Repository: eriklindernoren/PyTorch-YOLOv3
Branch: master
Commit: 1d621c8489e2
Files: 29
Total size: 142.9 KB
Directory structure:
gitextract_qej34q3x/
├── .github/
│ ├── ISSUE_TEMPLATE/
│ │ ├── 1_bug_report.md
│ │ ├── 2_need_help.md
│ │ └── 3_feature_request.md
│ ├── dependabot.yml
│ ├── pull_request_template.md
│ └── workflows/
│ └── main.yml
├── .gitignore
├── LICENSE
├── README.md
├── config/
│ ├── coco.data
│ ├── create_custom_model.sh
│ ├── custom.data
│ ├── yolov3-tiny.cfg
│ └── yolov3.cfg
├── pyproject.toml
├── pytorchyolo/
│ ├── __init__.py
│ ├── detect.py
│ ├── models.py
│ ├── test.py
│ ├── train.py
│ └── utils/
│ ├── __init__.py
│ ├── augmentations.py
│ ├── datasets.py
│ ├── logger.py
│ ├── loss.py
│ ├── parse_config.py
│ ├── transforms.py
│ └── utils.py
└── weights/
└── download_weights.sh
================================================
FILE CONTENTS
================================================
================================================
FILE: .github/ISSUE_TEMPLATE/1_bug_report.md
================================================
---
name: "\U0001F41B Bug report"
about: Report a bug, crash or some misbehavior
title: ''
labels: 'bug'
assignees: ''
---
<!--- Provide a general summary of the issue in the title above -->
## Context
<!--- Please provide context, as this streamlines the debugging process. Mark the correct cases and follow the instructions. -->
- [ ] I have installed this repo manually and the issue occurred on this commit:
<!--- Get the current commit hash either from the first printout of the program or by executing the following command: 'git rev-parse --short HEAD' -->
- [ ] I have installed this repo via `PIP` and the issue occurred on version: <!--- Get the current version number by executing the following command: 'pip show pytorchyolo' -->
- [ ] The issue occurred when using the following .cfg model:
- [ ] `yolov3`
- [ ] `yolov3-tiny`
- [ ] `CUSTOM`
## Necessary Checks
<!--- Please ensure, you have completed the following checks. This helps to give insight into the issue and prevent already resolved issues. -->
- [ ] The issue occurred on the newest version
<!--- If installed manually, run: 'git pull && poetry install' -->
<!--- If installed via PIP, run: 'pip install --upgrade pytorchyolo' -->
- [ ] I couldn't find a similar issue here on this project's github repo
- [ ] If the issue is CUDA related (CUDA error), I have tested and provided the traceback also when CUDA is turned off <!--- For linux, rerun your steps with the prefix CUDA_VISIBLE_DEVICES="" -->
- [ ] I have provided all tracebacks or printouts in ```Text Form``` <!--- This makes it easier to search for errors. -->
- [ ] In case, the issue occurred on a custom .cfg model, I have provided the model down below
## Expected behavior
<!--- Describe what you expected to happen -->
## Current behavior
<!--- Describe what actually happened instead of the expected behavior -->
## Steps to Reproduce
<!--- An unambiguous set of steps to reproduce this bug. -->
<!--- Code-snippets, screenshots ot other details are welcome if needed. -->
1.
2.
3.
...
## Possible Solution
<!--- If you already have an idea, you can suggest a fix/reason for the bug. This is not obligatory. -->
<!--- Please remove the following block, if this does not apply to you issue. -->
### Custom `.cfg`
<!--- Please paste your custom .cfg model below. -->
<details><summary>Custom .cfg</summary>
<p>
<!--- YOUR CUSTOM .CFG HERE -->
</p>
</details>
================================================
FILE: .github/ISSUE_TEMPLATE/2_need_help.md
================================================
---
name: "⁉️ Need help?"
about: "Get help with using or improving our software"
title: ''
labels: ''
assignees: ''
---
## What I'm trying to do
<!--- Please describe what you're trying to do so we know what your problem is about. -->
## What I've tried
<!--- If you tell us, what you already tried or what documentation you already read, we are able to help you better by not pointing to information you already know. -->
## Additional context
<!--- If there's more to say, feel free to do so :) -->
================================================
FILE: .github/ISSUE_TEMPLATE/3_feature_request.md
================================================
---
name: "\U0001F680 Feature request"
about: Suggest an idea for this project
labels: 'enhancement'
---
<!--
Thank you for suggesting an idea to make us better.
Please fill in as much of the template below as you're able.
-->
## Is your feature request related to a problem? Please describe.
<!-- Please describe the problem you are trying to solve. -->
## Describe the solution you'd like
<!-- Please describe the desired behavior. -->
## Describe alternatives you've considered
<!-- Please describe alternative solutions or features you have considered. -->
<!-- This is not strictly necessary but helps all of us get a different point-of-view -->
================================================
FILE: .github/dependabot.yml
================================================
# To get started with Dependabot version updates, you'll need to specify which
# package ecosystems to update and where the package manifests are located.
# Please see the documentation for all configuration options:
# https://help.github.com/github/administering-a-repository/configuration-options-for-dependency-updates
version: 2
updates:
- package-ecosystem: "pip" # See documentation for possible values
directory: "/" # Location of package manifests
schedule:
interval: "daily"
================================================
FILE: .github/pull_request_template.md
================================================
## Proposed changes
<!--- Describe your changes and why they are necessary. -->
## Related issues
<!--- Mention (link) related issues. -->
<!--- If you suggest a new feature, please discuss it in an issue first. -->
<!--- If fixing a bug, there should be an issue describing it with steps to reproduce -->
## Necessary checks
- [ ] Update poetry package version [semantically](https://semver.org/)
- [ ] Write documentation
- [ ] Create issues for future work
- [ ] Test on your machine
================================================
FILE: .github/workflows/main.yml
================================================
name: CI
on: [pull_request, workflow_dispatch]
jobs:
main:
runs-on: ${{ matrix.os }}
strategy:
matrix:
os: [ubuntu-22.04, ubuntu-20.04, windows-latest]
steps:
- uses: actions/checkout@v2
- name: Set up Python
uses: actions/setup-python@v1
with:
python-version: 3.8
- name: Upgrade pip
run: python3 -m pip install --upgrade pip
- name: Install Poetry
run: pip3 install poetry --user
- name: Install Dependencies
run: poetry install
# Prints the help pages of all scripts to see if the imports etc. work
- name: Test the help pages
run: |
poetry run yolo-train -h
poetry run yolo-test -h
poetry run yolo-detect -h
- name: Demo Training
run: poetry run yolo-train --data config/custom.data --model config/yolov3.cfg --epochs 30
- name: Demo Evaluate
run: poetry run yolo-test --data config/custom.data --model config/yolov3.cfg --weights checkpoints/yolov3_ckpt_29.pth
- name: Demo Detect
run: poetry run yolo-detect --batch_size 2 --weights checkpoints/yolov3_ckpt_29.pth
linter:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v2
- name: Flake8
uses: TrueBrain/actions-flake8@master
with:
only_warn: 1
max_line_length: 150
path: pytorchyolo
================================================
FILE: .gitignore
================================================
.DS_Store
build
.git
*.egg-info
dist
output/
data/*
backup
weights/*.weights
weights/*.conv.*
__pycache__
checkpoints/
.vscode/
logs/
.python-version
================================================
FILE: LICENSE
================================================
GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
the GNU General Public License is intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users. We, the Free Software Foundation, use the
GNU General Public License for most of our software; it applies also to
any other work released this way by its authors. You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you
these rights or asking you to surrender the rights. Therefore, you have
certain responsibilities if you distribute copies of the software, or if
you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must pass on to the recipients the same
freedoms that you received. You must make sure that they, too, receive
or can get the source code. And you must show them these terms so they
know their rights.
Developers that use the GNU GPL protect your rights with two steps:
(1) assert copyright on the software, and (2) offer you this License
giving you legal permission to copy, distribute and/or modify it.
For the developers' and authors' protection, the GPL clearly explains
that there is no warranty for this free software. For both users' and
authors' sake, the GPL requires that modified versions be marked as
changed, so that their problems will not be attributed erroneously to
authors of previous versions.
Some devices are designed to deny users access to install or run
modified versions of the software inside them, although the manufacturer
can do so. This is fundamentally incompatible with the aim of
protecting users' freedom to change the software. The systematic
pattern of such abuse occurs in the area of products for individuals to
use, which is precisely where it is most unacceptable. Therefore, we
have designed this version of the GPL to prohibit the practice for those
products. If such problems arise substantially in other domains, we
stand ready to extend this provision to those domains in future versions
of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents.
States should not allow patents to restrict development and use of
software on general-purpose computers, but in those that do, we wish to
avoid the special danger that patents applied to a free program could
make it effectively proprietary. To prevent this, the GPL assures that
patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU Affero General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License,
section 13, concerning interaction through a network will apply to the
combination as such.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
<program> Copyright (C) <year> <name of author>
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, your program's commands
might be different; for a GUI interface, you would use an "about box".
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU GPL, see
<http://www.gnu.org/licenses/>.
The GNU General Public License does not permit incorporating your program
into proprietary programs. If your program is a subroutine library, you
may consider it more useful to permit linking proprietary applications with
the library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License. But first, please read
<http://www.gnu.org/philosophy/why-not-lgpl.html>.
================================================
FILE: README.md
================================================
# PyTorch YOLO
A minimal PyTorch implementation of YOLOv3, with support for training, inference and evaluation.
YOLOv4 and YOLOv7 weights are also compatible with this implementation.
[](https://github.com/eriklindernoren/PyTorch-YOLOv3/actions/workflows/main.yml) [](https://pypi.python.org/pypi/pytorchyolo/) [](LICENSE)
## Installation
### Installing from source
For normal training and evaluation we recommend installing the package from source using a poetry virtual environment.
```bash
git clone https://github.com/eriklindernoren/PyTorch-YOLOv3
cd PyTorch-YOLOv3/
pip3 install poetry --user
poetry install
```
You need to join the virtual environment by running `poetry shell` in this directory before running any of the following commands without the `poetry run` prefix.
Also have a look at the other installing method, if you want to use the commands everywhere without opening a poetry-shell.
#### Download pretrained weights
```bash
./weights/download_weights.sh
```
#### Download COCO
```bash
./data/get_coco_dataset.sh
```
### Install via pip
This installation method is recommended, if you want to use this package as a dependency in another python project.
This method only includes the code, is less isolated and may conflict with other packages.
Weights and the COCO dataset need to be downloaded as stated above.
See __API__ for further information regarding the packages API.
It also enables the CLI tools `yolo-detect`, `yolo-train`, and `yolo-test` everywhere without any additional commands.
```bash
pip3 install pytorchyolo --user
```
## Test
Evaluates the model on COCO test dataset.
To download this dataset as well as weights, see above.
```bash
poetry run yolo-test --weights weights/yolov3.weights
```
| Model | mAP (min. 50 IoU) |
| ----------------------- |:-----------------:|
| YOLOv3 608 (paper) | 57.9 |
| YOLOv3 608 (this impl.) | 57.3 |
| YOLOv3 416 (paper) | 55.3 |
| YOLOv3 416 (this impl.) | 55.5 |
## Inference
Uses pretrained weights to make predictions on images. Below table displays the inference times when using as inputs images scaled to 256x256. The ResNet backbone measurements are taken from the YOLOv3 paper. The Darknet-53 measurement marked shows the inference time of this implementation on my 1080ti card.
| Backbone | GPU | FPS |
| ----------------------- |:--------:|:--------:|
| ResNet-101 | Titan X | 53 |
| ResNet-152 | Titan X | 37 |
| Darknet-53 (paper) | Titan X | 76 |
| Darknet-53 (this impl.) | 1080ti | 74 |
```bash
poetry run yolo-detect --images data/samples/
```
<p align="center"><img src="https://github.com/eriklindernoren/PyTorch-YOLOv3/raw/master/assets/giraffe.png" width="480"\></p>
<p align="center"><img src="https://github.com/eriklindernoren/PyTorch-YOLOv3/raw/master/assets/dog.png" width="480"\></p>
<p align="center"><img src="https://github.com/eriklindernoren/PyTorch-YOLOv3/raw/master/assets/traffic.png" width="480"\></p>
<p align="center"><img src="https://github.com/eriklindernoren/PyTorch-YOLOv3/raw/master/assets/messi.png" width="480"\></p>
## Train
For argument descriptions have a look at `poetry run yolo-train --help`
#### Example (COCO)
To train on COCO using a Darknet-53 backend pretrained on ImageNet run:
```bash
poetry run yolo-train --data config/coco.data --pretrained_weights weights/darknet53.conv.74
```
#### Tensorboard
Track training progress in Tensorboard:
* Initialize training
* Run the command below
* Go to http://localhost:6006/
```bash
poetry run tensorboard --logdir='logs' --port=6006
```
Storing the logs on a slow drive possibly leads to a significant training speed decrease.
You can adjust the log directory using `--logdir <path>` when running `tensorboard` and `yolo-train`.
## Train on Custom Dataset
#### Custom model
Run the commands below to create a custom model definition, replacing `<num-classes>` with the number of classes in your dataset.
```bash
cd config
./create_custom_model.sh <num-classes> # Will create custom model 'yolov3-custom.cfg'
```
#### Classes
Add class names to `data/custom/classes.names`. This file should have one row per class name.
#### Image Folder
Move the images of your dataset to `data/custom/images/`.
#### Annotation Folder
Move your annotations to `data/custom/labels/`. The dataloader expects that the annotation file corresponding to the image `data/custom/images/train.jpg` has the path `data/custom/labels/train.txt`. Each row in the annotation file should define one bounding box, using the syntax `label_idx x_center y_center width height`. The coordinates should be scaled `[0, 1]`, and the `label_idx` should be zero-indexed and correspond to the row number of the class name in `data/custom/classes.names`.
#### Define Train and Validation Sets
In `data/custom/train.txt` and `data/custom/valid.txt`, add paths to images that will be used as train and validation data respectively.
#### Train
To train on the custom dataset run:
```bash
poetry run yolo-train --model config/yolov3-custom.cfg --data config/custom.data
```
Add `--pretrained_weights weights/darknet53.conv.74` to train using a backend pretrained on ImageNet.
## API
You are able to import the modules of this repo in your own project if you install the pip package `pytorchyolo`.
An example prediction call from a simple OpenCV python script would look like this:
```python
import cv2
from pytorchyolo import detect, models
# Load the YOLO model
model = models.load_model(
"<PATH_TO_YOUR_CONFIG_FOLDER>/yolov3.cfg",
"<PATH_TO_YOUR_WEIGHTS_FOLDER>/yolov3.weights")
# Load the image as a numpy array
img = cv2.imread("<PATH_TO_YOUR_IMAGE>")
# Convert OpenCV bgr to rgb
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# Runs the YOLO model on the image
boxes = detect.detect_image(model, img)
print(boxes)
# Output will be a numpy array in the following format:
# [[x1, y1, x2, y2, confidence, class]]
```
For more advanced usage look at the method's doc strings.
## Credit
### YOLOv3: An Incremental Improvement
_Joseph Redmon, Ali Farhadi_ <br>
**Abstract** <br>
We present some updates to YOLO! We made a bunch
of little design changes to make it better. We also trained
this new network that’s pretty swell. It’s a little bigger than
last time but more accurate. It’s still fast though, don’t
worry. At 320 × 320 YOLOv3 runs in 22 ms at 28.2 mAP,
as accurate as SSD but three times faster. When we look
at the old .5 IOU mAP detection metric YOLOv3 is quite
good. It achieves 57.9 AP50 in 51 ms on a Titan X, compared
to 57.5 AP50 in 198 ms by RetinaNet, similar performance
but 3.8× faster. As always, all the code is online at
https://pjreddie.com/yolo/.
[[Paper]](https://pjreddie.com/media/files/papers/YOLOv3.pdf) [[Project Webpage]](https://pjreddie.com/darknet/yolo/) [[Authors' Implementation]](https://github.com/pjreddie/darknet)
```
@article{yolov3,
title={YOLOv3: An Incremental Improvement},
author={Redmon, Joseph and Farhadi, Ali},
journal = {arXiv},
year={2018}
}
```
## Other
### YOEO — You Only Encode Once
[YOEO](https://github.com/bit-bots/YOEO) extends this repo with the ability to train an additional semantic segmentation decoder. The lightweight example model is mainly targeted towards embedded real-time applications.
================================================
FILE: config/coco.data
================================================
classes= 80
train=data/coco/trainvalno5k.txt
valid=data/coco/5k.txt
names=data/coco.names
backup=backup/
eval=coco
================================================
FILE: config/create_custom_model.sh
================================================
#!/bin/bash
NUM_CLASSES=$1
echo "
[net]
# Testing
#batch=1
#subdivisions=1
# Training
batch=16
subdivisions=1
width=416
height=416
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1
learning_rate=0.001
burn_in=1000
max_batches = 500200
policy=steps
steps=400000,450000
scales=.1,.1
[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=leaky
# Downsample
[convolutional]
batch_normalize=1
filters=64
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=32
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=128
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=64
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=64
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=256
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=512
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
######################
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=$(expr 3 \* $(expr $NUM_CLASSES \+ 5))
activation=linear
[yolo]
mask = 6,7,8
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=$NUM_CLASSES
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
[route]
layers = -4
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[upsample]
stride=2
[route]
layers = -1, 61
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=$(expr 3 \* $(expr $NUM_CLASSES \+ 5))
activation=linear
[yolo]
mask = 3,4,5
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=$NUM_CLASSES
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
[route]
layers = -4
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[upsample]
stride=2
[route]
layers = -1, 36
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=$(expr 3 \* $(expr $NUM_CLASSES \+ 5))
activation=linear
[yolo]
mask = 0,1,2
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=$NUM_CLASSES
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
" >> yolov3-custom.cfg
================================================
FILE: config/custom.data
================================================
classes= 1
train=data/custom/train.txt
valid=data/custom/valid.txt
names=data/custom/classes.names
================================================
FILE: config/yolov3-tiny.cfg
================================================
[net]
# Testing
#batch=1
#subdivisions=1
# Training
batch=64
subdivisions=2
width=416
height=416
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1
learning_rate=0.0001
burn_in=1000
max_batches = 500200
policy=steps
steps=400000,450000
scales=.1,.1
# 0
[convolutional]
batch_normalize=1
filters=16
size=3
stride=1
pad=1
activation=leaky
# 1
[maxpool]
size=2
stride=2
# 2
[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=leaky
# 3
[maxpool]
size=2
stride=2
# 4
[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky
# 5
[maxpool]
size=2
stride=2
# 6
[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky
# 7
[maxpool]
size=2
stride=2
# 8
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
# 9
[maxpool]
size=2
stride=2
# 10
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
# 11
[maxpool]
size=2
stride=1
# 12
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
###########
# 13
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
# 14
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
# 15
[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear
# 16
[yolo]
mask = 3,4,5
anchors = 10,14, 23,27, 37,58, 81,82, 135,169, 344,319
classes=80
num=6
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
# 17
[route]
layers = -4
# 18
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
# 19
[upsample]
stride=2
# 20
[route]
layers = -1, 8
# 21
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
# 22
[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear
# 23
[yolo]
mask = 1,2,3
anchors = 10,14, 23,27, 37,58, 81,82, 135,169, 344,319
classes=80
num=6
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
================================================
FILE: config/yolov3.cfg
================================================
[net]
# Testing
#batch=1
#subdivisions=1
# Training
batch=16
subdivisions=1
width=416
height=416
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1
learning_rate=0.0001
burn_in=1000
max_batches = 500200
policy=steps
steps=400000,450000
scales=.1,.1
[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=leaky
# Downsample
[convolutional]
batch_normalize=1
filters=64
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=32
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=128
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=64
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=64
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=256
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=512
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
######################
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear
[yolo]
mask = 6,7,8
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=80
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
[route]
layers = -4
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[upsample]
stride=2
[route]
layers = -1, 61
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear
[yolo]
mask = 3,4,5
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=80
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
[route]
layers = -4
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[upsample]
stride=2
[route]
layers = -1, 36
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear
[yolo]
mask = 0,1,2
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=80
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
================================================
FILE: pyproject.toml
================================================
[tool.poetry]
name = "PyTorchYolo"
version = "1.8.0"
readme = "README.md"
repository = "https://github.com/eriklindernoren/PyTorch-YOLOv3"
description = "Minimal PyTorch implementation of YOLO"
authors = ["Florian Vahl <git@flova.de>", "Erik Linder-Noren <eriklindernoren@gmail.com>"]
license = "GPL-3.0"
[tool.poetry.dependencies]
python = ">=3.8,<4.0"
torch = ">=1.10.1, < 1.13.0"
torchvision = ">=0.13.1"
matplotlib = "^3.3.3"
tensorboard = "^2.10.0"
terminaltables = "^3.1.0"
Pillow = "^9.1.0"
tqdm = "^4.64.1"
urllib3 = [
{version = "<=1.22", python = ">=3.8,<3.9"},
{version = "^1.23", python = ">=3.9"}
] # Temp pin because of crash issue
scipy = [
{version = "<=1.6", python = ">=3.8,<3.9"},
{version = "^1.9", python = ">=3.9,<4.0"}
]
imgaug = "^0.4.0"
torchsummary = "^1.5.1"
numpy = "^1.23.4"
[tool.poetry.dev-dependencies]
profilehooks = "^1.12.0"
[build-system]
requires = ["poetry-core>=1.0.0"]
build-backend = "poetry.core.masonry.api"
[tool.poetry.scripts]
yolo-detect = "pytorchyolo.detect:run"
yolo-train = "pytorchyolo.train:run"
yolo-test = "pytorchyolo.test:run"
================================================
FILE: pytorchyolo/__init__.py
================================================
================================================
FILE: pytorchyolo/detect.py
================================================
#! /usr/bin/env python3
from __future__ import division
import os
import argparse
import tqdm
import random
import numpy as np
from PIL import Image
import torch
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
from torch.autograd import Variable
from pytorchyolo.models import load_model
from pytorchyolo.utils.utils import load_classes, rescale_boxes, non_max_suppression, print_environment_info
from pytorchyolo.utils.datasets import ImageFolder
from pytorchyolo.utils.transforms import Resize, DEFAULT_TRANSFORMS
import matplotlib.pyplot as plt
import matplotlib.patches as patches
from matplotlib.ticker import NullLocator
def detect_directory(model_path, weights_path, img_path, classes, output_path,
batch_size=8, img_size=416, n_cpu=8, conf_thres=0.5, nms_thres=0.5):
"""Detects objects on all images in specified directory and saves output images with drawn detections.
:param model_path: Path to model definition file (.cfg)
:type model_path: str
:param weights_path: Path to weights or checkpoint file (.weights or .pth)
:type weights_path: str
:param img_path: Path to directory with images to inference
:type img_path: str
:param classes: List of class names
:type classes: [str]
:param output_path: Path to output directory
:type output_path: str
:param batch_size: Size of each image batch, defaults to 8
:type batch_size: int, optional
:param img_size: Size of each image dimension for yolo, defaults to 416
:type img_size: int, optional
:param n_cpu: Number of cpu threads to use during batch generation, defaults to 8
:type n_cpu: int, optional
:param conf_thres: Object confidence threshold, defaults to 0.5
:type conf_thres: float, optional
:param nms_thres: IOU threshold for non-maximum suppression, defaults to 0.5
:type nms_thres: float, optional
"""
dataloader = _create_data_loader(img_path, batch_size, img_size, n_cpu)
model = load_model(model_path, weights_path)
img_detections, imgs = detect(
model,
dataloader,
output_path,
conf_thres,
nms_thres)
_draw_and_save_output_images(
img_detections, imgs, img_size, output_path, classes)
print(f"---- Detections were saved to: '{output_path}' ----")
def detect_image(model, image, img_size=416, conf_thres=0.5, nms_thres=0.5):
"""Inferences one image with model.
:param model: Model for inference
:type model: models.Darknet
:param image: Image to inference
:type image: nd.array
:param img_size: Size of each image dimension for yolo, defaults to 416
:type img_size: int, optional
:param conf_thres: Object confidence threshold, defaults to 0.5
:type conf_thres: float, optional
:param nms_thres: IOU threshold for non-maximum suppression, defaults to 0.5
:type nms_thres: float, optional
:return: Detections on image with each detection in the format: [x1, y1, x2, y2, confidence, class]
:rtype: nd.array
"""
model.eval() # Set model to evaluation mode
# Configure input
input_img = transforms.Compose([
DEFAULT_TRANSFORMS,
Resize(img_size)])(
(image, np.zeros((1, 5))))[0].unsqueeze(0)
if torch.cuda.is_available():
input_img = input_img.to("cuda")
# Get detections
with torch.no_grad():
detections = model(input_img)
detections = non_max_suppression(detections, conf_thres, nms_thres)
detections = rescale_boxes(detections[0], img_size, image.shape[:2])
return detections.numpy()
def detect(model, dataloader, output_path, conf_thres, nms_thres):
"""Inferences images with model.
:param model: Model for inference
:type model: models.Darknet
:param dataloader: Dataloader provides the batches of images to inference
:type dataloader: DataLoader
:param output_path: Path to output directory
:type output_path: str
:param conf_thres: Object confidence threshold, defaults to 0.5
:type conf_thres: float, optional
:param nms_thres: IOU threshold for non-maximum suppression, defaults to 0.5
:type nms_thres: float, optional
:return: List of detections. The coordinates are given for the padded image that is provided by the dataloader.
Use `utils.rescale_boxes` to transform them into the desired input image coordinate system before its transformed by the dataloader),
List of input image paths
:rtype: [Tensor], [str]
"""
# Create output directory, if missing
os.makedirs(output_path, exist_ok=True)
model.eval() # Set model to evaluation mode
Tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.FloatTensor
img_detections = [] # Stores detections for each image index
imgs = [] # Stores image paths
for (img_paths, input_imgs) in tqdm.tqdm(dataloader, desc="Detecting"):
# Configure input
input_imgs = Variable(input_imgs.type(Tensor))
# Get detections
with torch.no_grad():
detections = model(input_imgs)
detections = non_max_suppression(detections, conf_thres, nms_thres)
# Store image and detections
img_detections.extend(detections)
imgs.extend(img_paths)
return img_detections, imgs
def _draw_and_save_output_images(img_detections, imgs, img_size, output_path, classes):
"""Draws detections in output images and stores them.
:param img_detections: List of detections
:type img_detections: [Tensor]
:param imgs: List of paths to image files
:type imgs: [str]
:param img_size: Size of each image dimension for yolo
:type img_size: int
:param output_path: Path of output directory
:type output_path: str
:param classes: List of class names
:type classes: [str]
"""
# Iterate through images and save plot of detections
for (image_path, detections) in zip(imgs, img_detections):
print(f"Image {image_path}:")
_draw_and_save_output_image(
image_path, detections, img_size, output_path, classes)
def _draw_and_save_output_image(image_path, detections, img_size, output_path, classes):
"""Draws detections in output image and stores this.
:param image_path: Path to input image
:type image_path: str
:param detections: List of detections on image
:type detections: [Tensor]
:param img_size: Size of each image dimension for yolo
:type img_size: int
:param output_path: Path of output directory
:type output_path: str
:param classes: List of class names
:type classes: [str]
"""
# Create plot
img = np.array(Image.open(image_path))
plt.figure()
fig, ax = plt.subplots(1)
ax.imshow(img)
# Rescale boxes to original image
detections = rescale_boxes(detections, img_size, img.shape[:2])
unique_labels = detections[:, -1].cpu().unique()
n_cls_preds = len(unique_labels)
# Bounding-box colors
cmap = plt.get_cmap("tab20b")
colors = [cmap(i) for i in np.linspace(0, 1, n_cls_preds)]
bbox_colors = random.sample(colors, n_cls_preds)
for x1, y1, x2, y2, conf, cls_pred in detections:
print(f"\t+ Label: {classes[int(cls_pred)]} | Confidence: {conf.item():0.4f}")
box_w = x2 - x1
box_h = y2 - y1
color = bbox_colors[int(np.where(unique_labels == int(cls_pred))[0])]
# Create a Rectangle patch
bbox = patches.Rectangle((x1, y1), box_w, box_h, linewidth=2, edgecolor=color, facecolor="none")
# Add the bbox to the plot
ax.add_patch(bbox)
# Add label
plt.text(
x1,
y1,
s=f"{classes[int(cls_pred)]}: {conf:.2f}",
color="white",
verticalalignment="top",
bbox={"color": color, "pad": 0})
# Save generated image with detections
plt.axis("off")
plt.gca().xaxis.set_major_locator(NullLocator())
plt.gca().yaxis.set_major_locator(NullLocator())
filename = os.path.basename(image_path).split(".")[0]
output_path = os.path.join(output_path, f"{filename}.png")
plt.savefig(output_path, bbox_inches="tight", pad_inches=0.0)
plt.close()
def _create_data_loader(img_path, batch_size, img_size, n_cpu):
"""Creates a DataLoader for inferencing.
:param img_path: Path to file containing all paths to validation images.
:type img_path: str
:param batch_size: Size of each image batch
:type batch_size: int
:param img_size: Size of each image dimension for yolo
:type img_size: int
:param n_cpu: Number of cpu threads to use during batch generation
:type n_cpu: int
:return: Returns DataLoader
:rtype: DataLoader
"""
dataset = ImageFolder(
img_path,
transform=transforms.Compose([DEFAULT_TRANSFORMS, Resize(img_size)]))
dataloader = DataLoader(
dataset,
batch_size=batch_size,
shuffle=False,
num_workers=n_cpu,
pin_memory=True)
return dataloader
def run():
print_environment_info()
parser = argparse.ArgumentParser(description="Detect objects on images.")
parser.add_argument("-m", "--model", type=str, default="config/yolov3.cfg", help="Path to model definition file (.cfg)")
parser.add_argument("-w", "--weights", type=str, default="weights/yolov3.weights", help="Path to weights or checkpoint file (.weights or .pth)")
parser.add_argument("-i", "--images", type=str, default="data/samples", help="Path to directory with images to inference")
parser.add_argument("-c", "--classes", type=str, default="data/coco.names", help="Path to classes label file (.names)")
parser.add_argument("-o", "--output", type=str, default="output", help="Path to output directory")
parser.add_argument("-b", "--batch_size", type=int, default=1, help="Size of each image batch")
parser.add_argument("--img_size", type=int, default=416, help="Size of each image dimension for yolo")
parser.add_argument("--n_cpu", type=int, default=8, help="Number of cpu threads to use during batch generation")
parser.add_argument("--conf_thres", type=float, default=0.5, help="Object confidence threshold")
parser.add_argument("--nms_thres", type=float, default=0.4, help="IOU threshold for non-maximum suppression")
args = parser.parse_args()
print(f"Command line arguments: {args}")
# Extract class names from file
classes = load_classes(args.classes) # List of class names
detect_directory(
args.model,
args.weights,
args.images,
classes,
args.output,
batch_size=args.batch_size,
img_size=args.img_size,
n_cpu=args.n_cpu,
conf_thres=args.conf_thres,
nms_thres=args.nms_thres)
if __name__ == '__main__':
run()
================================================
FILE: pytorchyolo/models.py
================================================
from __future__ import division
import os
from itertools import chain
from typing import List, Tuple
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from pytorchyolo.utils.parse_config import parse_model_config
from pytorchyolo.utils.utils import weights_init_normal
def create_modules(module_defs: List[dict]) -> Tuple[dict, nn.ModuleList]:
"""
Constructs module list of layer blocks from module configuration in module_defs
:param module_defs: List of dictionaries with module definitions
:return: Hyperparameters and pytorch module list
"""
hyperparams = module_defs.pop(0)
hyperparams.update({
'batch': int(hyperparams['batch']),
'subdivisions': int(hyperparams['subdivisions']),
'width': int(hyperparams['width']),
'height': int(hyperparams['height']),
'channels': int(hyperparams['channels']),
'optimizer': hyperparams.get('optimizer'),
'momentum': float(hyperparams['momentum']),
'decay': float(hyperparams['decay']),
'learning_rate': float(hyperparams['learning_rate']),
'burn_in': int(hyperparams['burn_in']),
'max_batches': int(hyperparams['max_batches']),
'policy': hyperparams['policy'],
'lr_steps': list(zip(map(int, hyperparams["steps"].split(",")),
map(float, hyperparams["scales"].split(","))))
})
assert hyperparams["height"] == hyperparams["width"], \
"Height and width should be equal! Non square images are padded with zeros."
output_filters = [hyperparams["channels"]]
module_list = nn.ModuleList()
for module_i, module_def in enumerate(module_defs):
modules = nn.Sequential()
if module_def["type"] == "convolutional":
bn = int(module_def["batch_normalize"])
filters = int(module_def["filters"])
kernel_size = int(module_def["size"])
pad = (kernel_size - 1) // 2
modules.add_module(
f"conv_{module_i}",
nn.Conv2d(
in_channels=output_filters[-1],
out_channels=filters,
kernel_size=kernel_size,
stride=int(module_def["stride"]),
padding=pad,
bias=not bn,
),
)
if bn:
modules.add_module(f"batch_norm_{module_i}",
nn.BatchNorm2d(filters, momentum=0.1, eps=1e-5))
if module_def["activation"] == "leaky":
modules.add_module(f"leaky_{module_i}", nn.LeakyReLU(0.1))
elif module_def["activation"] == "mish":
modules.add_module(f"mish_{module_i}", nn.Mish())
elif module_def["activation"] == "logistic":
modules.add_module(f"sigmoid_{module_i}", nn.Sigmoid())
elif module_def["activation"] == "swish":
modules.add_module(f"swish_{module_i}", nn.SiLU())
elif module_def["type"] == "maxpool":
kernel_size = int(module_def["size"])
stride = int(module_def["stride"])
if kernel_size == 2 and stride == 1:
modules.add_module(f"_debug_padding_{module_i}", nn.ZeroPad2d((0, 1, 0, 1)))
maxpool = nn.MaxPool2d(kernel_size=kernel_size, stride=stride,
padding=int((kernel_size - 1) // 2))
modules.add_module(f"maxpool_{module_i}", maxpool)
elif module_def["type"] == "upsample":
upsample = Upsample(scale_factor=int(module_def["stride"]), mode="nearest")
modules.add_module(f"upsample_{module_i}", upsample)
elif module_def["type"] == "route":
layers = [int(x) for x in module_def["layers"].split(",")]
filters = sum([output_filters[1:][i] for i in layers]) // int(module_def.get("groups", 1))
modules.add_module(f"route_{module_i}", nn.Sequential())
elif module_def["type"] == "shortcut":
filters = output_filters[1:][int(module_def["from"])]
modules.add_module(f"shortcut_{module_i}", nn.Sequential())
elif module_def["type"] == "yolo":
anchor_idxs = [int(x) for x in module_def["mask"].split(",")]
# Extract anchors
anchors = [int(x) for x in module_def["anchors"].split(",")]
anchors = [(anchors[i], anchors[i + 1]) for i in range(0, len(anchors), 2)]
anchors = [anchors[i] for i in anchor_idxs]
num_classes = int(module_def["classes"])
new_coords = bool(module_def.get("new_coords", False))
# Define detection layer
yolo_layer = YOLOLayer(anchors, num_classes, new_coords)
modules.add_module(f"yolo_{module_i}", yolo_layer)
# Register module list and number of output filters
module_list.append(modules)
output_filters.append(filters)
return hyperparams, module_list
class Upsample(nn.Module):
""" nn.Upsample is deprecated """
def __init__(self, scale_factor, mode: str = "nearest"):
super(Upsample, self).__init__()
self.scale_factor = scale_factor
self.mode = mode
def forward(self, x):
x = F.interpolate(x, scale_factor=self.scale_factor, mode=self.mode)
return x
class YOLOLayer(nn.Module):
"""Detection layer"""
def __init__(self, anchors: List[Tuple[int, int]], num_classes: int, new_coords: bool):
"""
Create a YOLO layer
:param anchors: List of anchors
:param num_classes: Number of classes
:param new_coords: Whether to use the new coordinate format from YOLO V7
"""
super(YOLOLayer, self).__init__()
self.num_anchors = len(anchors)
self.num_classes = num_classes
self.new_coords = new_coords
self.mse_loss = nn.MSELoss()
self.bce_loss = nn.BCELoss()
self.no = num_classes + 5 # number of outputs per anchor
self.grid = torch.zeros(1) # TODO
anchors = torch.tensor(list(chain(*anchors))).float().view(-1, 2)
self.register_buffer('anchors', anchors)
self.register_buffer(
'anchor_grid', anchors.clone().view(1, -1, 1, 1, 2))
self.stride = None
def forward(self, x: torch.Tensor, img_size: int) -> torch.Tensor:
"""
Forward pass of the YOLO layer
:param x: Input tensor
:param img_size: Size of the input image
"""
stride = img_size // x.size(2)
self.stride = stride
bs, _, ny, nx = x.shape # x(bs,255,20,20) to x(bs,3,20,20,85)
x = x.view(bs, self.num_anchors, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
if not self.training: # inference
if self.grid.shape[2:4] != x.shape[2:4]:
self.grid = self._make_grid(nx, ny).to(x.device)
if self.new_coords:
x[..., 0:2] = (x[..., 0:2] + self.grid) * stride # xy
x[..., 2:4] = x[..., 2:4] ** 2 * (4 * self.anchor_grid) # wh
else:
x[..., 0:2] = (x[..., 0:2].sigmoid() + self.grid) * stride # xy
x[..., 2:4] = torch.exp(x[..., 2:4]) * self.anchor_grid # wh
x[..., 4:] = x[..., 4:].sigmoid() # conf, cls
x = x.view(bs, -1, self.no)
return x
@staticmethod
def _make_grid(nx: int = 20, ny: int = 20) -> torch.Tensor:
"""
Create a grid of (x, y) coordinates
:param nx: Number of x coordinates
:param ny: Number of y coordinates
"""
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)], indexing='ij')
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
class Darknet(nn.Module):
"""YOLOv3 object detection model"""
def __init__(self, config_path):
super(Darknet, self).__init__()
self.module_defs = parse_model_config(config_path)
self.hyperparams, self.module_list = create_modules(self.module_defs)
self.yolo_layers = [layer[0]
for layer in self.module_list if isinstance(layer[0], YOLOLayer)]
self.seen = 0
self.header_info = np.array([0, 0, 0, self.seen, 0], dtype=np.int32)
def forward(self, x):
img_size = x.size(2)
layer_outputs, yolo_outputs = [], []
for i, (module_def, module) in enumerate(zip(self.module_defs, self.module_list)):
if module_def["type"] in ["convolutional", "upsample", "maxpool"]:
x = module(x)
elif module_def["type"] == "route":
combined_outputs = torch.cat([layer_outputs[int(layer_i)] for layer_i in module_def["layers"].split(",")], 1)
group_size = combined_outputs.shape[1] // int(module_def.get("groups", 1))
group_id = int(module_def.get("group_id", 0))
x = combined_outputs[:, group_size * group_id : group_size * (group_id + 1)] # Slice groupings used by yolo v4
elif module_def["type"] == "shortcut":
layer_i = int(module_def["from"])
x = layer_outputs[-1] + layer_outputs[layer_i]
elif module_def["type"] == "yolo":
x = module[0](x, img_size)
yolo_outputs.append(x)
layer_outputs.append(x)
return yolo_outputs if self.training else torch.cat(yolo_outputs, 1)
def load_darknet_weights(self, weights_path):
"""Parses and loads the weights stored in 'weights_path'"""
# Open the weights file
with open(weights_path, "rb") as f:
# First five are header values
header = np.fromfile(f, dtype=np.int32, count=5)
self.header_info = header # Needed to write header when saving weights
self.seen = header[3] # number of images seen during training
weights = np.fromfile(f, dtype=np.float32) # The rest are weights
# Establish cutoff for loading backbone weights
cutoff = None
# If the weights file has a cutoff, we can find out about it by looking at the filename
# examples: darknet53.conv.74 -> cutoff is 74
filename = os.path.basename(weights_path)
if ".conv." in filename:
try:
cutoff = int(filename.split(".")[-1]) # use last part of filename
except ValueError:
pass
ptr = 0
for i, (module_def, module) in enumerate(zip(self.module_defs, self.module_list)):
if i == cutoff:
break
if module_def["type"] == "convolutional":
conv_layer = module[0]
if module_def["batch_normalize"]:
# Load BN bias, weights, running mean and running variance
bn_layer = module[1]
num_b = bn_layer.bias.numel() # Number of biases
# Bias
bn_b = torch.from_numpy(
weights[ptr: ptr + num_b]).view_as(bn_layer.bias)
bn_layer.bias.data.copy_(bn_b)
ptr += num_b
# Weight
bn_w = torch.from_numpy(
weights[ptr: ptr + num_b]).view_as(bn_layer.weight)
bn_layer.weight.data.copy_(bn_w)
ptr += num_b
# Running Mean
bn_rm = torch.from_numpy(
weights[ptr: ptr + num_b]).view_as(bn_layer.running_mean)
bn_layer.running_mean.data.copy_(bn_rm)
ptr += num_b
# Running Var
bn_rv = torch.from_numpy(
weights[ptr: ptr + num_b]).view_as(bn_layer.running_var)
bn_layer.running_var.data.copy_(bn_rv)
ptr += num_b
else:
# Load conv. bias
num_b = conv_layer.bias.numel()
conv_b = torch.from_numpy(
weights[ptr: ptr + num_b]).view_as(conv_layer.bias)
conv_layer.bias.data.copy_(conv_b)
ptr += num_b
# Load conv. weights
num_w = conv_layer.weight.numel()
conv_w = torch.from_numpy(
weights[ptr: ptr + num_w]).view_as(conv_layer.weight)
conv_layer.weight.data.copy_(conv_w)
ptr += num_w
def save_darknet_weights(self, path, cutoff=-1):
"""
@:param path - path of the new weights file
@:param cutoff - save layers between 0 and cutoff (cutoff = -1 -> all are saved)
"""
fp = open(path, "wb")
self.header_info[3] = self.seen
self.header_info.tofile(fp)
# Iterate through layers
for i, (module_def, module) in enumerate(zip(self.module_defs[:cutoff], self.module_list[:cutoff])):
if module_def["type"] == "convolutional":
conv_layer = module[0]
# If batch norm, load bn first
if module_def["batch_normalize"]:
bn_layer = module[1]
bn_layer.bias.data.cpu().numpy().tofile(fp)
bn_layer.weight.data.cpu().numpy().tofile(fp)
bn_layer.running_mean.data.cpu().numpy().tofile(fp)
bn_layer.running_var.data.cpu().numpy().tofile(fp)
# Load conv bias
else:
conv_layer.bias.data.cpu().numpy().tofile(fp)
# Load conv weights
conv_layer.weight.data.cpu().numpy().tofile(fp)
fp.close()
def load_model(model_path, weights_path=None):
"""Loads the yolo model from file.
:param model_path: Path to model definition file (.cfg)
:type model_path: str
:param weights_path: Path to weights or checkpoint file (.weights or .pth)
:type weights_path: str
:return: Returns model
:rtype: Darknet
"""
device = torch.device("cuda" if torch.cuda.is_available()
else "cpu") # Select device for inference
model = Darknet(model_path).to(device)
model.apply(weights_init_normal)
# If pretrained weights are specified, start from checkpoint or weight file
if weights_path:
if weights_path.endswith(".pth"):
# Load checkpoint weights
model.load_state_dict(torch.load(weights_path, map_location=device))
else:
# Load darknet weights
model.load_darknet_weights(weights_path)
return model
================================================
FILE: pytorchyolo/test.py
================================================
#! /usr/bin/env python3
from __future__ import division
import argparse
import tqdm
import numpy as np
from terminaltables import AsciiTable
import torch
from torch.utils.data import DataLoader
from torch.autograd import Variable
from pytorchyolo.models import load_model
from pytorchyolo.utils.utils import load_classes, ap_per_class, get_batch_statistics, non_max_suppression, to_cpu, xywh2xyxy, print_environment_info
from pytorchyolo.utils.datasets import ListDataset
from pytorchyolo.utils.transforms import DEFAULT_TRANSFORMS
from pytorchyolo.utils.parse_config import parse_data_config
def evaluate_model_file(model_path, weights_path, img_path, class_names, batch_size=8, img_size=416,
n_cpu=8, iou_thres=0.5, conf_thres=0.5, nms_thres=0.5, verbose=True):
"""Evaluate model on validation dataset.
:param model_path: Path to model definition file (.cfg)
:type model_path: str
:param weights_path: Path to weights or checkpoint file (.weights or .pth)
:type weights_path: str
:param img_path: Path to file containing all paths to validation images.
:type img_path: str
:param class_names: List of class names
:type class_names: [str]
:param batch_size: Size of each image batch, defaults to 8
:type batch_size: int, optional
:param img_size: Size of each image dimension for yolo, defaults to 416
:type img_size: int, optional
:param n_cpu: Number of cpu threads to use during batch generation, defaults to 8
:type n_cpu: int, optional
:param iou_thres: IOU threshold required to qualify as detected, defaults to 0.5
:type iou_thres: float, optional
:param conf_thres: Object confidence threshold, defaults to 0.5
:type conf_thres: float, optional
:param nms_thres: IOU threshold for non-maximum suppression, defaults to 0.5
:type nms_thres: float, optional
:param verbose: If True, prints stats of model, defaults to True
:type verbose: bool, optional
:return: Returns precision, recall, AP, f1, ap_class
"""
dataloader = _create_validation_data_loader(
img_path, batch_size, img_size, n_cpu)
model = load_model(model_path, weights_path)
metrics_output = _evaluate(
model,
dataloader,
class_names,
img_size,
iou_thres,
conf_thres,
nms_thres,
verbose)
return metrics_output
def print_eval_stats(metrics_output, class_names, verbose):
if metrics_output is not None:
precision, recall, AP, f1, ap_class = metrics_output
if verbose:
# Prints class AP and mean AP
ap_table = [["Index", "Class", "AP"]]
for i, c in enumerate(ap_class):
ap_table += [[c, class_names[c], "%.5f" % AP[i]]]
print(AsciiTable(ap_table).table)
print(f"---- mAP {AP.mean():.5f} ----")
else:
print("---- mAP not measured (no detections found by model) ----")
def _evaluate(model, dataloader, class_names, img_size, iou_thres, conf_thres, nms_thres, verbose):
"""Evaluate model on validation dataset.
:param model: Model to evaluate
:type model: models.Darknet
:param dataloader: Dataloader provides the batches of images with targets
:type dataloader: DataLoader
:param class_names: List of class names
:type class_names: [str]
:param img_size: Size of each image dimension for yolo
:type img_size: int
:param iou_thres: IOU threshold required to qualify as detected
:type iou_thres: float
:param conf_thres: Object confidence threshold
:type conf_thres: float
:param nms_thres: IOU threshold for non-maximum suppression
:type nms_thres: float
:param verbose: If True, prints stats of model
:type verbose: bool
:return: Returns precision, recall, AP, f1, ap_class
"""
model.eval() # Set model to evaluation mode
Tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.FloatTensor
labels = []
sample_metrics = [] # List of tuples (TP, confs, pred)
for _, imgs, targets in tqdm.tqdm(dataloader, desc="Validating"):
# Extract labels
labels += targets[:, 1].tolist()
# Rescale target
targets[:, 2:] = xywh2xyxy(targets[:, 2:])
targets[:, 2:] *= img_size
imgs = Variable(imgs.type(Tensor), requires_grad=False)
with torch.no_grad():
outputs = model(imgs)
outputs = non_max_suppression(outputs, conf_thres=conf_thres, iou_thres=nms_thres)
sample_metrics += get_batch_statistics(outputs, targets, iou_threshold=iou_thres)
if len(sample_metrics) == 0: # No detections over whole validation set.
print("---- No detections over whole validation set ----")
return None
# Concatenate sample statistics
true_positives, pred_scores, pred_labels = [
np.concatenate(x, 0) for x in list(zip(*sample_metrics))]
metrics_output = ap_per_class(
true_positives, pred_scores, pred_labels, labels)
print_eval_stats(metrics_output, class_names, verbose)
return metrics_output
def _create_validation_data_loader(img_path, batch_size, img_size, n_cpu):
"""
Creates a DataLoader for validation.
:param img_path: Path to file containing all paths to validation images.
:type img_path: str
:param batch_size: Size of each image batch
:type batch_size: int
:param img_size: Size of each image dimension for yolo
:type img_size: int
:param n_cpu: Number of cpu threads to use during batch generation
:type n_cpu: int
:return: Returns DataLoader
:rtype: DataLoader
"""
dataset = ListDataset(img_path, img_size=img_size, multiscale=False, transform=DEFAULT_TRANSFORMS)
dataloader = DataLoader(
dataset,
batch_size=batch_size,
shuffle=False,
num_workers=n_cpu,
pin_memory=True,
collate_fn=dataset.collate_fn)
return dataloader
def run():
print_environment_info()
parser = argparse.ArgumentParser(description="Evaluate validation data.")
parser.add_argument("-m", "--model", type=str, default="config/yolov3.cfg", help="Path to model definition file (.cfg)")
parser.add_argument("-w", "--weights", type=str, default="weights/yolov3.weights", help="Path to weights or checkpoint file (.weights or .pth)")
parser.add_argument("-d", "--data", type=str, default="config/coco.data", help="Path to data config file (.data)")
parser.add_argument("-b", "--batch_size", type=int, default=8, help="Size of each image batch")
parser.add_argument("-v", "--verbose", action='store_true', help="Makes the validation more verbose")
parser.add_argument("--img_size", type=int, default=416, help="Size of each image dimension for yolo")
parser.add_argument("--n_cpu", type=int, default=8, help="Number of cpu threads to use during batch generation")
parser.add_argument("--iou_thres", type=float, default=0.5, help="IOU threshold required to qualify as detected")
parser.add_argument("--conf_thres", type=float, default=0.01, help="Object confidence threshold")
parser.add_argument("--nms_thres", type=float, default=0.4, help="IOU threshold for non-maximum suppression")
args = parser.parse_args()
print(f"Command line arguments: {args}")
# Load configuration from data file
data_config = parse_data_config(args.data)
# Path to file containing all images for validation
valid_path = data_config["valid"]
class_names = load_classes(data_config["names"]) # List of class names
precision, recall, AP, f1, ap_class = evaluate_model_file(
args.model,
args.weights,
valid_path,
class_names,
batch_size=args.batch_size,
img_size=args.img_size,
n_cpu=args.n_cpu,
iou_thres=args.iou_thres,
conf_thres=args.conf_thres,
nms_thres=args.nms_thres,
verbose=True)
if __name__ == "__main__":
run()
================================================
FILE: pytorchyolo/train.py
================================================
#! /usr/bin/env python3
from __future__ import division
import os
import argparse
import tqdm
import torch
from torch.utils.data import DataLoader
import torch.optim as optim
from pytorchyolo.models import load_model
from pytorchyolo.utils.logger import Logger
from pytorchyolo.utils.utils import to_cpu, load_classes, print_environment_info, provide_determinism, worker_seed_set
from pytorchyolo.utils.datasets import ListDataset
from pytorchyolo.utils.augmentations import AUGMENTATION_TRANSFORMS
#from pytorchyolo.utils.transforms import DEFAULT_TRANSFORMS
from pytorchyolo.utils.parse_config import parse_data_config
from pytorchyolo.utils.loss import compute_loss
from pytorchyolo.test import _evaluate, _create_validation_data_loader
from terminaltables import AsciiTable
from torchsummary import summary
def _create_data_loader(img_path, batch_size, img_size, n_cpu, multiscale_training=False):
"""Creates a DataLoader for training.
:param img_path: Path to file containing all paths to training images.
:type img_path: str
:param batch_size: Size of each image batch
:type batch_size: int
:param img_size: Size of each image dimension for yolo
:type img_size: int
:param n_cpu: Number of cpu threads to use during batch generation
:type n_cpu: int
:param multiscale_training: Scale images to different sizes randomly
:type multiscale_training: bool
:return: Returns DataLoader
:rtype: DataLoader
"""
dataset = ListDataset(
img_path,
img_size=img_size,
multiscale=multiscale_training,
transform=AUGMENTATION_TRANSFORMS)
dataloader = DataLoader(
dataset,
batch_size=batch_size,
shuffle=True,
num_workers=n_cpu,
pin_memory=True,
collate_fn=dataset.collate_fn,
worker_init_fn=worker_seed_set)
return dataloader
def run():
print_environment_info()
parser = argparse.ArgumentParser(description="Trains the YOLO model.")
parser.add_argument("-m", "--model", type=str, default="config/yolov3.cfg", help="Path to model definition file (.cfg)")
parser.add_argument("-d", "--data", type=str, default="config/coco.data", help="Path to data config file (.data)")
parser.add_argument("-e", "--epochs", type=int, default=300, help="Number of epochs")
parser.add_argument("-v", "--verbose", action='store_true', help="Makes the training more verbose")
parser.add_argument("--n_cpu", type=int, default=8, help="Number of cpu threads to use during batch generation")
parser.add_argument("--pretrained_weights", type=str, help="Path to checkpoint file (.weights or .pth). Starts training from checkpoint model")
parser.add_argument("--checkpoint_interval", type=int, default=1, help="Interval of epochs between saving model weights")
parser.add_argument("--evaluation_interval", type=int, default=1, help="Interval of epochs between evaluations on validation set")
parser.add_argument("--multiscale_training", action="store_true", help="Allow multi-scale training")
parser.add_argument("--iou_thres", type=float, default=0.5, help="Evaluation: IOU threshold required to qualify as detected")
parser.add_argument("--conf_thres", type=float, default=0.1, help="Evaluation: Object confidence threshold")
parser.add_argument("--nms_thres", type=float, default=0.5, help="Evaluation: IOU threshold for non-maximum suppression")
parser.add_argument("--logdir", type=str, default="logs", help="Directory for training log files (e.g. for TensorBoard)")
parser.add_argument("--seed", type=int, default=-1, help="Makes results reproducable. Set -1 to disable.")
args = parser.parse_args()
print(f"Command line arguments: {args}")
if args.seed != -1:
provide_determinism(args.seed)
logger = Logger(args.logdir) # Tensorboard logger
# Create output directories if missing
os.makedirs("output", exist_ok=True)
os.makedirs("checkpoints", exist_ok=True)
# Get data configuration
data_config = parse_data_config(args.data)
train_path = data_config["train"]
valid_path = data_config["valid"]
class_names = load_classes(data_config["names"])
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# ############
# Create model
# ############
model = load_model(args.model, args.pretrained_weights)
# Print model
if args.verbose:
summary(model, input_size=(3, model.hyperparams['height'], model.hyperparams['height']))
mini_batch_size = model.hyperparams['batch'] // model.hyperparams['subdivisions']
# #################
# Create Dataloader
# #################
# Load training dataloader
dataloader = _create_data_loader(
train_path,
mini_batch_size,
model.hyperparams['height'],
args.n_cpu,
args.multiscale_training)
# Load validation dataloader
validation_dataloader = _create_validation_data_loader(
valid_path,
mini_batch_size,
model.hyperparams['height'],
args.n_cpu)
# ################
# Create optimizer
# ################
params = [p for p in model.parameters() if p.requires_grad]
if (model.hyperparams['optimizer'] in [None, "adam"]):
optimizer = optim.Adam(
params,
lr=model.hyperparams['learning_rate'],
weight_decay=model.hyperparams['decay'],
)
elif (model.hyperparams['optimizer'] == "sgd"):
optimizer = optim.SGD(
params,
lr=model.hyperparams['learning_rate'],
weight_decay=model.hyperparams['decay'],
momentum=model.hyperparams['momentum'])
else:
print("Unknown optimizer. Please choose between (adam, sgd).")
# skip epoch zero, because then the calculations for when to evaluate/checkpoint makes more intuitive sense
# e.g. when you stop after 30 epochs and evaluate every 10 epochs then the evaluations happen after: 10,20,30
# instead of: 0, 10, 20
for epoch in range(1, args.epochs+1):
print("\n---- Training Model ----")
model.train() # Set model to training mode
for batch_i, (_, imgs, targets) in enumerate(tqdm.tqdm(dataloader, desc=f"Training Epoch {epoch}")):
batches_done = len(dataloader) * epoch + batch_i
imgs = imgs.to(device, non_blocking=True)
targets = targets.to(device)
outputs = model(imgs)
loss, loss_components = compute_loss(outputs, targets, model)
loss.backward()
###############
# Run optimizer
###############
if batches_done % model.hyperparams['subdivisions'] == 0:
# Adapt learning rate
# Get learning rate defined in cfg
lr = model.hyperparams['learning_rate']
if batches_done < model.hyperparams['burn_in']:
# Burn in
lr *= (batches_done / model.hyperparams['burn_in'])
else:
# Set and parse the learning rate to the steps defined in the cfg
for threshold, value in model.hyperparams['lr_steps']:
if batches_done > threshold:
lr *= value
# Log the learning rate
logger.scalar_summary("train/learning_rate", lr, batches_done)
# Set learning rate
for g in optimizer.param_groups:
g['lr'] = lr
# Run optimizer
optimizer.step()
# Reset gradients
optimizer.zero_grad()
# ############
# Log progress
# ############
if args.verbose:
print(AsciiTable(
[
["Type", "Value"],
["IoU loss", float(loss_components[0])],
["Object loss", float(loss_components[1])],
["Class loss", float(loss_components[2])],
["Loss", float(loss_components[3])],
["Batch loss", to_cpu(loss).item()],
]).table)
# Tensorboard logging
tensorboard_log = [
("train/iou_loss", float(loss_components[0])),
("train/obj_loss", float(loss_components[1])),
("train/class_loss", float(loss_components[2])),
("train/loss", to_cpu(loss).item())]
logger.list_of_scalars_summary(tensorboard_log, batches_done)
model.seen += imgs.size(0)
# #############
# Save progress
# #############
# Save model to checkpoint file
if epoch % args.checkpoint_interval == 0:
checkpoint_path = f"checkpoints/yolov3_ckpt_{epoch}.pth"
print(f"---- Saving checkpoint to: '{checkpoint_path}' ----")
torch.save(model.state_dict(), checkpoint_path)
# ########
# Evaluate
# ########
if epoch % args.evaluation_interval == 0:
print("\n---- Evaluating Model ----")
# Evaluate the model on the validation set
metrics_output = _evaluate(
model,
validation_dataloader,
class_names,
img_size=model.hyperparams['height'],
iou_thres=args.iou_thres,
conf_thres=args.conf_thres,
nms_thres=args.nms_thres,
verbose=args.verbose
)
if metrics_output is not None:
precision, recall, AP, f1, ap_class = metrics_output
evaluation_metrics = [
("validation/precision", precision.mean()),
("validation/recall", recall.mean()),
("validation/mAP", AP.mean()),
("validation/f1", f1.mean())]
logger.list_of_scalars_summary(evaluation_metrics, epoch)
if __name__ == "__main__":
run()
================================================
FILE: pytorchyolo/utils/__init__.py
================================================
================================================
FILE: pytorchyolo/utils/augmentations.py
================================================
import imgaug.augmenters as iaa
from torchvision import transforms
from pytorchyolo.utils.transforms import ToTensor, PadSquare, RelativeLabels, AbsoluteLabels, ImgAug
class DefaultAug(ImgAug):
def __init__(self, ):
self.augmentations = iaa.Sequential([
iaa.Sharpen((0.0, 0.1)),
iaa.Affine(rotate=(-0, 0), translate_percent=(-0.1, 0.1), scale=(0.8, 1.5)),
iaa.AddToBrightness((-60, 40)),
iaa.AddToHue((-10, 10)),
iaa.Fliplr(0.5),
])
class StrongAug(ImgAug):
def __init__(self, ):
self.augmentations = iaa.Sequential([
iaa.Dropout([0.0, 0.01]),
iaa.Sharpen((0.0, 0.1)),
iaa.Affine(rotate=(-10, 10), translate_percent=(-0.1, 0.1), scale=(0.8, 1.5)),
iaa.AddToBrightness((-60, 40)),
iaa.AddToHue((-20, 20)),
iaa.Fliplr(0.5),
])
AUGMENTATION_TRANSFORMS = transforms.Compose([
AbsoluteLabels(),
DefaultAug(),
PadSquare(),
RelativeLabels(),
ToTensor(),
])
================================================
FILE: pytorchyolo/utils/datasets.py
================================================
from torch.utils.data import Dataset
import torch.nn.functional as F
import torch
import glob
import random
import os
import warnings
import numpy as np
from PIL import Image
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
def pad_to_square(img, pad_value):
c, h, w = img.shape
dim_diff = np.abs(h - w)
# (upper / left) padding and (lower / right) padding
pad1, pad2 = dim_diff // 2, dim_diff - dim_diff // 2
# Determine padding
pad = (0, 0, pad1, pad2) if h <= w else (pad1, pad2, 0, 0)
# Add padding
img = F.pad(img, pad, "constant", value=pad_value)
return img, pad
def resize(image, size):
image = F.interpolate(image.unsqueeze(0), size=size, mode="nearest").squeeze(0)
return image
class ImageFolder(Dataset):
def __init__(self, folder_path, transform=None):
self.files = sorted(glob.glob("%s/*.*" % folder_path))
self.transform = transform
def __getitem__(self, index):
img_path = self.files[index % len(self.files)]
img = np.array(
Image.open(img_path).convert('RGB'),
dtype=np.uint8)
# Label Placeholder
boxes = np.zeros((1, 5))
# Apply transforms
if self.transform:
img, _ = self.transform((img, boxes))
return img_path, img
def __len__(self):
return len(self.files)
class ListDataset(Dataset):
def __init__(self, list_path, img_size=416, multiscale=True, transform=None):
with open(list_path, "r") as file:
self.img_files = file.readlines()
self.label_files = []
for path in self.img_files:
image_dir = os.path.dirname(path)
label_dir = "labels".join(image_dir.rsplit("images", 1))
assert label_dir != image_dir, \
f"Image path must contain a folder named 'images'! \n'{image_dir}'"
label_file = os.path.join(label_dir, os.path.basename(path))
label_file = os.path.splitext(label_file)[0] + '.txt'
self.label_files.append(label_file)
self.img_size = img_size
self.max_objects = 100
self.multiscale = multiscale
self.min_size = self.img_size - 3 * 32
self.max_size = self.img_size + 3 * 32
self.batch_count = 0
self.transform = transform
def __getitem__(self, index):
# ---------
# Image
# ---------
try:
img_path = self.img_files[index % len(self.img_files)].rstrip()
img = np.array(Image.open(img_path).convert('RGB'), dtype=np.uint8)
except Exception:
print(f"Could not read image '{img_path}'.")
return
# ---------
# Label
# ---------
try:
label_path = self.label_files[index % len(self.img_files)].rstrip()
# Ignore warning if file is empty
with warnings.catch_warnings():
warnings.simplefilter("ignore")
boxes = np.loadtxt(label_path).reshape(-1, 5)
except Exception:
print(f"Could not read label '{label_path}'.")
return
# -----------
# Transform
# -----------
if self.transform:
try:
img, bb_targets = self.transform((img, boxes))
except Exception:
print("Could not apply transform.")
return
return img_path, img, bb_targets
def collate_fn(self, batch):
self.batch_count += 1
# Drop invalid images
batch = [data for data in batch if data is not None]
paths, imgs, bb_targets = list(zip(*batch))
# Selects new image size every tenth batch
if self.multiscale and self.batch_count % 10 == 0:
self.img_size = random.choice(
range(self.min_size, self.max_size + 1, 32))
# Resize images to input shape
imgs = torch.stack([resize(img, self.img_size) for img in imgs])
# Add sample index to targets
for i, boxes in enumerate(bb_targets):
boxes[:, 0] = i
bb_targets = torch.cat(bb_targets, 0)
return paths, imgs, bb_targets
def __len__(self):
return len(self.img_files)
================================================
FILE: pytorchyolo/utils/logger.py
================================================
import os
import datetime
from torch.utils.tensorboard import SummaryWriter
class Logger(object):
def __init__(self, log_dir, log_hist=True):
"""Create a summary writer logging to log_dir."""
if log_hist: # Check a new folder for each log should be dreated
log_dir = os.path.join(
log_dir,
datetime.datetime.now().strftime("%Y_%m_%d__%H_%M_%S"))
self.writer = SummaryWriter(log_dir)
def scalar_summary(self, tag, value, step):
"""Log a scalar variable."""
self.writer.add_scalar(tag, value, step)
def list_of_scalars_summary(self, tag_value_pairs, step):
"""Log scalar variables."""
for tag, value in tag_value_pairs:
self.writer.add_scalar(tag, value, step)
================================================
FILE: pytorchyolo/utils/loss.py
================================================
import math
import torch
import torch.nn as nn
from .utils import to_cpu
# This new loss function is based on https://github.com/ultralytics/yolov3/blob/master/utils/loss.py
def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-9):
# Returns the IoU of box1 to box2. box1 is 4, box2 is nx4
box2 = box2.T
# Get the coordinates of bounding boxes
if x1y1x2y2: # x1, y1, x2, y2 = box1
b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]
b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]
else: # transform from xywh to xyxy
b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2
b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2
b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2
b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2
# Intersection area
inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \
(torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)
# Union Area
w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps
union = w1 * h1 + w2 * h2 - inter + eps
iou = inter / union
if GIoU or DIoU or CIoU:
# convex (smallest enclosing box) width
cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1)
ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1) # convex height
if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared
rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 +
(b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center distance squared
if DIoU:
return iou - rho2 / c2 # DIoU
elif CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
v = (4 / math.pi ** 2) * \
torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2)
with torch.no_grad():
alpha = v / ((1 + eps) - iou + v)
return iou - (rho2 / c2 + v * alpha) # CIoU
else: # GIoU https://arxiv.org/pdf/1902.09630.pdf
c_area = cw * ch + eps # convex area
return iou - (c_area - union) / c_area # GIoU
else:
return iou # IoU
def compute_loss(predictions, targets, model):
# Check which device was used
device = targets.device
# Add placeholder varables for the different losses
lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
# Build yolo targets
tcls, tbox, indices, anchors = build_targets(predictions, targets, model) # targets
# Define different loss functions classification
BCEcls = nn.BCEWithLogitsLoss(
pos_weight=torch.tensor([1.0], device=device))
BCEobj = nn.BCEWithLogitsLoss(
pos_weight=torch.tensor([1.0], device=device))
# Calculate losses for each yolo layer
for layer_index, layer_predictions in enumerate(predictions):
# Get image ids, anchors, grid index i and j for each target in the current yolo layer
b, anchor, grid_j, grid_i = indices[layer_index]
# Build empty object target tensor with the same shape as the object prediction
tobj = torch.zeros_like(layer_predictions[..., 0], device=device) # target obj
# Get the number of targets for this layer.
# Each target is a label box with some scaling and the association of an anchor box.
# Label boxes may be associated to 0 or multiple anchors. So they are multiple times or not at all in the targets.
num_targets = b.shape[0]
# Check if there are targets for this batch
if num_targets:
# Load the corresponding values from the predictions for each of the targets
ps = layer_predictions[b, anchor, grid_j, grid_i]
# Regression of the box
# Apply sigmoid to xy offset predictions in each cell that has a target
pxy = ps[:, :2].sigmoid()
# Apply exponent to wh predictions and multiply with the anchor box that matched best with the label for each cell that has a target
pwh = torch.exp(ps[:, 2:4]) * anchors[layer_index]
# Build box out of xy and wh
pbox = torch.cat((pxy, pwh), 1)
# Calculate CIoU or GIoU for each target with the predicted box for its cell + anchor
iou = bbox_iou(pbox.T, tbox[layer_index], x1y1x2y2=False, CIoU=True)
# We want to minimize our loss so we and the best possible IoU is 1 so we take 1 - IoU and reduce it with a mean
lbox += (1.0 - iou).mean() # iou loss
# Classification of the objectness
# Fill our empty object target tensor with the IoU we just calculated for each target at the targets position
tobj[b, anchor, grid_j, grid_i] = iou.detach().clamp(0).type(tobj.dtype) # Use cells with iou > 0 as object targets
# Classification of the class
# Check if we need to do a classification (number of classes > 1)
if ps.size(1) - 5 > 1:
# Hot one class encoding
t = torch.zeros_like(ps[:, 5:], device=device) # targets
t[range(num_targets), tcls[layer_index]] = 1
# Use the tensor to calculate the BCE loss
lcls += BCEcls(ps[:, 5:], t) # BCE
# Classification of the objectness the sequel
# Calculate the BCE loss between the on the fly generated target and the network prediction
lobj += BCEobj(layer_predictions[..., 4], tobj) # obj loss
lbox *= 0.05
lobj *= 1.0
lcls *= 0.5
# Merge losses
loss = lbox + lobj + lcls
return loss, to_cpu(torch.cat((lbox, lobj, lcls, loss)))
def build_targets(p, targets, model):
# Build targets for compute_loss(), input targets(image,class,x,y,w,h)
na, nt = 3, targets.shape[0] # number of anchors, targets #TODO
tcls, tbox, indices, anch = [], [], [], []
gain = torch.ones(7, device=targets.device) # normalized to gridspace gain
# Make a tensor that iterates 0-2 for 3 anchors and repeat that as many times as we have target boxes
ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt)
# Copy target boxes anchor size times and append an anchor index to each copy the anchor index is also expressed by the new first dimension
targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2)
for i, yolo_layer in enumerate(model.yolo_layers):
# Scale anchors by the yolo grid cell size so that an anchor with the size of the cell would result in 1
anchors = yolo_layer.anchors / yolo_layer.stride
# Add the number of yolo cells in this layer the gain tensor
# The gain tensor matches the collums of our targets (img id, class, x, y, w, h, anchor id)
gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain
# Scale targets by the number of yolo layer cells, they are now in the yolo cell coordinate system
t = targets * gain
# Check if we have targets
if nt:
# Calculate ration between anchor and target box for both width and height
r = t[:, :, 4:6] / anchors[:, None]
# Select the ratios that have the highest divergence in any axis and check if the ratio is less than 4
j = torch.max(r, 1. / r).max(2)[0] < 4 # compare #TODO
# Only use targets that have the correct ratios for their anchors
# That means we only keep ones that have a matching anchor and we loose the anchor dimension
# The anchor id is still saved in the 7th value of each target
t = t[j]
else:
t = targets[0]
# Extract image id in batch and class id
b, c = t[:, :2].long().T
# We isolate the target cell associations.
# x, y, w, h are allready in the cell coordinate system meaning an x = 1.2 would be 1.2 times cellwidth
gxy = t[:, 2:4]
gwh = t[:, 4:6] # grid wh
# Cast to int to get an cell index e.g. 1.2 gets associated to cell 1
gij = gxy.long()
# Isolate x and y index dimensions
gi, gj = gij.T # grid xy indices
# Convert anchor indexes to int
a = t[:, 6].long()
# Add target tensors for this yolo layer to the output lists
# Add to index list and limit index range to prevent out of bounds
indices.append((b, a, gj.clamp_(0, gain[3].long() - 1), gi.clamp_(0, gain[2].long() - 1)))
# Add to target box list and convert box coordinates from global grid coordinates to local offsets in the grid cell
tbox.append(torch.cat((gxy - gij, gwh), 1)) # box
# Add correct anchor for each target to the list
anch.append(anchors[a])
# Add class for each target to the list
tcls.append(c)
return tcls, tbox, indices, anch
================================================
FILE: pytorchyolo/utils/parse_config.py
================================================
def parse_model_config(path):
"""Parses the yolo-v3 layer configuration file and returns module definitions"""
file = open(path, 'r')
lines = file.read().split('\n')
lines = [x for x in lines if x and not x.startswith('#')]
lines = [x.rstrip().lstrip() for x in lines] # get rid of fringe whitespaces
module_defs = []
for line in lines:
if line.startswith('['): # This marks the start of a new block
module_defs.append({})
module_defs[-1]['type'] = line[1:-1].rstrip()
if module_defs[-1]['type'] == 'convolutional':
module_defs[-1]['batch_normalize'] = 0
else:
key, value = line.split("=")
value = value.strip()
module_defs[-1][key.rstrip()] = value.strip()
return module_defs
def parse_data_config(path):
"""Parses the data configuration file"""
options = dict()
options['gpus'] = '0,1,2,3'
options['num_workers'] = '10'
with open(path, 'r') as fp:
lines = fp.readlines()
for line in lines:
line = line.strip()
if line == '' or line.startswith('#'):
continue
key, value = line.split('=')
options[key.strip()] = value.strip()
return options
================================================
FILE: pytorchyolo/utils/transforms.py
================================================
import torch
import torch.nn.functional as F
import numpy as np
import imgaug.augmenters as iaa
from imgaug.augmentables.bbs import BoundingBox, BoundingBoxesOnImage
from .utils import xywh2xyxy_np
import torchvision.transforms as transforms
class ImgAug(object):
def __init__(self, augmentations=[]):
self.augmentations = augmentations
def __call__(self, data):
# Unpack data
img, boxes = data
# Convert xywh to xyxy
boxes = np.array(boxes)
boxes[:, 1:] = xywh2xyxy_np(boxes[:, 1:])
# Convert bounding boxes to imgaug
bounding_boxes = BoundingBoxesOnImage(
[BoundingBox(*box[1:], label=box[0]) for box in boxes],
shape=img.shape)
# Apply augmentations
img, bounding_boxes = self.augmentations(
image=img,
bounding_boxes=bounding_boxes)
# Clip out of image boxes
bounding_boxes = bounding_boxes.clip_out_of_image()
# Convert bounding boxes back to numpy
boxes = np.zeros((len(bounding_boxes), 5))
for box_idx, box in enumerate(bounding_boxes):
# Extract coordinates for unpadded + unscaled image
x1 = box.x1
y1 = box.y1
x2 = box.x2
y2 = box.y2
# Returns (x, y, w, h)
boxes[box_idx, 0] = box.label
boxes[box_idx, 1] = ((x1 + x2) / 2)
boxes[box_idx, 2] = ((y1 + y2) / 2)
boxes[box_idx, 3] = (x2 - x1)
boxes[box_idx, 4] = (y2 - y1)
return img, boxes
class RelativeLabels(object):
def __init__(self, ):
pass
def __call__(self, data):
img, boxes = data
h, w, _ = img.shape
boxes[:, [1, 3]] /= w
boxes[:, [2, 4]] /= h
return img, boxes
class AbsoluteLabels(object):
def __init__(self, ):
pass
def __call__(self, data):
img, boxes = data
h, w, _ = img.shape
boxes[:, [1, 3]] *= w
boxes[:, [2, 4]] *= h
return img, boxes
class PadSquare(ImgAug):
def __init__(self, ):
self.augmentations = iaa.Sequential([
iaa.PadToAspectRatio(
1.0,
position="center-center").to_deterministic()
])
class ToTensor(object):
def __init__(self, ):
pass
def __call__(self, data):
img, boxes = data
# Extract image as PyTorch tensor
img = transforms.ToTensor()(img)
bb_targets = torch.zeros((len(boxes), 6))
bb_targets[:, 1:] = transforms.ToTensor()(boxes)
return img, bb_targets
class Resize(object):
def __init__(self, size):
self.size = size
def __call__(self, data):
img, boxes = data
img = F.interpolate(img.unsqueeze(0), size=self.size, mode="nearest").squeeze(0)
return img, boxes
DEFAULT_TRANSFORMS = transforms.Compose([
AbsoluteLabels(),
PadSquare(),
RelativeLabels(),
ToTensor(),
])
================================================
FILE: pytorchyolo/utils/utils.py
================================================
from __future__ import division
import time
import platform
import tqdm
import torch
import torch.nn as nn
import torchvision
import numpy as np
import subprocess
import random
import imgaug as ia
def provide_determinism(seed=42):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
ia.seed(seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
def worker_seed_set(worker_id):
# See for details of numpy:
# https://github.com/pytorch/pytorch/issues/5059#issuecomment-817392562
# See for details of random:
# https://pytorch.org/docs/stable/notes/randomness.html#dataloader
# NumPy
uint64_seed = torch.initial_seed()
ss = np.random.SeedSequence([uint64_seed])
np.random.seed(ss.generate_state(4))
# random
worker_seed = torch.initial_seed() % 2**32
random.seed(worker_seed)
def to_cpu(tensor):
return tensor.detach().cpu()
def load_classes(path):
"""
Loads class labels at 'path'
"""
with open(path, "r") as fp:
names = fp.read().splitlines()
return names
def weights_init_normal(m):
classname = m.__class__.__name__
if classname.find("Conv") != -1:
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif classname.find("BatchNorm2d") != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0.0)
def rescale_boxes(boxes, current_dim, original_shape):
"""
Rescales bounding boxes to the original shape
"""
orig_h, orig_w = original_shape
# The amount of padding that was added
pad_x = max(orig_h - orig_w, 0) * (current_dim / max(original_shape))
pad_y = max(orig_w - orig_h, 0) * (current_dim / max(original_shape))
# Image height and width after padding is removed
unpad_h = current_dim - pad_y
unpad_w = current_dim - pad_x
# Rescale bounding boxes to dimension of original image
boxes[:, 0] = ((boxes[:, 0] - pad_x // 2) / unpad_w) * orig_w
boxes[:, 1] = ((boxes[:, 1] - pad_y // 2) / unpad_h) * orig_h
boxes[:, 2] = ((boxes[:, 2] - pad_x // 2) / unpad_w) * orig_w
boxes[:, 3] = ((boxes[:, 3] - pad_y // 2) / unpad_h) * orig_h
return boxes
def xywh2xyxy(x):
y = x.new(x.shape)
y[..., 0] = x[..., 0] - x[..., 2] / 2
y[..., 1] = x[..., 1] - x[..., 3] / 2
y[..., 2] = x[..., 0] + x[..., 2] / 2
y[..., 3] = x[..., 1] + x[..., 3] / 2
return y
def xywh2xyxy_np(x):
y = np.zeros_like(x)
y[..., 0] = x[..., 0] - x[..., 2] / 2
y[..., 1] = x[..., 1] - x[..., 3] / 2
y[..., 2] = x[..., 0] + x[..., 2] / 2
y[..., 3] = x[..., 1] + x[..., 3] / 2
return y
def ap_per_class(tp, conf, pred_cls, target_cls):
""" Compute the average precision, given the recall and precision curves.
Source: https://github.com/rafaelpadilla/Object-Detection-Metrics.
# Arguments
tp: True positives (list).
conf: Objectness value from 0-1 (list).
pred_cls: Predicted object classes (list).
target_cls: True object classes (list).
# Returns
The average precision as computed in py-faster-rcnn.
"""
# Sort by objectness
i = np.argsort(-conf)
tp, conf, pred_cls = tp[i], conf[i], pred_cls[i]
# Find unique classes
unique_classes = np.unique(target_cls)
# Create Precision-Recall curve and compute AP for each class
ap, p, r = [], [], []
for c in tqdm.tqdm(unique_classes, desc="Computing AP"):
i = pred_cls == c
n_gt = (target_cls == c).sum() # Number of ground truth objects
n_p = i.sum() # Number of predicted objects
if n_p == 0 and n_gt == 0:
continue
elif n_p == 0 or n_gt == 0:
ap.append(0)
r.append(0)
p.append(0)
else:
# Accumulate FPs and TPs
fpc = (1 - tp[i]).cumsum()
tpc = (tp[i]).cumsum()
# Recall
recall_curve = tpc / (n_gt + 1e-16)
r.append(recall_curve[-1])
# Precision
precision_curve = tpc / (tpc + fpc)
p.append(precision_curve[-1])
# AP from recall-precision curve
ap.append(compute_ap(recall_curve, precision_curve))
# Compute F1 score (harmonic mean of precision and recall)
p, r, ap = np.array(p), np.array(r), np.array(ap)
f1 = 2 * p * r / (p + r + 1e-16)
return p, r, ap, f1, unique_classes.astype("int32")
def compute_ap(recall, precision):
""" Compute the average precision, given the recall and precision curves.
Code originally from https://github.com/rbgirshick/py-faster-rcnn.
# Arguments
recall: The recall curve (list).
precision: The precision curve (list).
# Returns
The average precision as computed in py-faster-rcnn.
"""
# correct AP calculation
# first append sentinel values at the end
mrec = np.concatenate(([0.0], recall, [1.0]))
mpre = np.concatenate(([0.0], precision, [0.0]))
# compute the precision envelope
for i in range(mpre.size - 1, 0, -1):
mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])
# to calculate area under PR curve, look for points
# where X axis (recall) changes value
i = np.where(mrec[1:] != mrec[:-1])[0]
# and sum (\Delta recall) * prec
ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])
return ap
def get_batch_statistics(outputs, targets, iou_threshold):
""" Compute true positives, predicted scores and predicted labels per sample """
batch_metrics = []
for sample_i in range(len(outputs)):
if outputs[sample_i] is None:
continue
output = outputs[sample_i]
pred_boxes = output[:, :4]
pred_scores = output[:, 4]
pred_labels = output[:, -1]
true_positives = np.zeros(pred_boxes.shape[0])
annotations = targets[targets[:, 0] == sample_i][:, 1:]
target_labels = annotations[:, 0] if len(annotations) else []
if len(annotations):
detected_boxes = []
target_boxes = annotations[:, 1:]
for pred_i, (pred_box, pred_label) in enumerate(zip(pred_boxes, pred_labels)):
# If targets are found break
if len(detected_boxes) == len(annotations):
break
# Ignore if label is not one of the target labels
if pred_label not in target_labels:
continue
# Filter target_boxes by pred_label so that we only match against boxes of our own label
filtered_target_position, filtered_targets = zip(*filter(lambda x: target_labels[x[0]] == pred_label, enumerate(target_boxes)))
# Find the best matching target for our predicted box
iou, box_filtered_index = bbox_iou(pred_box.unsqueeze(0), torch.stack(filtered_targets)).max(0)
# Remap the index in the list of filtered targets for that label to the index in the list with all targets.
box_index = filtered_target_position[box_filtered_index]
# Check if the iou is above the min treshold and i
if iou >= iou_threshold and box_index not in detected_boxes:
true_positives[pred_i] = 1
detected_boxes += [box_index]
batch_metrics.append([true_positives, pred_scores, pred_labels])
return batch_metrics
def bbox_wh_iou(wh1, wh2):
wh2 = wh2.t()
w1, h1 = wh1[0], wh1[1]
w2, h2 = wh2[0], wh2[1]
inter_area = torch.min(w1, w2) * torch.min(h1, h2)
union_area = (w1 * h1 + 1e-16) + w2 * h2 - inter_area
return inter_area / union_area
def bbox_iou(box1, box2, x1y1x2y2=True):
"""
Returns the IoU of two bounding boxes
"""
if not x1y1x2y2:
# Transform from center and width to exact coordinates
b1_x1, b1_x2 = box1[:, 0] - box1[:, 2] / 2, box1[:, 0] + box1[:, 2] / 2
b1_y1, b1_y2 = box1[:, 1] - box1[:, 3] / 2, box1[:, 1] + box1[:, 3] / 2
b2_x1, b2_x2 = box2[:, 0] - box2[:, 2] / 2, box2[:, 0] + box2[:, 2] / 2
b2_y1, b2_y2 = box2[:, 1] - box2[:, 3] / 2, box2[:, 1] + box2[:, 3] / 2
else:
# Get the coordinates of bounding boxes
b1_x1, b1_y1, b1_x2, b1_y2 = \
box1[:, 0], box1[:, 1], box1[:, 2], box1[:, 3]
b2_x1, b2_y1, b2_x2, b2_y2 = \
box2[:, 0], box2[:, 1], box2[:, 2], box2[:, 3]
# get the corrdinates of the intersection rectangle
inter_rect_x1 = torch.max(b1_x1, b2_x1)
inter_rect_y1 = torch.max(b1_y1, b2_y1)
inter_rect_x2 = torch.min(b1_x2, b2_x2)
inter_rect_y2 = torch.min(b1_y2, b2_y2)
# Intersection area
inter_area = torch.clamp(inter_rect_x2 - inter_rect_x1 + 1, min=0) * torch.clamp(
inter_rect_y2 - inter_rect_y1 + 1, min=0
)
# Union Area
b1_area = (b1_x2 - b1_x1 + 1) * (b1_y2 - b1_y1 + 1)
b2_area = (b2_x2 - b2_x1 + 1) * (b2_y2 - b2_y1 + 1)
iou = inter_area / (b1_area + b2_area - inter_area + 1e-16)
return iou
def box_iou(box1, box2):
# https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
"""
Return intersection-over-union (Jaccard index) of boxes.
Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
Arguments:
box1 (Tensor[N, 4])
box2 (Tensor[M, 4])
Returns:
iou (Tensor[N, M]): the NxM matrix containing the pairwise
IoU values for every element in boxes1 and boxes2
"""
def box_area(box):
# box = 4xn
return (box[2] - box[0]) * (box[3] - box[1])
area1 = box_area(box1.T)
area2 = box_area(box2.T)
# inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) -
torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
# iou = inter / (area1 + area2 - inter)
return inter / (area1[:, None] + area2 - inter)
def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None):
"""Performs Non-Maximum Suppression (NMS) on inference results
Returns:
detections with shape: nx6 (x1, y1, x2, y2, conf, cls)
"""
nc = prediction.shape[2] - 5 # number of classes
# Settings
# (pixels) minimum and maximum box width and height
max_wh = 4096
max_det = 300 # maximum number of detections per image
max_nms = 30000 # maximum number of boxes into torchvision.ops.nms()
time_limit = 1.0 # seconds to quit after
multi_label = nc > 1 # multiple labels per box (adds 0.5ms/img)
t = time.time()
output = [torch.zeros((0, 6), device="cpu")] * prediction.shape[0]
for xi, x in enumerate(prediction): # image index, image inference
# Apply constraints
# x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height
x = x[x[..., 4] > conf_thres] # confidence
# If none remain process next image
if not x.shape[0]:
continue
# Compute conf
x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf
# Box (center x, center y, width, height) to (x1, y1, x2, y2)
box = xywh2xyxy(x[:, :4])
# Detections matrix nx6 (xyxy, conf, cls)
if multi_label:
i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
else: # best class only
conf, j = x[:, 5:].max(1, keepdim=True)
x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
# Filter by class
if classes is not None:
x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
# Check shape
n = x.shape[0] # number of boxes
if not n: # no boxes
continue
elif n > max_nms: # excess boxes
# sort by confidence
x = x[x[:, 4].argsort(descending=True)[:max_nms]]
# Batched NMS
c = x[:, 5:6] * max_wh # classes
# boxes (offset by class), scores
boxes, scores = x[:, :4] + c, x[:, 4]
i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS
if i.shape[0] > max_det: # limit detections
i = i[:max_det]
output[xi] = to_cpu(x[i])
if (time.time() - t) > time_limit:
print(f'WARNING: NMS time limit {time_limit}s exceeded')
break # time limit exceeded
return output
def print_environment_info():
"""
Prints infos about the environment and the system.
This should help when people make issues containg the printout.
"""
print("Environment information:")
# Print OS information
print(f"System: {platform.system()} {platform.release()}")
# Print poetry package version
try:
print(f"Current Version: {subprocess.check_output(['poetry', 'version'], stderr=subprocess.DEVNULL).decode('ascii').strip()}")
except (subprocess.CalledProcessError, FileNotFoundError):
print("Not using the poetry package")
# Print commit hash if possible
try:
print(f"Current Commit Hash: {subprocess.check_output(['git', 'rev-parse', '--short', 'HEAD'], stderr=subprocess.DEVNULL).decode('ascii').strip()}")
except (subprocess.CalledProcessError, FileNotFoundError):
print("No git or repo found")
================================================
FILE: weights/download_weights.sh
================================================
#!/bin/bash
# Download weights for vanilla YOLOv3
wget -c "https://pjreddie.com/media/files/yolov3.weights" --header "Referer: pjreddie.com"
# # Download weights for tiny YOLOv3
wget -c "https://pjreddie.com/media/files/yolov3-tiny.weights" --header "Referer: pjreddie.com"
# Download weights for backbone network
wget -c "https://pjreddie.com/media/files/darknet53.conv.74" --header "Referer: pjreddie.com"
gitextract_qej34q3x/
├── .github/
│ ├── ISSUE_TEMPLATE/
│ │ ├── 1_bug_report.md
│ │ ├── 2_need_help.md
│ │ └── 3_feature_request.md
│ ├── dependabot.yml
│ ├── pull_request_template.md
│ └── workflows/
│ └── main.yml
├── .gitignore
├── LICENSE
├── README.md
├── config/
│ ├── coco.data
│ ├── create_custom_model.sh
│ ├── custom.data
│ ├── yolov3-tiny.cfg
│ └── yolov3.cfg
├── pyproject.toml
├── pytorchyolo/
│ ├── __init__.py
│ ├── detect.py
│ ├── models.py
│ ├── test.py
│ ├── train.py
│ └── utils/
│ ├── __init__.py
│ ├── augmentations.py
│ ├── datasets.py
│ ├── logger.py
│ ├── loss.py
│ ├── parse_config.py
│ ├── transforms.py
│ └── utils.py
└── weights/
└── download_weights.sh
SYMBOL INDEX (85 symbols across 11 files)
FILE: pytorchyolo/detect.py
function detect_directory (line 28) | def detect_directory(model_path, weights_path, img_path, classes, output...
function detect_image (line 67) | def detect_image(model, image, img_size=416, conf_thres=0.5, nms_thres=0...
function detect (line 102) | def detect(model, dataloader, output_path, conf_thres, nms_thres):
function _draw_and_save_output_images (line 145) | def _draw_and_save_output_images(img_detections, imgs, img_size, output_...
function _draw_and_save_output_image (line 167) | def _draw_and_save_output_image(image_path, detections, img_size, output...
function _create_data_loader (line 225) | def _create_data_loader(img_path, batch_size, img_size, n_cpu):
function run (line 251) | def run():
FILE: pytorchyolo/models.py
function create_modules (line 16) | def create_modules(module_defs: List[dict]) -> Tuple[dict, nn.ModuleList]:
class Upsample (line 115) | class Upsample(nn.Module):
method __init__ (line 118) | def __init__(self, scale_factor, mode: str = "nearest"):
method forward (line 123) | def forward(self, x):
class YOLOLayer (line 128) | class YOLOLayer(nn.Module):
method __init__ (line 131) | def __init__(self, anchors: List[Tuple[int, int]], num_classes: int, n...
method forward (line 154) | def forward(self, x: torch.Tensor, img_size: int) -> torch.Tensor:
method _make_grid (line 182) | def _make_grid(nx: int = 20, ny: int = 20) -> torch.Tensor:
class Darknet (line 193) | class Darknet(nn.Module):
method __init__ (line 196) | def __init__(self, config_path):
method forward (line 205) | def forward(self, x):
method load_darknet_weights (line 225) | def load_darknet_weights(self, weights_path):
method save_darknet_weights (line 291) | def save_darknet_weights(self, path, cutoff=-1):
function load_model (line 320) | def load_model(model_path, weights_path=None):
FILE: pytorchyolo/test.py
function evaluate_model_file (line 22) | def evaluate_model_file(model_path, weights_path, img_path, class_names,...
function print_eval_stats (line 65) | def print_eval_stats(metrics_output, class_names, verbose):
function _evaluate (line 79) | def _evaluate(model, dataloader, class_names, img_size, iou_thres, conf_...
function _create_validation_data_loader (line 136) | def _create_validation_data_loader(img_path, batch_size, img_size, n_cpu):
function run (line 162) | def run():
FILE: pytorchyolo/train.py
function _create_data_loader (line 28) | def _create_data_loader(img_path, batch_size, img_size, n_cpu, multiscal...
function run (line 60) | def run():
FILE: pytorchyolo/utils/augmentations.py
class DefaultAug (line 6) | class DefaultAug(ImgAug):
method __init__ (line 7) | def __init__(self, ):
class StrongAug (line 17) | class StrongAug(ImgAug):
method __init__ (line 18) | def __init__(self, ):
FILE: pytorchyolo/utils/datasets.py
function pad_to_square (line 15) | def pad_to_square(img, pad_value):
function resize (line 28) | def resize(image, size):
class ImageFolder (line 33) | class ImageFolder(Dataset):
method __init__ (line 34) | def __init__(self, folder_path, transform=None):
method __getitem__ (line 38) | def __getitem__(self, index):
method __len__ (line 54) | def __len__(self):
class ListDataset (line 58) | class ListDataset(Dataset):
method __init__ (line 59) | def __init__(self, list_path, img_size=416, multiscale=True, transform...
method __getitem__ (line 81) | def __getitem__(self, index):
method collate_fn (line 121) | def collate_fn(self, batch):
method __len__ (line 144) | def __len__(self):
FILE: pytorchyolo/utils/logger.py
class Logger (line 6) | class Logger(object):
method __init__ (line 7) | def __init__(self, log_dir, log_hist=True):
method scalar_summary (line 15) | def scalar_summary(self, tag, value, step):
method list_of_scalars_summary (line 19) | def list_of_scalars_summary(self, tag_value_pairs, step):
FILE: pytorchyolo/utils/loss.py
function bbox_iou (line 11) | def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=Fal...
function compute_loss (line 58) | def compute_loss(predictions, targets, model):
function build_targets (line 128) | def build_targets(p, targets, model):
FILE: pytorchyolo/utils/parse_config.py
function parse_model_config (line 3) | def parse_model_config(path):
function parse_data_config (line 24) | def parse_data_config(path):
FILE: pytorchyolo/utils/transforms.py
class ImgAug (line 12) | class ImgAug(object):
method __init__ (line 13) | def __init__(self, augmentations=[]):
method __call__ (line 16) | def __call__(self, data):
class RelativeLabels (line 56) | class RelativeLabels(object):
method __init__ (line 57) | def __init__(self, ):
method __call__ (line 60) | def __call__(self, data):
class AbsoluteLabels (line 68) | class AbsoluteLabels(object):
method __init__ (line 69) | def __init__(self, ):
method __call__ (line 72) | def __call__(self, data):
class PadSquare (line 80) | class PadSquare(ImgAug):
method __init__ (line 81) | def __init__(self, ):
class ToTensor (line 89) | class ToTensor(object):
method __init__ (line 90) | def __init__(self, ):
method __call__ (line 93) | def __call__(self, data):
class Resize (line 104) | class Resize(object):
method __init__ (line 105) | def __init__(self, size):
method __call__ (line 108) | def __call__(self, data):
FILE: pytorchyolo/utils/utils.py
function provide_determinism (line 15) | def provide_determinism(seed=42):
function worker_seed_set (line 26) | def worker_seed_set(worker_id):
function to_cpu (line 42) | def to_cpu(tensor):
function load_classes (line 46) | def load_classes(path):
function weights_init_normal (line 55) | def weights_init_normal(m):
function rescale_boxes (line 64) | def rescale_boxes(boxes, current_dim, original_shape):
function xywh2xyxy (line 86) | def xywh2xyxy(x):
function xywh2xyxy_np (line 95) | def xywh2xyxy_np(x):
function ap_per_class (line 104) | def ap_per_class(tp, conf, pred_cls, target_cls):
function compute_ap (line 159) | def compute_ap(recall, precision):
function get_batch_statistics (line 187) | def get_batch_statistics(outputs, targets, iou_threshold):
function bbox_wh_iou (line 235) | def bbox_wh_iou(wh1, wh2):
function bbox_iou (line 244) | def bbox_iou(box1, box2, x1y1x2y2=True):
function box_iou (line 279) | def box_iou(box1, box2):
function non_max_suppression (line 306) | def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, cla...
function print_environment_info (line 377) | def print_environment_info():
Condensed preview — 29 files, each showing path, character count, and a content snippet. Download the .json file or copy for the full structured content (154K chars).
[
{
"path": ".github/ISSUE_TEMPLATE/1_bug_report.md",
"chars": 2420,
"preview": "---\nname: \"\\U0001F41B Bug report\"\nabout: Report a bug, crash or some misbehavior\ntitle: ''\nlabels: 'bug'\nassignees: ''\n-"
},
{
"path": ".github/ISSUE_TEMPLATE/2_need_help.md",
"chars": 504,
"preview": "---\nname: \"⁉️ Need help?\"\nabout: \"Get help with using or improving our software\"\ntitle: ''\nlabels: ''\nassignees: ''\n---\n"
},
{
"path": ".github/ISSUE_TEMPLATE/3_feature_request.md",
"chars": 656,
"preview": "---\nname: \"\\U0001F680 Feature request\"\nabout: Suggest an idea for this project\nlabels: 'enhancement'\n---\n\n<!--\nThank you"
},
{
"path": ".github/dependabot.yml",
"chars": 501,
"preview": "# To get started with Dependabot version updates, you'll need to specify which\n# package ecosystems to update and where "
},
{
"path": ".github/pull_request_template.md",
"chars": 489,
"preview": "## Proposed changes\n<!--- Describe your changes and why they are necessary. -->\n\n## Related issues\n<!--- Mention (link) "
},
{
"path": ".github/workflows/main.yml",
"chars": 1441,
"preview": "name: CI\n\non: [pull_request, workflow_dispatch]\n\njobs:\n main:\n runs-on: ${{ matrix.os }}\n strategy:\n matri"
},
{
"path": ".gitignore",
"chars": 153,
"preview": "\n.DS_Store\nbuild\n.git\n*.egg-info\ndist\noutput/\ndata/*\nbackup\nweights/*.weights\nweights/*.conv.*\n__pycache__\ncheckpoints/\n"
},
{
"path": "LICENSE",
"chars": 35127,
"preview": "GNU GENERAL PUBLIC LICENSE\n Version 3, 29 June 2007\n\n Copyright (C) 2007 Free Software Foundation,"
},
{
"path": "README.md",
"chars": 7651,
"preview": "# PyTorch YOLO\nA minimal PyTorch implementation of YOLOv3, with support for training, inference and evaluation.\n\nYOLOv4 "
},
{
"path": "config/coco.data",
"chars": 115,
"preview": "classes= 80\ntrain=data/coco/trainvalno5k.txt\nvalid=data/coco/5k.txt\nnames=data/coco.names\nbackup=backup/\neval=coco\n"
},
{
"path": "config/create_custom_model.sh",
"chars": 8532,
"preview": "#!/bin/bash\n\nNUM_CLASSES=$1\n\necho \"\n[net]\n# Testing\n#batch=1\n#subdivisions=1\n# Training\nbatch=16\nsubdivisions=1\nwidth=41"
},
{
"path": "config/custom.data",
"chars": 99,
"preview": "classes= 1\ntrain=data/custom/train.txt\nvalid=data/custom/valid.txt\nnames=data/custom/classes.names\n"
},
{
"path": "config/yolov3-tiny.cfg",
"chars": 2024,
"preview": "[net]\n# Testing\n#batch=1\n#subdivisions=1\n# Training\nbatch=64\nsubdivisions=2\nwidth=416\nheight=416\nchannels=3\nmomentum=0.9"
},
{
"path": "config/yolov3.cfg",
"chars": 8339,
"preview": "[net]\n# Testing\n#batch=1\n#subdivisions=1\n# Training\nbatch=16\nsubdivisions=1\nwidth=416\nheight=416\nchannels=3\nmomentum=0.9"
},
{
"path": "pyproject.toml",
"chars": 1105,
"preview": "[tool.poetry]\nname = \"PyTorchYolo\"\nversion = \"1.8.0\"\nreadme = \"README.md\"\nrepository = \"https://github.com/eriklindernor"
},
{
"path": "pytorchyolo/__init__.py",
"chars": 0,
"preview": ""
},
{
"path": "pytorchyolo/detect.py",
"chars": 10872,
"preview": "#! /usr/bin/env python3\n\nfrom __future__ import division\n\nimport os\nimport argparse\nimport tqdm\nimport random\nimport num"
},
{
"path": "pytorchyolo/models.py",
"chars": 14731,
"preview": "from __future__ import division\n\nimport os\nfrom itertools import chain\nfrom typing import List, Tuple\n\nimport numpy as n"
},
{
"path": "pytorchyolo/test.py",
"chars": 7996,
"preview": "#! /usr/bin/env python3\n\nfrom __future__ import division\n\nimport argparse\nimport tqdm\nimport numpy as np\n\nfrom terminalt"
},
{
"path": "pytorchyolo/train.py",
"chars": 10137,
"preview": "#! /usr/bin/env python3\n\nfrom __future__ import division\n\nimport os\nimport argparse\nimport tqdm\n\nimport torch\nfrom torch"
},
{
"path": "pytorchyolo/utils/__init__.py",
"chars": 0,
"preview": ""
},
{
"path": "pytorchyolo/utils/augmentations.py",
"chars": 1048,
"preview": "import imgaug.augmenters as iaa\nfrom torchvision import transforms\nfrom pytorchyolo.utils.transforms import ToTensor, Pa"
},
{
"path": "pytorchyolo/utils/datasets.py",
"chars": 4276,
"preview": "from torch.utils.data import Dataset\nimport torch.nn.functional as F\nimport torch\nimport glob\nimport random\nimport os\nim"
},
{
"path": "pytorchyolo/utils/logger.py",
"chars": 790,
"preview": "import os\nimport datetime\nfrom torch.utils.tensorboard import SummaryWriter\n\n\nclass Logger(object):\n def __init__(sel"
},
{
"path": "pytorchyolo/utils/loss.py",
"chars": 9145,
"preview": "import math\n\nimport torch\nimport torch.nn as nn\n\nfrom .utils import to_cpu\n\n# This new loss function is based on https:/"
},
{
"path": "pytorchyolo/utils/parse_config.py",
"chars": 1264,
"preview": "\n\ndef parse_model_config(path):\n \"\"\"Parses the yolo-v3 layer configuration file and returns module definitions\"\"\"\n "
},
{
"path": "pytorchyolo/utils/transforms.py",
"chars": 3005,
"preview": "import torch\nimport torch.nn.functional as F\nimport numpy as np\n\nimport imgaug.augmenters as iaa\nfrom imgaug.augmentable"
},
{
"path": "pytorchyolo/utils/utils.py",
"chars": 13466,
"preview": "from __future__ import division\n\nimport time\nimport platform\nimport tqdm\nimport torch\nimport torch.nn as nn\nimport torch"
},
{
"path": "weights/download_weights.sh",
"chars": 408,
"preview": "#!/bin/bash\n# Download weights for vanilla YOLOv3\nwget -c \"https://pjreddie.com/media/files/yolov3.weights\" --header \"Re"
}
]
About this extraction
This page contains the full source code of the eriklindernoren/PyTorch-YOLOv3 GitHub repository, extracted and formatted as plain text for AI agents and large language models (LLMs). The extraction includes 29 files (142.9 KB), approximately 39.1k tokens, and a symbol index with 85 extracted functions, classes, methods, constants, and types. Use this with OpenClaw, Claude, ChatGPT, Cursor, Windsurf, or any other AI tool that accepts text input. You can copy the full output to your clipboard or download it as a .txt file.
Extracted by GitExtract — free GitHub repo to text converter for AI. Built by Nikandr Surkov.