fix a confusing message for `freeze_bn_stats` by Jiefei-Wang · Pull Request #3166 · DeepLabCut/DeepLabCut

Jiefei-Wang · 2026-01-06T02:37:26Z

This pull request fixes a very confusing message: According to your model configuration, you're training with batch size 1 and/or freeze_bn_stats=false. This is not an optimal setting if you have powerful GPUs.

The condition that triggers this message is loader.model_cfg["model"].get("freeze_bn_stats", False). Therefore, the user can see this message only when freeze_bn_stats=true. This pull request flips the boolean value to give the correct message.

DeepLabCut/deeplabcut/pose_estimation_pytorch/apis/training.py

Lines 170 to 188 in ee866cc

    
           if ( 
        
               loader.model_cfg["model"].get("freeze_bn_stats", False) 
        
               or loader.model_cfg["model"].get("backbone", {}).get("freeze_bn_stats", False) 
        
               or batch_size == 1 
        
           ): 
        
               logging.info( 
        
                   "\nNote: According to your model configuration, you're training with batch " 
        
                   "size 1 and/or ``freeze_bn_stats=false``. This is not an optimal setting " 
        
                   "if you have powerful GPUs.\n" 
        
                   "This is good for small batch sizes (e.g., when training on a CPU), where " 
        
                   "you should keep ``freeze_bn_stats=true``.\n" 
        
                   "If you're using a GPU to train, you can obtain faster performance by " 
        
                   "setting a larger batch size (the biggest power of 2 where you don't get" 
        
                   "a CUDA out-of-memory error, such as 8, 16, 32 or 64 depending on the " 
        
                   "model, size of your images, and GPU memory) and ``freeze_bn_stats=false`` " 
        
                   "for the backbone of your model. \n" 
        
                   "This also allows you to increase the learning rate (empirically you can " 
        
                   "scale the learning rate by sqrt(batch_size) times).\n" 
        
               )

…true`

Copilot

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

deruyter92

Well spotted! Thanks for your contribution.

fix a confusing message: freeze_bn_stats=false -> `freeze_bn_stats=…

67a6440

…true`

MMathisLab requested a review from Copilot January 9, 2026 14:49

Merge branch 'main' into fix

9cdb4d9

MMathisLab added contributions-welcome documentation documentation updates/comments labels Jan 9, 2026

Copilot started reviewing on behalf of MMathisLab January 9, 2026 14:50 View session

Copilot AI reviewed Jan 9, 2026

View reviewed changes

deruyter92 approved these changes Jan 9, 2026

View reviewed changes

deruyter92 merged commit 8fd2f05 into DeepLabCut:main Jan 9, 2026
9 of 10 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix a confusing message for `freeze_bn_stats`#3166

fix a confusing message for `freeze_bn_stats`#3166
deruyter92 merged 2 commits intoDeepLabCut:mainfrom
Jiefei-Wang:fix

Jiefei-Wang commented Jan 6, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

deruyter92 left a comment •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

	if (
	loader.model_cfg["model"].get("freeze_bn_stats", False)
	or loader.model_cfg["model"].get("backbone", {}).get("freeze_bn_stats", False)
	or batch_size == 1
	):
	logging.info(
	"\nNote: According to your model configuration, you're training with batch "
	"size 1 and/or ``freeze_bn_stats=false``. This is not an optimal setting "
	"if you have powerful GPUs.\n"
	"This is good for small batch sizes (e.g., when training on a CPU), where "
	"you should keep ``freeze_bn_stats=true``.\n"
	"If you're using a GPU to train, you can obtain faster performance by "
	"setting a larger batch size (the biggest power of 2 where you don't get"
	"a CUDA out-of-memory error, such as 8, 16, 32 or 64 depending on the "
	"model, size of your images, and GPU memory) and ``freeze_bn_stats=false`` "
	"for the backbone of your model. \n"
	"This also allows you to increase the learning rate (empirically you can "
	"scale the learning rate by sqrt(batch_size) times).\n"
	)

Uh oh!

Conversation

Jiefei-Wang commented Jan 6, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

deruyter92 left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

deruyter92 left a comment •

edited

Loading