Set `TORCH_CPP_LOG_LEVEL=ERROR` for reduced verbosity

ultralytics / yolov5

YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite

https://docs.ultralytics.com

GNU Affero General Public License v3.0

47.78k stars 15.73k forks source link

Set `TORCH_CPP_LOG_LEVEL=ERROR` for reduced verbosity #12989

Closed inisis closed 2 weeks ago

inisis commented 3 weeks ago

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced Logging Control in YOLOv5 Export Script ✨

📊 Key Changes

Added a line to set the environment variable TORCH_CPP_LOG_LEVEL to "ERROR" in export.py.

🎯 Purpose & Impact

Purpose: To reduce the verbosity of log messages from the PyTorch C++ backend during the export process. This change specifically filters out messages less severe than errors, making significant messages more noticeable.
Impact: Users exporting models will experience a cleaner console output, making it easier to spot and address actual errors. This can lead to a more user-friendly experience during model export, particularly for those troubleshooting export issues. 🚀

github-actions[bot] commented 3 weeks ago

CLA Assistant Lite bot All Contributors have signed the CLA. ✅

inisis commented 3 weeks ago

I have read the CLA Document and I sign the CLA

glenn-jocher commented 3 weeks ago

@inisis thanks for the PR! What type of output are you seeing that will help suppress?

BTW we have an issue with ultralytics that we'd like to switch to PyTorch's built-in profiler but we have a similar problem that it is too verbose. Do you have any idea how we can silence these logging messages?

If we can figure out how to quiet the native FLOPs function down then we could migrate to it and ditch thop alltogether (would be awesome to eliminate an unmaintained dependency). The too-verbose output looks like this:

inisis commented 3 weeks ago

@glenn-jocher for you first question, here are the screen shot. and my pr can help fix it.

and for your second question, we can add export KINETO_LOG_LEVEL=5 to supress it, I can help pr it.

glenn-jocher commented 3 weeks ago

Thanks for the screenshots, @inisis! Your PR looks promising for addressing the log issue in the first screenshot. 😊 For suppressing verbose output from PyTorch's built-in profiler as indicated in the second screenshot, setting KINETO_LOG_LEVEL=5 sounds like a great approach. It would be fantastic if you could submit a PR for that as well! 🚀

inisis commented 3 weeks ago

@glenn-jocher can you cc me in that issue in ultralytics , i searched but didn't find it. and I think this pr can be merged and closed.

glenn-jocher commented 2 weeks ago

Hi @inisis! I'll make sure to mention you in the relevant issue on ultralytics once it's up for easier tracking. Regarding this PR, we'll review it shortly and merge if everything checks out. Thanks for your contributions! 😊👍