👋 Hello @glenn-jocher, thank you for submitting an Ultralytics pull request! 🚀 To ensure a seamless integration of your changes into the ultralytics/ultralytics repository, please check the following:

• ✅ Define a Purpose: Clearly explain the purpose of your changes in your PR description. You've done a great job highlighting the enhancements with the UV toolset! If this PR relates to any existing issues, please link them in your description.
• ✅ Synchronize with Source: Make sure your PR is up-to-date with the main branch of the ultralytics/ultralytics repository. If necessary, update it by clicking the 'Update branch' button or by using git pull and git merge main locally.
• ✅ Ensure CI Checks Pass: Check that all Ultralytics CI checks are successful. If any issues arise, addressing them promptly will help keep the merge process smooth.
• ✅ Update Documentation: If your changes alter any functionalities, consider updating the documentation to reflect your improvements.
• ✅ Add Tests: Where relevant, include new tests or update existing ones to verify your changes, and confirm all tests pass successfully.
• ✅ Sign the CLA: If this is your first contribution to Ultralytics, please ensure you've signed the Contributor License Agreement by stating "I have read the CLA Document and I sign the CLA" in the comments.
• ✅ Minimize Changes: As Bruce Lee wisely said, "It is not daily increase but daily decrease, hack away the unessential."

For additional guidance, you can refer to our Contributing Guide. Feel free to ask if you have any questions or need assistance. An Ultralytics engineer will review your PR soon. Thanks for your valuable contribution to Ultralytics! 🌟

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 74.01%. Comparing base (69c12a4) to head (c8f77e6).
Report is 1 commits behind head on main.

@@            Coverage Diff             @@
##             main   #17749      +/-   ##
==========================================
- Coverage   78.64%   74.01%   -4.64%     
==========================================
  Files         127      127              
  Lines       17099    17099              
==========================================
- Hits        13447    12655     -792     
- Misses       3652     4444     +792     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

Hi, @glenn-jocher what's the advantage of using uv?

Faster, but it looks like it's causing problems here unfortunately.

@glenn-jocher After several tries I figured CI failing mostly because ofuv package has a different strategy(walking through all the dependencies and find proper versions for all the packages and install them) of handling dependencies. And it selected tensorflowjs==3.9.0 based on our dependency of tensorflowjs==3.9.0, however actually tensorflowjs==3.9.0 is not compatible with numpy>=1.20 hence the error. The original pip somehow would select tensorflowjs==4.15.0 which is a version compatible with numpy>=1.20 hence all CI tests passing.

I found the versions of tensorflowjs that compatible with numpy>=1.20 and tried pining tensorflowjs>=4.5.0 but then strange errors thorown from uv.

  × No solution found when resolving dependencies:
  ╰─▶ Because tensorflowjs>=4.5.0,<=4.10.0 depends on packaging>=20.9,<21.dev0
      and only the following versions of tensorflowjs are available:
          tensorflowjs<=4.11.0
          tensorflowjs==4.12.0
          tensorflowjs>=4.13.0
      we can conclude that all of:
          tensorflowjs>=4.5.0,<4.11.0
          tensorflowjs>4.11.0,<4.12.0
          tensorflowjs>4.12.0,<4.13.0
      depend on packaging>=20.9,<21.dev0.
      And because all of:
          tensorflowjs==4.11.0
          tensorflowjs==4.12.0
          tensorflowjs>=4.13.0
      depend on packaging>=23.1,<24.dev0 and only packaging>=22.0 is
      available, we can conclude that tensorflowjs>=4.5.0 depends on
      packaging>=23.1,<24.dev0.
      And because only the following versions of packaging are available:
          packaging<23.1
          packaging>24.dev0
      and ultralytics[export]==8.3.38 depends on tensorflowjs>=4.5.0, we can
      conclude that ultralytics[export]==8.3.38 cannot be used.
      And because only ultralytics[export]==8.3.38 is available and you
      require ultralytics[export], we can conclude that your requirements
      are unsatisfiable.

      hint: Pre-releases are available for tensorflowjs in the requested
      range (e.g., 4.13.0rc0), but pre-releases weren't enabled (try:
      `--prerelease=allow`)

      hint: An index URL (https://download.pytorch.org/whl/cpu) could not
      be queried due to a lack of valid authentication credentials (403
      Forbidden).

      hint: packaging was requested with a pre-release marker (e.g.,
      packaging>=20.9,<21.dev0), but pre-releases weren't enabled (try:
      `--prerelease=allow`)

      hint: `packaging` was found on https://download.pytorch.org/whl/cpu,
      but not at the requested version (packaging>=20.9,<21.dev0). A
      compatible version may be available on a subsequent index (e.g.,
      https://pypi.org/simple). By default, uv will only consider versions
      that are published on the first index that contains a given package, to
      avoid dependency confusion attacks. If all indexes are equally trusted,
      use `--index-strategy unsafe-best-match` to consider all versions from
      all indexes, regardless of the order in which they were defined.

@glenn-jocher ok reported a issue to the official uv repo and seek for help: astral-sh/uv#9462

@glenn-jocher @Laughing-q GPU CI is also not working correctly right now. It passes, but it's running tests on the container installed Ultralytics package instead of the local/PR changes. Visible from the logs that shows the install type to be pip instead of git.

A fix was tested on this issue: #17781 (comment)

@Y-T-G @Skillnoob Hey I saw the error in that PR and thanks guys for the fix there! Probably it'd be better to open another PR to fix the issue there instead of directly modifying the user's PR since it's a common issue than a PR issue. Can you guys open another PR for it? THanks

@Laughing-q Created #17883

@Skillnoob nice work! Thanks!

@glenn-jocher After several tries I figured CI failing mostly because ofuv package has a different strategy(walking through all the dependencies and find proper versions for all the packages and install them) of handling dependencies. And it selected tensorflowjs==3.9.0 based on our dependency of tensorflowjs==3.9.0, however actually tensorflowjs==3.9.0 is not compatible with numpy>=1.20 hence the error. The original pip somehow would select tensorflowjs==4.15.0 which is a version compatible with numpy>=1.20 hence all CI tests passing.

I found the versions of tensorflowjs that compatible with numpy>=1.20 and tried pining tensorflowjs>=4.5.0 but then strange errors thorown from uv.

  × No solution found when resolving dependencies:
  ╰─▶ Because tensorflowjs>=4.5.0,<=4.10.0 depends on packaging>=20.9,<21.dev0
      and only the following versions of tensorflowjs are available:
          tensorflowjs<=4.11.0
          tensorflowjs==4.12.0
          tensorflowjs>=4.13.0
      we can conclude that all of:
          tensorflowjs>=4.5.0,<4.11.0
          tensorflowjs>4.11.0,<4.12.0
          tensorflowjs>4.12.0,<4.13.0
      depend on packaging>=20.9,<21.dev0.
      And because all of:
          tensorflowjs==4.11.0
          tensorflowjs==4.12.0
          tensorflowjs>=4.13.0
      depend on packaging>=23.1,<24.dev0 and only packaging>=22.0 is
      available, we can conclude that tensorflowjs>=4.5.0 depends on
      packaging>=23.1,<24.dev0.
      And because only the following versions of packaging are available:
          packaging<23.1
          packaging>24.dev0
      and ultralytics[export]==8.3.38 depends on tensorflowjs>=4.5.0, we can
      conclude that ultralytics[export]==8.3.38 cannot be used.
      And because only ultralytics[export]==8.3.38 is available and you
      require ultralytics[export], we can conclude that your requirements
      are unsatisfiable.

      hint: Pre-releases are available for tensorflowjs in the requested
      range (e.g., 4.13.0rc0), but pre-releases weren't enabled (try:
      `--prerelease=allow`)

      hint: An index URL (https://download.pytorch.org/whl/cpu) could not
      be queried due to a lack of valid authentication credentials (403
      Forbidden).

      hint: packaging was requested with a pre-release marker (e.g.,
      packaging>=20.9,<21.dev0), but pre-releases weren't enabled (try:
      `--prerelease=allow`)

      hint: `packaging` was found on https://download.pytorch.org/whl/cpu,
      but not at the requested version (packaging>=20.9,<21.dev0). A
      compatible version may be available on a subsequent index (e.g.,
      https://pypi.org/simple). By default, uv will only consider versions
      that are published on the first index that contains a given package, to
      avoid dependency confusion attacks. If all indexes are equally trusted,
      use `--index-strategy unsafe-best-match` to consider all versions from
      all indexes, regardless of the order in which they were defined.

@glenn-jocher uv currently is working properly with our benchmark now. I solved the issue by adding --index-strategy unsafe-first-match, based on the suggestion from the uv repo maintainer. It's because uv has different strategy when it works with --extra-index-url, see https://docs.astral.sh/uv/pip/compatibility/#packages-that-exist-on-multiple-indexes

@Skillnoob @Y-T-G #17883 merged, updating this one now, nice work!

@Laughing-q PR merged, great work!