UnboundLocalError: local variable 'wandb_logger' referenced before assignment

[wq9]

🐛 Bug

Multi-GPU Training does not work with latest commit (e8fc97a Improved W&B integration)

To Reproduce (REQUIRED)

python -m torch.distributed.launch --nproc_per_node 4 train.py --data coco128.yaml --cfg yolov5s.yaml --weights '' --batch-size 16

Output:

Traceback (most recent call last):
  File "train.py", line 537, in <module>
Traceback (most recent call last):
  File "train.py", line 537, in <module>
    train(hyp, opt, device, tb_writer)
  File "train.py", line 76, in train
    train(hyp, opt, device, tb_writer)
  File "train.py", line 76, in train
    loggers = {'wandb': wandb_logger.wandb}  # loggers dict
UnboundLocalError: local variable 'wandb_logger' referenced before assignment
    loggers = {'wandb': wandb_logger.wandb}  # loggers dict
UnboundLocalError: local variable 'wandb_logger' referenced before assignment
Traceback (most recent call last):
  File "train.py", line 537, in <module>
    train(hyp, opt, device, tb_writer)
  File "train.py", line 76, in train
    loggers = {'wandb': wandb_logger.wandb}  # loggers dict
UnboundLocalError: local variable 'wandb_logger' referenced before assignment

Expected behavior

Normal Multi-GPU Training

Environment

N/A

Additional context

wandb_logger is not defined for child processes, so child processes terminates prematurely (e8fc97a#diff-ed183d67207df065a11e1289f19d34cc2abbc5448dea952683cfe9728c342b95R76)

[github-actions]

👋 Hello @wq9, thank you for your interest in 🚀 YOLOv5! Please visit our ⭐️ Tutorials to get started, where you can find quickstart guides for simple tasks like Custom Data Training all the way to advanced concepts like Hyperparameter Evolution.

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 training ❓ Question, please provide as much information as possible, including dataset images, training logs, screenshots, and a public link to online W&B logging if available.

For business inquiries or professional support requests please visit https://www.ultralytics.com or email Glenn Jocher at [email protected].

Requirements

Python 3.8 or later with all requirements.txt dependencies installed, including torch>=1.7. To install run:

$ pip install -r requirements.txt

Environments

YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Google Colab and Kaggle notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all YOLOv5 GitHub Actions Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training (train.py), testing (test.py), inference (detect.py) and export (export.py) on MacOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

@wq9 thank you for the bug report!

@AyushExel do you think there's a quick fix for this (i.e. next few hours)? If not we should probably revert #2125 until we have a chance to debug further.

[AyushExel]

@glenn-jocher Yes I'm looking into this now. Any idea what's the difference in execution in multi-GPU case?

[AyushExel]

@glenn-jocher Okay I think I found the solution. Testing it in a bit. I'll let you know

[glenn-jocher]

@AyushExel I'm not sure what it could be. Multi-GPU is always tricky. Rank 0 and -1 processes are intended for the main process/GPU, while all other parts of the code will run on all processes.

Do you have access to a Multi-GPU instance to test on?

[AyushExel]

@glenn-jocher Yes, I'm testing on multi-GPUS. I usually work on single GPU so this test didn't cross my mind. In the previous integration, wandb logging worked only with the single GPU case right? Juding by the previous integration, here:

if rank in [-1, 0] and wandb and wandb.run is None:

[glenn-jocher]

@AyushExel oh, no, master before PR performed W&B logging with Multi-GPU well, i.e.:

See Multi-GPU Tutorial for DP and DDP commands (we need to test both). DDP is strongly preferred though due to it's better performance.

[AyushExel]

@glenn-jocher thanks for the info. I'm working on a fix.

[AyushExel]

@glenn-jocher Okay this is fixed. I'm testing in a different environment without wandb installed. Expect a PR in a few minutes

[AyushExel]

@glenn-jocher I've tested on my end and the fix seems to work in all cases. CI tests are running on the PR now. Hopefully, I haven't introduced any CI typos.
Sorry about this bug though. I'll include DDP in my test cases.

[glenn-jocher]

@wq9 thanks for identifying this problem! We've pushed a fix for this now in PR #2574, and made sure to test Multi-GPU training as well as resuming Multi-GPU training. Please git pull or clone again to receive this update, and let us know if you have any further problems!