Training bug when using --sync_bn

[twotwoiscute]

❔Question

Hi I have a problem with using convert_sync_batchnorm ,When I was trying to use DDP everything works fine ,but when I turn on the sync_bn mode ,the training process start and get stuck right away…

## Additional context
Here’s some info :
torch                         1.8.1                   
torchaudio                0.7.0                    
torchvision               0.9.1              

#How I run the script : 
python -m torch.distributed.launch  \
--nproc_per_node 4  \ 
--master_addr $master_addr  \ 
--master_port $port  train.py \ 
--train.py
--batch 256  --weights yolov5s.pt --device 0,1,2,3  \
--sync_bn

#How I init :
#note : opt.local_rank is -1 here 
    if opt.local_rank != -1:
        assert torch.cuda.device_count() > opt.local_rank
        torch.cuda.set_device(opt.local_rank)
        device = torch.device('cuda', opt.local_rank)
        dist.init_process_group(backend='nccl', init_method='env://')

Thanks

[github-actions]

👋 Hello @twotwoiscute, 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 glenn.jocher@ultralytics.com.

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]

@twotwoiscute your code is out of date, the opt class no longer carries DDP variables. To update your code:

• Git – git pull from within your yolov5/ directory or git clone https://github.com/ultralytics/yolov5 again
• PyTorch Hub – Force-reload with model = torch.hub.load('ultralytics/yolov5', 'yolov5s', force_reload=True)
• Notebooks – View updated notebooks
• Docker – sudo docker pull ultralytics/yolov5:latest to update your image

[twotwoiscute]

@glenn-jocher Thanks for the suggestion
However , I clone the latest version , However it still get stucked at forward pass ,Here's what I did :

with amp.autocast(enabled=cuda):
    try:
        print("pred")
        pred = model(imgs)  # forward
        print("done pred")
        loss, loss_items = compute_loss(pred, targets.to(device))  # loss scaled by batch_size
        if RANK != -1:
            loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
        if opt.quad:
            loss *= 4.
    except Exception as error :
             print(error)

I found that if I use 4 gpus and the first 3 gpus would have message print("pred") on the screen and the last gpu never print the message "pred" ,futhermore the first 3 gpus never enter to the code `print("done pred") which means the forward pass never complete in any of gpus.

And this happen after first iteration completed.By the way,sync_bn would only fail if I use multiple gpus, with single gpu everything works perfectly.

[glenn-jocher]

@twotwoiscute ok understood, I will try to reproduce this today.

[glenn-jocher]

@twotwoiscute yes I can reproduce this. I'm not sure what the cause is. In looking at the documentation perhaps we are missing a process_group input to the convert function.
https://pytorch.org/docs/stable/generated/torch.nn.SyncBatchNorm.html

The convert function is here:

yolov5/train.py

Lines 217 to 221 in 3749573

	# SyncBatchNorm
	if opt.sync_bn and cuda and RANK != -1:
	model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
	logger.info('Using SyncBatchNorm()')

[twotwoiscute]

I follow the script you mentioned somehing like :

r1 = [ 0,1,2,3 ]
process_groups = [torch.distributed.new_group(pids) for pids in [r1] ] 
process_group = process_groups[0]

However ,it still does not work ...

[glenn-jocher]

@twotwoiscute I'm not sure if this is due to torch 1.9.0 or our own code, but I don't see anything wrong with the implementation.

I would actually not use --sync if I were you though. It mainly helps in early training but we found it to have little to no effect on final mAP (provided you train long enough, i.e. 300 COCO epochs).

[twotwoiscute]

In my case , I have 64 batch size per gpu , I think it's enough for calculating running mean & var correctly . However,for completeness,maybe to let pytorch team to have a look at this issue ?

[glenn-jocher]

@twotwoiscute yes, that's a good idea. I would raise a bug report on the pytorch repo linking to this issue.

[glenn-jocher]

@twotwoiscute --sync known issue PR in #4032 to alert future users to the existing problem

[github-actions]

👋 Hello, this issue has been automatically marked as stale because it has not had recent activity. Please note it will be closed if no further activity occurs.

Access additional YOLOv5 🚀 resources:

• Wiki – https://github.com/ultralytics/yolov5/wiki
• Tutorials – https://docs.ultralytics.com/yolov5
• Docs – https://docs.ultralytics.com

Access additional Ultralytics ⚡ resources:

• Ultralytics HUB – https://ultralytics.com
• Vision API – https://ultralytics.com/yolov5
• About Us – https://ultralytics.com/about
• Join Our Team – https://ultralytics.com/work
• Contact Us – https://ultralytics.com/contact

Feel free to inform us of any other issues you discover or feature requests that come to mind in the future. Pull Requests (PRs) are also always welcomed!

Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐!

[twotwoiscute]

@twotwoiscute --sync known issue PR in #4032 to alert future users to the existing problem

Have you get any response from pytorch team? I posted the similiar question in PyTorch Forums However, does not get any response..

[glenn-jocher]

@twotwoiscute we still have no resolution on this issue. If you find anything or hear from the PyTorch team please update here, thank you!

[twotwoiscute]

@twotwoiscute we still have no resolution on this issue. If you find anything or hear from the PyTorch team please update here, thank you!

I think this issue would be solved in current version since the way I solve this issue by commenting out :

Thread(target=plot_images, args=(imgs, targets, paths, f), daemon=True).start()

And sync-bn works perfectly fine.

[glenn-jocher]

@twotwoiscute do you mean you comment out L81 in loggers/init.py?

yolov5/utils/loggers/__init__.py

Lines 79 to 81 in b894e69

	if ni < 3:
	f = self.save_dir / f'train_batch{ni}.jpg' # filename
	Thread(target=plot_images, args=(imgs, targets, paths, f), daemon=True).start()

[glenn-jocher]

@twotwoiscute what about the daemon Thread plots in val.py here?

yolov5/val.py

Lines 221 to 227 in b894e69

	# Plot images
	if plots and batch_i < 3:
	f = save_dir / f'val_batch{batch_i}_labels.jpg' # labels
	Thread(target=plot_images, args=(img, targets, paths, f, names), daemon=True).start()
	f = save_dir / f'val_batch{batch_i}_pred.jpg' # predictions
	Thread(target=plot_images, args=(img, output_to_target(out), paths, f, names), daemon=True).start()

[twotwoiscute]

@glenn-jocher Ops! I think i made a wrong statement.. actually I comment out this part of code :

if tb_writer:
    tb_writer.add_graph(torch.jit.trace(de_parallel(model), imgs, strict=False), [])  # model graph
    # tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) 

but keep the ploting part .

[glenn-jocher]

@twotwoiscute I don't think this line correlates with the issue at all. It doesn't actually exist anymore, it's been replaced with loggers/init.py L137 that only runs on the on_train_end() callback, which will not have any effect on starting training with or without --sync:

yolov5/utils/loggers/__init__.py

Lines 126 to 138 in b894e69

	def on_train_end(self, last, best, plots, epoch):
	# Callback runs on training end
	if plots:
	plot_results(file=self.save_dir / 'results.csv') # save results.png
	files = ['results.png', 'confusion_matrix.png', *[f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R')]]
	files = [(self.save_dir / f) for f in files if (self.save_dir / f).exists()] # filter
	if self.tb:
	from PIL import Image
	import numpy as np
	for f in files:
	self.tb.add_image(f.stem, np.asarray(Image.open(f)), epoch, dataformats='HWC')

[glenn-jocher]

@twotwoiscute wait are you saying the line you commented in your comment is uncommented, this line?
This line does run at train start, it's been moved to loggers/init.py L78. I can try commenting it, thanks for the tip!

yolov5/utils/loggers/__init__.py

Lines 72 to 78 in b894e69

	def on_train_batch_end(self, ni, model, imgs, targets, paths, plots):
	# Callback runs on train batch end
	if plots:
	if ni == 0:
	with warnings.catch_warnings():
	warnings.simplefilter('ignore') # suppress jit trace warning
	self.tb.add_graph(torch.jit.trace(de_parallel(model), imgs[0:1], strict=False), [])

[twotwoiscute]

@glenn-jocher okay,last weekend I tried to comment out plot_image part and it works ,but today it gets stuck again...but when I comment out self.tb.add_graph(torch.jit.trace(de_parallel(model), imgs[0:1], strict=False), []) and works again ...

[glenn-jocher]

@twotwoiscute ok got it, thanks for the feedback! These are two very different things, add images is just adding images to TensorBoard, add_graph is adding an interactive YOLOv5 model visualizer (below), which is a much more complicated operation. I don't know what it has to do with --sync but if it's causing the hang we can simply not add graphs on --sync trainings as a workaround.

[glenn-jocher]

@twotwoiscute good news 😃! Your original issue may now be fixed ✅ in PR #4615. We discovered of the DPP --sync-bn issue was caused by TensorBoard add_graph() logging (used for visualizing the model interactively, example below). I don't know the exact cause and thus did not implement a fix, instead I implemented a workaround to avoid TensorBoard model visualization when --sync-bn is used.

This means DDP training now works without issue with or without --sync-bn, but --sync-bn runs will not show a model visualization component in TensorBoard.

To receive this update:

• Git – git pull from within your yolov5/ directory or git clone https://github.com/ultralytics/yolov5 again
• PyTorch Hub – Force-reload with model = torch.hub.load('ultralytics/yolov5', 'yolov5s', force_reload=True)
• Notebooks – View updated notebooks
• Docker – sudo docker pull ultralytics/yolov5:latest to update your image

Thank you for spotting this issue and informing us of the problem. Please let us know if this update resolves the issue for you, and feel free to inform us of any other issues you discover or feature requests that come to mind. Happy trainings with YOLOv5 🚀!