[BUG] AttributeError: module 'timm.models._registry' has no attribute 'get_pretrained_cfgs_for_arch' (deprecated import paths)

[Benjamin-Tan]

Describe the bug
As per title.

To Reproduce
Steps to reproduce the behavior:

1. pip3 install timm
2. python3 -c "from timm.models.registry import register_model"

Expected behavior
No error and register_model is imported without any error

Screenshots

Desktop (please complete the following information):

• OS: Ubuntu 22.04
• timm version: 1.0.10
• PyTorch version w/ CUDA/cuDNN : torch version: 2.4.0+cu118

Additional context
Add any other context about the problem here.

[rwightman]

@Benjamin-Tan hmm, there is a small problem in the _all_ for the registry, but it's being triggered by using a deprecated import path, timm.models.registry was deprecated as of two years ago https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/registry.py

it should be from timm.models import register_model

[rwightman]

I'll fix it soon and maybe do another release, but could you tell me what's using the old path, do you have any ability to fix it? I will be removing those import hacks in some future release as they are 2 years old now as mentioned.

[rwightman]

I'll do a 1.0.11 release with the fix (on main now) for this. I'm hoping to get one other small addition in first so probably tomorrow.

I did expose a lot of uses of the long deprecated import path, sigh. Probably by 2025 I will be removing the backwards compatible import paths like this one with some new, significant additions to timm...

[Benjamin-Tan]

@Benjamin-Tan hmm, there is a small problem in the _all_ for the registry, but it's being triggered by using a deprecated import path, timm.models.registry was deprecated as of two years ago https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/registry.py

it should be from timm.models import register_model

this fixes it. Thanks for the update. I did not see the warning message while running. I could remove the .registry. Not sure how many code out there is using the deprecated import path. Another workaround was removing the 'get_pretrained_cfgs_for_arch' from __all__.

[rwightman]

1.0.11 will be out in another 1-2 hours that will fix the issue.

I'll move this to discussions once fixed so people have visibility, to re-iterate what I had at the top of the README (but removed recently since it was there so long)

from timm.models.layers -> from timm.layers
from timm.models.fx_features -> from timm.models
from timm.models.features -> from timm.models
from timm.models.helpers -> from timm.models
from timm.models.hub -> from timm.models
from timm.models.registry -> from timm.models

I am going to enable a warning when the deprecated import paths are used though. I should only fire on the first use of each path.

so

from timm.models.registry import register_model

will spit out this to stderr:

.../timm/models/registry.py:4: FutureWarning: Importing from timm.models.registry is deprecated, please import via timm.models
  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.models", FutureWarning)

Other common ones like

from timm.models.layers import DropPath

will spit out this to stderr:

.../timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)