4 years ago · 1e84a23f38
--- a/.dockerignore
+++ b/.dockerignore
@@ -0,0 +1,217 @@
 # Repo-specific DockerIgnore -------------------------------------------------------------------------------------------
 # .git
 .cache
 .idea
 runs
 output
 coco
 storage.googleapis.com

 data/samples/*
 !data/samples/zidane.jpg
 !data/samples/bus.jpg
 **/results*.txt
 *.jpg

 # Neural Network weights -----------------------------------------------------------------------------------------------
 **/*.weights
 **/*.pt
 **/*.onnx
 **/*.mlmodel
 **/darknet53.conv.74
 **/yolov3-tiny.conv.15


 # Below Copied From .gitignore -----------------------------------------------------------------------------------------
 # Below Copied From .gitignore -----------------------------------------------------------------------------------------


 # GitHub Python GitIgnore ----------------------------------------------------------------------------------------------
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]
 *$py.class

 # C extensions
 *.so

 # Distribution / packaging
 .Python
 env/
 build/
 develop-eggs/
 dist/
 downloads/
 eggs/
 .eggs/
 lib/
 lib64/
 parts/
 sdist/
 var/
 wheels/
 *.egg-info/
 .installed.cfg
 *.egg

 # PyInstaller
 #  Usually these files are written by a python script from a template
 #  before PyInstaller builds the exe, so as to inject date/other infos into it.
 *.manifest
 *.spec

 # Installer logs
 pip-log.txt
 pip-delete-this-directory.txt

 # Unit test / coverage reports
 htmlcov/
 .tox/
 .coverage
 .coverage.*
 .cache
 nosetests.xml
 coverage.xml
 *.cover
 .hypothesis/

 # Translations
 *.mo
 *.pot

 # Django stuff:
 *.log
 local_settings.py

 # Flask stuff:
 instance/
 .webassets-cache

 # Scrapy stuff:
 .scrapy

 # Sphinx documentation
 docs/_build/

 # PyBuilder
 target/

 # Jupyter Notebook
 .ipynb_checkpoints

 # pyenv
 .python-version

 # celery beat schedule file
 celerybeat-schedule

 # SageMath parsed files
 *.sage.py

 # dotenv
 .env

 # virtualenv
 .venv
 venv/
 ENV/

 # Spyder project settings
 .spyderproject
 .spyproject

 # Rope project settings
 .ropeproject

 # mkdocs documentation
 /site

 # mypy
 .mypy_cache/


 # https://github.com/github/gitignore/blob/master/Global/macOS.gitignore -----------------------------------------------

 # General
 .DS_Store
 .AppleDouble
 .LSOverride

 # Icon must end with two \r
 Icon
 Icon?

 # Thumbnails
 ._*

 # Files that might appear in the root of a volume
 .DocumentRevisions-V100
 .fseventsd
 .Spotlight-V100
 .TemporaryItems
 .Trashes
 .VolumeIcon.icns
 .com.apple.timemachine.donotpresent

 # Directories potentially created on remote AFP share
 .AppleDB
 .AppleDesktop
 Network Trash Folder
 Temporary Items
 .apdisk


 # https://github.com/github/gitignore/blob/master/Global/JetBrains.gitignore
 # Covers JetBrains IDEs: IntelliJ, RubyMine, PhpStorm, AppCode, PyCharm, CLion, Android Studio and WebStorm
 # Reference: https://intellij-support.jetbrains.com/hc/en-us/articles/206544839

 # User-specific stuff:
 .idea/*
 .idea/**/workspace.xml
 .idea/**/tasks.xml
 .idea/dictionaries
 .html  # Bokeh Plots
 .pg  # TensorFlow Frozen Graphs
 .avi # videos

 # Sensitive or high-churn files:
 .idea/**/dataSources/
 .idea/**/dataSources.ids
 .idea/**/dataSources.local.xml
 .idea/**/sqlDataSources.xml
 .idea/**/dynamic.xml
 .idea/**/uiDesigner.xml

 # Gradle:
 .idea/**/gradle.xml
 .idea/**/libraries

 # CMake
 cmake-build-debug/
 cmake-build-release/

 # Mongo Explorer plugin:
 .idea/**/mongoSettings.xml

 ## File-based project format:
 *.iws

 ## Plugin-specific files:

 # IntelliJ
 out/

 # mpeltonen/sbt-idea plugin
 .idea_modules/

 # JIRA plugin
 atlassian-ide-plugin.xml

 # Cursive Clojure plugin
 .idea/replstate.xml

 # Crashlytics plugin (for Android Studio and IntelliJ)
 com_crashlytics_export_strings.xml
 crashlytics.properties
 crashlytics-build.properties
 fabric.properties
--- a/.github/ISSUE_TEMPLATE/--bug-report.md
+++ b/.github/ISSUE_TEMPLATE/--bug-report.md
@@ -0,0 +1,41 @@
 ---
 name: "\U0001F41BBug report"
 about: Create a report to help us improve
 title: ''
 labels: bug
 assignees: ''

 ---

 Before submitting a bug report, please ensure that you are using the latest versions of:
 - Python
 - PyTorch
 - This repository (run `git fetch && git status -uno` to check and `git pull` to update)
 
 **Your issue must be reproducible on a public dataset (i.e COCO) using the latest version of the repository, and you must supply code to reproduce, or we can not help you.**

 If this is a custom training question we suggest you include your `train*.jpg`, `test*.jpg` and `results.png` figures.


 ## 🐛 Bug
 A clear and concise description of what the bug is.

 ## To Reproduce
 **REQUIRED**: Code to reproduce your issue below
 ```
 python train.py ...
 ```


 ## Expected behavior
 A clear and concise description of what you expected to happen.

 ## Environment
 If applicable, add screenshots to help explain your problem.

 - OS: [e.g. Ubuntu]
 - GPU [e.g. 2080 Ti]


 ## Additional context
 Add any other context about the problem here.
--- a/.github/ISSUE_TEMPLATE/--feature-request.md
+++ b/.github/ISSUE_TEMPLATE/--feature-request.md
@@ -0,0 +1,27 @@
 ---
 name: "\U0001F680Feature request"
 about: Suggest an idea for this project
 title: ''
 labels: enhancement
 assignees: ''

 ---

 ## 🚀 Feature
 <!-- A clear and concise description of the feature proposal -->

 ## Motivation

 <!-- Please outline the motivation for the proposal. Is your feature request related to a problem? e.g., I'm always frustrated when [...]. If this is related to another GitHub issue, please link here too -->

 ## Pitch

 <!-- A clear and concise description of what you want to happen. -->

 ## Alternatives

 <!-- A clear and concise description of any alternative solutions or features you've considered, if any. -->

 ## Additional context

 <!-- Add any other context or screenshots about the feature request here. -->
--- a/.github/workflows/greetings.yml
+++ b/.github/workflows/greetings.yml
@@ -0,0 +1,23 @@
 name: Greetings

 on: [pull_request, issues]

 jobs:
  greeting:
    runs-on: ubuntu-latest
    steps:
    - uses: actions/first-interaction@v1
      with:
        repo-token: ${{ secrets.GITHUB_TOKEN }}
        pr-message: 'Hello @${{ github.actor }}, thank you for submitting a PR! We will respond as soon as possible.'
        issue-message: >
          Hello @${{ github.actor }}, thank you for your interest in our work! Please visit our [Custom Training Tutorial](https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data) to get started, and see our [Google Colab Notebook](https://github.com/ultralytics/yolov5/blob/master/tutorial.ipynb), [Docker Image](https://hub.docker.com/r/ultralytics/yolov5), and [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart) for example environments.
            
          If this is a bug report, please provide screenshots and **minimum viable code to reproduce your issue**, otherwise we can not help you.

          If this is a custom model or data training question, please note that Ultralytics does **not** provide free personal support. As a leader in vision ML and AI, we do offer professional consulting, from simple expert advice up to delivery of fully customized, end-to-end production solutions for our clients, such as:
          - **Cloud-based AI** surveillance systems operating on **hundreds of HD video streams in realtime.**
          - **Edge AI** integrated into custom iOS and Android apps for realtime **30 FPS video inference.**
          - **Custom data training**, hyperparameter evolution, and model exportation to any destination.

          To discuss your business requirements in greater detail please visit us at https://www.ultralytics.com.
--- a/.github/workflows/stale.yml
+++ b/.github/workflows/stale.yml
@@ -0,0 +1,17 @@
 name: Close stale issues
 on:
  schedule:
  - cron: "0 0 * * *"

 jobs:
  stale:
    runs-on: ubuntu-latest
    steps:
    - uses: actions/stale@v1
      with:
        repo-token: ${{ secrets.GITHUB_TOKEN }}
        stale-issue-message: 'This issue is stale because it has been open 30 days with no activity. Remove Stale label or comment or this will be closed in 5 days.'
        stale-pr-message: 'This pull request is stale because it has been open 30 days with no activity. Remove Stale label or comment or this will be closed in 5 days.'
        days-before-stale: 30
        days-before-close: 5
        exempt-issue-label: 'tutorial'
--- a/.gitignore
+++ b/.gitignore
@@ -0,0 +1,244 @@
 # Repo-specific GitIgnore ----------------------------------------------------------------------------------------------
 *.jpg
 *.jpeg
 *.png
 *.bmp
 *.tif
 *.tiff
 *.heic
 *.JPG
 *.JPEG
 *.PNG
 *.BMP
 *.TIF
 *.TIFF
 *.HEIC
 *.mp4
 *.mov
 *.MOV
 *.avi
 *.data
 *.json

 *.cfg
 !cfg/yolov3*.cfg

 storage.googleapis.com
 runs/*
 data/*
 !data/samples/zidane.jpg
 !data/samples/bus.jpg
 !data/coco.names
 !data/coco_paper.names
 !data/coco.data
 !data/coco_*.data
 !data/coco_*.txt
 !data/trainvalno5k.shapes
 !data/*.sh

 pycocotools/*
 results*.txt
 gcp_test*.sh

 # MATLAB GitIgnore -----------------------------------------------------------------------------------------------------
 *.m~
 *.mat
 !targets*.mat

 # Neural Network weights -----------------------------------------------------------------------------------------------
 *.weights
 *.pt
 *.onnx
 *.mlmodel
 darknet53.conv.74
 yolov3-tiny.conv.15

 # GitHub Python GitIgnore ----------------------------------------------------------------------------------------------
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]
 *$py.class

 # C extensions
 *.so

 # Distribution / packaging
 .Python
 env/
 build/
 develop-eggs/
 dist/
 downloads/
 eggs/
 .eggs/
 lib/
 lib64/
 parts/
 sdist/
 var/
 wheels/
 *.egg-info/
 .installed.cfg
 *.egg

 # PyInstaller
 #  Usually these files are written by a python script from a template
 #  before PyInstaller builds the exe, so as to inject date/other infos into it.
 *.manifest
 *.spec

 # Installer logs
 pip-log.txt
 pip-delete-this-directory.txt

 # Unit test / coverage reports
 htmlcov/
 .tox/
 .coverage
 .coverage.*
 .cache
 nosetests.xml
 coverage.xml
 *.cover
 .hypothesis/

 # Translations
 *.mo
 *.pot

 # Django stuff:
 *.log
 local_settings.py

 # Flask stuff:
 instance/
 .webassets-cache

 # Scrapy stuff:
 .scrapy

 # Sphinx documentation
 docs/_build/

 # PyBuilder
 target/

 # Jupyter Notebook
 .ipynb_checkpoints

 # pyenv
 .python-version

 # celery beat schedule file
 celerybeat-schedule

 # SageMath parsed files
 *.sage.py

 # dotenv
 .env

 # virtualenv
 .venv
 venv/
 ENV/

 # Spyder project settings
 .spyderproject
 .spyproject

 # Rope project settings
 .ropeproject

 # mkdocs documentation
 /site

 # mypy
 .mypy_cache/


 # https://github.com/github/gitignore/blob/master/Global/macOS.gitignore -----------------------------------------------

 # General
 .DS_Store
 .AppleDouble
 .LSOverride

 # Icon must end with two \r
 Icon
 Icon?

 # Thumbnails
 ._*

 # Files that might appear in the root of a volume
 .DocumentRevisions-V100
 .fseventsd
 .Spotlight-V100
 .TemporaryItems
 .Trashes
 .VolumeIcon.icns
 .com.apple.timemachine.donotpresent

 # Directories potentially created on remote AFP share
 .AppleDB
 .AppleDesktop
 Network Trash Folder
 Temporary Items
 .apdisk


 # https://github.com/github/gitignore/blob/master/Global/JetBrains.gitignore
 # Covers JetBrains IDEs: IntelliJ, RubyMine, PhpStorm, AppCode, PyCharm, CLion, Android Studio and WebStorm
 # Reference: https://intellij-support.jetbrains.com/hc/en-us/articles/206544839

 # User-specific stuff:
 .idea/*
 .idea/**/workspace.xml
 .idea/**/tasks.xml
 .idea/dictionaries
 .html  # Bokeh Plots
 .pg  # TensorFlow Frozen Graphs
 .avi # videos

 # Sensitive or high-churn files:
 .idea/**/dataSources/
 .idea/**/dataSources.ids
 .idea/**/dataSources.local.xml
 .idea/**/sqlDataSources.xml
 .idea/**/dynamic.xml
 .idea/**/uiDesigner.xml

 # Gradle:
 .idea/**/gradle.xml
 .idea/**/libraries

 # CMake
 cmake-build-debug/
 cmake-build-release/

 # Mongo Explorer plugin:
 .idea/**/mongoSettings.xml

 ## File-based project format:
 *.iws

 ## Plugin-specific files:

 # IntelliJ
 out/

 # mpeltonen/sbt-idea plugin
 .idea_modules/

 # JIRA plugin
 atlassian-ide-plugin.xml

 # Cursive Clojure plugin
 .idea/replstate.xml

 # Crashlytics plugin (for Android Studio and IntelliJ)
 com_crashlytics_export_strings.xml
 crashlytics.properties
 crashlytics-build.properties
 fabric.properties
--- a/Dockerfile
+++ b/Dockerfile
@@ -0,0 +1,49 @@
 # Start FROM Nvidia PyTorch image https://ngc.nvidia.com/catalog/containers/nvidia:pytorch
 FROM nvcr.io/nvidia/pytorch:20.03-py3

 # Install dependencies (pip or conda)
 RUN pip install -U gsutil thop
 # RUN pip install -U -r requirements.txt

 # Create working directory
 RUN mkdir -p /usr/src/app
 WORKDIR /usr/src/app

 # Copy contents
 COPY . /usr/src/app

 # Copy weights
 #RUN python3 -c "from models import *; \
 #attempt_download('weights/yolov5s.pt'); \
 #attempt_download('weights/yolov5m.pt'); \
 #attempt_download('weights/yolov5l.pt')"


 # ---------------------------------------------------  Extras Below  ---------------------------------------------------

 # Build and Push
 # t=ultralytics/yolov5:latest && sudo docker build -t $t . && sudo docker push $t

 # Pull and Run
 # t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host $t bash

 # Pull and Run with local directory access
 # t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t bash

 # Kill all
 # sudo docker kill "$(sudo docker ps -q)"

 # Kill all image-based
 # sudo docker kill $(sudo docker ps -a -q --filter ancestor=ultralytics/yolov5:latest)

 # Run bash for loop
 # sudo docker run --gpus all --ipc=host ultralytics/yolov5:latest while true; do python3 train.py --evolve; done

 # Bash in running container
 # sudo docker container exec -it 97919ad657de /bin/bash

 # Bash last stopped container
 # python -c "from utils.utils import *; create_backbone('weights/best.pt')" && gsutil cp weights/backbone.pt gs://ult/coco/yolov5s.pt

 # Clean up
 # docker system prune -a --volumes
--- a/LICENSE
+++ b/LICENSE
@@ -0,0 +1,674 @@
 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>.
--- a/README.md
+++ b/README.md
@@ -0,0 +1,110 @@
 <a href="https://apps.apple.com/app/id1452689527" target="_blank">
 <img src="https://user-images.githubusercontent.com/26833433/82944393-f7644d80-9f4f-11ea-8b87-1a5b04f555f1.jpg" width="1000"></a>
 &nbsp

 This repository represents Ultralytics open-source research into future object detection methods, and incorporates our lessons learned and best practices evolved over training thousands of models on custom client datasets with our previous YOLO repository https://github.com/ultralytics/yolov3. **All code and models are under active development, and are subject to modification or deletion without notice.** Use at your own risk.

 Updates:
 - **May 27, 2020**: Public release of repo. yolov3-spp (this repo) is SOTA among all known yolo implementations, yolov5 family will be undergoing architecture research and development over Q2/Q3 2020 to increase performance. Updates may include CSP bottlenecks from yolov4, as well as PANet or BiFPN head features.
 - **May 24, 2020**: Training yolov5s/x and yolov3-spp. yolov5m/l suffered early overfitting and also code 137 early docker terminations, cause unknown. yolov5l underperforms yolov3-spp due to earlier overfitting, cause unknown.
 - **April 1, 2020**: Begin development of a 100% pytorch scaleable yolov3/4-based group of future models, in small, medium, large and extra large sizes, collectively known as yolov5. Models will be defined by new user-friendly yaml-based configuration files for ease of construction and modification. Datasets will likewise use yaml configuration files. New training platform will be simpler use, harder to break, and more robust to training a wider variety of custom dataset.


 ## Ultralytics Professional Support

 Ultralytics is a U.S.-based particle physics and AI startup with over 6 years of expertise supporting government, academic and business clients. We offer a wide range of vision AI services, spanning from simple expert advice up to delivery of fully customized, end-to-end production solutions, including:
 - **Cloud-based AI** surveillance systems operating on **hundreds of HD video streams in realtime.**
 - **Edge AI** integrated into custom iOS and Android apps for realtime **30 FPS video inference.**
 - **Custom data training**, hyperparameter evolution, and model exportation to any destination.

 For business inquiries and professional support requests please visit us at https://www.ultralytics.com. 


 ## Pretrained Checkpoints

 |       Model    |  AP<sup>val</sup> | AP<sup>test</sup>    |  AP<sub>50</sub> | Latency<sub>GPU</sub> | FPS<sub>GPU</sub>  | | params | FLOPs |
 |----------     |------ |------ |------ | -------- | ------| ------ |------  |  :------: |
 |     YOLOv5-s ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 33.1 | 33.0 | 53.3 | **3.3ms** | **303** | | 7.0B   | 14.0
 |     YOLOv5-m ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 41.5 | 41.5 | 61.5 | 5.5ms | 182 | | 25.2B  | 50.2
 |     YOLOv5-l ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 44.2 | 44.5 | 64.3 | 9.7ms | 103 | | 61.8B  | 123.1
 |     YOLOv5-x ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | **47.1** | **47.2** | **66.7** | 15.8ms | 63 | | 123.1B | 245.7
 |     YOLOv3-SPP ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))  | 45.5 | 45.4 | 65.2 | 8.9ms | 112 | | 63.0B  | 118.0

 ** AP<sup>test</sup> denotes COCO [test-dev2017](http://cocodataset.org/#upload) server results, all other AP results in the table denote val2017 accuracy.  
 ** All accuracy numbers are for single-model single-scale without ensemble or test-time augmentation. Reproduce by  `python test.py --img-size 736 --conf_thres 0.001`  
 ** Latency<sub>GPU</sub> measures end-to-end latency per image averaged over 5000 COCO val2017 images using a V100 GPU and includes image preprocessing, inference, postprocessing and NMS. Average NMS time included in this chart is 1.6ms/image.  Reproduce by `python test.py --img-size 640 --conf_thres 0.1 --batch-size 16`  
 ** All checkpoints are trained to 300 epochs with default settings and hyperparameters (no autoaugmentation). 


 ## Requirements

 Python 3.7 or later with all `requirements.txt` dependencies installed, including `torch >= 1.5`. To install run:
 ```bash
 $ pip install -U -r requirements.txt
 ```


 ## Tutorials

 * [Train Custom Data](https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data)
 * [Google Colab Notebook](https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb) with training, testing and testing examples
 * [GCP Quickstart](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart)
 * [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) 


 ## Inference

 Inference can be run on most common media formats. Model [checkpoints](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J) are downloaded automatically if available. Results are saved to `./inference/output`.
 ```bash
 $ python detect.py --source file.jpg  # image 
                            file.mp4  # video
                            ./dir  # directory
                            0  # webcam
                            rtsp://170.93.143.139/rtplive/470011e600ef003a004ee33696235daa  # rtsp stream
                            http://112.50.243.8/PLTV/88888888/224/3221225900/1.m3u8  # http stream
 ```

 To run inference on examples in the `./inference/images` folder:

 ```bash
 $ python detect.py --source ./inference/images/ --weights yolov5s.pt --conf 0.4

 Namespace(agnostic_nms=False, augment=False, classes=None, conf_thres=0.4, device='', fourcc='mp4v', half=False, img_size=640, iou_thres=0.5, output='inference/output', save_txt=False, source='./inference/images/', view_img=False, weights='yolov5s.pt')
 Using CUDA device0 _CudaDeviceProperties(name='Tesla P100-PCIE-16GB', total_memory=16280MB)

 Downloading https://drive.google.com/uc?export=download&id=1R5T6rIyy3lLwgFXNms8whc-387H0tMQO as yolov5s.pt... Done (2.6s)

 image 1/2 inference/images/bus.jpg: 640x512 3 persons, 1 buss, Done. (0.009s)
 image 2/2 inference/images/zidane.jpg: 384x640 2 persons, 2 ties, Done. (0.009s)
 Results saved to /content/yolov5/inference/output
 ```

 <img src="https://user-images.githubusercontent.com/26833433/83082816-59e54880-a039-11ea-8abe-ab90cc1ec4b0.jpeg" width="500">  


 ## Reproduce Our Training

 Run commands below. Training takes a few days for yolov5s, to a few weeks for yolov5x on a 2080Ti GPU.
 ```bash
 $ python train.py --data coco.yaml --cfg yolov5s.yaml --weights '' --batch-size 16 
 ```
 <img src="https://user-images.githubusercontent.com/26833433/82960433-5a191180-9f6f-11ea-85cc-c49dbd1555e1.png" width="900">


 ## Reproduce Our Environment

 To access an up-to-date working environment (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled), consider a:

 - **GCP** Deep Learning VM with $300 free credit offer: See our [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart) 
 - **Google Colab Notebook** with 12 hours of free GPU time: [Google Colab Notebook](https://colab.sandbox.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb)
 - **Docker Image** from https://hub.docker.com/r/ultralytics/yolov5. See [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) 


 ## Citation

 [![DOI](https://zenodo.org/badge/146165888.svg)](https://zenodo.org/badge/latestdoi/146165888)


 ## Contact

 **Issues should be raised directly in the repository.** For business inquiries or professional support requests please visit us at https://www.ultralytics.com.
--- a/data/coco.yaml
+++ b/data/coco.yaml
@@ -0,0 +1,33 @@
 # COCO 2017 dataset http://cocodataset.org
 # Download command: bash yolov5/data/get_coco2017.sh
 # Train command: python train.py --data ./data/coco.yaml
 # Dataset should be placed next to yolov5 folder:
 #   /parent_folder
 #     /coco
 #     /yolov5


 # train and val datasets (image directory or *.txt file with image paths)
 train: ../coco/train2017.txt  # 118k images
 val: ../coco/val2017.txt  # 5k images
 test: ../coco/test-dev2017.txt  # 20k images for submission to https://competitions.codalab.org/competitions/20794

 # number of classes
 nc: 80

 # class names
 names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
        'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
        'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
        'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
        'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
        'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
        'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
        'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
        'hair drier', 'toothbrush']

 # Print classes
 # with open('data/coco.yaml') as f:
 #   d = yaml.load(f, Loader=yaml.FullLoader)  # dict
 #   for i, x in enumerate(d['names']):
 #     print(i, x)
--- a/data/coco128.yaml
+++ b/data/coco128.yaml
@@ -0,0 +1,26 @@
 # COCO 2017 dataset http://cocodataset.org - first 128 training images
 # Download command:  python -c "from yolov5.utils.google_utils import gdrive_download; gdrive_download('1n_oKgR81BJtqk75b00eAjdv03qVCQn2f','coco128.zip')"
 # Train command: python train.py --data ./data/coco128.yaml
 # Dataset should be placed next to yolov5 folder:
 #   /parent_folder
 #     /coco128
 #     /yolov5


 # train and val datasets (image directory or *.txt file with image paths)
 train: ../coco128/images/train2017/
 val: ../coco128/images/train2017/

 # number of classes
 nc: 80

 # class names
 names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
        'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
        'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
        'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
        'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
        'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
        'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
        'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
        'hair drier', 'toothbrush']
--- a/data/get_coco2017.sh
+++ b/data/get_coco2017.sh
@@ -0,0 +1,25 @@
 #!/bin/bash
 # Zip coco folder
 # zip -r coco.zip coco
 # tar -czvf coco.tar.gz coco

 # Download labels from Google Drive, accepting presented query
 filename="coco2017labels.zip"
 fileid="1cXZR_ckHki6nddOmcysCuuJFM--T-Q6L"
 curl -c ./cookie -s -L "https://drive.google.com/uc?export=download&id=${fileid}" > /dev/null
 curl -Lb ./cookie "https://drive.google.com/uc?export=download&confirm=`awk '/download/ {print $NF}' ./cookie`&id=${fileid}" -o ${filename}
 rm ./cookie

 # Unzip labels
 unzip -q ${filename}  # for coco.zip
 # tar -xzf ${filename}  # for coco.tar.gz
 rm ${filename}

 # Download and unzip images
 cd coco/images
 f="train2017.zip" && curl http://images.cocodataset.org/zips/$f -o $f && unzip -q $f && rm $f  # 19G, 118k images
 f="val2017.zip" && curl http://images.cocodataset.org/zips/$f -o $f && unzip -q $f && rm $f  # 1G, 5k images
 # f="test2017.zip" && curl http://images.cocodataset.org/zips/$f -o $f && unzip -q $f && rm $f  # 7G,  41k images

 # cd out
 cd ../..
--- a/detect.py
+++ b/detect.py
@@ -0,0 +1,182 @@
 import argparse

 from utils.datasets import *
 from utils.utils import *

 ONNX_EXPORT = False


 def detect(save_img=False):
    imgsz = (320, 192) if ONNX_EXPORT else opt.img_size  # (320, 192) or (416, 256) or (608, 352) for (height, width)
    out, source, weights, half, view_img, save_txt = opt.output, opt.source, opt.weights, opt.half, opt.view_img, opt.save_txt
    webcam = source == '0' or source.startswith('rtsp') or source.startswith('http') or source.endswith('.txt')

    # Initialize
    device = torch_utils.select_device(device='cpu' if ONNX_EXPORT else opt.device)
    if os.path.exists(out):
        shutil.rmtree(out)  # delete output folder
    os.makedirs(out)  # make new output folder

    # Load model
    google_utils.attempt_download(weights)
    model = torch.load(weights, map_location=device)['model']
    # torch.save(torch.load(weights, map_location=device), weights)  # update model if SourceChangeWarning

    # Second-stage classifier
    classify = False
    if classify:
        modelc = torch_utils.load_classifier(name='resnet101', n=2)  # initialize
        modelc.load_state_dict(torch.load('weights/resnet101.pt', map_location=device)['model'])  # load weights
        modelc.to(device).eval()

    # Eval mode
    model.to(device).eval()

    # Fuse Conv2d + BatchNorm2d layers
    # model.fuse()

    # Export mode
    if ONNX_EXPORT:
        model.fuse()
        img = torch.zeros((1, 3) + imgsz)  # (1, 3, 320, 192)
        f = opt.weights.replace(opt.weights.split('.')[-1], 'onnx')  # *.onnx filename
        torch.onnx.export(model, img, f, verbose=False, opset_version=11)

        # Validate exported model
        import onnx
        model = onnx.load(f)  # Load the ONNX model
        onnx.checker.check_model(model)  # Check that the IR is well formed
        print(onnx.helper.printable_graph(model.graph))  # Print a human readable representation of the graph
        return

    # Half precision
    half = half and device.type != 'cpu'  # half precision only supported on CUDA
    if half:
        model.half()

    # Set Dataloader
    vid_path, vid_writer = None, None
    if webcam:
        view_img = True
        torch.backends.cudnn.benchmark = True  # set True to speed up constant image size inference
        dataset = LoadStreams(source, img_size=imgsz)
    else:
        save_img = True
        dataset = LoadImages(source, img_size=imgsz)

    # Get names and colors
    names = model.names
    colors = [[random.randint(0, 255) for _ in range(3)] for _ in range(len(names))]

    # Run inference
    t0 = time.time()
    img = torch.zeros((1, 3, imgsz, imgsz), device=device)  # init img
    _ = model(img.half() if half else img.float()) if device.type != 'cpu' else None  # run once
    for path, img, im0s, vid_cap in dataset:
        img = torch.from_numpy(img).to(device)
        img = img.half() if half else img.float()  # uint8 to fp16/32
        img /= 255.0  # 0 - 255 to 0.0 - 1.0
        if img.ndimension() == 3:
            img = img.unsqueeze(0)

        # Inference
        t1 = torch_utils.time_synchronized()
        pred = model(img, augment=opt.augment)[0]
        t2 = torch_utils.time_synchronized()

        # to float
        if half:
            pred = pred.float()

        # Apply NMS
        pred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres,
                                   multi_label=False, classes=opt.classes, agnostic=opt.agnostic_nms)

        # Apply Classifier
        if classify:
            pred = apply_classifier(pred, modelc, img, im0s)

        # Process detections
        for i, det in enumerate(pred):  # detections per image
            if webcam:  # batch_size >= 1
                p, s, im0 = path[i], '%g: ' % i, im0s[i]
            else:
                p, s, im0 = path, '', im0s

            save_path = str(Path(out) / Path(p).name)
            s += '%gx%g ' % img.shape[2:]  # print string
            gn = torch.tensor(im0.shape)[[1, 0, 1, 0]]  #  normalization gain whwh
            if det is not None and len(det):
                # Rescale boxes from img_size to im0 size
                det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round()

                # Print results
                for c in det[:, -1].unique():
                    n = (det[:, -1] == c).sum()  # detections per class
                    s += '%g %ss, ' % (n, names[int(c)])  # add to string

                # Write results
                for *xyxy, conf, cls in det:
                    if save_txt:  # Write to file
                        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
                        with open(save_path[:save_path.rfind('.')] + '.txt', 'a') as file:
                            file.write(('%g ' * 5 + '\n') % (cls, *xywh))  # label format

                    if save_img or view_img:  # Add bbox to image
                        label = '%s %.2f' % (names[int(cls)], conf)
                        plot_one_box(xyxy, im0, label=label, color=colors[int(cls)], line_thickness=3)

            # Print time (inference + NMS)
            print('%sDone. (%.3fs)' % (s, t2 - t1))

            # Stream results
            if view_img:
                cv2.imshow(p, im0)
                if cv2.waitKey(1) == ord('q'):  # q to quit
                    raise StopIteration

            # Save results (image with detections)
            if save_img:
                if dataset.mode == 'images':
                    cv2.imwrite(save_path, im0)
                else:
                    if vid_path != save_path:  # new video
                        vid_path = save_path
                        if isinstance(vid_writer, cv2.VideoWriter):
                            vid_writer.release()  # release previous video writer

                        fps = vid_cap.get(cv2.CAP_PROP_FPS)
                        w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
                        h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
                        vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*opt.fourcc), fps, (w, h))
                    vid_writer.write(im0)

    if save_txt or save_img:
        print('Results saved to %s' % os.getcwd() + os.sep + out)
        if platform == 'darwin':  # MacOS
            os.system('open ' + save_path)

    print('Done. (%.3fs)' % (time.time() - t0))


 if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', type=str, default='weights/yolov5s.pt', help='model.pt path')
    parser.add_argument('--source', type=str, default='inference/images', help='source')  # file/folder, 0 for webcam
    parser.add_argument('--output', type=str, default='inference/output', help='output folder')  # output folder
    parser.add_argument('--img-size', type=int, default=640, help='inference size (pixels)')
    parser.add_argument('--conf-thres', type=float, default=0.4, help='object confidence threshold')
    parser.add_argument('--iou-thres', type=float, default=0.5, help='IOU threshold for NMS')
    parser.add_argument('--fourcc', type=str, default='mp4v', help='output video codec (verify ffmpeg support)')
    parser.add_argument('--half', action='store_true', help='half precision FP16 inference')
    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    parser.add_argument('--view-img', action='store_true', help='display results')
    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
    parser.add_argument('--classes', nargs='+', type=int, help='filter by class')
    parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')
    parser.add_argument('--augment', action='store_true', help='augmented inference')
    opt = parser.parse_args()
    print(opt)

    with torch.no_grad():
        detect()
--- a/inference/images/bus.jpg
+++ b/inference/images/bus.jpg
--- a/inference/images/zidane.jpg
+++ b/inference/images/zidane.jpg
--- a/models/common.py
+++ b/models/common.py
@@ -0,0 +1,174 @@
 # This file contains modules common to various models


 import torch.nn.functional as F

 from utils.utils import *


 def DWConv(c1, c2, k=1, s=1, act=True):  # depthwise convolution
    return Conv(c1, c2, k, s, g=math.gcd(c1, c2), act=act)


 class Conv(nn.Module):  # standard convolution
    def __init__(self, c1, c2, k=1, s=1, g=1, act=True):  # ch_in, ch_out, kernel, stride, groups
        super(Conv, self).__init__()
        self.conv = nn.Conv2d(c1, c2, k, s, k // 2, groups=g, bias=False)
        self.bn = nn.BatchNorm2d(c2)
        self.act = nn.LeakyReLU(0.1, inplace=True) if act else nn.Identity()

    def forward(self, x):
        return self.act(self.bn(self.conv(x)))


 class Bottleneck(nn.Module):
    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
        super(Bottleneck, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_, c2, 3, 1, g=g)
        self.add = shortcut and c1 == c2

    def forward(self, x):
        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))


 class BottleneckLight(nn.Module):
    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
        super(BottleneckLight, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = nn.Conv2d(c_, c2, 3, 1, 3 // 2, groups=g, bias=False)
        self.bn = nn.BatchNorm2d(c2)
        self.act = nn.LeakyReLU(0.1, inplace=True)
        self.add = shortcut and c1 == c2

    def forward(self, x):
        return self.act(self.bn(x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))))


 class BottleneckCSP(nn.Module):
    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
        super(BottleneckCSP, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
        self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
        self.cv4 = Conv(c2, c2, 1, 1)
        self.bn = nn.BatchNorm2d(2 * c_)  # applied to cat(cv2, cv3)
        self.act = nn.LeakyReLU(0.1, inplace=True)
        self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])

    def forward(self, x):
        y1 = self.cv3(self.m(self.cv1(x)))
        y2 = self.cv2(x)
        return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))


 class BottleneckCSPF(nn.Module):
    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
        super(BottleneckCSPF, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
        self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
        self.cv4 = Conv(c2, c2, 1, 1)
        self.bn = nn.BatchNorm2d(2 * c_)  # applied to cat(cv2, cv3)
        self.act = nn.LeakyReLU(0.1, inplace=True)
        self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])

    def forward(self, x):
        y1 = self.cv3(self.m(self.cv1(x)))
        y2 = self.cv2(x)
        return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))


 class Narrow(nn.Module):
    def __init__(self, c1, c2, shortcut=True, g=1):  # ch_in, ch_out, shortcut, groups
        super(Narrow, self).__init__()
        c_ = c2 // 2  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_, c2, 3, 1, g=g)
        self.add = shortcut and c1 == c2

    def forward(self, x):
        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))


 class Origami(nn.Module):  # 5-side layering
    def forward(self, x):
        y = F.pad(x, [1, 1, 1, 1])
        return torch.cat([x, y[..., :-2, 1:-1], y[..., 1:-1, :-2], y[..., 2:, 1:-1], y[..., 1:-1, 2:]], 1)


 class ConvPlus(nn.Module):  # standard convolution
    def __init__(self, c1, c2, k=3, s=1, g=1, bias=True):  # ch_in, ch_out, kernel, stride, groups
        super(ConvPlus, self).__init__()
        self.cv1 = nn.Conv2d(c1, c2, (k, 1), s, (k // 2, 0), groups=g, bias=bias)
        self.cv2 = nn.Conv2d(c1, c2, (1, k), s, (0, k // 2), groups=g, bias=bias)

    def forward(self, x):
        return self.cv1(x) + self.cv2(x)


 class SPP(nn.Module):  # Spatial pyramid pooling layer used in YOLOv3-SPP
    def __init__(self, c1, c2, k=(5, 9, 13)):
        super(SPP, self).__init__()
        c_ = c1 // 2  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1)
        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])

    def forward(self, x):
        x = self.cv1(x)
        return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))


 class Flatten(nn.Module):
    # Use after nn.AdaptiveAvgPool2d(1) to remove last 2 dimensions
    def forward(self, x):
        return x.view(x.size(0), -1)


 class Focus(nn.Module):
    # Focus wh information into c-space
    def __init__(self, c1, c2, k=1):
        super(Focus, self).__init__()
        self.conv = Conv(c1 * 4, c2, k, 1)

    def forward(self, x):  # x(b,c,w,h) -> y(b,4c,w/2,h/2)
        return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))


 class Concat(nn.Module):
    # Concatenate a list of tensors along dimension
    def __init__(self, dimension=1):
        super(Concat, self).__init__()
        self.d = dimension

    def forward(self, x):
        return torch.cat(x, self.d)


 class MixConv2d(nn.Module):
    # Mixed Depthwise Conv https://arxiv.org/abs/1907.09595
    def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True):
        super(MixConv2d, self).__init__()
        groups = len(k)
        if equal_ch:  # equal c_ per group
            i = torch.linspace(0, groups - 1E-6, c2).floor()  # c2 indices
            c_ = [(i == g).sum() for g in range(groups)]  # intermediate channels
        else:  # equal weight.numel() per group
            b = [c2] + [0] * groups
            a = np.eye(groups + 1, groups, k=-1)
            a -= np.roll(a, 1, axis=1)
            a *= np.array(k) ** 2
            a[0] = 1
            c_ = np.linalg.lstsq(a, b, rcond=None)[0].round()  # solve for equal weight indices, ax = b

        self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)])
        self.bn = nn.BatchNorm2d(c2)
        self.act = nn.LeakyReLU(0.1, inplace=True)

    def forward(self, x):
        return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1)))
--- a/models/experimental.py
+++ b/models/experimental.py
@@ -0,0 +1,48 @@
 from models.common import *


 class Sum(nn.Module):
    # weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070
    def __init__(self, n, weight=False):  # n: number of inputs
        super(Sum, self).__init__()
        self.weight = weight  # apply weights boolean
        self.iter = range(n - 1)  # iter object
        if weight:
            self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True)  # layer weights

    def forward(self, x):
        y = x[0]  # no weight
        if self.weight:
            w = torch.sigmoid(self.w) * 2
            for i in self.iter:
                y = y + x[i + 1] * w[i]
        else:
            for i in self.iter:
                y = y + x[i + 1]
        return y


 class GhostConv(nn.Module):
    def __init__(self, c1, c2, k=1, s=1, g=1, act=True):  # ch_in, ch_out, kernel, stride, groups
        super(GhostConv, self).__init__()
        c_ = c2 // 2  # hidden channels
        self.cv1 = Conv(c1, c_, k, s, g, act)
        self.cv2 = Conv(c_, c_, 5, 1, c_, act)

    def forward(self, x):
        y = self.cv1(x)
        return torch.cat([y, self.cv2(y)], 1)


 class GhostBottleneck(nn.Module):
    def __init__(self, c1, c2, k, s):
        super(GhostBottleneck, self).__init__()
        c_ = c2 // 2
        self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1),  # pw
                                  DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(),  # dw
                                  GhostConv(c_, c2, 1, 1, act=False))  # pw-linear
        self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False),
                                      Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity()

    def forward(self, x):
        return self.conv(x) + self.shortcut(x)
--- a/models/yolo.py
+++ b/models/yolo.py
@@ -0,0 +1,197 @@
 import argparse

 import yaml

 from models.common import *


 class Detect(nn.Module):
    def __init__(self, nc=80, anchors=()):  # detection layer
        super(Detect, self).__init__()
        self.stride = None  # strides computed during build
        self.nc = nc  # number of classes
        self.no = nc + 5  # number of outputs per anchor
        self.nl = len(anchors)  # number of detection layers
        self.na = len(anchors[0]) // 2  # number of anchors
        self.grid = [torch.zeros(1)] * self.nl  # init grid
        a = torch.tensor(anchors).float().view(self.nl, -1, 2)
        self.register_buffer('anchors', a)  # shape(nl,na,2)
        self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2))  # shape(nl,1,na,1,1,2)
        self.export = False  # onnx export

    def forward(self, x):
        x = x.copy()
        z = []  # inference output
        self.training |= self.export
        for i in range(self.nl):
            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()

            if not self.training:  # inference
                if self.grid[i].shape[2:4] != x[i].shape[2:4]:
                    self.grid[i] = self._make_grid(nx, ny).to(x[i].device)

                y = x[i].sigmoid()
                y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
                y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
                z.append(y.view(bs, -1, self.no))

        return x if self.training else (torch.cat(z, 1), x)

    @staticmethod
    def _make_grid(nx=20, ny=20):
        yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
        return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()


 class Model(nn.Module):
    def __init__(self, model_yaml='yolov5s.yaml'):  # cfg, number of classes, depth-width gains
        super(Model, self).__init__()
        with open(model_yaml) as f:
            self.md = yaml.load(f, Loader=yaml.FullLoader)  # model dict

        # Define model
        self.model, self.save, ch = parse_model(self.md, ch=[3])  # model, savelist, ch_out
        # print([x.shape for x in self.forward(torch.zeros(1, 3, 64, 64))])

        # Build strides, anchors
        m = self.model[-1]  # Detect()
        m.stride = torch.tensor([64 / x.shape[-2] for x in self.forward(torch.zeros(1, 3, 64, 64))])  # forward
        m.anchors /= m.stride.view(-1, 1, 1)
        self.stride = m.stride

        # Init weights, biases
        torch_utils.initialize_weights(self)
        self._initialize_biases()  # only run once
        torch_utils.model_info(self)
        print('')

    def forward(self, x, augment=False, profile=False):
        y, ts = [], 0  # outputs
        for m in self.model:
            if m.f != -1:  # if not from previous layer
                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers

            if profile:
                import thop
                o = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2  # FLOPS
                t = torch_utils.time_synchronized()
                for _ in range(10):
                    _ = m(x)
                dt = torch_utils.time_synchronized() - t
                ts += dt
                print('%10.1f%10.0f%10.1fms %-40s' % (o, m.np, dt * 100, m.type))

            x = m(x)  # run
            y.append(x if m.i in self.save else None)  # save output

        if profile:
            print(ts * 100)
        return x

    def _initialize_biases(self, cf=None):  # initialize biases into Detect(), cf is class frequency
        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
        m = self.model[-1]  # Detect() module
        for f, s in zip(m.f, m.stride):  #  from
            mi = self.model[f % m.i]
            b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
            b[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
            b[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
            mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)

    def _print_biases(self):
        m = self.model[-1]  # Detect() module
        for f in sorted([x % m.i for x in m.f]):  #  from
            b = self.model[f].bias.detach().view(m.na, -1).T  # conv.bias(255) to (3,85)
            print(('%g Conv2d.bias:' + '%10.3g' * 6) % (f, *b[:5].mean(1).tolist(), b[5:].mean()))


 def parse_model(md, ch):  # model_dict, input_channels(3)
    print('\n%3s%15s%3s%10s  %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments'))
    anchors, nc, gd, gw = md['anchors'], md['nc'], md['depth_multiple'], md['width_multiple']
    na = (len(anchors[0]) // 2)  # number of anchors
    no = na * (nc + 5)  # number of outputs = anchors * (classes + 5)

    layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out
    for i, (f, n, m, args) in enumerate(md['backbone'] + md['head']):  # from, number, module, args
        m = eval(m) if isinstance(m, str) else m  # eval strings
        for j, a in enumerate(args):
            try:
                args[j] = eval(a) if isinstance(a, str) else a  # eval strings
            except:
                pass

        n = max(round(n * gd), 1) if n > 1 else n  # depth gain
        if m in [nn.Conv2d, Conv, Bottleneck, SPP, DWConv, MixConv2d, Focus, ConvPlus, BottleneckCSP, BottleneckCSPF, BottleneckLight]:
            c1, c2 = ch[f], args[0]

            # Normal
            # if i > 0 and args[0] != no:  # channel expansion factor
            #     ex = 1.75  # exponential (default 2.0)
            #     e = math.log(c2 / ch[1]) / math.log(2)
            #     c2 = int(ch[1] * ex ** e)
            # if m != Focus:
            c2 = make_divisible(c2 * gw, 8) if c2 != no else c2

            # Experimental
            # if i > 0 and args[0] != no:  # channel expansion factor
            #     ex = 1 + gw  # exponential (default 2.0)
            #     ch1 = 32  # ch[1]
            #     e = math.log(c2 / ch1) / math.log(2)  # level 1-n
            #     c2 = int(ch1 * ex ** e)
            # if m != Focus:
            #     c2 = make_divisible(c2, 8) if c2 != no else c2

            args = [c1, c2, *args[1:]]
            if m in [BottleneckCSP, BottleneckCSPF]:
                args += [n]
                n = 1
        elif m is nn.BatchNorm2d:
            args = [ch[f]]
        elif m is Concat:
            c2 = sum([ch[x] for x in f])
        elif m is Origami:
            c2 = ch[f] * 5
        elif m is Detect:
            f = f or list(reversed([(-1 if j == i else j - 1) for j, x in enumerate(ch) if x == no]))
        else:
            c2 = ch[f]

        m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args)  # module
        t = str(m)[8:-2].replace('__main__.', '')  # module type
        np = sum([x.numel() for x in m_.parameters()])  # number params
        m_.i, m_.f, m_.type, m_.np = i, f, t, np  # attach index, 'from' index, type, number params
        print('%3s%15s%3s%10.0f  %-40s%-30s' % (i, f, n, np, t, args))  # print
        save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist
        layers.append(m_)
        ch.append(c2)
    return nn.Sequential(*layers), sorted(save), ch


 if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml')
    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    opt = parser.parse_args()
    opt.cfg = glob.glob('./**/' + opt.cfg, recursive=True)[0]  # find file
    device = torch_utils.select_device(opt.device)

    # Create model
    model = Model(opt.cfg).to(device)
    model.train()

    # Profile
    # img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 640, 640).to(device)
    # y = model(img, profile=True)
    # print([y[0].shape] + [x.shape for x in y[1]])

    # ONNX export
    # model.model[-1].export = True
    # torch.onnx.export(model, img, f.replace('.yaml', '.onnx'), verbose=True, opset_version=11)

    # Tensorboard
    # from torch.utils.tensorboard import SummaryWriter
    # tb_writer = SummaryWriter()
    # print("Run 'tensorboard --logdir=models/runs' to view tensorboard at http://localhost:6006/")
    # tb_writer.add_graph(model.model, img)  # add model to tensorboard
    # tb_writer.add_image('test', img[0], dataformats='CWH')  # add model to tensorboard
--- a/models/yolov3-spp.yaml
+++ b/models/yolov3-spp.yaml
@@ -0,0 +1,55 @@
 # parameters
 nc: 80  # number of classes
 depth_multiple: 1.0  # expand model depth
 width_multiple: 1.0  # expand layer channels

 # anchors
 anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 # darknet53 backbone
 backbone:
  # [from, number, module, args]
  [[-1, 1, Conv, [32, 3, 1]],  # 0
   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
   [-1, 1, Bottleneck, [64]],
   [-1, 1, Conv, [128, 3, 2]],  # 3-P2/4
   [-1, 2, Bottleneck, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 5-P3/8
   [-1, 8, Bottleneck, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 7-P4/16
   [-1, 8, Bottleneck, [512]],
   [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32
   [-1, 4, Bottleneck, [1024]],  # 10
  ]

 # yolov3-spp head
 # na = len(anchors[0])
 head:
  [[-1, 1, Bottleneck, [1024, False]],  # 11
   [-1, 1, SPP, [512, [5, 9, 13]]],
   [-1, 1, Conv, [1024, 3, 1]],
   [-1, 1, Conv, [512, 1, 1]],
   [-1, 1, Conv, [1024, 3, 1]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 16 (P5/32-large)

   [-3, 1, Conv, [256, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 8], 1, Concat, [1]],  # cat backbone P4
   [-1, 1, Bottleneck, [512, False]],
   [-1, 1, Bottleneck, [512, False]],
   [-1, 1, Conv, [256, 1, 1]],
   [-1, 1, Conv, [512, 3, 1]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 24 (P4/16-medium)

   [-3, 1, Conv, [128, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P3
   [-1, 1, Bottleneck, [256, False]],
   [-1, 2, Bottleneck, [256, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 30 (P3/8-small)

   [[], 1, Detect, [nc, anchors]],   # Detect(P3, P4, P5)
  ]
--- a/models/yolov3-spp_csp.yaml
+++ b/models/yolov3-spp_csp.yaml
@@ -0,0 +1,55 @@
 # parameters
 nc: 80  # number of classes
 depth_multiple: 1.0  # expand model depth
 width_multiple: 1.0  # expand layer channels

 # anchors
 anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 # darknet53 backbone
 backbone:
  # [from, number, module, args]
  [[-1, 1, Conv, [32, 3, 1]],  # 0
   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
   [-1, 1, BottleneckCSP, [64]],
   [-1, 1, Conv, [128, 3, 2]],  # 3-P2/4
   [-1, 2, BottleneckCSP, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 5-P3/8
   [-1, 8, BottleneckCSP, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 7-P4/16
   [-1, 8, BottleneckCSP, [512]],
   [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32
   [-1, 4, BottleneckCSP, [1024]],  # 10
  ]

 # yolov3-spp head
 # na = len(anchors[0])
 head:
  [[-1, 1, Bottleneck, [1024, False]],  # 11
   [-1, 1, SPP, [512, [5, 9, 13]]],
   [-1, 1, Conv, [1024, 3, 1]],
   [-1, 1, Conv, [512, 1, 1]],
   [-1, 1, Conv, [1024, 3, 1]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 16 (P5/32-large)

   [-3, 1, Conv, [256, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 8], 1, Concat, [1]],  # cat backbone P4
   [-1, 1, Bottleneck, [512, False]],
   [-1, 1, Bottleneck, [512, False]],
   [-1, 1, Conv, [256, 1, 1]],
   [-1, 1, Conv, [512, 3, 1]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 24 (P4/16-medium)

   [-3, 1, Conv, [128, 1, 1]],
   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P3
   [-1, 1, Bottleneck, [256, False]],
   [-1, 2, Bottleneck, [256, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 30 (P3/8-small)

   [[], 1, Detect, [nc, anchors]],   # Detect(P3, P4, P5)
  ]
--- a/models/yolov5l.yaml
+++ b/models/yolov5l.yaml
@@ -0,0 +1,45 @@
 # parameters
 nc: 80  # number of classes
 depth_multiple: 1.0  # model depth multiple
 width_multiple: 1.0  # layer channel multiple

 # anchors
 anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 # yolov5 backbone
 backbone:
  # [from, number, module, args]
  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
   [-1, 3, Bottleneck, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
   [-1, 9, Bottleneck, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
   [-1, 9, Bottleneck, [512]],
   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
   [-1, 1, SPP, [1024, [5, 9, 13]]],
   [-1, 3, Bottleneck, [1024]],  # 10
  ]

 # yolov5 head
 head:
  [[-1, 3, Bottleneck, [1024, False]],  # 11
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 12 (P5/32-large)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
   [-1, 1, Conv, [512, 1, 1]],
   [-1, 3, Bottleneck, [512, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 17 (P4/16-medium)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
   [-1, 1, Conv, [256, 1, 1]],
   [-1, 3, Bottleneck, [256, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 22 (P3/8-small)

   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]
--- a/models/yolov5m.yaml
+++ b/models/yolov5m.yaml
@@ -0,0 +1,46 @@
 # parameters
 nc: 80  # number of classes
 depth_multiple: 0.67  # model depth multiple
 width_multiple: 0.75  # layer channel multiple

 # anchors
 anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 # yolov5 backbone
 backbone:
  # [from, number, module, args]
  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
   [-1, 3, Bottleneck, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
   [-1, 9, Bottleneck, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
   [-1, 9, Bottleneck, [512]],
   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
   [-1, 1, SPP, [1024, [5, 9, 13]]],
   [-1, 3, Bottleneck, [1024]],  # 10
  ]

 # yolov5 head
 head:
  [[-1, 3, Bottleneck, [1024, False]],  # 11
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 12 (P5/32-large)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
   [-1, 1, Conv, [512, 1, 1]],
   [-1, 3, Bottleneck, [512, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 17 (P4/16-medium)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
   [-1, 1, Conv, [256, 1, 1]],
   [-1, 3, Bottleneck, [256, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 22 (P3/8-small)

   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]

--- a/models/yolov5s.yaml
+++ b/models/yolov5s.yaml
@@ -0,0 +1,45 @@
 # parameters
 nc: 80  # number of classes
 depth_multiple: 0.33  # model depth multiple
 width_multiple: 0.50  # layer channel multiple

 # anchors
 anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 # yolov5 backbone
 backbone:
  # [from, number, module, args]
  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
   [-1, 3, Bottleneck, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
   [-1, 9, Bottleneck, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
   [-1, 9, Bottleneck, [512]],
   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
   [-1, 1, SPP, [1024, [5, 9, 13]]],
   [-1, 3, Bottleneck, [1024]],  # 10
  ]

 # yolov5 head
 head:
  [[-1, 3, Bottleneck, [1024, False]],  # 11
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 12 (P5/32-large)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
   [-1, 1, Conv, [512, 1, 1]],
   [-1, 3, Bottleneck, [512, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 17 (P4/16-medium)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
   [-1, 1, Conv, [256, 1, 1]],
   [-1, 3, Bottleneck, [256, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 22 (P3/8-small)

   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]
--- a/models/yolov5s_csp.yaml
+++ b/models/yolov5s_csp.yaml
@@ -0,0 +1,46 @@
 # parameters
 nc: 80  # number of classes
 depth_multiple: 0.33  # model depth multiple
 width_multiple: 0.50  # layer channel multiple

 # anchors
 anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 # yolov5 backbone
 backbone:
  # [from, number, module, args]
  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
   [-1, 3, BottleneckCSP, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
   [-1, 9, BottleneckCSP, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
   [-1, 9, BottleneckCSP, [512]],
   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
   [-1, 1, SPP, [1024, [5, 9, 13]]],
   [-1, 3, BottleneckCSP, [1024]],  # 10
  ]

 # yolov5 head
 head:
  [[-1, 3, Bottleneck, [1024, False]],  # 11
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 12 (P5/32-large)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
   [-1, 1, Conv, [512, 1, 1]],
   [-1, 3, Bottleneck, [512, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 17 (P4/16-medium)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
   [-1, 1, Conv, [256, 1, 1]],
   [-1, 3, Bottleneck, [256, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 22 (P3/8-small)

   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]

--- a/models/yolov5s_csp1.yaml
+++ b/models/yolov5s_csp1.yaml
@@ -0,0 +1,46 @@
 # parameters
 nc: 80  # number of classes
 depth_multiple: 0.33  # model depth multiple
 width_multiple: 0.50  # layer channel multiple

 # anchors
 anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 # yolov5 backbone
 backbone:
  # [from, number, module, args]
  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
   [-1, 3, BottleneckCSP, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
   [-1, 9, BottleneckCSP, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
   [-1, 9, BottleneckCSP, [512]],
   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
   [-1, 1, SPP, [1024, [5, 9, 13]]],
   [-1, 6, BottleneckCSP, [1024]],  # 10
  ]

 # yolov5 head
 head:
  [[-1, 3, Bottleneck, [1024, False]],  # 11
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 12 (P5/32-large)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
   [-1, 1, Conv, [512, 1, 1]],
   [-1, 3, Bottleneck, [512, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 17 (P4/16-medium)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
   [-1, 1, Conv, [256, 1, 1]],
   [-1, 3, Bottleneck, [256, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 22 (P3/8-small)

   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]

--- a/models/yolov5s_csp2.yaml
+++ b/models/yolov5s_csp2.yaml
@@ -0,0 +1,46 @@
 # parameters
 nc: 80  # number of classes
 depth_multiple: 0.33  # model depth multiple
 width_multiple: 0.50  # layer channel multiple

 # anchors
 anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 # yolov5 backbone
 backbone:
  # [from, number, module, args]
  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
   [-1, 3, BottleneckCSP, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
   [-1, 9, BottleneckCSP, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
   [-1, 9, BottleneckCSP, [512]],
   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
   [-1, 1, SPP, [1024, [5, 9, 13]]],
   [-1, 6, BottleneckCSP, [1024]],  # 10
  ]

 # yolov5 head
 head:
  [[-1, 3, BottleneckCSPF, [1024]],  # 11
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 12 (P5/32-large)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
   [-1, 1, Conv, [512, 1, 1]],
   [-1, 3, BottleneckCSPF, [512]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 17 (P4/16-medium)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
   [-1, 1, Conv, [256, 1, 1]],
   [-1, 3, BottleneckCSPF, [256]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 22 (P3/8-small)

   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]

--- a/models/yolov5x.yaml
+++ b/models/yolov5x.yaml
@@ -0,0 +1,45 @@
 # parameters
 nc: 80  # number of classes
 depth_multiple: 1.33  # model depth multiple
 width_multiple: 1.25  # layer channel multiple

 # anchors
 anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 # yolov5 backbone
 backbone:
  # [from, number, module, args]
  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
   [-1, 3, Bottleneck, [128]],
   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
   [-1, 9, Bottleneck, [256]],
   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
   [-1, 9, Bottleneck, [512]],
   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
   [-1, 1, SPP, [1024, [5, 9, 13]]],
   [-1, 3, Bottleneck, [1024]],  # 10
  ]

 # yolov5 head
 head:
  [[-1, 3, Bottleneck, [1024]],  # 11
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 12 (P5/32-large)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
   [-1, 1, Conv, [512, 1, 1]],
   [-1, 3, Bottleneck, [512, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 17 (P4/16-medium)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
   [-1, 1, Conv, [256, 1, 1]],
   [-1, 3, Bottleneck, [256, False]],
   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1, 0]],  # 22 (P3/8-small)

   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]
--- a/requirements.txt
+++ b/requirements.txt
@@ -0,0 +1,22 @@
 # pip install -U -r requirements.txt
 numpy==1.17
 opencv-python
 torch >= 1.5
 matplotlib
 pycocotools
 tqdm
 pillow
 tensorboard
 pyyaml


 # Nvidia Apex (optional) for mixed precision training --------------------------
 # git clone https://github.com/NVIDIA/apex && cd apex && pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" . --user && cd .. && rm -rf apex

 # Conda commands (in place of pip) ---------------------------------------------
 # conda update -yn base -c defaults conda
 # conda install -yc anaconda numpy opencv matplotlib tqdm pillow ipython
 # conda install -yc conda-forge scikit-image pycocotools tensorboard
 # conda install -yc spyder-ide spyder-line-profiler
 # conda install -yc pytorch pytorch torchvision
 # conda install -yc conda-forge protobuf numpy && pip install onnx  # https://github.com/onnx/onnx#linux-and-macos
--- a/test.py
+++ b/test.py
@@ -0,0 +1,268 @@
 import argparse
 import json

 import yaml
 from torch.utils.data import DataLoader

 from utils.datasets import *
 from utils.utils import *


 def test(data,
         weights=None,
         batch_size=16,
         imgsz=640,
         conf_thres=0.001,
         iou_thres=0.6,  # for nms
         save_json=False,
         single_cls=False,
         augment=False,
         model=None,
         dataloader=None,
         multi_label=True,
         verbose=False):  # 0 fast, 1 accurate
    # Initialize/load model and set device
    if model is None:
        device = torch_utils.select_device(opt.device, batch_size=batch_size)

        # Remove previous
        for f in glob.glob('test_batch*.jpg'):
            os.remove(f)

        # Load model
        google_utils.attempt_download(weights)
        model = torch.load(weights, map_location=device)['model']
        torch_utils.model_info(model)

        # Fuse
        # model.fuse()
        model.to(device)

        if device.type != 'cpu' and torch.cuda.device_count() > 1:
            model = nn.DataParallel(model)

        training = False
    else:  # called by train.py
        device = next(model.parameters()).device  # get model device
        training = True

    # Configure run
    with open(data) as f:
        data = yaml.load(f, Loader=yaml.FullLoader)  # model dict
    nc = 1 if single_cls else int(data['nc'])  # number of classes
    iouv = torch.linspace(0.5, 0.95, 10).to(device)  # iou vector for mAP@0.5:0.95
    # iouv = iouv[0].view(1)  # comment for mAP@0.5:0.95
    niou = iouv.numel()

    # Dataloader
    if dataloader is None:
        path = data['test'] if opt.task == 'test' else data['val']  # path to val/test images
        dataset = LoadImagesAndLabels(path, imgsz, batch_size, rect=True, single_cls=opt.single_cls)
        batch_size = min(batch_size, len(dataset))
        dataloader = DataLoader(dataset,
                                batch_size=batch_size,
                                num_workers=min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8]),
                                pin_memory=True,
                                collate_fn=dataset.collate_fn)

    seen = 0
    model.eval()
    _ = model(torch.zeros((1, 3, imgsz, imgsz), device=device)) if device.type != 'cpu' else None  # run once
    coco91class = coco80_to_coco91_class()
    s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', 'mAP@.5', 'mAP@.5:.95')
    p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0.
    loss = torch.zeros(3, device=device)
    jdict, stats, ap, ap_class = [], [], [], []
    for batch_i, (imgs, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)):
        imgs = imgs.to(device).float() / 255.0  # uint8 to float32, 0 - 255 to 0.0 - 1.0
        targets = targets.to(device)
        nb, _, height, width = imgs.shape  # batch size, channels, height, width
        whwh = torch.Tensor([width, height, width, height]).to(device)

        # Disable gradients
        with torch.no_grad():
            # Run model
            t = torch_utils.time_synchronized()
            inf_out, train_out = model(imgs, augment=augment)  # inference and training outputs
            t0 += torch_utils.time_synchronized() - t

            # Compute loss
            if training:  # if model has loss hyperparameters
                loss += compute_loss(train_out, targets, model)[1][:3]  # GIoU, obj, cls

            # Run NMS
            t = torch_utils.time_synchronized()
            output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, multi_label=multi_label)
            t1 += torch_utils.time_synchronized() - t

        # Statistics per image
        for si, pred in enumerate(output):
            labels = targets[targets[:, 0] == si, 1:]
            nl = len(labels)
            tcls = labels[:, 0].tolist() if nl else []  # target class
            seen += 1

            if pred is None:
                if nl:
                    stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls))
                continue

            # Append to text file
            # with open('test.txt', 'a') as file:
            #    [file.write('%11.5g' * 7 % tuple(x) + '\n') for x in pred]

            # Clip boxes to image bounds
            clip_coords(pred, (height, width))

            # Append to pycocotools JSON dictionary
            if save_json:
                # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ...
                image_id = int(Path(paths[si]).stem.split('_')[-1])
                box = pred[:, :4].clone()  # xyxy
                scale_coords(imgs[si].shape[1:], box, shapes[si][0], shapes[si][1])  # to original shape
                box = xyxy2xywh(box)  # xywh
                box[:, :2] -= box[:, 2:] / 2  # xy center to top-left corner
                for p, b in zip(pred.tolist(), box.tolist()):
                    jdict.append({'image_id': image_id,
                                  'category_id': coco91class[int(p[5])],
                                  'bbox': [round(x, 3) for x in b],
                                  'score': round(p[4], 5)})

            # Assign all predictions as incorrect
            correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device)
            if nl:
                detected = []  # target indices
                tcls_tensor = labels[:, 0]

                # target boxes
                tbox = xywh2xyxy(labels[:, 1:5]) * whwh

                # Per target class
                for cls in torch.unique(tcls_tensor):
                    ti = (cls == tcls_tensor).nonzero().view(-1)  # prediction indices
                    pi = (cls == pred[:, 5]).nonzero().view(-1)  # target indices

                    # Search for detections
                    if pi.shape[0]:
                        # Prediction to target ious
                        ious, i = box_iou(pred[pi, :4], tbox[ti]).max(1)  # best ious, indices

                        # Append detections
                        for j in (ious > iouv[0]).nonzero():
                            d = ti[i[j]]  # detected target
                            if d not in detected:
                                detected.append(d)
                                correct[pi[j]] = ious[j] > iouv  # iou_thres is 1xn
                                if len(detected) == nl:  # all targets already located in image
                                    break

            # Append statistics (correct, conf, pcls, tcls)
            stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls))

        # Plot images
        if batch_i < 1:
            f = 'test_batch%g_gt.jpg' % batch_i  # filename
            plot_images(imgs, targets, paths, f, model.names)  # ground truth
            f = 'test_batch%g_pred.jpg' % batch_i
            plot_images(imgs, output_to_target(output, width, height), paths, f, model.names)  # predictions

    # Compute statistics
    stats = [np.concatenate(x, 0) for x in zip(*stats)]  # to numpy
    if len(stats):
        p, r, ap, f1, ap_class = ap_per_class(*stats)
        p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean(1)  # [P, R, AP@0.5, AP@0.5:0.95]
        mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean()
        nt = np.bincount(stats[3].astype(np.int64), minlength=nc)  # number of targets per class
    else:
        nt = torch.zeros(1)

    # Print results
    pf = '%20s' + '%12.3g' * 6  # print format
    print(pf % ('all', seen, nt.sum(), mp, mr, map50, map))

    # Print results per class
    if verbose and nc > 1 and len(stats):
        for i, c in enumerate(ap_class):
            print(pf % (model.names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i]))

    # Print speeds
    t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size)  # tuple
    if not training:
        print('Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t)

    # Save JSON
    if save_json and map50 and len(jdict):
        imgIds = [int(Path(x).stem.split('_')[-1]) for x in dataloader.dataset.img_files]
        f = 'detections_val2017_%s_results.json' % \
            (weights.split(os.sep)[-1].replace('.pt', '') if weights else '')  # filename
        print('\nCOCO mAP with pycocotools... saving %s...' % f)
        with open(f, 'w') as file:
            json.dump(jdict, file)

        try:
            from pycocotools.coco import COCO
            from pycocotools.cocoeval import COCOeval

            # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
            cocoGt = COCO(glob.glob('../coco/annotations/instances_val*.json')[0])  # initialize COCO ground truth api
            cocoDt = cocoGt.loadRes(f)  # initialize COCO pred api

            cocoEval = COCOeval(cocoGt, cocoDt, 'bbox')
            cocoEval.params.imgIds = imgIds  # [:32]  # only evaluate these images
            cocoEval.evaluate()
            cocoEval.accumulate()
            cocoEval.summarize()
            map, map50 = cocoEval.stats[:2]  # update to pycocotools results (mAP@0.5:0.95, mAP@0.5)
        except:
            print('WARNING: pycocotools must be installed with numpy==1.17 to run correctly. '
                  'See https://github.com/cocodataset/cocoapi/issues/356')

    # Return results
    maps = np.zeros(nc) + map
    for i, c in enumerate(ap_class):
        maps[c] = ap[i]
    return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t


 if __name__ == '__main__':
    parser = argparse.ArgumentParser(prog='test.py')
    parser.add_argument('--weights', type=str, default='weights/yolov5s.pt', help='model.pt path')
    parser.add_argument('--data', type=str, default='data/coco.yaml', help='*.data path')
    parser.add_argument('--batch-size', type=int, default=16, help='size of each image batch')
    parser.add_argument('--img-size', type=int, default=640, help='inference size (pixels)')
    parser.add_argument('--conf-thres', type=float, default=0.001, help='object confidence threshold')
    parser.add_argument('--iou-thres', type=float, default=0.65, help='IOU threshold for NMS')
    parser.add_argument('--save-json', action='store_true', help='save a cocoapi-compatible JSON results file')
    parser.add_argument('--task', default='val', help="'val', 'test', 'study'")
    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset')
    parser.add_argument('--augment', action='store_true', help='augmented inference')
    parser.add_argument('--verbose', action='store_true', help='report mAP by class')
    opt = parser.parse_args()
    opt.save_json = opt.save_json or opt.data.endswith(os.sep + 'coco.yaml')
    opt.data = glob.glob('./**/' + opt.data, recursive=True)[0]  # find file
    print(opt)

    # task = 'val', 'test', 'study'
    if opt.task == 'val':  # (default) run normally
        test(opt.data,
             opt.weights,
             opt.batch_size,
             opt.img_size,
             opt.conf_thres,
             opt.iou_thres,
             opt.save_json,
             opt.single_cls,
             opt.augment)

    elif opt.task == 'study':  # run over a range of settings and save/plot
        for weights in ['yolov5s.pt', 'yolov5m.pt', 'yolovl.p5', 'yolov5x.pt', 'yolov3-spp.pt']:
            f = 'study_%s_%s.txt' % (Path(opt.data).stem, Path(weights).stem)  # filename to save to
            x = list(range(256, 1024, 32))  # x axis
            y = []  # y axis
            for i in x:  # img-size
                print('\nRunning %s point %s...' % (f, i))
                r, _, t = test(opt.data, weights, opt.batch_size, i, opt.conf_thres, opt.iou_thres, opt.save_json)
                y.append(r + t)  # results and times
            np.savetxt(f, y, fmt='%10.4g')  # save
            plot_study_txt(f, x)  # plot
--- a/train.py
+++ b/train.py
@@ -0,0 +1,443 @@
 import argparse

 import torch.distributed as dist
 import torch.nn.functional as F
 import torch.optim as optim
 import torch.optim.lr_scheduler as lr_scheduler
 import yaml
 from torch.utils.tensorboard import SummaryWriter

 import test  # import test.py to get mAP after each epoch
 from models.yolo import Model
 from utils.datasets import *
 from utils.utils import *

 mixed_precision = True
 try:  # Mixed precision training https://github.com/NVIDIA/apex
    from apex import amp
 except:
    print('Apex recommended for faster mixed precision training: https://github.com/NVIDIA/apex')
    mixed_precision = False  # not installed

 wdir = 'weights' + os.sep  # weights dir
 last = wdir + 'last.pt'
 best = wdir + 'best.pt'
 results_file = 'results.txt'

 # Hyperparameters
 hyp = {'lr0': 0.01,  # initial learning rate (SGD=1E-2, Adam=1E-3)
       'momentum': 0.937,  # SGD momentum
       'weight_decay': 5e-4,  # optimizer weight decay
       'giou': 0.05,  # giou loss gain
       'cls': 0.58,  # cls loss gain
       'cls_pw': 1.0,  # cls BCELoss positive_weight
       'obj': 1.0,  # obj loss gain (*=img_size/320 if img_size != 320)
       'obj_pw': 1.0,  # obj BCELoss positive_weight
       'iou_t': 0.20,  # iou training threshold
       'anchor_t': 4.0,  # anchor-multiple threshold
       'fl_gamma': 0.0,  # focal loss gamma (efficientDet default is gamma=1.5)
       'hsv_h': 0.014,  # image HSV-Hue augmentation (fraction)
       'hsv_s': 0.68,  # image HSV-Saturation augmentation (fraction)
       'hsv_v': 0.36,  # image HSV-Value augmentation (fraction)
       'degrees': 0.0,  # image rotation (+/- deg)
       'translate': 0.0,  # image translation (+/- fraction)
       'scale': 0.5,  # image scale (+/- gain)
       'shear': 0.0}  # image shear (+/- deg)
 print(hyp)

 # Overwrite hyp with hyp*.txt (optional)
 f = glob.glob('hyp*.txt')
 if f:
    print('Using %s' % f[0])
    for k, v in zip(hyp.keys(), np.loadtxt(f[0])):
        hyp[k] = v

 # Print focal loss if gamma > 0
 if hyp['fl_gamma']:
    print('Using FocalLoss(gamma=%g)' % hyp['fl_gamma'])


 def train(hyp):
    epochs = opt.epochs  # 300
    batch_size = opt.batch_size  # 64
    weights = opt.weights  # initial training weights

    # Configure
    init_seeds()
    with open(opt.data) as f:
        data_dict = yaml.load(f, Loader=yaml.FullLoader)  # model dict
    train_path = data_dict['train']
    test_path = data_dict['val']
    nc = 1 if opt.single_cls else int(data_dict['nc'])  # number of classes

    # Remove previous results
    for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
        os.remove(f)

    # Create model
    model = Model(opt.cfg).to(device)

    # Image sizes
    gs = int(max(model.stride))  # grid size (max stride)
    if any(x % gs != 0 for x in opt.img_size):
        print('WARNING: --img-size %g,%g must be multiple of %s max stride %g' % (*opt.img_size, opt.cfg, gs))
    imgsz, imgsz_test = [make_divisible(x, gs) for x in opt.img_size]  # image sizes (train, test)

    # Optimizer
    nbs = 64  # nominal batch size
    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
    pg0, pg1, pg2 = [], [], []  # optimizer parameter groups
    for k, v in model.named_parameters():
        if v.requires_grad:
            if '.bias' in k:
                pg2.append(v)  # biases
            elif '.weight' in k and '.bn' not in k:
                pg1.append(v)  # apply weight decay
            else:
                pg0.append(v)  # all else

    optimizer = optim.Adam(pg0, lr=hyp['lr0']) if opt.adam else \
        optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
    optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']})  # add pg1 with weight_decay
    optimizer.add_param_group({'params': pg2})  # add pg2 (biases)
    print('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0)))
    del pg0, pg1, pg2

    # Load Model
    google_utils.attempt_download(weights)
    start_epoch, best_fitness = 0, 0.0
    if weights.endswith('.pt'):  # pytorch format
        chkpt = torch.load(weights, map_location=device)

        # load model
        try:
            chkpt['model'] = \
                {k: v for k, v in chkpt['model'].state_dict().items() if model.state_dict()[k].numel() == v.numel()}
            model.load_state_dict(chkpt['model'], strict=False)
        except KeyError as e:
            s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s." \
                % (opt.weights, opt.cfg, opt.weights)
            raise KeyError(s) from e

        # load optimizer
        if chkpt['optimizer'] is not None:
            optimizer.load_state_dict(chkpt['optimizer'])
            best_fitness = chkpt['best_fitness']

        # load results
        if chkpt.get('training_results') is not None:
            with open(results_file, 'w') as file:
                file.write(chkpt['training_results'])  # write results.txt

        start_epoch = chkpt['epoch'] + 1
        del chkpt

    # Mixed precision training https://github.com/NVIDIA/apex
    if mixed_precision:
        model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)

    # Scheduler https://arxiv.org/pdf/1812.01187.pdf
    lf = lambda x: (((1 + math.cos(x * math.pi / epochs)) / 2) ** 1.0) * 0.9 + 0.1  # cosine
    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
    scheduler.last_epoch = start_epoch - 1  # do not move
    # https://discuss.pytorch.org/t/a-problem-occured-when-resuming-an-optimizer/28822
    # plot_lr_scheduler(optimizer, scheduler, epochs)

    # Initialize distributed training
    if device.type != 'cpu' and torch.cuda.device_count() > 1 and torch.distributed.is_available():
        dist.init_process_group(backend='nccl',  # distributed backend
                                init_method='tcp://127.0.0.1:9999',  # init method
                                world_size=1,  # number of nodes
                                rank=0)  # node rank
        model = torch.nn.parallel.DistributedDataParallel(model)

    # Dataset
    dataset = LoadImagesAndLabels(train_path, imgsz, batch_size,
                                  augment=True,
                                  hyp=hyp,  # augmentation hyperparameters
                                  rect=opt.rect,  # rectangular training
                                  cache_images=opt.cache_images,
                                  single_cls=opt.single_cls)
    mlc = np.concatenate(dataset.labels, 0)[:, 0].max()  # max label class
    assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Correct your labels or your model.' % (mlc, nc, opt.cfg)

    # Dataloader
    batch_size = min(batch_size, len(dataset))
    nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
    dataloader = torch.utils.data.DataLoader(dataset,
                                             batch_size=batch_size,
                                             num_workers=nw,
                                             shuffle=not opt.rect,  # Shuffle=True unless rectangular training is used
                                             pin_memory=True,
                                             collate_fn=dataset.collate_fn)

    # Testloader
    testloader = torch.utils.data.DataLoader(LoadImagesAndLabels(test_path, imgsz_test, batch_size,
                                                                 hyp=hyp,
                                                                 rect=True,
                                                                 cache_images=opt.cache_images,
                                                                 single_cls=opt.single_cls),
                                             batch_size=batch_size,
                                             num_workers=nw,
                                             pin_memory=True,
                                             collate_fn=dataset.collate_fn)

    # Model parameters
    hyp['cls'] *= nc / 80.  # scale coco-tuned hyp['cls'] to current dataset
    model.nc = nc  # attach number of classes to model
    model.hyp = hyp  # attach hyperparameters to model
    model.gr = 1.0  # giou loss ratio (obj_loss = 1.0 or giou)
    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device)  # attach class weights
    model.names = data_dict['names']

    # class frequency
    labels = np.concatenate(dataset.labels, 0)
    c = torch.tensor(labels[:, 0])  # classes
    # cf = torch.bincount(c.long(), minlength=nc) + 1.
    # model._initialize_biases(cf.to(device))
    plot_labels(labels)
    tb_writer.add_histogram('classes', c, 0)

    # Exponential moving average
    ema = torch_utils.ModelEMA(model)

    # Start training
    t0 = time.time()
    nb = len(dataloader)  # number of batches
    n_burn = max(3 * nb, 1e3)  # burn-in iterations, max(3 epochs, 1k iterations)
    maps = np.zeros(nc)  # mAP per class
    results = (0, 0, 0, 0, 0, 0, 0)  # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
    print('Image sizes %g train, %g test' % (imgsz, imgsz_test))
    print('Using %g dataloader workers' % nw)
    print('Starting training for %g epochs...' % epochs)
    # torch.autograd.set_detect_anomaly(True)
    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
        model.train()

        # Update image weights (optional)
        if dataset.image_weights:
            w = model.class_weights.cpu().numpy() * (1 - maps) ** 2  # class weights
            image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w)
            dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n)  # rand weighted idx

        mloss = torch.zeros(4, device=device)  # mean losses
        print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size'))
        pbar = tqdm(enumerate(dataloader), total=nb)  # progress bar
        for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------
            ni = i + nb * epoch  # number integrated batches (since train start)
            imgs = imgs.to(device).float() / 255.0  # uint8 to float32, 0 - 255 to 0.0 - 1.0

            # Burn-in
            if ni <= n_burn:
                xi = [0, n_burn]  # x interp
                # model.gr = np.interp(ni, xi, [0.0, 1.0])  # giou loss ratio (obj_loss = 1.0 or giou)
                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
                for j, x in enumerate(optimizer.param_groups):
                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
                    x['lr'] = np.interp(ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
                    if 'momentum' in x:
                        x['momentum'] = np.interp(ni, xi, [0.9, hyp['momentum']])

            # Multi-scale
            if True:
                imgsz = random.randrange(640, 640 + gs) // gs * gs
                sf = imgsz / max(imgs.shape[2:])  # scale factor
                if sf != 1:
                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
                    imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)

            # Forward
            pred = model(imgs)

            # Loss
            loss, loss_items = compute_loss(pred, targets.to(device), model)
            if not torch.isfinite(loss):
                print('WARNING: non-finite loss, ending training ', loss_items)
                return results

            # Backward
            if mixed_precision:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
            else:
                loss.backward()

            # Optimize
            if ni % accumulate == 0:
                optimizer.step()
                optimizer.zero_grad()
                ema.update(model)

            # Print
            mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
            mem = '%.3gG' % (torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available() else 0)  # (GB)
            s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgsz)
            pbar.set_description(s)

            # Plot
            if ni < 3:
                f = 'train_batch%g.jpg' % i  # filename
                res = plot_images(images=imgs, targets=targets, paths=paths, fname=f)
                if tb_writer:
                    tb_writer.add_image(f, res, dataformats='HWC', global_step=epoch)
                    # tb_writer.add_graph(model, imgs)  # add model to tensorboard

            # end batch ------------------------------------------------------------------------------------------------

        # Scheduler
        scheduler.step()

        # mAP
        ema.update_attr(model)
        final_epoch = epoch + 1 == epochs
        if not opt.notest or final_epoch:  # Calculate mAP
            results, maps, times = test.test(opt.data,
                                      batch_size=batch_size,
                                      imgsz=imgsz_test,
                                      save_json=final_epoch and opt.data.endswith(os.sep + 'coco.yaml'),
                                      model=ema.ema,
                                      single_cls=opt.single_cls,
                                      dataloader=testloader,
                                      multi_label=ni > n_burn)

        # Write
        with open(results_file, 'a') as f:
            f.write(s + '%10.4g' * 7 % results + '\n')  # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
        if len(opt.name) and opt.bucket:
            os.system('gsutil cp results.txt gs://%s/results/results%s.txt' % (opt.bucket, opt.name))

        # Tensorboard
        if tb_writer:
            tags = ['train/giou_loss', 'train/obj_loss', 'train/cls_loss',
                    'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/F1',
                    'val/giou_loss', 'val/obj_loss', 'val/cls_loss']
            for x, tag in zip(list(mloss[:-1]) + list(results), tags):
                tb_writer.add_scalar(tag, x, epoch)

        # Update best mAP
        fi = fitness(np.array(results).reshape(1, -1))  # fitness_i = weighted combination of [P, R, mAP, F1]
        if fi > best_fitness:
            best_fitness = fi

        # Save model
        save = (not opt.nosave) or (final_epoch and not opt.evolve)
        if save:
            with open(results_file, 'r') as f:  # create checkpoint
                chkpt = {'epoch': epoch,
                         'best_fitness': best_fitness,
                         'training_results': f.read(),
                         'model': ema.ema.module if hasattr(model, 'module') else ema.ema,
                         'optimizer': None if final_epoch else optimizer.state_dict()}

            # Save last, best and delete
            torch.save(chkpt, last)
            if (best_fitness == fi) and not final_epoch:
                torch.save(chkpt, best)
            del chkpt

        # end epoch ----------------------------------------------------------------------------------------------------
    # end training

    n = opt.name
    if len(n):
        n = '_' + n if not n.isnumeric() else n
        fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n
        for f1, f2 in zip([wdir + 'last.pt', wdir + 'best.pt', 'results.txt'], [flast, fbest, fresults]):
            if os.path.exists(f1):
                os.rename(f1, f2)  # rename
                ispt = f2.endswith('.pt')  # is *.pt
                strip_optimizer(f2) if ispt else None  # strip optimizer
                os.system('gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket and ispt else None  # upload

    if not opt.evolve:
        plot_results()  # save as results.png
    print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
    dist.destroy_process_group() if torch.cuda.device_count() > 1 else None
    torch.cuda.empty_cache()
    return results


 if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--epochs', type=int, default=300)
    parser.add_argument('--batch-size', type=int, default=16)
    parser.add_argument('--cfg', type=str, default='models/yolov5s.yaml', help='*.cfg path')
    parser.add_argument('--data', type=str, default='data/coco128.yaml', help='*.data path')
    parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='train,test sizes')
    parser.add_argument('--rect', action='store_true', help='rectangular training')
    parser.add_argument('--resume', action='store_true', help='resume training from last.pt')
    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
    parser.add_argument('--notest', action='store_true', help='only test final epoch')
    parser.add_argument('--evolve', action='store_true', help='evolve hyperparameters')
    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
    parser.add_argument('--cache-images', action='store_true', help='cache images for faster training')
    parser.add_argument('--weights', type=str, default='', help='initial weights path')
    parser.add_argument('--name', default='', help='renames results.txt to results_name.txt if supplied')
    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    parser.add_argument('--adam', action='store_true', help='use adam optimizer')
    parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset')
    opt = parser.parse_args()
    opt.weights = last if opt.resume else opt.weights
    opt.cfg = glob.glob('./**/' + opt.cfg, recursive=True)[0]  # find file
    opt.data = glob.glob('./**/' + opt.data, recursive=True)[0]  # find file
    print(opt)
    opt.img_size.extend([opt.img_size[-1]] * (2 - len(opt.img_size)))  # extend to 2 sizes (train, test)
    device = torch_utils.select_device(opt.device, apex=mixed_precision, batch_size=opt.batch_size)
    # check_git_status()
    if device.type == 'cpu':
        mixed_precision = False

    # Train
    if not opt.evolve:
        tb_writer = SummaryWriter(comment=opt.name)
        print('Start Tensorboard with "tensorboard --logdir=runs", view at http://localhost:6006/')
        train(hyp)

    # Evolve hyperparameters (optional)
    else:
        tb_writer = None
        opt.notest, opt.nosave = True, True  # only test/save final epoch
        if opt.bucket:
            os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket)  # download evolve.txt if exists

        for _ in range(10):  # generations to evolve
            if os.path.exists('evolve.txt'):  # if evolve.txt exists: select best hyps and mutate
                # Select parent(s)
                parent = 'single'  # parent selection method: 'single' or 'weighted'
                x = np.loadtxt('evolve.txt', ndmin=2)
                n = min(5, len(x))  # number of previous results to consider
                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
                w = fitness(x) - fitness(x).min()  # weights
                if parent == 'single' or len(x) == 1:
                    # x = x[random.randint(0, n - 1)]  # random selection
                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
                elif parent == 'weighted':
                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination

                # Mutate
                mp, s = 0.9, 0.2  # mutation probability, sigma
                npr = np.random
                npr.seed(int(time.time()))
                g = np.array([1, 1, 1, 1, 1, 1, 1, 0, .1, 1, 0, 1, 1, 1, 1, 1, 1, 1])  # gains
                ng = len(g)
                v = np.ones(ng)
                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
                    hyp[k] = x[i + 7] * v[i]  # mutate

            # Clip to limits
            keys = ['lr0', 'iou_t', 'momentum', 'weight_decay', 'hsv_s', 'hsv_v', 'translate', 'scale', 'fl_gamma']
            limits = [(1e-5, 1e-2), (0.00, 0.70), (0.60, 0.98), (0, 0.001), (0, .9), (0, .9), (0, .9), (0, .9), (0, 3)]
            for k, v in zip(keys, limits):
                hyp[k] = np.clip(hyp[k], v[0], v[1])

            # Train mutation
            results = train(hyp.copy())

            # Write mutation results
            print_mutation(hyp, results, opt.bucket)

            # Plot results
            # plot_evolution_results(hyp)
--- a/tutorial.ipynb
+++ b/tutorial.ipynb
--- a/utils/__init__.py
+++ b/utils/__init__.py
--- a/utils/activations.py
+++ b/utils/activations.py
@@ -0,0 +1,62 @@
 import torch
 import torch.functional as F
 import torch.nn as nn


 # Swish ------------------------------------------------------------------------
 class SwishImplementation(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x):
        ctx.save_for_backward(x)
        return x * torch.sigmoid(x)

    @staticmethod
    def backward(ctx, grad_output):
        x = ctx.saved_tensors[0]
        sx = torch.sigmoid(x)
        return grad_output * (sx * (1 + x * (1 - sx)))


 class MemoryEfficientSwish(nn.Module):
    @staticmethod
    def forward(x):
        return SwishImplementation.apply(x)


 class HardSwish(nn.Module):  # https://arxiv.org/pdf/1905.02244.pdf
    @staticmethod
    def forward(x):
        return x * F.hardtanh(x + 3, 0., 6., True) / 6.


 class Swish(nn.Module):
    @staticmethod
    def forward(x):
        return x * torch.sigmoid(x)


 # Mish ------------------------------------------------------------------------
 class MishImplementation(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x):
        ctx.save_for_backward(x)
        return x.mul(torch.tanh(F.softplus(x)))  # x * tanh(ln(1 + exp(x)))

    @staticmethod
    def backward(ctx, grad_output):
        x = ctx.saved_tensors[0]
        sx = torch.sigmoid(x)
        fx = F.softplus(x).tanh()
        return grad_output * (fx + x * sx * (1 - fx * fx))


 class MemoryEfficientMish(nn.Module):
    @staticmethod
    def forward(x):
        return MishImplementation.apply(x)


 class Mish(nn.Module):  # https://github.com/digantamisra98/Mish
    @staticmethod
    def forward(x):
        return x * F.softplus(x).tanh()
--- a/utils/datasets.py
+++ b/utils/datasets.py
@@ -0,0 +1,842 @@
 import glob
 import math
 import os
 import random
 import shutil
 import time
 from pathlib import Path
 from threading import Thread

 import cv2
 import numpy as np
 import torch
 from PIL import Image, ExifTags
 from torch.utils.data import Dataset
 from tqdm import tqdm

 from utils.utils import xyxy2xywh, xywh2xyxy

 help_url = 'https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data'
 img_formats = ['.bmp', '.jpg', '.jpeg', '.png', '.tif', '.dng']
 vid_formats = ['.mov', '.avi', '.mp4']

 # Get orientation exif tag
 for orientation in ExifTags.TAGS.keys():
    if ExifTags.TAGS[orientation] == 'Orientation':
        break


 def exif_size(img):
    # Returns exif-corrected PIL size
    s = img.size  # (width, height)
    try:
        rotation = dict(img._getexif().items())[orientation]
        if rotation == 6:  # rotation 270
            s = (s[1], s[0])
        elif rotation == 8:  # rotation 90
            s = (s[1], s[0])
    except:
        pass

    return s


 class LoadImages:  # for inference
    def __init__(self, path, img_size=416):
        path = str(Path(path))  # os-agnostic
        files = []
        if os.path.isdir(path):
            files = sorted(glob.glob(os.path.join(path, '*.*')))
        elif os.path.isfile(path):
            files = [path]

        images = [x for x in files if os.path.splitext(x)[-1].lower() in img_formats]
        videos = [x for x in files if os.path.splitext(x)[-1].lower() in vid_formats]
        nI, nV = len(images), len(videos)

        self.img_size = img_size
        self.files = images + videos
        self.nF = nI + nV  # number of files
        self.video_flag = [False] * nI + [True] * nV
        self.mode = 'images'
        if any(videos):
            self.new_video(videos[0])  # new video
        else:
            self.cap = None
        assert self.nF > 0, 'No images or videos found in ' + path

    def __iter__(self):
        self.count = 0
        return self

    def __next__(self):
        if self.count == self.nF:
            raise StopIteration
        path = self.files[self.count]

        if self.video_flag[self.count]:
            # Read video
            self.mode = 'video'
            ret_val, img0 = self.cap.read()
            if not ret_val:
                self.count += 1
                self.cap.release()
                if self.count == self.nF:  # last video
                    raise StopIteration
                else:
                    path = self.files[self.count]
                    self.new_video(path)
                    ret_val, img0 = self.cap.read()

            self.frame += 1
            print('video %g/%g (%g/%g) %s: ' % (self.count + 1, self.nF, self.frame, self.nframes, path), end='')

        else:
            # Read image
            self.count += 1
            img0 = cv2.imread(path)  # BGR
            assert img0 is not None, 'Image Not Found ' + path
            print('image %g/%g %s: ' % (self.count, self.nF, path), end='')

        # Padded resize
        img = letterbox(img0, new_shape=self.img_size)[0]

        # Convert
        img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
        img = np.ascontiguousarray(img)

        # cv2.imwrite(path + '.letterbox.jpg', 255 * img.transpose((1, 2, 0))[:, :, ::-1])  # save letterbox image
        return path, img, img0, self.cap

    def new_video(self, path):
        self.frame = 0
        self.cap = cv2.VideoCapture(path)
        self.nframes = int(self.cap.get(cv2.CAP_PROP_FRAME_COUNT))

    def __len__(self):
        return self.nF  # number of files


 class LoadWebcam:  # for inference
    def __init__(self, pipe=0, img_size=416):
        self.img_size = img_size

        if pipe == '0':
            pipe = 0  # local camera
        # pipe = 'rtsp://192.168.1.64/1'  # IP camera
        # pipe = 'rtsp://username:password@192.168.1.64/1'  # IP camera with login
        # pipe = 'rtsp://170.93.143.139/rtplive/470011e600ef003a004ee33696235daa'  # IP traffic camera
        # pipe = 'http://wmccpinetop.axiscam.net/mjpg/video.mjpg'  # IP golf camera

        # https://answers.opencv.org/question/215996/changing-gstreamer-pipeline-to-opencv-in-pythonsolved/
        # pipe = '"rtspsrc location="rtsp://username:password@192.168.1.64/1" latency=10 ! appsink'  # GStreamer

        # https://answers.opencv.org/question/200787/video-acceleration-gstremer-pipeline-in-videocapture/
        # https://stackoverflow.com/questions/54095699/install-gstreamer-support-for-opencv-python-package  # install help
        # pipe = "rtspsrc location=rtsp://root:root@192.168.0.91:554/axis-media/media.amp?videocodec=h264&resolution=3840x2160 protocols=GST_RTSP_LOWER_TRANS_TCP ! rtph264depay ! queue ! vaapih264dec ! videoconvert ! appsink"  # GStreamer

        self.pipe = pipe
        self.cap = cv2.VideoCapture(pipe)  # video capture object
        self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 3)  # set buffer size

    def __iter__(self):
        self.count = -1
        return self

    def __next__(self):
        self.count += 1
        if cv2.waitKey(1) == ord('q'):  # q to quit
            self.cap.release()
            cv2.destroyAllWindows()
            raise StopIteration

        # Read frame
        if self.pipe == 0:  # local camera
            ret_val, img0 = self.cap.read()
            img0 = cv2.flip(img0, 1)  # flip left-right
        else:  # IP camera
            n = 0
            while True:
                n += 1
                self.cap.grab()
                if n % 30 == 0:  # skip frames
                    ret_val, img0 = self.cap.retrieve()
                    if ret_val:
                        break

        # Print
        assert ret_val, 'Camera Error %s' % self.pipe
        img_path = 'webcam.jpg'
        print('webcam %g: ' % self.count, end='')

        # Padded resize
        img = letterbox(img0, new_shape=self.img_size)[0]

        # Convert
        img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
        img = np.ascontiguousarray(img)

        return img_path, img, img0, None

    def __len__(self):
        return 0


 class LoadStreams:  # multiple IP or RTSP cameras
    def __init__(self, sources='streams.txt', img_size=416):
        self.mode = 'images'
        self.img_size = img_size

        if os.path.isfile(sources):
            with open(sources, 'r') as f:
                sources = [x.strip() for x in f.read().splitlines() if len(x.strip())]
        else:
            sources = [sources]

        n = len(sources)
        self.imgs = [None] * n
        self.sources = sources
        for i, s in enumerate(sources):
            # Start the thread to read frames from the video stream
            print('%g/%g: %s... ' % (i + 1, n, s), end='')
            cap = cv2.VideoCapture(0 if s == '0' else s)
            assert cap.isOpened(), 'Failed to open %s' % s
            w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
            h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
            fps = cap.get(cv2.CAP_PROP_FPS) % 100
            _, self.imgs[i] = cap.read()  # guarantee first frame
            thread = Thread(target=self.update, args=([i, cap]), daemon=True)
            print(' success (%gx%g at %.2f FPS).' % (w, h, fps))
            thread.start()
        print('')  # newline

        # check for common shapes
        s = np.stack([letterbox(x, new_shape=self.img_size)[0].shape for x in self.imgs], 0)  # inference shapes
        self.rect = np.unique(s, axis=0).shape[0] == 1  # rect inference if all shapes equal
        if not self.rect:
            print('WARNING: Different stream shapes detected. For optimal performance supply similarly-shaped streams.')

    def update(self, index, cap):
        # Read next stream frame in a daemon thread
        n = 0
        while cap.isOpened():
            n += 1
            # _, self.imgs[index] = cap.read()
            cap.grab()
            if n == 4:  # read every 4th frame
                _, self.imgs[index] = cap.retrieve()
                n = 0
            time.sleep(0.01)  # wait time

    def __iter__(self):
        self.count = -1
        return self

    def __next__(self):
        self.count += 1
        img0 = self.imgs.copy()
        if cv2.waitKey(1) == ord('q'):  # q to quit
            cv2.destroyAllWindows()
            raise StopIteration

        # Letterbox
        img = [letterbox(x, new_shape=self.img_size, auto=self.rect)[0] for x in img0]

        # Stack
        img = np.stack(img, 0)

        # Convert
        img = img[:, :, :, ::-1].transpose(0, 3, 1, 2)  # BGR to RGB, to bsx3x416x416
        img = np.ascontiguousarray(img)

        return self.sources, img, img0, None

    def __len__(self):
        return 0  # 1E12 frames = 32 streams at 30 FPS for 30 years


 class LoadImagesAndLabels(Dataset):  # for training/testing
    def __init__(self, path, img_size=416, batch_size=16, augment=False, hyp=None, rect=False, image_weights=False,
                 cache_images=False, single_cls=False):
        try:
            path = str(Path(path))  # os-agnostic
            parent = str(Path(path).parent) + os.sep
            if os.path.isfile(path):  # file
                with open(path, 'r') as f:
                    f = f.read().splitlines()
                    f = [x.replace('./', parent) if x.startswith('./') else x for x in f]  # local to global path
            elif os.path.isdir(path):  # folder
                f = glob.iglob(path + os.sep + '*.*')
            else:
                raise Exception('%s does not exist' % path)
            self.img_files = [x.replace('/', os.sep) for x in f if os.path.splitext(x)[-1].lower() in img_formats]
        except:
            raise Exception('Error loading data from %s. See %s' % (path, help_url))

        n = len(self.img_files)
        assert n > 0, 'No images found in %s. See %s' % (path, help_url)
        bi = np.floor(np.arange(n) / batch_size).astype(np.int)  # batch index
        nb = bi[-1] + 1  # number of batches

        self.n = n  # number of images
        self.batch = bi  # batch index of image
        self.img_size = img_size
        self.augment = augment
        self.hyp = hyp
        self.image_weights = image_weights
        self.rect = False if image_weights else rect
        self.mosaic = self.augment and not self.rect  # load 4 images at a time into a mosaic (only during training)

        # Define labels
        self.label_files = [x.replace('images', 'labels').replace(os.path.splitext(x)[-1], '.txt')
                            for x in self.img_files]

        # Rectangular Training  https://github.com/ultralytics/yolov3/issues/232
        if self.rect:
            # Read image shapes (wh)
            sp = path.replace('.txt', '') + '.shapes'  # shapefile path
            try:
                with open(sp, 'r') as f:  # read existing shapefile
                    s = [x.split() for x in f.read().splitlines()]
                    assert len(s) == n, 'Shapefile out of sync'
            except:
                s = [exif_size(Image.open(f)) for f in tqdm(self.img_files, desc='Reading image shapes')]
                np.savetxt(sp, s, fmt='%g')  # overwrites existing (if any)

            # Sort by aspect ratio
            s = np.array(s, dtype=np.float64)
            ar = s[:, 1] / s[:, 0]  # aspect ratio
            irect = ar.argsort()
            self.img_files = [self.img_files[i] for i in irect]
            self.label_files = [self.label_files[i] for i in irect]
            self.shapes = s[irect]  # wh
            ar = ar[irect]

            # Set training image shapes
            shapes = [[1, 1]] * nb
            for i in range(nb):
                ari = ar[bi == i]
                mini, maxi = ari.min(), ari.max()
                if maxi < 1:
                    shapes[i] = [maxi, 1]
                elif mini > 1:
                    shapes[i] = [1, 1 / mini]

            self.batch_shapes = np.ceil(np.array(shapes) * img_size / 64.).astype(np.int) * 64

        # Cache labels
        self.imgs = [None] * n
        self.labels = [np.zeros((0, 5), dtype=np.float32)] * n
        create_datasubset, extract_bounding_boxes, labels_loaded = False, False, False
        nm, nf, ne, ns, nd = 0, 0, 0, 0, 0  # number missing, found, empty, datasubset, duplicate
        np_labels_path = str(Path(self.label_files[0]).parent) + '.npy'  # saved labels in *.npy file
        if os.path.isfile(np_labels_path):
            s = np_labels_path  # print string
            x = np.load(np_labels_path, allow_pickle=True)
            if len(x) == n:
                self.labels = x
                labels_loaded = True
        else:
            s = path.replace('images', 'labels')

        pbar = tqdm(self.label_files)
        for i, file in enumerate(pbar):
            if labels_loaded:
                l = self.labels[i]
                # np.savetxt(file, l, '%g')  # save *.txt from *.npy file
            else:
                try:
                    with open(file, 'r') as f:
                        l = np.array([x.split() for x in f.read().splitlines()], dtype=np.float32)
                except:
                    nm += 1  # print('missing labels for image %s' % self.img_files[i])  # file missing
                    continue

            if l.shape[0]:
                assert l.shape[1] == 5, '> 5 label columns: %s' % file
                assert (l >= 0).all(), 'negative labels: %s' % file
                assert (l[:, 1:] <= 1).all(), 'non-normalized or out of bounds coordinate labels: %s' % file
                if np.unique(l, axis=0).shape[0] < l.shape[0]:  # duplicate rows
                    nd += 1  # print('WARNING: duplicate rows in %s' % self.label_files[i])  # duplicate rows
                if single_cls:
                    l[:, 0] = 0  # force dataset into single-class mode
                self.labels[i] = l
                nf += 1  # file found

                # Create subdataset (a smaller dataset)
                if create_datasubset and ns < 1E4:
                    if ns == 0:
                        create_folder(path='./datasubset')
                        os.makedirs('./datasubset/images')
                    exclude_classes = 43
                    if exclude_classes not in l[:, 0]:
                        ns += 1
                        # shutil.copy(src=self.img_files[i], dst='./datasubset/images/')  # copy image
                        with open('./datasubset/images.txt', 'a') as f:
                            f.write(self.img_files[i] + '\n')

                # Extract object detection boxes for a second stage classifier
                if extract_bounding_boxes:
                    p = Path(self.img_files[i])
                    img = cv2.imread(str(p))
                    h, w = img.shape[:2]
                    for j, x in enumerate(l):
                        f = '%s%sclassifier%s%g_%g_%s' % (p.parent.parent, os.sep, os.sep, x[0], j, p.name)
                        if not os.path.exists(Path(f).parent):
                            os.makedirs(Path(f).parent)  # make new output folder

                        b = x[1:] * [w, h, w, h]  # box
                        b[2:] = b[2:].max()  # rectangle to square
                        b[2:] = b[2:] * 1.3 + 30  # pad
                        b = xywh2xyxy(b.reshape(-1, 4)).ravel().astype(np.int)

                        b[[0, 2]] = np.clip(b[[0, 2]], 0, w)  # clip boxes outside of image
                        b[[1, 3]] = np.clip(b[[1, 3]], 0, h)
                        assert cv2.imwrite(f, img[b[1]:b[3], b[0]:b[2]]), 'Failure extracting classifier boxes'
            else:
                ne += 1  # print('empty labels for image %s' % self.img_files[i])  # file empty
                # os.system("rm '%s' '%s'" % (self.img_files[i], self.label_files[i]))  # remove

            pbar.desc = 'Caching labels %s (%g found, %g missing, %g empty, %g duplicate, for %g images)' % (
                s, nf, nm, ne, nd, n)
        assert nf > 0 or n == 20288, 'No labels found in %s. See %s' % (os.path.dirname(file) + os.sep, help_url)
        if not labels_loaded and n > 1000:
            print('Saving labels to %s for faster future loading' % np_labels_path)
            np.save(np_labels_path, self.labels)  # save for next time

        # Cache images into memory for faster training (WARNING: large datasets may exceed system RAM)
        if cache_images:  # if training
            gb = 0  # Gigabytes of cached images
            pbar = tqdm(range(len(self.img_files)), desc='Caching images')
            self.img_hw0, self.img_hw = [None] * n, [None] * n
            for i in pbar:  # max 10k images
                self.imgs[i], self.img_hw0[i], self.img_hw[i] = load_image(self, i)  # img, hw_original, hw_resized
                gb += self.imgs[i].nbytes
                pbar.desc = 'Caching images (%.1fGB)' % (gb / 1E9)

        # Detect corrupted images https://medium.com/joelthchao/programmatically-detect-corrupted-image-8c1b2006c3d3
        detect_corrupted_images = False
        if detect_corrupted_images:
            from skimage import io  # conda install -c conda-forge scikit-image
            for file in tqdm(self.img_files, desc='Detecting corrupted images'):
                try:
                    _ = io.imread(file)
                except:
                    print('Corrupted image detected: %s' % file)

    def __len__(self):
        return len(self.img_files)

    # def __iter__(self):
    #     self.count = -1
    #     print('ran dataset iter')
    #     #self.shuffled_vector = np.random.permutation(self.nF) if self.augment else np.arange(self.nF)
    #     return self

    def __getitem__(self, index):
        if self.image_weights:
            index = self.indices[index]

        hyp = self.hyp
        if self.mosaic:
            # Load mosaic
            img, labels = load_mosaic(self, index)
            shapes = None

        else:
            # Load image
            img, (h0, w0), (h, w) = load_image(self, index)

            # Letterbox
            shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size  # final letterboxed shape
            img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment)
            shapes = (h0, w0), ((h / h0, w / w0), pad)  # for COCO mAP rescaling

            # Load labels
            labels = []
            x = self.labels[index]
            if x.size > 0:
                # Normalized xywh to pixel xyxy format
                labels = x.copy()
                labels[:, 1] = ratio[0] * w * (x[:, 1] - x[:, 3] / 2) + pad[0]  # pad width
                labels[:, 2] = ratio[1] * h * (x[:, 2] - x[:, 4] / 2) + pad[1]  # pad height
                labels[:, 3] = ratio[0] * w * (x[:, 1] + x[:, 3] / 2) + pad[0]
                labels[:, 4] = ratio[1] * h * (x[:, 2] + x[:, 4] / 2) + pad[1]

        if self.augment:
            # Augment imagespace
            if not self.mosaic:
                img, labels = random_affine(img, labels,
                                            degrees=hyp['degrees'],
                                            translate=hyp['translate'],
                                            scale=hyp['scale'],
                                            shear=hyp['shear'])

            # Augment colorspace
            augment_hsv(img, hgain=hyp['hsv_h'], sgain=hyp['hsv_s'], vgain=hyp['hsv_v'])

            # Apply cutouts
            # if random.random() < 0.9:
            #     labels = cutout(img, labels)

        nL = len(labels)  # number of labels
        if nL:
            # convert xyxy to xywh
            labels[:, 1:5] = xyxy2xywh(labels[:, 1:5])

            # Normalize coordinates 0 - 1
            labels[:, [2, 4]] /= img.shape[0]  # height
            labels[:, [1, 3]] /= img.shape[1]  # width

        if self.augment:
            # random left-right flip
            lr_flip = True
            if lr_flip and random.random() < 0.5:
                img = np.fliplr(img)
                if nL:
                    labels[:, 1] = 1 - labels[:, 1]

            # random up-down flip
            ud_flip = False
            if ud_flip and random.random() < 0.5:
                img = np.flipud(img)
                if nL:
                    labels[:, 2] = 1 - labels[:, 2]

        labels_out = torch.zeros((nL, 6))
        if nL:
            labels_out[:, 1:] = torch.from_numpy(labels)

        # Convert
        img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
        img = np.ascontiguousarray(img)

        return torch.from_numpy(img), labels_out, self.img_files[index], shapes

    @staticmethod
    def collate_fn(batch):
        img, label, path, shapes = zip(*batch)  # transposed
        for i, l in enumerate(label):
            l[:, 0] = i  # add target image index for build_targets()
        return torch.stack(img, 0), torch.cat(label, 0), path, shapes


 def load_image(self, index):
    # loads 1 image from dataset, returns img, original hw, resized hw
    img = self.imgs[index]
    if img is None:  # not cached
        path = self.img_files[index]
        img = cv2.imread(path)  # BGR
        assert img is not None, 'Image Not Found ' + path
        h0, w0 = img.shape[:2]  # orig hw
        r = self.img_size / max(h0, w0)  # resize image to img_size
        if r != 1:  # always resize down, only resize up if training with augmentation
            interp = cv2.INTER_AREA if r < 1 and not self.augment else cv2.INTER_LINEAR
            img = cv2.resize(img, (int(w0 * r), int(h0 * r)), interpolation=interp)
        return img, (h0, w0), img.shape[:2]  # img, hw_original, hw_resized
    else:
        return self.imgs[index], self.img_hw0[index], self.img_hw[index]  # img, hw_original, hw_resized


 def augment_hsv(img, hgain=0.5, sgain=0.5, vgain=0.5):
    r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1  # random gains
    hue, sat, val = cv2.split(cv2.cvtColor(img, cv2.COLOR_BGR2HSV))
    dtype = img.dtype  # uint8

    x = np.arange(0, 256, dtype=np.int16)
    lut_hue = ((x * r[0]) % 180).astype(dtype)
    lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
    lut_val = np.clip(x * r[2], 0, 255).astype(dtype)

    img_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val))).astype(dtype)
    cv2.cvtColor(img_hsv, cv2.COLOR_HSV2BGR, dst=img)  # no return needed

    # Histogram equalization
    # if random.random() < 0.2:
    #     for i in range(3):
    #         img[:, :, i] = cv2.equalizeHist(img[:, :, i])


 def load_mosaic(self, index):
    # loads images in a mosaic

    labels4 = []
    s = self.img_size
    xc, yc = [int(random.uniform(s * 0.5, s * 1.5)) for _ in range(2)]  # mosaic center x, y
    indices = [index] + [random.randint(0, len(self.labels) - 1) for _ in range(3)]  # 3 additional image indices
    for i, index in enumerate(indices):
        # Load image
        img, _, (h, w) = load_image(self, index)

        # place img in img4
        if i == 0:  # top left
            img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
            x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc  # xmin, ymin, xmax, ymax (large image)
            x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h  # xmin, ymin, xmax, ymax (small image)
        elif i == 1:  # top right
            x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
            x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
        elif i == 2:  # bottom left
            x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
            x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, max(xc, w), min(y2a - y1a, h)
        elif i == 3:  # bottom right
            x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
            x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)

        img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b]  # img4[ymin:ymax, xmin:xmax]
        padw = x1a - x1b
        padh = y1a - y1b

        # Labels
        x = self.labels[index]
        labels = x.copy()
        if x.size > 0:  # Normalized xywh to pixel xyxy format
            labels[:, 1] = w * (x[:, 1] - x[:, 3] / 2) + padw
            labels[:, 2] = h * (x[:, 2] - x[:, 4] / 2) + padh
            labels[:, 3] = w * (x[:, 1] + x[:, 3] / 2) + padw
            labels[:, 4] = h * (x[:, 2] + x[:, 4] / 2) + padh
        labels4.append(labels)

    # Concat/clip labels
    if len(labels4):
        labels4 = np.concatenate(labels4, 0)
        # np.clip(labels4[:, 1:] - s / 2, 0, s, out=labels4[:, 1:])  # use with center crop
        np.clip(labels4[:, 1:], 0, 2 * s, out=labels4[:, 1:])  # use with random_affine

    # Augment
    # img4 = img4[s // 2: int(s * 1.5), s // 2:int(s * 1.5)]  # center crop (WARNING, requires box pruning)
    img4, labels4 = random_affine(img4, labels4,
                                  degrees=self.hyp['degrees'],
                                  translate=self.hyp['translate'],
                                  scale=self.hyp['scale'],
                                  shear=self.hyp['shear'],
                                  border=-s // 2)  # border to remove

    return img4, labels4


 def letterbox(img, new_shape=(416, 416), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True):
    # Resize image to a 32-pixel-multiple rectangle https://github.com/ultralytics/yolov3/issues/232
    shape = img.shape[:2]  # current shape [height, width]
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    if not scaleup:  # only scale down, do not scale up (for better test mAP)
        r = min(r, 1.0)

    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
    if auto:  # minimum rectangle
        dw, dh = np.mod(dw, 64), np.mod(dh, 64)  # wh padding
    elif scaleFill:  # stretch
        dw, dh = 0.0, 0.0
        new_unpad = new_shape
        ratio = new_shape[0] / shape[1], new_shape[1] / shape[0]  # width, height ratios

    dw /= 2  # divide padding into 2 sides
    dh /= 2

    if shape[::-1] != new_unpad:  # resize
        img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    return img, ratio, (dw, dh)


 def random_affine(img, targets=(), degrees=10, translate=.1, scale=.1, shear=10, border=0):
    # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))
    # https://medium.com/uruvideo/dataset-augmentation-with-random-homographies-a8f4b44830d4
    # targets = [cls, xyxy]

    height = img.shape[0] + border * 2
    width = img.shape[1] + border * 2

    # Rotation and Scale
    R = np.eye(3)
    a = random.uniform(-degrees, degrees)
    # a += random.choice([-180, -90, 0, 90])  # add 90deg rotations to small rotations
    s = random.uniform(1 - scale, 1 + scale)
    # s = 2 ** random.uniform(-scale, scale)
    R[:2] = cv2.getRotationMatrix2D(angle=a, center=(img.shape[1] / 2, img.shape[0] / 2), scale=s)

    # Translation
    T = np.eye(3)
    T[0, 2] = random.uniform(-translate, translate) * img.shape[0] + border  # x translation (pixels)
    T[1, 2] = random.uniform(-translate, translate) * img.shape[1] + border  # y translation (pixels)

    # Shear
    S = np.eye(3)
    S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # x shear (deg)
    S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # y shear (deg)

    # Combined rotation matrix
    M = S @ T @ R  # ORDER IS IMPORTANT HERE!!
    if (border != 0) or (M != np.eye(3)).any():  # image changed
        img = cv2.warpAffine(img, M[:2], dsize=(width, height), flags=cv2.INTER_LINEAR, borderValue=(114, 114, 114))

    # Transform label coordinates
    n = len(targets)
    if n:
        # warp points
        xy = np.ones((n * 4, 3))
        xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2)  # x1y1, x2y2, x1y2, x2y1
        xy = (xy @ M.T)[:, :2].reshape(n, 8)

        # create new boxes
        x = xy[:, [0, 2, 4, 6]]
        y = xy[:, [1, 3, 5, 7]]
        xy = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T

        # # apply angle-based reduction of bounding boxes
        # radians = a * math.pi / 180
        # reduction = max(abs(math.sin(radians)), abs(math.cos(radians))) ** 0.5
        # x = (xy[:, 2] + xy[:, 0]) / 2
        # y = (xy[:, 3] + xy[:, 1]) / 2
        # w = (xy[:, 2] - xy[:, 0]) * reduction
        # h = (xy[:, 3] - xy[:, 1]) * reduction
        # xy = np.concatenate((x - w / 2, y - h / 2, x + w / 2, y + h / 2)).reshape(4, n).T

        # reject warped points outside of image
        xy[:, [0, 2]] = xy[:, [0, 2]].clip(0, width)
        xy[:, [1, 3]] = xy[:, [1, 3]].clip(0, height)
        w = xy[:, 2] - xy[:, 0]
        h = xy[:, 3] - xy[:, 1]
        area = w * h
        area0 = (targets[:, 3] - targets[:, 1]) * (targets[:, 4] - targets[:, 2])
        ar = np.maximum(w / (h + 1e-16), h / (w + 1e-16))  # aspect ratio
        i = (w > 4) & (h > 4) & (area / (area0 * s + 1e-16) > 0.2) & (ar < 10)

        targets = targets[i]
        targets[:, 1:5] = xy[i]

    return img, targets


 def cutout(image, labels):
    # https://arxiv.org/abs/1708.04552
    # https://github.com/hysts/pytorch_cutout/blob/master/dataloader.py
    # https://towardsdatascience.com/when-conventional-wisdom-fails-revisiting-data-augmentation-for-self-driving-cars-4831998c5509
    h, w = image.shape[:2]

    def bbox_ioa(box1, box2):
        # Returns the intersection over box2 area given box1, box2. box1 is 4, box2 is nx4. boxes are x1y1x2y2
        box2 = box2.transpose()

        # 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]

        # Intersection area
        inter_area = (np.minimum(b1_x2, b2_x2) - np.maximum(b1_x1, b2_x1)).clip(0) * \
                     (np.minimum(b1_y2, b2_y2) - np.maximum(b1_y1, b2_y1)).clip(0)

        # box2 area
        box2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1) + 1e-16

        # Intersection over box2 area
        return inter_area / box2_area

    # create random masks
    scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16  # image size fraction
    for s in scales:
        mask_h = random.randint(1, int(h * s))
        mask_w = random.randint(1, int(w * s))

        # box
        xmin = max(0, random.randint(0, w) - mask_w // 2)
        ymin = max(0, random.randint(0, h) - mask_h // 2)
        xmax = min(w, xmin + mask_w)
        ymax = min(h, ymin + mask_h)

        # apply random color mask
        image[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)]

        # return unobscured labels
        if len(labels) and s > 0.03:
            box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32)
            ioa = bbox_ioa(box, labels[:, 1:5])  # intersection over area
            labels = labels[ioa < 0.60]  # remove >60% obscured labels

    return labels


 def reduce_img_size(path='../data/sm4/images', img_size=1024):  # from utils.datasets import *; reduce_img_size()
    # creates a new ./images_reduced folder with reduced size images of maximum size img_size
    path_new = path + '_reduced'  # reduced images path
    create_folder(path_new)
    for f in tqdm(glob.glob('%s/*.*' % path)):
        try:
            img = cv2.imread(f)
            h, w = img.shape[:2]
            r = img_size / max(h, w)  # size ratio
            if r < 1.0:
                img = cv2.resize(img, (int(w * r), int(h * r)), interpolation=cv2.INTER_AREA)  # _LINEAR fastest
            fnew = f.replace(path, path_new)  # .replace(Path(f).suffix, '.jpg')
            cv2.imwrite(fnew, img)
        except:
            print('WARNING: image failure %s' % f)


 def convert_images2bmp():  # from utils.datasets import *; convert_images2bmp()
    # Save images
    formats = [x.lower() for x in img_formats] + [x.upper() for x in img_formats]
    # for path in ['../coco/images/val2014', '../coco/images/train2014']:
    for path in ['../data/sm4/images', '../data/sm4/background']:
        create_folder(path + 'bmp')
        for ext in formats:  # ['.bmp', '.jpg', '.jpeg', '.png', '.tif', '.dng']
            for f in tqdm(glob.glob('%s/*%s' % (path, ext)), desc='Converting %s' % ext):
                cv2.imwrite(f.replace(ext.lower(), '.bmp').replace(path, path + 'bmp'), cv2.imread(f))

    # Save labels
    # for path in ['../coco/trainvalno5k.txt', '../coco/5k.txt']:
    for file in ['../data/sm4/out_train.txt', '../data/sm4/out_test.txt']:
        with open(file, 'r') as f:
            lines = f.read()
            # lines = f.read().replace('2014/', '2014bmp/')  # coco
            lines = lines.replace('/images', '/imagesbmp')
            lines = lines.replace('/background', '/backgroundbmp')
        for ext in formats:
            lines = lines.replace(ext, '.bmp')
        with open(file.replace('.txt', 'bmp.txt'), 'w') as f:
            f.write(lines)


 def recursive_dataset2bmp(dataset='../data/sm4_bmp'):  # from utils.datasets import *; recursive_dataset2bmp()
    # Converts dataset to bmp (for faster training)
    formats = [x.lower() for x in img_formats] + [x.upper() for x in img_formats]
    for a, b, files in os.walk(dataset):
        for file in tqdm(files, desc=a):
            p = a + '/' + file
            s = Path(file).suffix
            if s == '.txt':  # replace text
                with open(p, 'r') as f:
                    lines = f.read()
                for f in formats:
                    lines = lines.replace(f, '.bmp')
                with open(p, 'w') as f:
                    f.write(lines)
            elif s in formats:  # replace image
                cv2.imwrite(p.replace(s, '.bmp'), cv2.imread(p))
                if s != '.bmp':
                    os.system("rm '%s'" % p)


 def imagelist2folder(path='data/coco_64img.txt'):  # from utils.datasets import *; imagelist2folder()
    # Copies all the images in a text file (list of images) into a folder
    create_folder(path[:-4])
    with open(path, 'r') as f:
        for line in f.read().splitlines():
            os.system('cp "%s" %s' % (line, path[:-4]))
            print(line)


 def create_folder(path='./new_folder'):
    # Create folder
    if os.path.exists(path):
        shutil.rmtree(path)  # delete output folder
    os.makedirs(path)  # make new output folder
--- a/utils/gcp.sh
+++ b/utils/gcp.sh
@@ -0,0 +1,181 @@
 #!/usr/bin/env bash

 # New VM
 if [ ! -d ./coco ]
 then
  echo "COCO folder not found. Running startup script."
  git clone https://github.com/ultralytics/yolov5
  # git clone -b test --depth 1 https://github.com/ultralytics/yolov5 test  # branch
  # sudo apt-get install zip
  # git clone https://github.com/NVIDIA/apex && cd apex && pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" . --user && cd .. && rm -rf apex
  # sudo conda install -yc conda-forge scikit-image pycocotools
  python3 -c "from yolov5.utils.google_utils import gdrive_download; gdrive_download('1rrL-Jbc68iHiGjXOYc8u9tKfFiOX21Tn','coco2017.zip')"
  python3 -c "from yolov5.utils.google_utils import gdrive_download; gdrive_download('1Y6Kou6kEB0ZEMCCpJSKStCor4KAReE43','coco2017.zip')"
  sudo docker pull ultralytics/coco:198

  # Add 64GB swap
  sudo fallocate -l 64G /swapfile
  sudo chmod 600 /swapfile
  sudo mkswap /swapfile
  sudo swapon /swapfile
  free -h  # check memory

  # sudo reboot now
 fi
 n=198 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s_csp2.yaml --bucket ult/coco --name $n --data data/coco.yaml


 # Evolve coco
 sudo -s
 t=ultralytics/yolov3:evolve
 # docker kill $(docker ps -a -q --filter ancestor=$t)
 for i in 0 1 6 7
 do
  docker pull $t && docker run --gpus all -d --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t bash utils/evolve.sh $i
  sleep 30
 done


 # kmean_anchors(path='../coco/train2017.txt', n=12, img_size=(256, 1024), thr=0.10, gen=10000)
 # 0.10 iou_thr: 1.000 best possible recall, 6.15 anchors > thr
 # n=12, img_size=(256, 1024), IoU_all=0.188/0.641-mean/best, IoU>thr=0.338-mean: 7,9,  13,17,  30,21,  18,38,  31,63,  54,38,  52,110,  90,69,  98,187,  164,116,  218,255,  448,414

 # from yolov4l_10iou
 # computed with utils.kmean_anchors(path='../coco/train2017.txt', n=12, img_size=(320, 1024), thr=0.10, gen=1000)
 # Evolving anchors: 100%|█████████████████| 1000/1000 [39:57<00:00,  2.40s/it]
 # 0.10 iou_thr: 0.998 best possible recall, 6.07 anchors > thr
 # n=12, img_size=(320, 1024), IoU_all=0.187/0.635-mean/best, IoU>thr=0.339-mean: 9,13,  20,21,  21,50,  43,34,  44,89,  84,59,  76,164,  133,105,  216,176,  153,267,  312,331,  623,467

 # from yolov4l_10iou_9anchors
 # computed with utils.kmean_anchors(path='../coco/train2017.txt', n=9, img_size=(320, 1024), thr=0.10, gen=1000)
 # Evolving anchors: 100%|█████████████████| 1000/1000 [31:26<00:00,  1.89s/it]
 # 0.10 iou_thr: 0.998 best possible recall, 4.60 anchors > thr
 # n=9, img_size=(320, 1024), IoU_all=0.190/0.604-mean/best, IoU>thr=0.342-mean: 9,13,  20,26,  29,58,  55,37,  57,115,  105,77,  133,165,  254,280,  508,450

 # ar < 5
 #Evolving anchors: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1000/1000 [14:56<00:00,  1.00it/s]
 #0.10 iou_thr: 0.999 best possible recall, 4.63 anchors > thr
 #n=9, img_size=(320, 1024), IoU_all=0.187/0.601-mean/best, IoU>thr=0.334-mean: 9,11,  15,35,  36,21,  34,61,  86,55,  66,145,  164,110,  185,262,  450,369

 #Evolving anchors:  10%|████████████                                                                                                           | 1015/10000 [15:59<1:56:24,  1.29it/s]
 #5.00 iou_thr: 1.000 best possible recall, 4.79 anchors > thr
 #n=9, img_size=(320, 1024), IoU_all=9.193/1.535-mean/best, IoU>thr=2.745-mean: 8,8,  14,24,  41,21,  24,56,  64,57,  60,140,  154,102,  174,262,  443,340

 #Evolving anchors: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10000/10000 [2:15:31<00:00,  1.25it/s]
 #5.00 iou_thr: 1.000 best possible recall, 4.80 anchors > thr
 #n=9, img_size=(320, 1024), IoU_all=9.106/1.531-mean/best, IoU>thr=2.744-mean: 8,8,  12,24,  33,19,  25,53,  64,48,  58,128,  145,93,  164,240,  419,349


 Evolving anchors: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10000/10000 [21:39<00:00,  7.83it/s]
 0.20 iou_thr: 0.992 best possible recall, 3.44 anchors > thr
 n=9, img_size=(640, 640), IoU_all=0.199/0.614-mean/best, IoU>thr=0.421-mean: 9,11,  16,25,  38,25,  26,54,  70,50,  51,102,  113,109,  161,218,  366,343

 Evolving anchors: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10000/10000 [16:34<00:00, 10.05it/s]
 0.10 iou_thr: 0.999 best possible recall, 4.87 anchors > thr
 n=9, img_size=(640, 640), IoU_all=0.200/0.614-mean/best, IoU>thr=0.342-mean: 10,12,  16,27,  37,25,  27,56,  69,51,  53,108,  116,107,  163,216,  364,343

  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

 # coco (small cancelled, large, medium, small, yolov3)
 n=129 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov4.yaml --bucket ult/coco --name $n && sudo shutdown
 n=130 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 32 --weights '' --cfg models/yolov4.yaml --bucket ult/coco --name $n && sudo shutdown
 n=133 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 16 --weights '' --cfg models/yolov4.yaml --bucket ult/coco --name $n && sudo shutdown
 n=134 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov4s.yaml --bucket ult/coco --name $n && sudo shutdown
 n=135 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 32 --weights '' --cfg models/yolov3-spp.yaml --bucket ult/coco --name $n && sudo shutdown
 n=136 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 16 --weights '' --cfg models/yolov3-spp.yaml --bucket ult/coco --name $n && sudo shutdown


 n=138 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 24 --weights '' --cfg models/yolov3-spp.yaml --bucket ult/coco --name $n && sudo shutdown
 n=139 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 24 --weights '' --cfg models/yolov3-spp.yaml --bucket ult/coco --name $n && sudo shutdown
 n=140 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 24 --weights '' --cfg models/yolov3-spp.yaml --bucket ult/coco --name $n && sudo shutdown

 n=141 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown
 n=142 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown
 n=143 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown
 n=144 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown
 n=145 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 12 --weights '' --cfg models/yolov5l.yaml --bucket ult/coco --name $n && sudo shutdown
 n=146 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown

 n=147 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown
 n=148 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown
 n=149 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 832 --batch 32 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown

 n=150 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 32 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown
 n=151 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 320 640 --batch 32 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown

 n=152 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n && sudo shutdown
 n=154 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 32 --weights '' --cfg models/yolov5s_exp.yaml --bucket ult/coco --name $n && sudo shutdown

 n=153 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_focus1.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=155 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_focus2.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=156 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_exp_focus0.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown

 n=157 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_focus3.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=158 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_focus4.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=159 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_focus5.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown

 n=160 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5_final.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=161 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5_final.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown

 n=162 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5_final.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=163 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5_final.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown

 n=164 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5_final.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=165 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_origami.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=166 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_k3.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=167 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_plus.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=168 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown

 n=169 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown

 n=170 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 24 --weights '' --cfg models/yolov3-spp.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown

 n=171 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_k3_spp5913.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=172 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown

 n=173 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 24 --weights '' --cfg models/yolov3-spp.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=174 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 24 --weights '' --cfg models/yolov3-spp.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown


 n=178 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 96 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=179 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 48 --weights '' --cfg models/yolov5m.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=180 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 28 --weights '' --cfg models/yolov5l.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=181 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 16 --weights '' --cfg models/yolov5x.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=182 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 24 --weights '' --cfg models/yolov3-spp.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=183 && t=ultralytics/coco:v178 && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown

 n=184 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_csp.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=185 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 64 --weights '' --cfg models/yolov5s_csp.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown


 n=186 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n --data data/coco.yaml
 n=187 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s_csp.yaml --bucket ult/coco --name $n --data data/coco.yaml
 n=188 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s_csp.yaml --bucket ult/coco --name $n --data data/coco.yaml

 n=189 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s_focus.yaml --bucket ult/coco --name $n --data data/coco.yaml
 n=190 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n --data data/coco.yaml
 n=191 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n --data data/coco.yaml

 n=192 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n --data data/coco.yaml
 n=193 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s.yaml --bucket ult/coco --name $n --data data/coco.yaml

 n=194 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s_csp.yaml --bucket ult/coco --name $n --data data/coco.yaml
 n=195 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s_csp.yaml --bucket ult/coco --name $n --data data/coco.yaml

 n=196 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s_csp.yaml --bucket ult/coco --name $n --data data/coco.yaml
 n=197 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s_csp1.yaml --bucket ult/coco --name $n --data data/coco.yaml
 n=198 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s_csp2.yaml --bucket ult/coco --name $n --data data/coco.yaml
 n=199 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --epochs 50 --batch 64 --weights '' --cfg models/yolov5s_csp2.yaml --bucket ult/coco --name $n --data data/coco.yaml



 n=201 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 48 --weights '' --cfg models/yolov5m_csp.yaml --bucket ult/coco --name $n --data data/coco.yaml
 n=204 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 24 --weights '' --cfg models/yolov3-spp.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown

 n=206 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 24 --weights '' --cfg models/yolov3-spp_csp.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown
 n=207 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t python3 train.py --img 640 640 --batch 24 --weights '' --cfg models/yolov3-spp_csp.yaml --bucket ult/coco --name $n --data data/coco.yaml && sudo shutdown


 n=205 && t=ultralytics/coco:v$n && sudo docker pull $t && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco $t bash
--- a/utils/google_utils.py
+++ b/utils/google_utils.py
@@ -0,0 +1,94 @@
 # This file contains google utils: https://cloud.google.com/storage/docs/reference/libraries
 # pip install --upgrade google-cloud-storage
 # from google.cloud import storage

 import os
 import time
 from pathlib import Path


 def attempt_download(weights):
    # Attempt to download pretrained weights if not found locally
    weights = weights.strip()
    msg = weights + ' missing, try downloading from https://drive.google.com/open?id=1LezFG5g3BCW6iYaV89B2i64cqEUZD7e0'

    r = 1
    if len(weights) > 0 and not os.path.isfile(weights):
        d = {'yolov3-spp.pt': '1mM67oNw4fZoIOL1c8M3hHmj66d8e-ni_',  # yolov3-spp.yaml
             'yolov5s.pt': '1R5T6rIyy3lLwgFXNms8whc-387H0tMQO',  # yolov5s.yaml
             'yolov5m.pt': '1vobuEExpWQVpXExsJ2w-Mbf3HJjWkQJr',  # yolov5m.yaml
             'yolov5l.pt': '1hrlqD1Wdei7UT4OgT785BEk1JwnSvNEV',  # yolov5l.yaml
             'yolov5x.pt': '1mM8aZJlWTxOg7BZJvNUMrTnA2AbeCVzS',  # yolov5x.yaml
             }

        file = Path(weights).name
        if file in d:
            r = gdrive_download(id=d[file], name=weights)

        # Error check
        if not (r == 0 and os.path.exists(weights) and os.path.getsize(weights) > 1E6):  # weights exist and > 1MB
            os.system('rm ' + weights)  # remove partial downloads
            raise Exception(msg)


 def gdrive_download(id='1HaXkef9z6y5l4vUnCYgdmEAj61c6bfWO', name='coco.zip'):
    # https://gist.github.com/tanaikech/f0f2d122e05bf5f971611258c22c110f
    # Downloads a file from Google Drive, accepting presented query
    # from utils.google_utils import *; gdrive_download()
    t = time.time()

    print('Downloading https://drive.google.com/uc?export=download&id=%s as %s... ' % (id, name), end='')
    os.remove(name) if os.path.exists(name) else None  # remove existing
    os.remove('cookie') if os.path.exists('cookie') else None

    # Attempt file download
    os.system("curl -c ./cookie -s -L \"https://drive.google.com/uc?export=download&id=%s\" > /dev/null" % id)
    if os.path.exists('cookie'):  # large file
        s = "curl -Lb ./cookie \"https://drive.google.com/uc?export=download&confirm=`awk '/download/ {print $NF}' ./cookie`&id=%s\" -o %s" % (
            id, name)
    else:  # small file
        s = "curl -s -L -o %s 'https://drive.google.com/uc?export=download&id=%s'" % (name, id)
    r = os.system(s)  # execute, capture return values
    os.remove('cookie') if os.path.exists('cookie') else None

    # Error check
    if r != 0:
        os.remove(name) if os.path.exists(name) else None  # remove partial
        print('Download error ')  # raise Exception('Download error')
        return r

    # Unzip if archive
    if name.endswith('.zip'):
        print('unzipping... ', end='')
        os.system('unzip -q %s' % name)  # unzip
        os.remove(name)  # remove zip to free space

    print('Done (%.1fs)' % (time.time() - t))
    return r

 # def upload_blob(bucket_name, source_file_name, destination_blob_name):
 #     # Uploads a file to a bucket
 #     # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python
 #
 #     storage_client = storage.Client()
 #     bucket = storage_client.get_bucket(bucket_name)
 #     blob = bucket.blob(destination_blob_name)
 #
 #     blob.upload_from_filename(source_file_name)
 #
 #     print('File {} uploaded to {}.'.format(
 #         source_file_name,
 #         destination_blob_name))
 #
 #
 # def download_blob(bucket_name, source_blob_name, destination_file_name):
 #     # Uploads a blob from a bucket
 #     storage_client = storage.Client()
 #     bucket = storage_client.get_bucket(bucket_name)
 #     blob = bucket.blob(source_blob_name)
 #
 #     blob.download_to_filename(destination_file_name)
 #
 #     print('Blob {} downloaded to {}.'.format(
 #         source_blob_name,
 #         destination_file_name))
--- a/utils/torch_utils.py
+++ b/utils/torch_utils.py
@@ -0,0 +1,194 @@
 import math
 import os
 import time
 from copy import deepcopy

 import torch
 import torch.backends.cudnn as cudnn
 import torch.nn as nn
 import torch.nn.functional as F


 def init_seeds(seed=0):
    torch.manual_seed(seed)

    # Reduce randomness (may be slower on Tesla GPUs) # https://pytorch.org/docs/stable/notes/randomness.html
    if seed == 0:
        cudnn.deterministic = False
        cudnn.benchmark = True


 def select_device(device='', apex=False, batch_size=None):
    # device = 'cpu' or '0' or '0,1,2,3'
    cpu_request = device.lower() == 'cpu'
    if device and not cpu_request:  # if device requested other than 'cpu'
        os.environ['CUDA_VISIBLE_DEVICES'] = device  # set environment variable
        assert torch.cuda.is_available(), 'CUDA unavailable, invalid device %s requested' % device  # check availablity

    cuda = False if cpu_request else torch.cuda.is_available()
    if cuda:
        c = 1024 ** 2  # bytes to MB
        ng = torch.cuda.device_count()
        if ng > 1 and batch_size:  # check that batch_size is compatible with device_count
            assert batch_size % ng == 0, 'batch-size %g not multiple of GPU count %g' % (batch_size, ng)
        x = [torch.cuda.get_device_properties(i) for i in range(ng)]
        s = 'Using CUDA ' + ('Apex ' if apex else '')  # apex for mixed precision https://github.com/NVIDIA/apex
        for i in range(0, ng):
            if i == 1:
                s = ' ' * len(s)
            print("%sdevice%g _CudaDeviceProperties(name='%s', total_memory=%dMB)" %
                  (s, i, x[i].name, x[i].total_memory / c))
    else:
        print('Using CPU')

    print('')  # skip a line
    return torch.device('cuda:0' if cuda else 'cpu')


 def time_synchronized():
    torch.cuda.synchronize() if torch.cuda.is_available() else None
    return time.time()


 def initialize_weights(model):
    for m in model.modules():
        t = type(m)
        if t is nn.Conv2d:
            pass  # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
        elif t is nn.BatchNorm2d:
            m.eps = 1e-4
            m.momentum = 0.03
        elif t in [nn.LeakyReLU, nn.ReLU, nn.ReLU6]:
            m.inplace = True


 def find_modules(model, mclass=nn.Conv2d):
    # finds layer indices matching module class 'mclass'
    return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)]


 def fuse_conv_and_bn(conv, bn):
    # https://tehnokv.com/posts/fusing-batchnorm-and-conv/
    with torch.no_grad():
        # init
        fusedconv = torch.nn.Conv2d(conv.in_channels,
                                    conv.out_channels,
                                    kernel_size=conv.kernel_size,
                                    stride=conv.stride,
                                    padding=conv.padding,
                                    bias=True)

        # prepare filters
        w_conv = conv.weight.clone().view(conv.out_channels, -1)
        w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
        fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.size()))

        # prepare spatial bias
        if conv.bias is not None:
            b_conv = conv.bias
        else:
            b_conv = torch.zeros(conv.weight.size(0))
        b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
        fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)

        return fusedconv


 def model_info(model, verbose=False):
    # Plots a line-by-line description of a PyTorch model
    n_p = sum(x.numel() for x in model.parameters())  # number parameters
    n_g = sum(x.numel() for x in model.parameters() if x.requires_grad)  # number gradients
    if verbose:
        print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma'))
        for i, (name, p) in enumerate(model.named_parameters()):
            name = name.replace('module_list.', '')
            print('%5g %40s %9s %12g %20s %10.3g %10.3g' %
                  (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))

    try:  # FLOPS
        from thop import profile
        macs, _ = profile(model, inputs=(torch.zeros(1, 3, 480, 640),), verbose=False)
        fs = ', %.1f GFLOPS' % (macs / 1E9 * 2)
    except:
        fs = ''

    print('Model Summary: %g layers, %g parameters, %g gradients%s' % (len(list(model.parameters())), n_p, n_g, fs))


 def load_classifier(name='resnet101', n=2):
    # Loads a pretrained model reshaped to n-class output
    import pretrainedmodels  # https://github.com/Cadene/pretrained-models.pytorch#torchvision
    model = pretrainedmodels.__dict__[name](num_classes=1000, pretrained='imagenet')

    # Display model properties
    for x in ['model.input_size', 'model.input_space', 'model.input_range', 'model.mean', 'model.std']:
        print(x + ' =', eval(x))

    # Reshape output to n classes
    filters = model.last_linear.weight.shape[1]
    model.last_linear.bias = torch.nn.Parameter(torch.zeros(n))
    model.last_linear.weight = torch.nn.Parameter(torch.zeros(n, filters))
    model.last_linear.out_features = n
    return model


 def scale_img(img, ratio=1.0, same_shape=True):  # img(16,3,256,416), r=ratio
    # scales img(bs,3,y,x) by ratio
    h, w = img.shape[2:]
    s = (int(h * ratio), int(w * ratio))  # new size
    img = F.interpolate(img, size=s, mode='bilinear', align_corners=False)  # resize
    if not same_shape:  # pad/crop img
        gs = 64  # (pixels) grid size
        h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)]
    return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447)  # value = imagenet mean


 class ModelEMA:
    """ Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models
    Keep a moving average of everything in the model state_dict (parameters and buffers).
    This is intended to allow functionality like
    https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage
    A smoothed version of the weights is necessary for some training schemes to perform well.
    E.g. Google's hyper-params for training MNASNet, MobileNet-V3, EfficientNet, etc that use
    RMSprop with a short 2.4-3 epoch decay period and slow LR decay rate of .96-.99 requires EMA
    smoothing of weights to match results. Pay attention to the decay constant you are using
    relative to your update count per epoch.
    To keep EMA from using GPU resources, set device='cpu'. This will save a bit of memory but
    disable validation of the EMA weights. Validation will have to be done manually in a separate
    process, or after the training stops converging.
    This class is sensitive where it is initialized in the sequence of model init,
    GPU assignment and distributed training wrappers.
    I've tested with the sequence in my own train.py for torch.DataParallel, apex.DDP, and single-GPU.
    """

    def __init__(self, model, decay=0.9999, device=''):
        # make a copy of the model for accumulating moving average of weights
        self.ema = deepcopy(model)
        self.ema.eval()
        self.updates = 0  # number of EMA updates
        self.decay = lambda x: decay * (1 - math.exp(-x / 2000))  # decay exponential ramp (to help early epochs)
        self.device = device  # perform ema on different device from model if set
        if device:
            self.ema.to(device=device)
        for p in self.ema.parameters():
            p.requires_grad_(False)

    def update(self, model):
        self.updates += 1
        d = self.decay(self.updates)
        with torch.no_grad():
            if type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel):
                msd, esd = model.module.state_dict(), self.ema.module.state_dict()
            else:
                msd, esd = model.state_dict(), self.ema.state_dict()

            for k, v in esd.items():
                if v.dtype.is_floating_point:
                    v *= d
                    v += (1. - d) * msd[k].detach()

    def update_attr(self, model):
        # Assign attributes (which may change during training)
        for k in model.__dict__.keys():
            if not k.startswith('_'):
                setattr(self.ema, k, getattr(model, k))
--- a/utils/utils.py
+++ b/utils/utils.py
--- a/weights/download_weights.sh
+++ b/weights/download_weights.sh
@@ -0,0 +1,7 @@
 #!/bin/bash
 # Download common models

 python3 -c "from models import *;
 attempt_download('weights/yolov5s.pt');
 attempt_download('weights/yolov5m.pt');
 attempt_download('weights/yolov5l.pt')"