration (float, optional): Minimum audio file duration, in seconds, if provided
|
282 |
audio shorter than this will be filtered out.
|
283 |
+
max_audio_duration (float, optional): Maximal audio file duration in seconds, if provided
|
284 |
audio longer than this will be filtered out.
|
285 |
+
shuffle_seed (int): can be used to further randomize
|
286 |
+
load_wav (bool): if False, skip loading the wav but returns a tensor of 0
|
287 |
+
with the expected segment_duration (which must be provided if load_wav is False).
|
288 |
+
permutation_on_files (bool): only if `sample_on_weight` and `sample_on_duration`
|
289 |
+
are False. Will ensure a permutation on files when going through the dataset.
|
290 |
+
In that case the epoch number must be provided in order for the model
|
291 |
+
to continue the permutation across epochs. In that case, it is assumed
|
292 |
+
that `num_samples = total_batch_size * num_updates_per_epoch`, with
|
293 |
+
`total_batch_size` the overall batch size accounting for all gpus.
|
294 |
"""
|
295 |
def __init__(self,
|
296 |
meta: tp.List[AudioMeta],
|
|
|
306 |
max_read_retry: int = 10,
|
307 |
return_info: bool = False,
|
308 |
min_audio_duration: tp.Optional[float] = None,
|
309 |
+
max_audio_duration: tp.Optional[float] = None,
|
310 |
+
shuffle_seed: int = 0,
|
311 |
+
load_wav: bool = True,
|
312 |
+
permutation_on_files: bool = False,
|
313 |
):
|
314 |
+
assert len(meta) > 0, "No audio meta provided to AudioDataset. Please check loading of audio meta."
|
315 |
assert segment_duration is None or segment_duration > 0
|
316 |
assert segment_duration is None or min_segment_ratio >= 0
|
|
|
|
|
|
|
|
|
|
|
317 |
self.segment_duration = segment_duration
|
318 |
self.min_segment_ratio = min_segment_ratio
|
319 |
self.max_audio_duration = max_audio_duration
|
|
|
336 |
self.sampling_probabilities = self._get_sampling_probabilities()
|
337 |
self.max_read_retry = max_read_retry
|
338 |
self.return_info = return_info
|
339 |
+
self.shuffle_seed = shuffle_seed
|
340 |
+
self.current_epoch: tp.Optional[int] = None
|
341 |
+
self.load_wav = load_wav
|
342 |
+
if not load_wav:
|
343 |
+
assert segment_duration is not None
|
344 |
+
self.permutation_on_files = permutation_on_files
|
345 |
+
if permutation_on_files:
|
346 |
+
assert not self.sample_on_duration
|
347 |
+
assert not self.sample_on_weight
|
348 |
+
assert self.shuffle
|
349 |
+
|
350 |
+
def start_epoch(self, epoch: int):
|
351 |
+
self.current_epoch = epoch
|
352 |
|
353 |
def __len__(self):
|
354 |
return self.num_samples
|
355 |
|
356 |
def _get_sampling_probabilities(self, normalized: bool = True):
|
357 |
+
"""Return the sampling probabilities for each file inside `self.meta`."""
|
|
|
358 |
scores: tp.List[float] = []
|
359 |
for file_meta in self.meta:
|
360 |
score = 1.
|
|
|
368 |
probabilities /= probabilities.sum()
|
369 |
return probabilities
|
370 |
|
371 |
+
@staticmethod
|
372 |
+
@lru_cache(16)
|
373 |
+
def _get_file_permutation(num_files: int, permutation_index: int, base_seed: int):
|
374 |
+
# Used to keep the most recent files permutation in memory implicitely.
|
375 |
+
# will work unless someone is using a lot of Datasets in parallel.
|
376 |
+
rng = torch.Generator()
|
377 |
+
rng.manual_seed(base_seed + permutation_index)
|
378 |
+
return torch.randperm(num_files, generator=rng)
|
379 |
+
|
380 |
+
def sample_file(self, index: int, rng: torch.Generator) -> AudioMeta:
|
381 |
+
"""Sample a given file from `self.meta`. Can be overridden in subclasses.
|
382 |
This is only called if `segment_duration` is not None.
|
383 |
|
384 |
You must use the provided random number generator `rng` for reproducibility.
|
385 |
+
You can further make use of the index accessed.
|
386 |
"""
|
387 |
+
if self.permutation_on_files:
|
388 |
+
assert self.current_epoch is not None
|
389 |
+
total_index = self.current_epoch * len(self) + index
|
390 |
+
permutation_index = total_index // len(self.meta)
|
391 |
+
relative_index = total_index % len(self.meta)
|
392 |
+
permutation = AudioDataset._get_file_permutation(
|
393 |
+
len(self.meta), permutation_index, self.shuffle_seed)
|
394 |
+
file_index = permutation[relative_index]
|
395 |
+
return self.meta[file_index]
|
396 |
+
|
397 |
if not self.sample_on_weight and not self.sample_on_duration:
|
398 |
file_index = int(torch.randint(len(self.sampling_probabilities), (1,), generator=rng).item())
|
399 |
else:
|
|
|
401 |
|
402 |
return self.meta[file_index]
|
403 |
|
404 |
+
def _audio_read(self, path: str, seek_time: float = 0, duration: float = -1):
|
405 |
+
# Override this method in subclass if needed.
|
406 |
+
if self.load_wav:
|
407 |
+
return audio_read(path, seek_time, duration, pad=False)
|
408 |
+
else:
|
409 |
+
assert self.segment_duration is not None
|
410 |
+
n_frames = int(self.sample_rate * self.segment_duration)
|
411 |
+
return torch.zeros(self.channels, n_frames), self.sample_rate
|
412 |
+
|
413 |
def __getitem__(self, index: int) -> tp.Union[torch.Tensor, tp.Tuple[torch.Tensor, SegmentInfo]]:
|
414 |
if self.segment_duration is None:
|
415 |
file_meta = self.meta[index]
|
|
|
417 |
out = convert_audio(out, sr, self.sample_rate, self.channels)
|
418 |
n_frames = out.shape[-1]
|
419 |
segment_info = SegmentInfo(file_meta, seek_time=0., n_frames=n_frames, total_frames=n_frames,
|
420 |
+
sample_rate=self.sample_rate, channels=out.shape[0])
|
421 |
else:
|
422 |
rng = torch.Generator()
|
423 |
if self.shuffle:
|
424 |
+
# We use index, plus extra randomness, either totally random if we don't know the epoch.
|
425 |
+
# otherwise we make use of the epoch number and optional shuffle_seed.
|
426 |
+
if self.current_epoch is None:
|
427 |
+
rng.manual_seed(index + self.num_samples * random.randint(0, 2**24))
|
428 |
+
else:
|
429 |
+
rng.manual_seed(index + self.num_samples * (self.current_epoch + self.shuffle_seed))
|
430 |
else:
|
431 |
# We only use index
|
432 |
rng.manual_seed(index)
|
433 |
|
434 |
for retry in range(self.max_read_retry):
|
435 |
+
file_meta = self.sample_file(index, rng)
|
436 |
# We add some variance in the file position even if audio file is smaller than segment
|
437 |
# without ending up with empty segments
|
438 |
max_seek = max(0, file_meta.duration - self.segment_duration * self.min_segment_ratio)
|
|
|
445 |
if self.pad:
|
446 |
out = F.pad(out, (0, target_frames - n_frames))
|
447 |
segment_info = SegmentInfo(file_meta, seek_time, n_frames=n_frames, total_frames=target_frames,
|
448 |
+
sample_rate=self.sample_rate, channels=out.shape[0])
|
449 |
except Exception as exc:
|
450 |
logger.warning("Error opening file %s: %r", file_meta.path, exc)
|
451 |
if retry == self.max_read_retry - 1:
|
|
|
487 |
if to_pad:
|
488 |
# Each wav could be of a different duration as they are not segmented.
|
489 |
for i in range(len(samples)):
|
490 |
+
# Determines the total length of the signal with padding, so we update here as we pad.
|
491 |
segment_infos[i].total_frames = max_len
|
492 |
wavs[i] = _pad_wav(wavs[i])
|
493 |
|
|
|
500 |
return torch.stack(samples)
|
501 |
|
502 |
def _filter_duration(self, meta: tp.List[AudioMeta]) -> tp.List[AudioMeta]:
|
503 |
+
"""Filters out audio files with audio durations that will not allow to sample examples from them."""
|
|
|
|
|
504 |
orig_len = len(meta)
|
505 |
|
506 |
# Filter data that is too short.
|
audiocraft/data/audio_utils.py
CHANGED
@@ -3,7 +3,8 @@
|
|
3 |
#
|
4 |
# This source code is licensed under the license found in the
|
5 |
# LICENSE file in the root directory of this source tree.
|
6 |
-
|
|
|
7 |
import sys
|
8 |
import typing as tp
|
9 |
|
@@ -47,8 +48,7 @@ def convert_audio_channels(wav: torch.Tensor, channels: int = 2) -> torch.Tensor
|
|
47 |
|
48 |
def convert_audio(wav: torch.Tensor, from_rate: float,
|
49 |
to_rate: float, to_channels: int) -> torch.Tensor:
|
50 |
-
"""Convert audio to new sample rate and number of audio channels.
|
51 |
-
"""
|
52 |
wav = julius.resample_frac(wav, int(from_rate), int(to_rate))
|
53 |
wav = convert_audio_channels(wav, to_channels)
|
54 |
return wav
|
@@ -66,7 +66,7 @@ def normalize_loudness(wav: torch.Tensor, sample_rate: int, loudness_headroom_db
|
|
66 |
loudness_compressor (bool): Uses tanh for soft clipping.
|
67 |
energy_floor (float): anything below that RMS level will not be rescaled.
|
68 |
Returns:
|
69 |
-
|
70 |
"""
|
71 |
energy = wav.pow(2).mean().sqrt().item()
|
72 |
if energy < energy_floor:
|
@@ -117,7 +117,7 @@ def normalize_audio(wav: torch.Tensor, normalize: bool = True,
|
|
117 |
log_clipping (bool): If True, basic logging on stderr when clipping still
|
118 |
occurs despite strategy (only for 'rms').
|
119 |
sample_rate (int): Sample rate for the audio data (required for loudness).
|
120 |
-
stem_name (
|
121 |
Returns:
|
122 |
torch.Tensor: Normalized audio.
|
123 |
"""
|
@@ -150,17 +150,19 @@ def f32_pcm(wav: torch.Tensor) -> torch.Tensor:
|
|
150 |
"""
|
151 |
if wav.dtype.is_floating_point:
|
152 |
return wav
|
153 |
-
|
154 |
-
assert wav.dtype == torch.int16
|
155 |
return wav.float() / 2**15
|
|
|
|
|
|
|
156 |
|
157 |
|
158 |
def i16_pcm(wav: torch.Tensor) -> torch.Tensor:
|
159 |
"""Convert audio to int 16 bits PCM format.
|
160 |
|
161 |
-
..Warning:: There exist many formula for doing this
|
162 |
-
due to the
|
163 |
-
or
|
164 |
it is possible that `i16_pcm(f32_pcm)) != Identity`.
|
165 |
"""
|
166 |
if wav.dtype.is_floating_point:
|
|
|
3 |
#
|
4 |
# This source code is licensed under the license found in the
|
5 |
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""Various utilities for audio convertion (pcm format, sample rate and channels),
|
7 |
+
and volume normalization."""
|
8 |
import sys
|
9 |
import typing as tp
|
10 |
|
|
|
48 |
|
49 |
def convert_audio(wav: torch.Tensor, from_rate: float,
|
50 |
to_rate: float, to_channels: int) -> torch.Tensor:
|
51 |
+
"""Convert audio to new sample rate and number of audio channels."""
|
|
|
52 |
wav = julius.resample_frac(wav, int(from_rate), int(to_rate))
|
53 |
wav = convert_audio_channels(wav, to_channels)
|
54 |
return wav
|
|
|
66 |
loudness_compressor (bool): Uses tanh for soft clipping.
|
67 |
energy_floor (float): anything below that RMS level will not be rescaled.
|
68 |
Returns:
|
69 |
+
torch.Tensor: Loudness normalized output data.
|
70 |
"""
|
71 |
energy = wav.pow(2).mean().sqrt().item()
|
72 |
if energy < energy_floor:
|
|
|
117 |
log_clipping (bool): If True, basic logging on stderr when clipping still
|
118 |
occurs despite strategy (only for 'rms').
|
119 |
sample_rate (int): Sample rate for the audio data (required for loudness).
|
120 |
+
stem_name (str, optional): Stem name for clipping logging.
|
121 |
Returns:
|
122 |
torch.Tensor: Normalized audio.
|
123 |
"""
|
|
|
150 |
"""
|
151 |
if wav.dtype.is_floating_point:
|
152 |
return wav
|
153 |
+
elif wav.dtype == torch.int16:
|
|
|
154 |
return wav.float() / 2**15
|
155 |
+
elif wav.dtype == torch.int32:
|
156 |
+
return wav.float() / 2**31
|
157 |
+
raise ValueError(f"Unsupported wav dtype: {wav.dtype}")
|
158 |
|
159 |
|
160 |
def i16_pcm(wav: torch.Tensor) -> torch.Tensor:
|
161 |
"""Convert audio to int 16 bits PCM format.
|
162 |
|
163 |
+
..Warning:: There exist many formula for doing this conversion. None are perfect
|
164 |
+
due to the asymmetry of the int16 range. One either have possible clipping, DC offset,
|
165 |
+
or inconsistencies with f32_pcm. If the given wav doesn't have enough headroom,
|
166 |
it is possible that `i16_pcm(f32_pcm)) != Identity`.
|
167 |
"""
|
168 |
if wav.dtype.is_floating_point:
|
audiocraft/data/info_audio_dataset.py
ADDED
@@ -0,0 +1,110 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""Base classes for the datasets that also provide non-audio metadata,
|
7 |
+
e.g. description, text transcription etc.
|
8 |
+
"""
|
9 |
+
from dataclasses import dataclass
|
10 |
+
import logging
|
11 |
+
import math
|
12 |
+
import re
|
13 |
+
import typing as tp
|
14 |
+
|
15 |
+
import torch
|
16 |
+
|
17 |
+
from .audio_dataset import AudioDataset, AudioMeta
|
18 |
+
from ..environment import AudioCraftEnvironment
|
19 |
+
from ..modules.conditioners import SegmentWithAttributes, ConditioningAttributes
|
20 |
+
|
21 |
+
|
22 |
+
logger = logging.getLogger(__name__)
|
23 |
+
|
24 |
+
|
25 |
+
def _clusterify_meta(meta: AudioMeta) -> AudioMeta:
|
26 |
+
"""Monkey-patch meta to match cluster specificities."""
|
27 |
+
meta.path = AudioCraftEnvironment.apply_dataset_mappers(meta.path)
|
28 |
+
if meta.info_path is not None:
|
29 |
+
meta.info_path.zip_path = AudioCraftEnvironment.apply_dataset_mappers(meta.info_path.zip_path)
|
30 |
+
return meta
|
31 |
+
|
32 |
+
|
33 |
+
def clusterify_all_meta(meta: tp.List[AudioMeta]) -> tp.List[AudioMeta]:
|
34 |
+
"""Monkey-patch all meta to match cluster specificities."""
|
35 |
+
return [_clusterify_meta(m) for m in meta]
|
36 |
+
|
37 |
+
|
38 |
+
@dataclass
|
39 |
+
class AudioInfo(SegmentWithAttributes):
|
40 |
+
"""Dummy SegmentInfo with empty attributes.
|
41 |
+
|
42 |
+
The InfoAudioDataset is expected to return metadata that inherits
|
43 |
+
from SegmentWithAttributes class and can return conditioning attributes.
|
44 |
+
|
45 |
+
This basically guarantees all datasets will be compatible with current
|
46 |
+
solver that contain conditioners requiring this.
|
47 |
+
"""
|
48 |
+
audio_tokens: tp.Optional[torch.Tensor] = None # populated when using cached batch for training a LM.
|
49 |
+
|
50 |
+
def to_condition_attributes(self) -> ConditioningAttributes:
|
51 |
+
return ConditioningAttributes()
|
52 |
+
|
53 |
+
|
54 |
+
class InfoAudioDataset(AudioDataset):
|
55 |
+
"""AudioDataset that always returns metadata as SegmentWithAttributes along with the audio waveform.
|
56 |
+
|
57 |
+
See `audiocraft.data.audio_dataset.AudioDataset` for initialization arguments.
|
58 |
+
"""
|
59 |
+
def __init__(self, meta: tp.List[AudioMeta], **kwargs):
|
60 |
+
super().__init__(clusterify_all_meta(meta), **kwargs)
|
61 |
+
|
62 |
+
def __getitem__(self, index: int) -> tp.Union[torch.Tensor, tp.Tuple[torch.Tensor, SegmentWithAttributes]]:
|
63 |
+
if not self.return_info:
|
64 |
+
wav = super().__getitem__(index)
|
65 |
+
assert isinstance(wav, torch.Tensor)
|
66 |
+
return wav
|
67 |
+
wav, meta = super().__getitem__(index)
|
68 |
+
return wav, AudioInfo(**meta.to_dict())
|
69 |
+
|
70 |
+
|
71 |
+
def get_keyword_or_keyword_list(value: tp.Optional[str]) -> tp.Union[tp.Optional[str], tp.Optional[tp.List[str]]]:
|
72 |
+
"""Preprocess a single keyword or possible a list of keywords."""
|
73 |
+
if isinstance(value, list):
|
74 |
+
return get_keyword_list(value)
|
75 |
+
else:
|
76 |
+
return get_keyword(value)
|
77 |
+
|
78 |
+
|
79 |
+
def get_string(value: tp.Optional[str]) -> tp.Optional[str]:
|
80 |
+
"""Preprocess a single keyword."""
|
81 |
+
if value is None or (not isinstance(value, str)) or len(value) == 0 or value == 'None':
|
82 |
+
return None
|
83 |
+
else:
|
84 |
+
return value.strip()
|
85 |
+
|
86 |
+
|
87 |
+
def get_keyword(value: tp.Optional[str]) -> tp.Optional[str]:
|
88 |
+
"""Preprocess a single keyword."""
|
89 |
+
if value is None or (not isinstance(value, str)) or len(value) == 0 or value == 'None':
|
90 |
+
return None
|
91 |
+
else:
|
92 |
+
return value.strip().lower()
|
93 |
+
|
94 |
+
|
95 |
+
def get_keyword_list(values: tp.Union[str, tp.List[str]]) -> tp.Optional[tp.List[str]]:
|
96 |
+
"""Preprocess a list of keywords."""
|
97 |
+
if isinstance(values, str):
|
98 |
+
values = [v.strip() for v in re.split(r'[,\s]', values)]
|
99 |
+
elif isinstance(values, float) and math.isnan(values):
|
100 |
+
values = []
|
101 |
+
if not isinstance(values, list):
|
102 |
+
logger.debug(f"Unexpected keyword list {values}")
|
103 |
+
values = [str(values)]
|
104 |
+
|
105 |
+
kws = [get_keyword(v) for v in values]
|
106 |
+
kw_list = [k for k in kws if k is not None]
|
107 |
+
if len(kw_list) == 0:
|
108 |
+
return None
|
109 |
+
else:
|
110 |
+
return kw_list
|
audiocraft/data/music_dataset.py
ADDED
@@ -0,0 +1,270 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""Dataset of music tracks with rich metadata.
|
7 |
+
"""
|
8 |
+
from dataclasses import dataclass, field, fields, replace
|
9 |
+
import gzip
|
10 |
+
import json
|
11 |
+
import logging
|
12 |
+
from pathlib import Path
|
13 |
+
import random
|
14 |
+
import typing as tp
|
15 |
+
|
16 |
+
import torch
|
17 |
+
|
18 |
+
from .info_audio_dataset import (
|
19 |
+
InfoAudioDataset,
|
20 |
+
AudioInfo,
|
21 |
+
get_keyword_list,
|
22 |
+
get_keyword,
|
23 |
+
get_string
|
24 |
+
)
|
25 |
+
from ..modules.conditioners import (
|
26 |
+
ConditioningAttributes,
|
27 |
+
JointEmbedCondition,
|
28 |
+
WavCondition,
|
29 |
+
)
|
30 |
+
from ..utils.utils import warn_once
|
31 |
+
|
32 |
+
|
33 |
+
logger = logging.getLogger(__name__)
|
34 |
+
|
35 |
+
|
36 |
+
@dataclass
|
37 |
+
class MusicInfo(AudioInfo):
|
38 |
+
"""Segment info augmented with music metadata.
|
39 |
+
"""
|
40 |
+
# music-specific metadata
|
41 |
+
title: tp.Optional[str] = None
|
42 |
+
artist: tp.Optional[str] = None # anonymized artist id, used to ensure no overlap between splits
|
43 |
+
key: tp.Optional[str] = None
|
44 |
+
bpm: tp.Optional[float] = None
|
45 |
+
genre: tp.Optional[str] = None
|
46 |
+
moods: tp.Optional[list] = None
|
47 |
+
keywords: tp.Optional[list] = None
|
48 |
+
description: tp.Optional[str] = None
|
49 |
+
name: tp.Optional[str] = None
|
50 |
+
instrument: tp.Optional[str] = None
|
51 |
+
# original wav accompanying the metadata
|
52 |
+
self_wav: tp.Optional[WavCondition] = None
|
53 |
+
# dict mapping attributes names to tuple of wav, text and metadata
|
54 |
+
joint_embed: tp.Dict[str, JointEmbedCondition] = field(default_factory=dict)
|
55 |
+
|
56 |
+
@property
|
57 |
+
def has_music_meta(self) -> bool:
|
58 |
+
return self.name is not None
|
59 |
+
|
60 |
+
def to_condition_attributes(self) -> ConditioningAttributes:
|
61 |
+
out = ConditioningAttributes()
|
62 |
+
for _field in fields(self):
|
63 |
+
key, value = _field.name, getattr(self, _field.name)
|
64 |
+
if key == 'self_wav':
|
65 |
+
out.wav[key] = value
|
66 |
+
elif key == 'joint_embed':
|
67 |
+
for embed_attribute, embed_cond in value.items():
|
68 |
+
out.joint_embed[embed_attribute] = embed_cond
|
69 |
+
else:
|
70 |
+
if isinstance(value, list):
|
71 |
+
value = ' '.join(value)
|
72 |
+
out.text[key] = value
|
73 |
+
return out
|
74 |
+
|
75 |
+
@staticmethod
|
76 |
+
def attribute_getter(attribute):
|
77 |
+
if attribute == 'bpm':
|
78 |
+
preprocess_func = get_bpm
|
79 |
+
elif attribute == 'key':
|
80 |
+
preprocess_func = get_musical_key
|
81 |
+
elif attribute in ['moods', 'keywords']:
|
82 |
+
preprocess_func = get_keyword_list
|
83 |
+
elif attribute in ['genre', 'name', 'instrument']:
|
84 |
+
preprocess_func = get_keyword
|
85 |
+
elif attribute in ['title', 'artist', 'description']:
|
86 |
+
preprocess_func = get_string
|
87 |
+
else:
|
88 |
+
preprocess_func = None
|
89 |
+
return preprocess_func
|
90 |
+
|
91 |
+
@classmethod
|
92 |
+
def from_dict(cls, dictionary: dict, fields_required: bool = False):
|
93 |
+
_dictionary: tp.Dict[str, tp.Any] = {}
|
94 |
+
|
95 |
+
# allow a subset of attributes to not be loaded from the dictionary
|
96 |
+
# these attributes may be populated later
|
97 |
+
post_init_attributes = ['self_wav', 'joint_embed']
|
98 |
+
optional_fields = ['keywords']
|
99 |
+
|
100 |
+
for _field in fields(cls):
|
101 |
+
if _field.name in post_init_attributes:
|
102 |
+
continue
|
103 |
+
elif _field.name not in dictionary:
|
104 |
+
if fields_required and _field.name not in optional_fields:
|
105 |
+
raise KeyError(f"Unexpected missing key: {_field.name}")
|
106 |
+
else:
|
107 |
+
preprocess_func: tp.Optional[tp.Callable] = cls.attribute_getter(_field.name)
|
108 |
+
value = dictionary[_field.name]
|
109 |
+
if preprocess_func:
|
110 |
+
value = preprocess_func(value)
|
111 |
+
_dictionary[_field.name] = value
|
112 |
+
return cls(**_dictionary)
|
113 |
+
|
114 |
+
|
115 |
+
def augment_music_info_description(music_info: MusicInfo, merge_text_p: float = 0.,
|
116 |
+
drop_desc_p: float = 0., drop_other_p: float = 0.) -> MusicInfo:
|
117 |
+
"""Augment MusicInfo description with additional metadata fields and potential dropout.
|
118 |
+
Additional textual attributes are added given probability 'merge_text_conditions_p' and
|
119 |
+
the original textual description is dropped from the augmented description given probability drop_desc_p.
|
120 |
+
|
121 |
+
Args:
|
122 |
+
music_info (MusicInfo): The music metadata to augment.
|
123 |
+
merge_text_p (float): Probability of merging additional metadata to the description.
|
124 |
+
If provided value is 0, then no merging is performed.
|
125 |
+
drop_desc_p (float): Probability of dropping the original description on text merge.
|
126 |
+
if provided value is 0, then no drop out is performed.
|
127 |
+
drop_other_p (float): Probability of dropping the other fields used for text augmentation.
|
128 |
+
Returns:
|
129 |
+
MusicInfo: The MusicInfo with augmented textual description.
|
130 |
+
"""
|
131 |
+
def is_valid_field(field_name: str, field_value: tp.Any) -> bool:
|
132 |
+
valid_field_name = field_name in ['key', 'bpm', 'genre', 'moods', 'instrument', 'keywords']
|
133 |
+
valid_field_value = field_value is not None and isinstance(field_value, (int, float, str, list))
|
134 |
+
keep_field = random.uniform(0, 1) < drop_other_p
|
135 |
+
return valid_field_name and valid_field_value and keep_field
|
136 |
+
|
137 |
+
def process_value(v: tp.Any) -> str:
|
138 |
+
if isinstance(v, (int, float, str)):
|
139 |
+
return str(v)
|
140 |
+
if isinstance(v, list):
|
141 |
+
return ", ".join(v)
|
142 |
+
else:
|
143 |
+
raise ValueError(f"Unknown type for text value! ({type(v), v})")
|
144 |
+
|
145 |
+
description = music_info.description
|
146 |
+
|
147 |
+
metadata_text = ""
|
148 |
+
if random.uniform(0, 1) < merge_text_p:
|
149 |
+
meta_pairs = [f'{_field.name}: {process_value(getattr(music_info, _field.name))}'
|
150 |
+
for _field in fields(music_info) if is_valid_field(_field.name, getattr(music_info, _field.name))]
|
151 |
+
random.shuffle(meta_pairs)
|
152 |
+
metadata_text = ". ".join(meta_pairs)
|
153 |
+
description = description if not random.uniform(0, 1) < drop_desc_p else None
|
154 |
+
logger.debug(f"Applying text augmentation on MMI info. description: {description}, metadata: {metadata_text}")
|
155 |
+
|
156 |
+
if description is None:
|
157 |
+
description = metadata_text if len(metadata_text) > 1 else None
|
158 |
+
else:
|
159 |
+
description = ". ".join([description.rstrip('.'), metadata_text])
|
160 |
+
description = description.strip() if description else None
|
161 |
+
|
162 |
+
music_info = replace(music_info)
|
163 |
+
music_info.description = description
|
164 |
+
return music_info
|
165 |
+
|
166 |
+
|
167 |
+
class Paraphraser:
|
168 |
+
def __init__(self, paraphrase_source: tp.Union[str, Path], paraphrase_p: float = 0.):
|
169 |
+
self.paraphrase_p = paraphrase_p
|
170 |
+
open_fn = gzip.open if str(paraphrase_source).lower().endswith('.gz') else open
|
171 |
+
with open_fn(paraphrase_source, 'rb') as f: # type: ignore
|
172 |
+
self.paraphrase_source = json.loads(f.read())
|
173 |
+
logger.info(f"loaded paraphrasing source from: {paraphrase_source}")
|
174 |
+
|
175 |
+
def sample_paraphrase(self, audio_path: str, description: str):
|
176 |
+
if random.random() >= self.paraphrase_p:
|
177 |
+
return description
|
178 |
+
info_path = Path(audio_path).with_suffix('.json')
|
179 |
+
if info_path not in self.paraphrase_source:
|
180 |
+
warn_once(logger, f"{info_path} not in paraphrase source!")
|
181 |
+
return description
|
182 |
+
new_desc = random.choice(self.paraphrase_source[info_path])
|
183 |
+
logger.debug(f"{description} -> {new_desc}")
|
184 |
+
return new_desc
|
185 |
+
|
186 |
+
|
187 |
+
class MusicDataset(InfoAudioDataset):
|
188 |
+
"""Music dataset is an AudioDataset with music-related metadata.
|
189 |
+
|
190 |
+
Args:
|
191 |
+
info_fields_required (bool): Whether to enforce having required fields.
|
192 |
+
merge_text_p (float): Probability of merging additional metadata to the description.
|
193 |
+
drop_desc_p (float): Probability of dropping the original description on text merge.
|
194 |
+
drop_other_p (float): Probability of dropping the other fields used for text augmentation.
|
195 |
+
joint_embed_attributes (list[str]): A list of attributes for which joint embedding metadata is returned.
|
196 |
+
paraphrase_source (str, optional): Path to the .json or .json.gz file containing the
|
197 |
+
paraphrases for the description. The json should be a dict with keys are the
|
198 |
+
original info path (e.g. track_path.json) and each value is a list of possible
|
199 |
+
paraphrased.
|
200 |
+
paraphrase_p (float): probability of taking a paraphrase.
|
201 |
+
|
202 |
+
See `audiocraft.data.info_audio_dataset.InfoAudioDataset` for full initialization arguments.
|
203 |
+
"""
|
204 |
+
def __init__(self, *args, info_fields_required: bool = True,
|
205 |
+
merge_text_p: float = 0., drop_desc_p: float = 0., drop_other_p: float = 0.,
|
206 |
+
joint_embed_attributes: tp.List[str] = [],
|
207 |
+
paraphrase_source: tp.Optional[str] = None, paraphrase_p: float = 0,
|
208 |
+
**kwargs):
|
209 |
+
kwargs['return_info'] = True # We require the info for each song of the dataset.
|
210 |
+
super().__init__(*args, **kwargs)
|
211 |
+
self.info_fields_required = info_fields_required
|
212 |
+
self.merge_text_p = merge_text_p
|
213 |
+
self.drop_desc_p = drop_desc_p
|
214 |
+
self.drop_other_p = drop_other_p
|
215 |
+
self.joint_embed_attributes = joint_embed_attributes
|
216 |
+
self.paraphraser = None
|
217 |
+
if paraphrase_source is not None:
|
218 |
+
self.paraphraser = Paraphraser(paraphrase_source, paraphrase_p)
|
219 |
+
|
220 |
+
def __getitem__(self, index):
|
221 |
+
wav, info = super().__getitem__(index)
|
222 |
+
info_data = info.to_dict()
|
223 |
+
music_info_path = Path(info.meta.path).with_suffix('.json')
|
224 |
+
|
225 |
+
if Path(music_info_path).exists():
|
226 |
+
with open(music_info_path, 'r') as json_file:
|
227 |
+
music_data = json.load(json_file)
|
228 |
+
music_data.update(info_data)
|
229 |
+
music_info = MusicInfo.from_dict(music_data, fields_required=self.info_fields_required)
|
230 |
+
if self.paraphraser is not None:
|
231 |
+
music_info.description = self.paraphraser.sample(music_info.meta.path, music_info.description)
|
232 |
+
if self.merge_text_p:
|
233 |
+
music_info = augment_music_info_description(
|
234 |
+
music_info, self.merge_text_p, self.drop_desc_p, self.drop_other_p)
|
235 |
+
else:
|
236 |
+
music_info = MusicInfo.from_dict(info_data, fields_required=False)
|
237 |
+
|
238 |
+
music_info.self_wav = WavCondition(
|
239 |
+
wav=wav[None], length=torch.tensor([info.n_frames]),
|
240 |
+
sample_rate=[info.sample_rate], path=[info.meta.path], seek_time=[info.seek_time])
|
241 |
+
|
242 |
+
for att in self.joint_embed_attributes:
|
243 |
+
att_value = getattr(music_info, att)
|
244 |
+
joint_embed_cond = JointEmbedCondition(
|
245 |
+
wav[None], [att_value], torch.tensor([info.n_frames]),
|
246 |
+
sample_rate=[info.sample_rate], path=[info.meta.path], seek_time=[info.seek_time])
|
247 |
+
music_info.joint_embed[att] = joint_embed_cond
|
248 |
+
|
249 |
+
return wav, music_info
|
250 |
+
|
251 |
+
|
252 |
+
def get_musical_key(value: tp.Optional[str]) -> tp.Optional[str]:
|
253 |
+
"""Preprocess key keywords, discarding them if there are multiple key defined."""
|
254 |
+
if value is None or (not isinstance(value, str)) or len(value) == 0 or value == 'None':
|
255 |
+
return None
|
256 |
+
elif ',' in value:
|
257 |
+
# For now, we discard when multiple keys are defined separated with comas
|
258 |
+
return None
|
259 |
+
else:
|
260 |
+
return value.strip().lower()
|
261 |
+
|
262 |
+
|
263 |
+
def get_bpm(value: tp.Optional[str]) -> tp.Optional[float]:
|
264 |
+
"""Preprocess to a float."""
|
265 |
+
if value is None:
|
266 |
+
return None
|
267 |
+
try:
|
268 |
+
return float(value)
|
269 |
+
except ValueError:
|
270 |
+
return None
|
audiocraft/data/sound_dataset.py
ADDED
@@ -0,0 +1,330 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""Dataset of audio with a simple description.
|
7 |
+
"""
|
8 |
+
|
9 |
+
from dataclasses import dataclass, fields, replace
|
10 |
+
import json
|
11 |
+
from pathlib import Path
|
12 |
+
import random
|
13 |
+
import typing as tp
|
14 |
+
|
15 |
+
import numpy as np
|
16 |
+
import torch
|
17 |
+
|
18 |
+
from .info_audio_dataset import (
|
19 |
+
InfoAudioDataset,
|
20 |
+
get_keyword_or_keyword_list
|
21 |
+
)
|
22 |
+
from ..modules.conditioners import (
|
23 |
+
ConditioningAttributes,
|
24 |
+
SegmentWithAttributes,
|
25 |
+
WavCondition,
|
26 |
+
)
|
27 |
+
|
28 |
+
|
29 |
+
EPS = torch.finfo(torch.float32).eps
|
30 |
+
TARGET_LEVEL_LOWER = -35
|
31 |
+
TARGET_LEVEL_UPPER = -15
|
32 |
+
|
33 |
+
|
34 |
+
@dataclass
|
35 |
+
class SoundInfo(SegmentWithAttributes):
|
36 |
+
"""Segment info augmented with Sound metadata.
|
37 |
+
"""
|
38 |
+
description: tp.Optional[str] = None
|
39 |
+
self_wav: tp.Optional[torch.Tensor] = None
|
40 |
+
|
41 |
+
@property
|
42 |
+
def has_sound_meta(self) -> bool:
|
43 |
+
return self.description is not None
|
44 |
+
|
45 |
+
def to_condition_attributes(self) -> ConditioningAttributes:
|
46 |
+
out = ConditioningAttributes()
|
47 |
+
|
48 |
+
for _field in fields(self):
|
49 |
+
key, value = _field.name, getattr(self, _field.name)
|
50 |
+
if key == 'self_wav':
|
51 |
+
out.wav[key] = value
|
52 |
+
else:
|
53 |
+
out.text[key] = value
|
54 |
+
return out
|
55 |
+
|
56 |
+
@staticmethod
|
57 |
+
def attribute_getter(attribute):
|
58 |
+
if attribute == 'description':
|
59 |
+
preprocess_func = get_keyword_or_keyword_list
|
60 |
+
else:
|
61 |
+
preprocess_func = None
|
62 |
+
return preprocess_func
|
63 |
+
|
64 |
+
@classmethod
|
65 |
+
def from_dict(cls, dictionary: dict, fields_required: bool = False):
|
66 |
+
_dictionary: tp.Dict[str, tp.Any] = {}
|
67 |
+
|
68 |
+
# allow a subset of attributes to not be loaded from the dictionary
|
69 |
+
# these attributes may be populated later
|
70 |
+
post_init_attributes = ['self_wav']
|
71 |
+
|
72 |
+
for _field in fields(cls):
|
73 |
+
if _field.name in post_init_attributes:
|
74 |
+
continue
|
75 |
+
elif _field.name not in dictionary:
|
76 |
+
if fields_required:
|
77 |
+
raise KeyError(f"Unexpected missing key: {_field.name}")
|
78 |
+
else:
|
79 |
+
preprocess_func: tp.Optional[tp.Callable] = cls.attribute_getter(_field.name)
|
80 |
+
value = dictionary[_field.name]
|
81 |
+
if preprocess_func:
|
82 |
+
value = preprocess_func(value)
|
83 |
+
_dictionary[_field.name] = value
|
84 |
+
return cls(**_dictionary)
|
85 |
+
|
86 |
+
|
87 |
+
class SoundDataset(InfoAudioDataset):
|
88 |
+
"""Sound audio dataset: Audio dataset with environmental sound-specific metadata.
|
89 |
+
|
90 |
+
Args:
|
91 |
+
info_fields_required (bool): Whether all the mandatory metadata fields should be in the loaded metadata.
|
92 |
+
external_metadata_source (tp.Optional[str]): Folder containing JSON metadata for the corresponding dataset.
|
93 |
+
The metadata files contained in this folder are expected to match the stem of the audio file with
|
94 |
+
a json extension.
|
95 |
+
aug_p (float): Probability of performing audio mixing augmentation on the batch.
|
96 |
+
mix_p (float): Proportion of batch items that are mixed together when applying audio mixing augmentation.
|
97 |
+
mix_snr_low (int): Lowerbound for SNR value sampled for mixing augmentation.
|
98 |
+
mix_snr_high (int): Upperbound for SNR value sampled for mixing augmentation.
|
99 |
+
mix_min_overlap (float): Minimum overlap between audio files when performing mixing augmentation.
|
100 |
+
kwargs: Additional arguments for AudioDataset.
|
101 |
+
|
102 |
+
See `audiocraft.data.info_audio_dataset.InfoAudioDataset` for full initialization arguments.
|
103 |
+
"""
|
104 |
+
def __init__(
|
105 |
+
self,
|
106 |
+
*args,
|
107 |
+
info_fields_required: bool = True,
|
108 |
+
external_metadata_source: tp.Optional[str] = None,
|
109 |
+
aug_p: float = 0.,
|
110 |
+
mix_p: float = 0.,
|
111 |
+
mix_snr_low: int = -5,
|
112 |
+
mix_snr_high: int = 5,
|
113 |
+
mix_min_overlap: float = 0.5,
|
114 |
+
**kwargs
|
115 |
+
):
|
116 |
+
kwargs['return_info'] = True # We require the info for each song of the dataset.
|
117 |
+
super().__init__(*args, **kwargs)
|
118 |
+
self.info_fields_required = info_fields_required
|
119 |
+
self.external_metadata_source = external_metadata_source
|
120 |
+
self.aug_p = aug_p
|
121 |
+
self.mix_p = mix_p
|
122 |
+
if self.aug_p > 0:
|
123 |
+
assert self.mix_p > 0, "Expecting some mixing proportion mix_p if aug_p > 0"
|
124 |
+
assert self.channels == 1, "SoundDataset with audio mixing considers only monophonic audio"
|
125 |
+
self.mix_snr_low = mix_snr_low
|
126 |
+
self.mix_snr_high = mix_snr_high
|
127 |
+
self.mix_min_overlap = mix_min_overlap
|
128 |
+
|
129 |
+
def _get_info_path(self, path: tp.Union[str, Path]) -> Path:
|
130 |
+
"""Get path of JSON with metadata (description, etc.).
|
131 |
+
If there exists a JSON with the same name as 'path.name', then it will be used.
|
132 |
+
Else, such JSON will be searched for in an external json source folder if it exists.
|
133 |
+
"""
|
134 |
+
info_path = Path(path).with_suffix('.json')
|
135 |
+
if Path(info_path).exists():
|
136 |
+
return info_path
|
137 |
+
elif self.external_metadata_source and (Path(self.external_metadata_source) / info_path.name).exists():
|
138 |
+
return Path(self.external_metadata_source) / info_path.name
|
139 |
+
else:
|
140 |
+
raise Exception(f"Unable to find a metadata JSON for path: {path}")
|
141 |
+
|
142 |
+
def __getitem__(self, index):
|
143 |
+
wav, info = super().__getitem__(index)
|
144 |
+
info_data = info.to_dict()
|
145 |
+
info_path = self._get_info_path(info.meta.path)
|
146 |
+
if Path(info_path).exists():
|
147 |
+
with open(info_path, 'r') as json_file:
|
148 |
+
sound_data = json.load(json_file)
|
149 |
+
sound_data.update(info_data)
|
150 |
+
sound_info = SoundInfo.from_dict(sound_data, fields_required=self.info_fields_required)
|
151 |
+
# if there are multiple descriptions, sample one randomly
|
152 |
+
if isinstance(sound_info.description, list):
|
153 |
+
sound_info.description = random.choice(sound_info.description)
|
154 |
+
else:
|
155 |
+
sound_info = SoundInfo.from_dict(info_data, fields_required=False)
|
156 |
+
|
157 |
+
sound_info.self_wav = WavCondition(
|
158 |
+
wav=wav[None], length=torch.tensor([info.n_frames]),
|
159 |
+
sample_rate=[sound_info.sample_rate], path=[info.meta.path], seek_time=[info.seek_time])
|
160 |
+
|
161 |
+
return wav, sound_info
|
162 |
+
|
163 |
+
def collater(self, samples):
|
164 |
+
# when training, audio mixing is performed in the collate function
|
165 |
+
wav, sound_info = super().collater(samples) # SoundDataset always returns infos
|
166 |
+
if self.aug_p > 0:
|
167 |
+
wav, sound_info = mix_samples(wav, sound_info, self.aug_p, self.mix_p,
|
168 |
+
snr_low=self.mix_snr_low, snr_high=self.mix_snr_high,
|
169 |
+
min_overlap=self.mix_min_overlap)
|
170 |
+
return wav, sound_info
|
171 |
+
|
172 |
+
|
173 |
+
def rms_f(x: torch.Tensor) -> torch.Tensor:
|
174 |
+
return (x ** 2).mean(1).pow(0.5)
|
175 |
+
|
176 |
+
|
177 |
+
def normalize(audio: torch.Tensor, target_level: int = -25) -> torch.Tensor:
|
178 |
+
"""Normalize the signal to the target level."""
|
179 |
+
rms = rms_f(audio)
|
180 |
+
scalar = 10 ** (target_level / 20) / (rms + EPS)
|
181 |
+
audio = audio * scalar.unsqueeze(1)
|
182 |
+
return audio
|
183 |
+
|
184 |
+
|
185 |
+
def is_clipped(audio: torch.Tensor, clipping_threshold: float = 0.99) -> torch.Tensor:
|
186 |
+
return (abs(audio) > clipping_threshold).any(1)
|
187 |
+
|
188 |
+
|
189 |
+
def mix_pair(src: torch.Tensor, dst: torch.Tensor, min_overlap: float) -> torch.Tensor:
|
190 |
+
start = random.randint(0, int(src.shape[1] * (1 - min_overlap)))
|
191 |
+
remainder = src.shape[1] - start
|
192 |
+
if dst.shape[1] > remainder:
|
193 |
+
src[:, start:] = src[:, start:] + dst[:, :remainder]
|
194 |
+
else:
|
195 |
+
src[:, start:start+dst.shape[1]] = src[:, start:start+dst.shape[1]] + dst
|
196 |
+
return src
|
197 |
+
|
198 |
+
|
199 |
+
def snr_mixer(clean: torch.Tensor, noise: torch.Tensor, snr: int, min_overlap: float,
|
200 |
+
target_level: int = -25, clipping_threshold: float = 0.99) -> torch.Tensor:
|
201 |
+
"""Function to mix clean speech and noise at various SNR levels.
|
202 |
+
|
203 |
+
Args:
|
204 |
+
clean (torch.Tensor): Clean audio source to mix, of shape [B, T].
|
205 |
+
noise (torch.Tensor): Noise audio source to mix, of shape [B, T].
|
206 |
+
snr (int): SNR level when mixing.
|
207 |
+
min_overlap (float): Minimum overlap between the two mixed sources.
|
208 |
+
target_level (int): Gain level in dB.
|
209 |
+
clipping_threshold (float): Threshold for clipping the audio.
|
210 |
+
Returns:
|
211 |
+
torch.Tensor: The mixed audio, of shape [B, T].
|
212 |
+
"""
|
213 |
+
if clean.shape[1] > noise.shape[1]:
|
214 |
+
noise = torch.nn.functional.pad(noise, (0, clean.shape[1] - noise.shape[1]))
|
215 |
+
else:
|
216 |
+
noise = noise[:, :clean.shape[1]]
|
217 |
+
|
218 |
+
# normalizing to -25 dB FS
|
219 |
+
clean = clean / (clean.max(1)[0].abs().unsqueeze(1) + EPS)
|
220 |
+
clean = normalize(clean, target_level)
|
221 |
+
rmsclean = rms_f(clean)
|
222 |
+
|
223 |
+
noise = noise / (noise.max(1)[0].abs().unsqueeze(1) + EPS)
|
224 |
+
noise = normalize(noise, target_level)
|
225 |
+
rmsnoise = rms_f(noise)
|
226 |
+
|
227 |
+
# set the noise level for a given SNR
|
228 |
+
noisescalar = (rmsclean / (10 ** (snr / 20)) / (rmsnoise + EPS)).unsqueeze(1)
|
229 |
+
noisenewlevel = noise * noisescalar
|
230 |
+
|
231 |
+
# mix noise and clean speech
|
232 |
+
noisyspeech = mix_pair(clean, noisenewlevel, min_overlap)
|
233 |
+
|
234 |
+
# randomly select RMS value between -15 dBFS and -35 dBFS and normalize noisyspeech with that value
|
235 |
+
# there is a chance of clipping that might happen with very less probability, which is not a major issue.
|
236 |
+
noisy_rms_level = np.random.randint(TARGET_LEVEL_LOWER, TARGET_LEVEL_UPPER)
|
237 |
+
rmsnoisy = rms_f(noisyspeech)
|
238 |
+
scalarnoisy = (10 ** (noisy_rms_level / 20) / (rmsnoisy + EPS)).unsqueeze(1)
|
239 |
+
noisyspeech = noisyspeech * scalarnoisy
|
240 |
+
clean = clean * scalarnoisy
|
241 |
+
noisenewlevel = noisenewlevel * scalarnoisy
|
242 |
+
|
243 |
+
# final check to see if there are any amplitudes exceeding +/- 1. If so, normalize all the signals accordingly
|
244 |
+
clipped = is_clipped(noisyspeech)
|
245 |
+
if clipped.any():
|
246 |
+
noisyspeech_maxamplevel = noisyspeech[clipped].max(1)[0].abs().unsqueeze(1) / (clipping_threshold - EPS)
|
247 |
+
noisyspeech[clipped] = noisyspeech[clipped] / noisyspeech_maxamplevel
|
248 |
+
|
249 |
+
return noisyspeech
|
250 |
+
|
251 |
+
|
252 |
+
def snr_mix(src: torch.Tensor, dst: torch.Tensor, snr_low: int, snr_high: int, min_overlap: float):
|
253 |
+
if snr_low == snr_high:
|
254 |
+
snr = snr_low
|
255 |
+
else:
|
256 |
+
snr = np.random.randint(snr_low, snr_high)
|
257 |
+
mix = snr_mixer(src, dst, snr, min_overlap)
|
258 |
+
return mix
|
259 |
+
|
260 |
+
|
261 |
+
def mix_text(src_text: str, dst_text: str):
|
262 |
+
"""Mix text from different sources by concatenating them."""
|
263 |
+
if src_text == dst_text:
|
264 |
+
return src_text
|
265 |
+
return src_text + " " + dst_text
|
266 |
+
|
267 |
+
|
268 |
+
def mix_samples(wavs: torch.Tensor, infos: tp.List[SoundInfo], aug_p: float, mix_p: float,
|
269 |
+
snr_low: int, snr_high: int, min_overlap: float):
|
270 |
+
"""Mix samples within a batch, summing the waveforms and concatenating the text infos.
|
271 |
+
|
272 |
+
Args:
|
273 |
+
wavs (torch.Tensor): Audio tensors of shape [B, C, T].
|
274 |
+
infos (list[SoundInfo]): List of SoundInfo items corresponding to the audio.
|
275 |
+
aug_p (float): Augmentation probability.
|
276 |
+
mix_p (float): Proportion of items in the batch to mix (and merge) together.
|
277 |
+
snr_low (int): Lowerbound for sampling SNR.
|
278 |
+
snr_high (int): Upperbound for sampling SNR.
|
279 |
+
min_overlap (float): Minimum overlap between mixed samples.
|
280 |
+
Returns:
|
281 |
+
tuple[torch.Tensor, list[SoundInfo]]: A tuple containing the mixed wavs
|
282 |
+
and mixed SoundInfo for the given batch.
|
283 |
+
"""
|
284 |
+
# no mixing to perform within the batch
|
285 |
+
if mix_p == 0:
|
286 |
+
return wavs, infos
|
287 |
+
|
288 |
+
if random.uniform(0, 1) < aug_p:
|
289 |
+
# perform all augmentations on waveforms as [B, T]
|
290 |
+
# randomly picking pairs of audio to mix
|
291 |
+
assert wavs.size(1) == 1, f"Mix samples requires monophonic audio but C={wavs.size(1)}"
|
292 |
+
wavs = wavs.mean(dim=1, keepdim=False)
|
293 |
+
B, T = wavs.shape
|
294 |
+
k = int(mix_p * B)
|
295 |
+
mixed_sources_idx = torch.randperm(B)[:k]
|
296 |
+
mixed_targets_idx = torch.randperm(B)[:k]
|
297 |
+
aug_wavs = snr_mix(
|
298 |
+
wavs[mixed_sources_idx],
|
299 |
+
wavs[mixed_targets_idx],
|
300 |
+
snr_low,
|
301 |
+
snr_high,
|
302 |
+
min_overlap,
|
303 |
+
)
|
304 |
+
# mixing textual descriptions in metadata
|
305 |
+
descriptions = [info.description for info in infos]
|
306 |
+
aug_infos = []
|
307 |
+
for i, j in zip(mixed_sources_idx, mixed_targets_idx):
|
308 |
+
text = mix_text(descriptions[i], descriptions[j])
|
309 |
+
m = replace(infos[i])
|
310 |
+
m.description = text
|
311 |
+
aug_infos.append(m)
|
312 |
+
|
313 |
+
# back to [B, C, T]
|
314 |
+
aug_wavs = aug_wavs.unsqueeze(1)
|
315 |
+
assert aug_wavs.shape[0] > 0, "Samples mixing returned empty batch."
|
316 |
+
assert aug_wavs.dim() == 3, f"Returned wav should be [B, C, T] but dim = {aug_wavs.dim()}"
|
317 |
+
assert aug_wavs.shape[0] == len(aug_infos), "Mismatch between number of wavs and infos in the batch"
|
318 |
+
|
319 |
+
return aug_wavs, aug_infos # [B, C, T]
|
320 |
+
else:
|
321 |
+
# randomly pick samples in the batch to match
|
322 |
+
# the batch size when performing audio mixing
|
323 |
+
B, C, T = wavs.shape
|
324 |
+
k = int(mix_p * B)
|
325 |
+
wav_idx = torch.randperm(B)[:k]
|
326 |
+
wavs = wavs[wav_idx]
|
327 |
+
infos = [infos[i] for i in wav_idx]
|
328 |
+
assert wavs.shape[0] == len(infos), "Mismatch between number of wavs and infos in the batch"
|
329 |
+
|
330 |
+
return wavs, infos # [B, C, T]
|
audiocraft/data/zip.py
CHANGED
@@ -3,6 +3,8 @@
|
|
3 |
#
|
4 |
# This source code is licensed under the license found in the
|
5 |
# LICENSE file in the root directory of this source tree.
|
|
|
|
|
6 |
|
7 |
import typing
|
8 |
import zipfile
|
@@ -18,13 +20,13 @@ MODE = Literal['r', 'w', 'x', 'a']
|
|
18 |
|
19 |
@dataclass(order=True)
|
20 |
class PathInZip:
|
21 |
-
"""
|
22 |
|
23 |
Args:
|
24 |
-
path: The convention is <path_to_zip>:<relative_path_inside_zip
|
25 |
Let's assume there is a zip file /some/location/foo.zip
|
26 |
and inside of it is a json file located at /data/file1.json,
|
27 |
-
Then we expect path = "/some/location/foo.zip:/data/file1.json"
|
28 |
"""
|
29 |
|
30 |
INFO_PATH_SEP = ':'
|
@@ -55,7 +57,7 @@ def set_zip_cache_size(max_size: int):
|
|
55 |
"""Sets the maximal LRU caching for zip file opening.
|
56 |
|
57 |
Args:
|
58 |
-
max_size: the maximal LRU cache.
|
59 |
"""
|
60 |
global _cached_open_zip
|
61 |
_cached_open_zip = lru_cache(max_size)(_open_zip)
|
@@ -65,8 +67,8 @@ def open_file_in_zip(path_in_zip: PathInZip, mode: str = 'r') -> typing.IO:
|
|
65 |
"""Opens a file stored inside a zip and returns a file-like object.
|
66 |
|
67 |
Args:
|
68 |
-
path_in_zip: A PathInZip object representing the file to return a file-like object of.
|
69 |
-
mode: The mode in which to open the file with.
|
70 |
Returns:
|
71 |
A file-like object for PathInZip.
|
72 |
"""
|
|
|
3 |
#
|
4 |
# This source code is licensed under the license found in the
|
5 |
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""Utility for reading some info from inside a zip file.
|
7 |
+
"""
|
8 |
|
9 |
import typing
|
10 |
import zipfile
|
|
|
20 |
|
21 |
@dataclass(order=True)
|
22 |
class PathInZip:
|
23 |
+
"""Hold a path of file within a zip file.
|
24 |
|
25 |
Args:
|
26 |
+
path (str): The convention is <path_to_zip>:<relative_path_inside_zip>.
|
27 |
Let's assume there is a zip file /some/location/foo.zip
|
28 |
and inside of it is a json file located at /data/file1.json,
|
29 |
+
Then we expect path = "/some/location/foo.zip:/data/file1.json".
|
30 |
"""
|
31 |
|
32 |
INFO_PATH_SEP = ':'
|
|
|
57 |
"""Sets the maximal LRU caching for zip file opening.
|
58 |
|
59 |
Args:
|
60 |
+
max_size (int): the maximal LRU cache.
|
61 |
"""
|
62 |
global _cached_open_zip
|
63 |
_cached_open_zip = lru_cache(max_size)(_open_zip)
|
|
|
67 |
"""Opens a file stored inside a zip and returns a file-like object.
|
68 |
|
69 |
Args:
|
70 |
+
path_in_zip (PathInZip): A PathInZip object representing the file to return a file-like object of.
|
71 |
+
mode (str): The mode in which to open the file with.
|
72 |
Returns:
|
73 |
A file-like object for PathInZip.
|
74 |
"""
|
audiocraft/environment.py
ADDED
@@ -0,0 +1,176 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
"""
|
8 |
+
Provides cluster and tools configuration across clusters (slurm, dora, utilities).
|
9 |
+
"""
|
10 |
+
|
11 |
+
import logging
|
12 |
+
import os
|
13 |
+
from pathlib import Path
|
14 |
+
import re
|
15 |
+
import typing as tp
|
16 |
+
|
17 |
+
import omegaconf
|
18 |
+
|
19 |
+
from .utils.cluster import _guess_cluster_type
|
20 |
+
|
21 |
+
|
22 |
+
logger = logging.getLogger(__name__)
|
23 |
+
|
24 |
+
|
25 |
+
class AudioCraftEnvironment:
|
26 |
+
"""Environment configuration for teams and clusters.
|
27 |
+
|
28 |
+
AudioCraftEnvironment picks compute cluster settings (slurm, dora) from the current running environment
|
29 |
+
or declared variable and the loaded team configuration. Additionally, the AudioCraftEnvironment
|
30 |
+
provides pointers to a reference folder resolved automatically across clusters that is shared across team members,
|
31 |
+
allowing to share sigs or other files to run jobs. Finally, it provides dataset mappers to automatically
|
32 |
+
map dataset file paths to new locations across clusters, allowing to use the same manifest of files across cluters.
|
33 |
+
|
34 |
+
The cluster type is identified automatically and base configuration file is read from config/teams.yaml.
|
35 |
+
Use the following environment variables to specify the cluster, team or configuration:
|
36 |
+
|
37 |
+
AUDIOCRAFT_CLUSTER (optional): Cluster type to enforce. Useful if the cluster type
|
38 |
+
cannot be inferred automatically.
|
39 |
+
AUDIOCRAFT_CONFIG (optional): Path to yaml config holding the teams configuration.
|
40 |
+
If not set, configuration is read from config/teams.yaml.
|
41 |
+
AUDIOCRAFT_TEAM (optional): Name of the team. Recommended to set to your own team.
|
42 |
+
Cluster configuration are shared across teams to match compute allocation,
|
43 |
+
specify your cluster configuration in the configuration file under a key mapping
|
44 |
+
your team name.
|
45 |
+
"""
|
46 |
+
_instance = None
|
47 |
+
DEFAULT_TEAM = "default"
|
48 |
+
|
49 |
+
def __init__(self) -> None:
|
50 |
+
"""Loads configuration."""
|
51 |
+
self.team: str = os.getenv("AUDIOCRAFT_TEAM", self.DEFAULT_TEAM)
|
52 |
+
cluster_type = _guess_cluster_type()
|
53 |
+
cluster = os.getenv(
|
54 |
+
"AUDIOCRAFT_CLUSTER", cluster_type.value
|
55 |
+
)
|
56 |
+
logger.info("Detecting cluster type %s", cluster_type)
|
57 |
+
|
58 |
+
self.cluster: str = cluster
|
59 |
+
|
60 |
+
config_path = os.getenv(
|
61 |
+
"AUDIOCRAFT_CONFIG",
|
62 |
+
Path(__file__)
|
63 |
+
.parent.parent.joinpath("config/teams", self.team)
|
64 |
+
.with_suffix(".yaml"),
|
65 |
+
)
|
66 |
+
self.config = omegaconf.OmegaConf.load(config_path)
|
67 |
+
self._dataset_mappers = []
|
68 |
+
cluster_config = self._get_cluster_config()
|
69 |
+
if "dataset_mappers" in cluster_config:
|
70 |
+
for pattern, repl in cluster_config["dataset_mappers"].items():
|
71 |
+
regex = re.compile(pattern)
|
72 |
+
self._dataset_mappers.append((regex, repl))
|
73 |
+
|
74 |
+
def _get_cluster_config(self) -> omegaconf.DictConfig:
|
75 |
+
assert isinstance(self.config, omegaconf.DictConfig)
|
76 |
+
return self.config[self.cluster]
|
77 |
+
|
78 |
+
@classmethod
|
79 |
+
def instance(cls):
|
80 |
+
if cls._instance is None:
|
81 |
+
cls._instance = cls()
|
82 |
+
return cls._instance
|
83 |
+
|
84 |
+
@classmethod
|
85 |
+
def reset(cls):
|
86 |
+
"""Clears the environment and forces a reload on next invocation."""
|
87 |
+
cls._instance = None
|
88 |
+
|
89 |
+
@classmethod
|
90 |
+
def get_team(cls) -> str:
|
91 |
+
"""Gets the selected team as dictated by the AUDIOCRAFT_TEAM env var.
|
92 |
+
If not defined, defaults to "labs".
|
93 |
+
"""
|
94 |
+
return cls.instance().team
|
95 |
+
|
96 |
+
@classmethod
|
97 |
+
def get_cluster(cls) -> str:
|
98 |
+
"""Gets the detected cluster.
|
99 |
+
This value can be overridden by the AUDIOCRAFT_CLUSTER env var.
|
100 |
+
"""
|
101 |
+
return cls.instance().cluster
|
102 |
+
|
103 |
+
@classmethod
|
104 |
+
def get_dora_dir(cls) -> Path:
|
105 |
+
"""Gets the path to the dora directory for the current team and cluster.
|
106 |
+
Value is overridden by the AUDIOCRAFT_DORA_DIR env var.
|
107 |
+
"""
|
108 |
+
cluster_config = cls.instance()._get_cluster_config()
|
109 |
+
dora_dir = os.getenv("AUDIOCRAFT_DORA_DIR", cluster_config["dora_dir"])
|
110 |
+
logger.warning(f"Dora directory: {dora_dir}")
|
111 |
+
return Path(dora_dir)
|
112 |
+
|
113 |
+
@classmethod
|
114 |
+
def get_reference_dir(cls) -> Path:
|
115 |
+
"""Gets the path to the reference directory for the current team and cluster.
|
116 |
+
Value is overridden by the AUDIOCRAFT_REFERENCE_DIR env var.
|
117 |
+
"""
|
118 |
+
cluster_config = cls.instance()._get_cluster_config()
|
119 |
+
return Path(os.getenv("AUDIOCRAFT_REFERENCE_DIR", cluster_config["reference_dir"]))
|
120 |
+
|
121 |
+
@classmethod
|
122 |
+
def get_slurm_exclude(cls) -> tp.Optional[str]:
|
123 |
+
"""Get the list of nodes to exclude for that cluster."""
|
124 |
+
cluster_config = cls.instance()._get_cluster_config()
|
125 |
+
return cluster_config.get("slurm_exclude")
|
126 |
+
|
127 |
+
@classmethod
|
128 |
+
def get_slurm_partitions(cls, partition_types: tp.Optional[tp.List[str]] = None) -> str:
|
129 |
+
"""Gets the requested partitions for the current team and cluster as a comma-separated string.
|
130 |
+
|
131 |
+
Args:
|
132 |
+
partition_types (list[str], optional): partition types to retrieve. Values must be
|
133 |
+
from ['global', 'team']. If not provided, the global partition is returned.
|
134 |
+
"""
|
135 |
+
if not partition_types:
|
136 |
+
partition_types = ["global"]
|
137 |
+
|
138 |
+
cluster_config = cls.instance()._get_cluster_config()
|
139 |
+
partitions = [
|
140 |
+
cluster_config["partitions"][partition_type]
|
141 |
+
for partition_type in partition_types
|
142 |
+
]
|
143 |
+
return ",".join(partitions)
|
144 |
+
|
145 |
+
@classmethod
|
146 |
+
def resolve_reference_path(cls, path: tp.Union[str, Path]) -> Path:
|
147 |
+
"""Converts reference placeholder in path with configured reference dir to resolve paths.
|
148 |
+
|
149 |
+
Args:
|
150 |
+
path (str or Path): Path to resolve.
|
151 |
+
Returns:
|
152 |
+
Path: Resolved path.
|
153 |
+
"""
|
154 |
+
path = str(path)
|
155 |
+
|
156 |
+
if path.startswith("//reference"):
|
157 |
+
reference_dir = cls.get_reference_dir()
|
158 |
+
logger.warn(f"Reference directory: {reference_dir}")
|
159 |
+
assert (
|
160 |
+
reference_dir.exists() and reference_dir.is_dir()
|
161 |
+
), f"Reference directory does not exist: {reference_dir}."
|
162 |
+
path = re.sub("^//reference", str(reference_dir), path)
|
163 |
+
|
164 |
+
return Path(path)
|
165 |
+
|
166 |
+
@classmethod
|
167 |
+
def apply_dataset_mappers(cls, path: str) -> str:
|
168 |
+
"""Applies dataset mapping regex rules as defined in the configuration.
|
169 |
+
If no rules are defined, the path is returned as-is.
|
170 |
+
"""
|
171 |
+
instance = cls.instance()
|
172 |
+
|
173 |
+
for pattern, repl in instance._dataset_mappers:
|
174 |
+
path = pattern.sub(repl, path)
|
175 |
+
|
176 |
+
return path
|
audiocraft/grids/__init__.py
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""Dora Grids."""
|
audiocraft/grids/_base_explorers.py
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from abc import ABC, abstractmethod
|
8 |
+
import time
|
9 |
+
import typing as tp
|
10 |
+
from dora import Explorer
|
11 |
+
import treetable as tt
|
12 |
+
|
13 |
+
|
14 |
+
def get_sheep_ping(sheep) -> tp.Optional[str]:
|
15 |
+
"""Return the amount of time since the Sheep made some update
|
16 |
+
to its log. Returns a str using the relevant time unit."""
|
17 |
+
ping = None
|
18 |
+
if sheep.log is not None and sheep.log.exists():
|
19 |
+
delta = time.time() - sheep.log.stat().st_mtime
|
20 |
+
if delta > 3600 * 24:
|
21 |
+
ping = f'{delta / (3600 * 24):.1f}d'
|
22 |
+
elif delta > 3600:
|
23 |
+
ping = f'{delta / (3600):.1f}h'
|
24 |
+
elif delta > 60:
|
25 |
+
ping = f'{delta / 60:.1f}m'
|
26 |
+
else:
|
27 |
+
ping = f'{delta:.1f}s'
|
28 |
+
return ping
|
29 |
+
|
30 |
+
|
31 |
+
class BaseExplorer(ABC, Explorer):
|
32 |
+
"""Base explorer for AudioCraft grids.
|
33 |
+
|
34 |
+
All task specific solvers are expected to implement the `get_grid_metrics`
|
35 |
+
method to specify logic about metrics to display for a given task.
|
36 |
+
|
37 |
+
If additional stages are used, the child explorer must define how to handle
|
38 |
+
these new stages in the `process_history` and `process_sheep` methods.
|
39 |
+
"""
|
40 |
+
def stages(self):
|
41 |
+
return ["train", "valid", "evaluate"]
|
42 |
+
|
43 |
+
def get_grid_meta(self):
|
44 |
+
"""Returns the list of Meta information to display for each XP/job.
|
45 |
+
"""
|
46 |
+
return [
|
47 |
+
tt.leaf("index", align=">"),
|
48 |
+
tt.leaf("name", wrap=140),
|
49 |
+
tt.leaf("state"),
|
50 |
+
tt.leaf("sig", align=">"),
|
51 |
+
tt.leaf("sid", align="<"),
|
52 |
+
]
|
53 |
+
|
54 |
+
@abstractmethod
|
55 |
+
def get_grid_metrics(self):
|
56 |
+
"""Return the metrics that should be displayed in the tracking table.
|
57 |
+
"""
|
58 |
+
...
|
59 |
+
|
60 |
+
def process_sheep(self, sheep, history):
|
61 |
+
train = {
|
62 |
+
"epoch": len(history),
|
63 |
+
}
|
64 |
+
parts = {"train": train}
|
65 |
+
for metrics in history:
|
66 |
+
for key, sub in metrics.items():
|
67 |
+
part = parts.get(key, {})
|
68 |
+
if 'duration' in sub:
|
69 |
+
# Convert to minutes for readability.
|
70 |
+
sub['duration'] = sub['duration'] / 60.
|
71 |
+
part.update(sub)
|
72 |
+
parts[key] = part
|
73 |
+
ping = get_sheep_ping(sheep)
|
74 |
+
if ping is not None:
|
75 |
+
for name in self.stages():
|
76 |
+
if name not in parts:
|
77 |
+
parts[name] = {}
|
78 |
+
# Add the ping to each part for convenience.
|
79 |
+
parts[name]['ping'] = ping
|
80 |
+
return parts
|
audiocraft/grids/audiogen/__init__.py
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""AudioGen grids."""
|
audiocraft/grids/audiogen/audiogen_base_16khz.py
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from ..musicgen._explorers import LMExplorer
|
8 |
+
from ...environment import AudioCraftEnvironment
|
9 |
+
|
10 |
+
|
11 |
+
@LMExplorer
|
12 |
+
def explorer(launcher):
|
13 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
14 |
+
launcher.slurm_(gpus=64, partition=partitions)
|
15 |
+
launcher.bind_(solver='audiogen/audiogen_base_16khz')
|
16 |
+
# replace this by the desired environmental sound dataset
|
17 |
+
launcher.bind_(dset='internal/sounds_16khz')
|
18 |
+
|
19 |
+
fsdp = {'autocast': False, 'fsdp.use': True}
|
20 |
+
medium = {'model/lm/model_scale': 'medium'}
|
21 |
+
|
22 |
+
launcher.bind_(fsdp)
|
23 |
+
launcher(medium)
|
audiocraft/grids/audiogen/audiogen_pretrained_16khz_eval.py
ADDED
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
"""
|
8 |
+
Evaluation with objective metrics for the pretrained AudioGen models.
|
9 |
+
This grid takes signature from the training grid and runs evaluation-only stage.
|
10 |
+
|
11 |
+
When running the grid for the first time, please use:
|
12 |
+
REGEN=1 dora grid audiogen.audiogen_pretrained_16khz_eval
|
13 |
+
and re-use the REGEN=1 option when the grid is changed to force regenerating it.
|
14 |
+
|
15 |
+
Note that you need the proper metrics external libraries setup to use all
|
16 |
+
the objective metrics activated in this grid. Refer to the README for more information.
|
17 |
+
"""
|
18 |
+
|
19 |
+
import os
|
20 |
+
|
21 |
+
from ..musicgen._explorers import GenerationEvalExplorer
|
22 |
+
from ...environment import AudioCraftEnvironment
|
23 |
+
from ... import train
|
24 |
+
|
25 |
+
|
26 |
+
def eval(launcher, batch_size: int = 32):
|
27 |
+
opts = {
|
28 |
+
'dset': 'audio/audiocaps_16khz',
|
29 |
+
'solver/audiogen/evaluation': 'objective_eval',
|
30 |
+
'execute_only': 'evaluate',
|
31 |
+
'+dataset.evaluate.batch_size': batch_size,
|
32 |
+
'+metrics.fad.tf.batch_size': 32,
|
33 |
+
}
|
34 |
+
# binary for FAD computation: replace this path with your own path
|
35 |
+
metrics_opts = {
|
36 |
+
'metrics.fad.tf.bin': '/data/home/jadecopet/local/usr/opt/google-research'
|
37 |
+
}
|
38 |
+
opt1 = {'generate.lm.use_sampling': True, 'generate.lm.top_k': 250, 'generate.lm.top_p': 0.}
|
39 |
+
opt2 = {'transformer_lm.two_step_cfg': True}
|
40 |
+
|
41 |
+
sub = launcher.bind(opts)
|
42 |
+
sub.bind_(metrics_opts)
|
43 |
+
|
44 |
+
# base objective metrics
|
45 |
+
sub(opt1, opt2)
|
46 |
+
|
47 |
+
|
48 |
+
@GenerationEvalExplorer
|
49 |
+
def explorer(launcher):
|
50 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
51 |
+
launcher.slurm_(gpus=4, partition=partitions)
|
52 |
+
|
53 |
+
if 'REGEN' not in os.environ:
|
54 |
+
folder = train.main.dora.dir / 'grids' / __name__.split('.', 2)[-1]
|
55 |
+
with launcher.job_array():
|
56 |
+
for sig in folder.iterdir():
|
57 |
+
if not sig.is_symlink():
|
58 |
+
continue
|
59 |
+
xp = train.main.get_xp_from_sig(sig.name)
|
60 |
+
launcher(xp.argv)
|
61 |
+
return
|
62 |
+
|
63 |
+
audiogen_base = launcher.bind(solver="audiogen/audiogen_base_16khz")
|
64 |
+
audiogen_base.bind_({'autocast': False, 'fsdp.use': True})
|
65 |
+
|
66 |
+
audiogen_base_medium = audiogen_base.bind({'continue_from': '//pretrained/facebook/audiogen-medium'})
|
67 |
+
audiogen_base_medium.bind_({'model/lm/model_scale': 'medium'})
|
68 |
+
eval(audiogen_base_medium, batch_size=128)
|
audiocraft/grids/compression/__init__.py
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""EnCodec grids."""
|
audiocraft/grids/compression/_explorers.py
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import treetable as tt
|
8 |
+
|
9 |
+
from .._base_explorers import BaseExplorer
|
10 |
+
|
11 |
+
|
12 |
+
class CompressionExplorer(BaseExplorer):
|
13 |
+
eval_metrics = ["sisnr", "visqol"]
|
14 |
+
|
15 |
+
def stages(self):
|
16 |
+
return ["train", "valid", "evaluate"]
|
17 |
+
|
18 |
+
def get_grid_meta(self):
|
19 |
+
"""Returns the list of Meta information to display for each XP/job.
|
20 |
+
"""
|
21 |
+
return [
|
22 |
+
tt.leaf("index", align=">"),
|
23 |
+
tt.leaf("name", wrap=140),
|
24 |
+
tt.leaf("state"),
|
25 |
+
tt.leaf("sig", align=">"),
|
26 |
+
]
|
27 |
+
|
28 |
+
def get_grid_metrics(self):
|
29 |
+
"""Return the metrics that should be displayed in the tracking table.
|
30 |
+
"""
|
31 |
+
return [
|
32 |
+
tt.group(
|
33 |
+
"train",
|
34 |
+
[
|
35 |
+
tt.leaf("epoch"),
|
36 |
+
tt.leaf("bandwidth", ".2f"),
|
37 |
+
tt.leaf("adv", ".4f"),
|
38 |
+
tt.leaf("d_loss", ".4f"),
|
39 |
+
],
|
40 |
+
align=">",
|
41 |
+
),
|
42 |
+
tt.group(
|
43 |
+
"valid",
|
44 |
+
[
|
45 |
+
tt.leaf("bandwidth", ".2f"),
|
46 |
+
tt.leaf("adv", ".4f"),
|
47 |
+
tt.leaf("msspec", ".4f"),
|
48 |
+
tt.leaf("sisnr", ".2f"),
|
49 |
+
],
|
50 |
+
align=">",
|
51 |
+
),
|
52 |
+
tt.group(
|
53 |
+
"evaluate", [tt.leaf(name, ".3f") for name in self.eval_metrics], align=">"
|
54 |
+
),
|
55 |
+
]
|
audiocraft/grids/compression/debug.py
ADDED
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
"""
|
8 |
+
Grid search file, simply list all the exp you want in `explorer`.
|
9 |
+
Any new exp added there will be scheduled.
|
10 |
+
You can cancel and experiment by commenting its line.
|
11 |
+
|
12 |
+
This grid is a minimal example for debugging compression task
|
13 |
+
and how to override parameters directly in a grid.
|
14 |
+
Learn more about dora grids: https://github.com/facebookresearch/dora
|
15 |
+
"""
|
16 |
+
|
17 |
+
from ._explorers import CompressionExplorer
|
18 |
+
from ...environment import AudioCraftEnvironment
|
19 |
+
|
20 |
+
|
21 |
+
@CompressionExplorer
|
22 |
+
def explorer(launcher):
|
23 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
24 |
+
launcher.slurm_(gpus=2, partition=partitions)
|
25 |
+
launcher.bind_(solver='compression/debug')
|
26 |
+
|
27 |
+
with launcher.job_array():
|
28 |
+
# base debug task using config from solver=compression/debug
|
29 |
+
launcher()
|
30 |
+
# we can override parameters in the grid to launch additional xps
|
31 |
+
launcher({'rvq.bins': 2048, 'rvq.n_q': 4})
|
audiocraft/grids/compression/encodec_audiogen_16khz.py
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
"""
|
8 |
+
Grid search file, simply list all the exp you want in `explorer`.
|
9 |
+
Any new exp added there will be scheduled.
|
10 |
+
You can cancel and experiment by commenting its line.
|
11 |
+
|
12 |
+
This grid shows how to train the new AudioGen EnCodec model at 16 kHz.
|
13 |
+
"""
|
14 |
+
|
15 |
+
from ._explorers import CompressionExplorer
|
16 |
+
from ...environment import AudioCraftEnvironment
|
17 |
+
|
18 |
+
|
19 |
+
@CompressionExplorer
|
20 |
+
def explorer(launcher):
|
21 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
22 |
+
launcher.slurm_(gpus=8, partition=partitions)
|
23 |
+
# use configuration for AudioGen's EnCodec model trained on monophonic audio sampled at 16 kHz
|
24 |
+
# AudioGen's EnCodec is trained with a total stride of 320 leading to a frame rate of 50 hz
|
25 |
+
launcher.bind_(solver='compression/encodec_audiogen_16khz')
|
26 |
+
# replace this by the desired sound dataset
|
27 |
+
launcher.bind_(dset='internal/sounds_16khz')
|
28 |
+
# launch xp
|
29 |
+
launcher()
|
audiocraft/grids/compression/encodec_base_24khz.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
"""
|
8 |
+
Grid search file, simply list all the exp you want in `explorer`.
|
9 |
+
Any new exp added there will be scheduled.
|
10 |
+
You can cancel and experiment by commenting its line.
|
11 |
+
|
12 |
+
This grid shows how to train a base causal EnCodec model at 24 kHz.
|
13 |
+
"""
|
14 |
+
|
15 |
+
from ._explorers import CompressionExplorer
|
16 |
+
from ...environment import AudioCraftEnvironment
|
17 |
+
|
18 |
+
|
19 |
+
@CompressionExplorer
|
20 |
+
def explorer(launcher):
|
21 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
22 |
+
launcher.slurm_(gpus=8, partition=partitions)
|
23 |
+
# base causal EnCodec trained on monophonic audio sampled at 24 kHz
|
24 |
+
launcher.bind_(solver='compression/encodec_base_24khz')
|
25 |
+
# replace this by the desired dataset
|
26 |
+
launcher.bind_(dset='audio/example')
|
27 |
+
# launch xp
|
28 |
+
launcher()
|
audiocraft/grids/compression/encodec_musicgen_32khz.py
ADDED
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
"""
|
8 |
+
Grid search file, simply list all the exp you want in `explorer`.
|
9 |
+
Any new exp added there will be scheduled.
|
10 |
+
You can cancel and experiment by commenting its line.
|
11 |
+
|
12 |
+
This grid shows how to train a MusicGen EnCodec model at 32 kHz.
|
13 |
+
"""
|
14 |
+
|
15 |
+
from ._explorers import CompressionExplorer
|
16 |
+
from ...environment import AudioCraftEnvironment
|
17 |
+
|
18 |
+
|
19 |
+
@CompressionExplorer
|
20 |
+
def explorer(launcher):
|
21 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
22 |
+
launcher.slurm_(gpus=8, partition=partitions)
|
23 |
+
# use configuration for MusicGen's EnCodec model trained on monophonic audio sampled at 32 kHz
|
24 |
+
# MusicGen's EnCodec is trained with a total stride of 640 leading to a frame rate of 50 hz
|
25 |
+
launcher.bind_(solver='compression/encodec_musicgen_32khz')
|
26 |
+
# replace this by the desired music dataset
|
27 |
+
launcher.bind_(dset='internal/music_400k_32khz')
|
28 |
+
# launch xp
|
29 |
+
launcher()
|
30 |
+
launcher({
|
31 |
+
'metrics.visqol.bin': '/data/home/jadecopet/local/usr/opt/visqol',
|
32 |
+
'label': 'visqol',
|
33 |
+
'evaluate.metrics.visqol': True
|
34 |
+
})
|
audiocraft/grids/diffusion/4_bands_base_32khz.py
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
"""
|
8 |
+
Training of the 4 diffusion models described in
|
9 |
+
"From Discrete Tokens to High-Fidelity Audio Using Multi-Band Diffusion"
|
10 |
+
(paper link).
|
11 |
+
"""
|
12 |
+
|
13 |
+
from ._explorers import DiffusionExplorer
|
14 |
+
|
15 |
+
|
16 |
+
@DiffusionExplorer
|
17 |
+
def explorer(launcher):
|
18 |
+
launcher.slurm_(gpus=4, partition='learnfair')
|
19 |
+
|
20 |
+
launcher.bind_({'solver': 'diffusion/default',
|
21 |
+
'dset': 'internal/music_10k_32khz'})
|
22 |
+
|
23 |
+
with launcher.job_array():
|
24 |
+
launcher({'filter.use': True, 'filter.idx_band': 0, "processor.use": False, 'processor.power_std': 0.4})
|
25 |
+
launcher({'filter.use': True, 'filter.idx_band': 1, "processor.use": False, 'processor.power_std': 0.4})
|
26 |
+
launcher({'filter.use': True, 'filter.idx_band': 2, "processor.use": True, 'processor.power_std': 0.4})
|
27 |
+
launcher({'filter.use': True, 'filter.idx_band': 3, "processor.use": True, 'processor.power_std': 0.75})
|
audiocraft/grids/diffusion/__init__.py
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""Diffusion grids."""
|
audiocraft/grids/diffusion/_explorers.py
ADDED
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import treetable as tt
|
8 |
+
|
9 |
+
from .._base_explorers import BaseExplorer
|
10 |
+
|
11 |
+
|
12 |
+
class DiffusionExplorer(BaseExplorer):
|
13 |
+
eval_metrics = ["sisnr", "visqol"]
|
14 |
+
|
15 |
+
def stages(self):
|
16 |
+
return ["train", "valid", "valid_ema", "evaluate", "evaluate_ema"]
|
17 |
+
|
18 |
+
def get_grid_meta(self):
|
19 |
+
"""Returns the list of Meta information to display for each XP/job.
|
20 |
+
"""
|
21 |
+
return [
|
22 |
+
tt.leaf("index", align=">"),
|
23 |
+
tt.leaf("name", wrap=140),
|
24 |
+
tt.leaf("state"),
|
25 |
+
tt.leaf("sig", align=">"),
|
26 |
+
]
|
27 |
+
|
28 |
+
def get_grid_metrics(self):
|
29 |
+
"""Return the metrics that should be displayed in the tracking table.
|
30 |
+
"""
|
31 |
+
return [
|
32 |
+
tt.group(
|
33 |
+
"train",
|
34 |
+
[
|
35 |
+
tt.leaf("epoch"),
|
36 |
+
tt.leaf("loss", ".3%"),
|
37 |
+
],
|
38 |
+
align=">",
|
39 |
+
),
|
40 |
+
tt.group(
|
41 |
+
"valid",
|
42 |
+
[
|
43 |
+
tt.leaf("loss", ".3%"),
|
44 |
+
# tt.leaf("loss_0", ".3%"),
|
45 |
+
],
|
46 |
+
align=">",
|
47 |
+
),
|
48 |
+
tt.group(
|
49 |
+
"valid_ema",
|
50 |
+
[
|
51 |
+
tt.leaf("loss", ".3%"),
|
52 |
+
# tt.leaf("loss_0", ".3%"),
|
53 |
+
],
|
54 |
+
align=">",
|
55 |
+
),
|
56 |
+
tt.group(
|
57 |
+
"evaluate", [tt.leaf("rvm", ".4f"), tt.leaf("rvm_0", ".4f"),
|
58 |
+
tt.leaf("rvm_1", ".4f"), tt.leaf("rvm_2", ".4f"),
|
59 |
+
tt.leaf("rvm_3", ".4f"), ], align=">"
|
60 |
+
),
|
61 |
+
tt.group(
|
62 |
+
"evaluate_ema", [tt.leaf("rvm", ".4f"), tt.leaf("rvm_0", ".4f"),
|
63 |
+
tt.leaf("rvm_1", ".4f"), tt.leaf("rvm_2", ".4f"),
|
64 |
+
tt.leaf("rvm_3", ".4f")], align=">"
|
65 |
+
),
|
66 |
+
]
|
audiocraft/grids/musicgen/__init__.py
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
"""MusicGen grids."""
|
audiocraft/grids/musicgen/_explorers.py
ADDED
@@ -0,0 +1,93 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
import typing as tp
|
8 |
+
|
9 |
+
import treetable as tt
|
10 |
+
|
11 |
+
from .._base_explorers import BaseExplorer
|
12 |
+
|
13 |
+
|
14 |
+
class LMExplorer(BaseExplorer):
|
15 |
+
eval_metrics: tp.List[str] = []
|
16 |
+
|
17 |
+
def stages(self) -> tp.List[str]:
|
18 |
+
return ['train', 'valid']
|
19 |
+
|
20 |
+
def get_grid_metrics(self):
|
21 |
+
"""Return the metrics that should be displayed in the tracking table."""
|
22 |
+
return [
|
23 |
+
tt.group(
|
24 |
+
'train',
|
25 |
+
[
|
26 |
+
tt.leaf('epoch'),
|
27 |
+
tt.leaf('duration', '.1f'), # duration in minutes
|
28 |
+
tt.leaf('ping'),
|
29 |
+
tt.leaf('ce', '.4f'), # cross entropy
|
30 |
+
tt.leaf("ppl", '.3f'), # perplexity
|
31 |
+
],
|
32 |
+
align='>',
|
33 |
+
),
|
34 |
+
tt.group(
|
35 |
+
'valid',
|
36 |
+
[
|
37 |
+
tt.leaf('ce', '.4f'),
|
38 |
+
tt.leaf('ppl', '.3f'),
|
39 |
+
tt.leaf('best_ppl', '.3f'),
|
40 |
+
],
|
41 |
+
align='>',
|
42 |
+
),
|
43 |
+
]
|
44 |
+
|
45 |
+
def process_sheep(self, sheep, history):
|
46 |
+
parts = super().process_sheep(sheep, history)
|
47 |
+
|
48 |
+
track_by = {'ppl': 'lower'} # values should be in ['lower', 'higher']
|
49 |
+
best_metrics = {k: (1 if v == 'lower' else -1) * float('inf') for k, v in track_by.items()}
|
50 |
+
|
51 |
+
def comparator(mode, a, b):
|
52 |
+
return a < b if mode == 'lower' else a > b
|
53 |
+
|
54 |
+
for metrics in history:
|
55 |
+
for key, sub in metrics.items():
|
56 |
+
for metric in track_by:
|
57 |
+
# for the validation set, keep track of best metrics (ppl in this example)
|
58 |
+
# this is so we can conveniently compare metrics between runs in the grid
|
59 |
+
if key == 'valid' and metric in sub and comparator(
|
60 |
+
track_by[metric], sub[metric], best_metrics[metric]
|
61 |
+
):
|
62 |
+
best_metrics[metric] = sub[metric]
|
63 |
+
|
64 |
+
if 'valid' in parts:
|
65 |
+
parts['valid'].update({f'best_{k}': v for k, v in best_metrics.items()})
|
66 |
+
return parts
|
67 |
+
|
68 |
+
|
69 |
+
class GenerationEvalExplorer(BaseExplorer):
|
70 |
+
eval_metrics: tp.List[str] = []
|
71 |
+
|
72 |
+
def stages(self) -> tp.List[str]:
|
73 |
+
return ['evaluate']
|
74 |
+
|
75 |
+
def get_grid_metrics(self):
|
76 |
+
"""Return the metrics that should be displayed in the tracking table."""
|
77 |
+
return [
|
78 |
+
tt.group(
|
79 |
+
'evaluate',
|
80 |
+
[
|
81 |
+
tt.leaf('epoch', '.3f'),
|
82 |
+
tt.leaf('duration', '.1f'),
|
83 |
+
tt.leaf('ping'),
|
84 |
+
tt.leaf('ce', '.4f'),
|
85 |
+
tt.leaf('ppl', '.3f'),
|
86 |
+
tt.leaf('fad', '.3f'),
|
87 |
+
tt.leaf('kld', '.3f'),
|
88 |
+
tt.leaf('text_consistency', '.3f'),
|
89 |
+
tt.leaf('chroma_cosine', '.3f'),
|
90 |
+
],
|
91 |
+
align='>',
|
92 |
+
),
|
93 |
+
]
|
audiocraft/grids/musicgen/musicgen_base_32khz.py
ADDED
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from ._explorers import LMExplorer
|
8 |
+
from ...environment import AudioCraftEnvironment
|
9 |
+
|
10 |
+
|
11 |
+
@LMExplorer
|
12 |
+
def explorer(launcher):
|
13 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
14 |
+
launcher.slurm_(gpus=32, partition=partitions)
|
15 |
+
launcher.bind_(solver='musicgen/musicgen_base_32khz')
|
16 |
+
# replace this by the desired music dataset
|
17 |
+
launcher.bind_(dset='internal/music_400k_32khz')
|
18 |
+
|
19 |
+
fsdp = {'autocast': False, 'fsdp.use': True}
|
20 |
+
medium = {'model/lm/model_scale': 'medium'}
|
21 |
+
large = {'model/lm/model_scale': 'large'}
|
22 |
+
|
23 |
+
cfg_low = {'classifier_free_guidance.training_dropout': 0.2}
|
24 |
+
wd_low = {'conditioners.description.t5.word_dropout': 0.2}
|
25 |
+
|
26 |
+
adam = {'optim.optimizer': 'adamw', 'optim.lr': 1e-4}
|
27 |
+
|
28 |
+
launcher.bind_(fsdp)
|
29 |
+
|
30 |
+
launcher.slurm_(gpus=32).bind_(label='32gpus')
|
31 |
+
with launcher.job_array():
|
32 |
+
sub = launcher.bind()
|
33 |
+
sub()
|
34 |
+
|
35 |
+
launcher.slurm_(gpus=64).bind_(label='64gpus')
|
36 |
+
with launcher.job_array():
|
37 |
+
sub = launcher.bind()
|
38 |
+
sub(medium, adam)
|
39 |
+
|
40 |
+
launcher.slurm_(gpus=96).bind_(label='96gpus')
|
41 |
+
with launcher.job_array():
|
42 |
+
sub = launcher.bind()
|
43 |
+
sub(large, cfg_low, wd_low, adam, {'optim.max_norm': 3})
|
audiocraft/grids/musicgen/musicgen_base_cached_32khz.py
ADDED
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from ._explorers import LMExplorer
|
8 |
+
from ...environment import AudioCraftEnvironment
|
9 |
+
|
10 |
+
|
11 |
+
@LMExplorer
|
12 |
+
def explorer(launcher):
|
13 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
14 |
+
launcher.slurm_(gpus=32, partition=partitions)
|
15 |
+
launcher.bind_(solver='musicgen/musicgen_base_32khz')
|
16 |
+
# replace this by the desired music dataset
|
17 |
+
launcher.bind_(dset='internal/music_400k_32khz')
|
18 |
+
|
19 |
+
fsdp = {'autocast': False, 'fsdp.use': True}
|
20 |
+
medium = {'model/lm/model_scale': 'medium'}
|
21 |
+
large = {'model/lm/model_scale': 'large'}
|
22 |
+
|
23 |
+
cfg_low = {'classifier_free_guidance.training_dropout': 0.2}
|
24 |
+
wd_low = {'conditioners.description.t5.word_dropout': 0.2}
|
25 |
+
|
26 |
+
adam = {'optim.optimizer': 'adamw', 'optim.lr': 1e-4}
|
27 |
+
|
28 |
+
# BEGINNING OF CACHE WRITING JOBS.
|
29 |
+
cache_write = {
|
30 |
+
'cache.path': '/fsx-codegen/defossez/cache/interleave_stereo_nv_32k',
|
31 |
+
'cache.write': True,
|
32 |
+
'generate.every': 500,
|
33 |
+
'evaluate.every': 500,
|
34 |
+
'logging.log_updates': 50,
|
35 |
+
}
|
36 |
+
|
37 |
+
cache_sub = launcher.bind({'model/lm/model_scale': 'xsmall', 'conditioner': 'none'})
|
38 |
+
cache_sub.bind_({'deadlock.use': True})
|
39 |
+
cache_sub.slurm_(gpus=8)
|
40 |
+
with launcher.job_array():
|
41 |
+
num_shards = 10 # total number of jobs running in parallel.
|
42 |
+
for shard in range(0, num_shards):
|
43 |
+
launcher(cache_write, {'cache.write_num_shards': num_shards, 'cache.write_shard': shard})
|
44 |
+
|
45 |
+
# REMOVE THE FOLLOWING RETURN STATEMENT ONCE THE ABOVE JOBS ARE DONE,
|
46 |
+
# OR SUFFICIENTLY AHEAD.
|
47 |
+
return
|
48 |
+
|
49 |
+
cache = {
|
50 |
+
'cache.path': '/fsx-codegen/defossez/cache/interleave_stereo_nv_32k',
|
51 |
+
}
|
52 |
+
launcher.bind_(fsdp, cache)
|
53 |
+
|
54 |
+
launcher.slurm_(gpus=32).bind_(label='32gpus')
|
55 |
+
with launcher.job_array():
|
56 |
+
sub = launcher.bind()
|
57 |
+
sub()
|
58 |
+
|
59 |
+
launcher.slurm_(gpus=64).bind_(label='64gpus')
|
60 |
+
with launcher.job_array():
|
61 |
+
sub = launcher.bind()
|
62 |
+
sub(medium, adam)
|
63 |
+
|
64 |
+
launcher.slurm_(gpus=96).bind_(label='96gpus')
|
65 |
+
with launcher.job_array():
|
66 |
+
sub = launcher.bind()
|
67 |
+
sub(large, cfg_low, wd_low, adam, {'optim.max_norm': 3})
|
audiocraft/grids/musicgen/musicgen_clapemb_32khz.py
ADDED
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from ._explorers import LMExplorer
|
8 |
+
from ...environment import AudioCraftEnvironment
|
9 |
+
|
10 |
+
|
11 |
+
@LMExplorer
|
12 |
+
def explorer(launcher):
|
13 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
14 |
+
launcher.slurm_(gpus=32, partition=partitions)
|
15 |
+
launcher.bind_(solver='musicgen/musicgen_base_32khz')
|
16 |
+
# replace this by the desired music dataset
|
17 |
+
launcher.bind_(dset='internal/music_400k_32khz')
|
18 |
+
launcher.bind_(conditioner='clapemb2music')
|
19 |
+
|
20 |
+
fsdp = {'autocast': False, 'fsdp.use': True}
|
21 |
+
cache_path = {'conditioners.description.clap.cache_path':
|
22 |
+
'/fsx-audio-craft-llm/jadecopet/experiments/audiocraft/caches/clap_embed_music'}
|
23 |
+
text_wav_training_opt = {'conditioners.description.clap.text_p': 0.5}
|
24 |
+
|
25 |
+
launcher.bind_(fsdp)
|
26 |
+
|
27 |
+
launcher.slurm_(gpus=32).bind_(label='32gpus')
|
28 |
+
with launcher.job_array():
|
29 |
+
launcher()
|
30 |
+
launcher(text_wav_training_opt)
|
31 |
+
launcher(cache_path)
|
32 |
+
launcher(cache_path, text_wav_training_opt)
|
audiocraft/grids/musicgen/musicgen_melody_32khz.py
ADDED
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from ._explorers import LMExplorer
|
8 |
+
from ...environment import AudioCraftEnvironment
|
9 |
+
|
10 |
+
|
11 |
+
@LMExplorer
|
12 |
+
def explorer(launcher):
|
13 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
14 |
+
launcher.slurm_(gpus=32, partition=partitions)
|
15 |
+
launcher.bind_(solver='musicgen/musicgen_melody_32khz')
|
16 |
+
# replace this by the desired music dataset
|
17 |
+
launcher.bind_(dset='internal/music_400k_32khz')
|
18 |
+
|
19 |
+
fsdp = {'autocast': False, 'fsdp.use': True}
|
20 |
+
medium = {'model/lm/model_scale': 'medium'}
|
21 |
+
large = {'model/lm/model_scale': 'large'}
|
22 |
+
|
23 |
+
cfg_low = {'classifier_free_guidance.training_dropout': 0.2}
|
24 |
+
wd_low = {'conditioners.description.t5.word_dropout': 0.2}
|
25 |
+
|
26 |
+
adam = {'optim.optimizer': 'adamw', 'optim.lr': 1e-4}
|
27 |
+
|
28 |
+
cache_path = {'conditioners.self_wav.chroma_stem.cache_path':
|
29 |
+
'/fsx-audio-craft-llm/jadecopet/experiments/audiocraft/caches/chroma_stem'}
|
30 |
+
|
31 |
+
# CACHE GENERATION JOBS
|
32 |
+
n_cache_gen_jobs = 4
|
33 |
+
gen_sub = launcher.slurm(gpus=1)
|
34 |
+
gen_sub.bind_(
|
35 |
+
cache_path, {
|
36 |
+
# the cache is always computed over the whole file, so duration doesn't matter here.
|
37 |
+
'dataset.segment_duration': 2.,
|
38 |
+
'dataset.batch_size': 8,
|
39 |
+
'dataset.train.permutation_on_files': True, # try to not repeat files.
|
40 |
+
'optim.epochs': 10,
|
41 |
+
'model/lm/model_scale': 'xsmall',
|
42 |
+
|
43 |
+
})
|
44 |
+
with gen_sub.job_array():
|
45 |
+
for gen_job in range(n_cache_gen_jobs):
|
46 |
+
gen_sub({'dataset.train.shuffle_seed': gen_job})
|
47 |
+
|
48 |
+
# ACTUAL TRAINING JOBS.
|
49 |
+
launcher.bind_(fsdp)
|
50 |
+
|
51 |
+
launcher.slurm_(gpus=32).bind_(label='32gpus')
|
52 |
+
with launcher.job_array():
|
53 |
+
sub = launcher.bind()
|
54 |
+
sub()
|
55 |
+
sub(cache_path)
|
56 |
+
|
57 |
+
launcher.slurm_(gpus=64).bind_(label='64gpus')
|
58 |
+
with launcher.job_array():
|
59 |
+
sub = launcher.bind()
|
60 |
+
sub(medium, adam)
|
61 |
+
|
62 |
+
launcher.slurm_(gpus=96).bind_(label='96gpus')
|
63 |
+
with launcher.job_array():
|
64 |
+
sub = launcher.bind()
|
65 |
+
sub(large, cfg_low, wd_low, adam, {'optim.max_norm': 3})
|
audiocraft/grids/musicgen/musicgen_pretrained_32khz_eval.py
ADDED
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
"""
|
8 |
+
Evaluation with objective metrics for the pretrained MusicGen models.
|
9 |
+
This grid takes signature from the training grid and runs evaluation-only stage.
|
10 |
+
|
11 |
+
When running the grid for the first time, please use:
|
12 |
+
REGEN=1 dora grid musicgen.musicgen_pretrained_32khz_eval
|
13 |
+
and re-use the REGEN=1 option when the grid is changed to force regenerating it.
|
14 |
+
|
15 |
+
Note that you need the proper metrics external libraries setup to use all
|
16 |
+
the objective metrics activated in this grid. Refer to the README for more information.
|
17 |
+
"""
|
18 |
+
|
19 |
+
import os
|
20 |
+
|
21 |
+
from ._explorers import GenerationEvalExplorer
|
22 |
+
from ...environment import AudioCraftEnvironment
|
23 |
+
from ... import train
|
24 |
+
|
25 |
+
|
26 |
+
def eval(launcher, batch_size: int = 32, eval_melody: bool = False):
|
27 |
+
opts = {
|
28 |
+
'dset': 'audio/musiccaps_32khz',
|
29 |
+
'solver/musicgen/evaluation': 'objective_eval',
|
30 |
+
'execute_only': 'evaluate',
|
31 |
+
'+dataset.evaluate.batch_size': batch_size,
|
32 |
+
'+metrics.fad.tf.batch_size': 16,
|
33 |
+
}
|
34 |
+
# chroma-specific evaluation
|
35 |
+
chroma_opts = {
|
36 |
+
'dset': 'internal/music_400k_32khz',
|
37 |
+
'dataset.evaluate.segment_duration': 30,
|
38 |
+
'dataset.evaluate.num_samples': 1000,
|
39 |
+
'evaluate.metrics.chroma_cosine': True,
|
40 |
+
'evaluate.metrics.fad': False,
|
41 |
+
'evaluate.metrics.kld': False,
|
42 |
+
'evaluate.metrics.text_consistency': False,
|
43 |
+
}
|
44 |
+
# binary for FAD computation: replace this path with your own path
|
45 |
+
metrics_opts = {
|
46 |
+
'metrics.fad.tf.bin': '/data/home/jadecopet/local/usr/opt/google-research'
|
47 |
+
}
|
48 |
+
opt1 = {'generate.lm.use_sampling': True, 'generate.lm.top_k': 250, 'generate.lm.top_p': 0.}
|
49 |
+
opt2 = {'transformer_lm.two_step_cfg': True}
|
50 |
+
|
51 |
+
sub = launcher.bind(opts)
|
52 |
+
sub.bind_(metrics_opts)
|
53 |
+
|
54 |
+
# base objective metrics
|
55 |
+
sub(opt1, opt2)
|
56 |
+
|
57 |
+
if eval_melody:
|
58 |
+
# chroma-specific metrics
|
59 |
+
sub(opt1, opt2, chroma_opts)
|
60 |
+
|
61 |
+
|
62 |
+
@GenerationEvalExplorer
|
63 |
+
def explorer(launcher):
|
64 |
+
partitions = AudioCraftEnvironment.get_slurm_partitions(['team', 'global'])
|
65 |
+
launcher.slurm_(gpus=4, partition=partitions)
|
66 |
+
|
67 |
+
if 'REGEN' not in os.environ:
|
68 |
+
folder = train.main.dora.dir / 'grids' / __name__.split('.', 2)[-1]
|
69 |
+
with launcher.job_array():
|
70 |
+
for sig in folder.iterdir():
|
71 |
+
if not sig.is_symlink():
|
72 |
+
continue
|
73 |
+
xp = train.main.get_xp_from_sig(sig.name)
|
74 |
+
launcher(xp.argv)
|
75 |
+
return
|
76 |
+
|
77 |
+
with launcher.job_array():
|
78 |
+
musicgen_base = launcher.bind(solver="musicgen/musicgen_base_32khz")
|
79 |
+
musicgen_base.bind_({'autocast': False, 'fsdp.use': True})
|
80 |
+
|
81 |
+
# base musicgen models
|
82 |
+
musicgen_base_small = musicgen_base.bind({'continue_from': '//pretrained/facebook/musicgen-small'})
|
83 |
+
eval(musicgen_base_small, batch_size=128)
|
84 |
+
|
85 |
+
musicgen_base_medium = musicgen_base.bind({'continue_from': '//pretrained/facebook/musicgen-medium'})
|
86 |
+
musicgen_base_medium.bind_({'model/lm/model_scale': 'medium'})
|
87 |
+
eval(musicgen_base_medium, batch_size=128)
|
88 |
+
|
89 |
+
musicgen_base_large = musicgen_base.bind({'continue_from': '//pretrained/facebook/musicgen-large'})
|
90 |
+
musicgen_base_large.bind_({'model/lm/model_scale': 'large'})
|
91 |
+
eval(musicgen_base_large, batch_size=128)
|
92 |
+
|
93 |
+
# melody musicgen model
|
94 |
+
musicgen_melody = launcher.bind(solver="musicgen/musicgen_melody_32khz")
|
95 |
+
musicgen_melody.bind_({'autocast': False, 'fsdp.use': True})
|
96 |
+
|
97 |
+
musicgen_melody_medium = musicgen_melody.bind({'continue_from': '//pretrained/facebook/musicgen-melody'})
|
98 |
+
musicgen_melody_medium.bind_({'model/lm/model_scale': 'medium'})
|
99 |
+
eval(musicgen_melody_medium, batch_size=128, eval_melody=True)
|