remove ultralytics dependency

[lachiewalker]

This change removes the direct dependency on ultralytics from the yolov5-pip package by in-lining the utility functions (and the handful of utils they depend on from the same ultralytics.utils module) which were previously imported from ultralytics.utils. No other changes were made to the codebase.

Summary of Changes

• Copied the ultralytics.utils functions actually used by yolov5-pip directly into this repo.
• Removed ultralytics>=8.0.100 entry in requirements.txt.

Impact & Compatibility

No behavior change: All functions and APIs remain identical. This change is completely backwards compatible.

Motivation & Next Steps

This is an incremental cleanup to shrink package size as I think the yolov5 package is very well written for training, inference, and model export but there is currently no way to opt out of GPU and training dependencies. As deploying onto AWS lambda functions is a frequent use case of mine, this is very inefficient from a storage and container image size perspective. This decoupling from ultralytics is intended as the precursor to a future separate PR which will introduce sub-module packaging (e.g. yolov5[train], yolov5[inference], etc.). For now, this patch focuses solely on removing ultralytics as a dependency without altering any APIs.

[gemini-code-assist]

Hello @lachiewalker, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

Summary of Changes

Hello team,

Gemini here, providing a summary of this pull request authored by @lachiewalker. The main objective of this PR is to remove the direct dependency on the ultralytics package from the yolov5-pip distribution. This is achieved by copying the specific utility functions from ultralytics.utils that were being used directly into the yolov5/utils/general.py file. The motivation behind this change, as stated in the description, is to reduce the overall package size, which is particularly beneficial for deployments in environments like AWS Lambda, and to pave the way for more granular, optional dependencies in the future (e.g., yolov5[train], yolov5[inference], etc.). The author emphasizes that this change is intended to be completely backwards compatible with no alteration to existing APIs or behavior.

Highlights

• Dependency Removal: The direct dependency on the ultralytics package has been removed from the requirements.txt file.
• Code In-lining: Necessary utility functions previously imported from ultralytics.utils have been copied directly into yolov5/utils/general.py. These functions include checks for online status, retries, requirement parsing, and version checking.
• Package Size Reduction: A primary goal of this change is to decrease the size of the yolov5-pip package by removing the large ultralytics dependency.
• Preparation for Modular Dependencies: This PR is described as a foundational step towards introducing optional, sub-module dependencies in future releases.
• No Behavior Change: The author asserts that this change maintains full backwards compatibility and does not alter the functionality or APIs of the package.

Changelog

• requirements.txt
  • Removed the ultralytics>=8.0.100 entry (line 18).
• yolov5/utils/general.py
  • Added imports for functools (line 7), importlib.metadata (line 24), and types.SimpleNamespace (line 30).
  • Removed the import from ultralytics.utils.checks import check_requirements (line 39).
  • Added a large block of code (lines 127-433) containing in-lined utility functions (is_online, Retry, parse_requirements, parse_version, check_uv, check_version, check_requirements) copied from the ultralytics repository, including license attribution comments.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

---

Dependency gone,
Copied code now stands alone,
Package size shrinks down.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request effectively removes the ultralytics dependency by vendoring the required utility functions. The motivation for reducing package size and preparing for sub-module packaging is well-explained. The changes are mostly straightforward, involving copying code and updating requirements.txt.

I've identified one critical issue related to Python version compatibility that needs to be addressed and one medium-severity security consideration for the copied code. Otherwise, the PR achieves its stated goal.

Summary of Findings

• Python Version Incompatibility: The code uses the walrus operator (:=), which is Python 3.8+, but setup.py claims Python 3.7+ support. This will break for Python 3.7 users.
• Potential Command Injection Risk: The copied attempt_install function uses subprocess.check_output with shell=True, which can be a security risk if input variables are not properly sanitized.

Merge Readiness

The pull request makes good progress towards decoupling from the ultralytics library. However, there is a critical Python version compatibility issue due to the use of the walrus operator, which will prevent users on Python 3.7 from using this package. This issue must be addressed before merging.

Additionally, there's a medium-severity security concern regarding the use of shell=True in the copied subprocess calls, which should be considered for future improvement.

I am not authorized to approve pull requests, but I recommend that the critical compatibility issue be resolved before this PR is merged. Other reviewers should also assess the changes.

[lachiewalker]

Pinned fire==0.6.0, torch==2.5.1, and torchvision==0.20.1 to address CI test failures

[tvasenin]

I assume this was done because torch 2.6.0+ has weights_only defaulting to True?

Instead of locking the torch version, it's better to add weights_only=False parameter to torch.load() calls, and bump minimal version to torch>=2.0.0 (where this option was first introduced).

Or just set env variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=yes in CI -- this would work with any PyTorch release without bumping the min version.

[lachiewalker]

Although I wanted this change to be minimal, and not alter minimum version requirements for dependencies, changing the CI as you proposed could result in bugs where the CI tests only pass because the TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=yes env var is present, whilst users who set up an environment with torch>=2.0.0 will see different behaviour due to the changed default of torch.load(weights_only=False->True). I thus opted to add weights_only=False to all torch.load() calls because, despite raising the minimum version for key dependencies, the changed logic is explicit.

Thanks for your continued feedback!

[Copilot]

Pull Request Overview

This PR removes the direct dependency on ultralytics by in-lining the required utility functions and updating torch.load calls to include the weights_only parameter.

• In-lined several utility functions from ultralytics into the repo (e.g., check_requirements, parse_requirements, check_version).
• Updated torch.load calls in multiple modules to explicitly pass weights_only=False.

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
yolov5/utils/general.py	In-lined ultralytics utilities and updated module imports for in-lined code.
yolov5/train.py	Updated torch.load call to include weights_only=False to load checkpoints correctly.
yolov5/segment/train.py	Similar torch.load update as in train.py for segmentation models.
yolov5/models/experimental.py	Updated torch.load call for ensemble model loading.
yolov5/hubconf.py	Updated torch.load call to use weights_only=False when loading weights.

Comments suppressed due to low confidence (5)

yolov5/utils/general.py:141

• The os module is used here but is not imported. Add 'import os' at the top of the file.

assert str(os.getenv("YOLO_OFFLINE", "")).lower() != "true"  # check if ENV var YOLO_OFFLINE="True"

yolov5/utils/general.py:196

• The time module is used for sleeping but is not imported. Please add 'import time' at the top.

time.sleep(self.delay * (2**self._attempts))  # exponential backoff delay

yolov5/utils/general.py:152

• The platform module is used here but is not imported; ensure to add 'import platform' at the top of the file.

ARM64 = platform.machine() in {"arm64", "aarch64"}  # ARM64 booleans

yolov5/utils/general.py:227

• The re module is used for regex matching but is not imported; please add 'import re' to the beginning of the file.

match = re.match(r"([a-zA-Z0-9-_]+)\s*([<>!=~]+.*)?", line)

yolov5/utils/general.py:256

• The subprocess module is used to run shell commands but is not imported; add 'import subprocess' to the file.

return subprocess.run(["uv", "-V"], capture_output=True).returncode == 0