[FEATURE] Consistent License Handling

[alexanderdann]

Is your feature request related to a problem? Please describe.

When using timm in commercial context it is very important to use models and weights which also have a viable license. Some models already have it in the config such as

model = timm.create_model("convnextv2_pico", pretrained=False)
print(model.default_cfg['license'])

>> cc-by-nc-4.0

However, now switching to the V1 I get an error

import timm

model = timm.create_model("convnext_pico", pretrained=False)
print(model.default_cfg['license'])

---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
Cell In[6], line 4
      1 import timm
      3 model = timm.create_model("convnext_pico", pretrained=False)
----> 4 print(model.default_cfg['license'])

KeyError: 'license'

and it is not really clear whether this refers to the model implementation or the weights. Looking in the source code is also ambiguous where it says:

# ConvNeXt
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the MIT license

# ConvNeXt-V2
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree (Attribution-NonCommercial 4.0 International (CC BY-NC 4.0))
# No code was used directly from ConvNeXt-V2, however the weights are CC BY-NC 4.0 so beware if using commercially.

So if no code from the official repo was used, then it should also only fall under the license of timm. Nonetheless, for me it is more about having the information easily accessible and consistent for each model and any changes to licenses need to be accessible as well, to ensure a previously commercially-viable model does not become an issue if the license becomes GPL or anything else which can become a problem.

Describe the solution you'd like

Instead of having to look into source code and look for every weight config at https://huggingface.co/timm/models it would make sense to make this information easily accessible through default_cfg with entries model_license & weights_license

For instance:

import timm

model = timm.create_model("convnextv2_pico", pretrained=True)

model.default_cfg["model_license"], model.default_cfg["weights_license"]

>> "cc-by-nc-4.0", "cc-by-nc-4.0"

or

import timm

model = timm.create_model("convnextv2_pico", pretrained=True)

model.licenses

>> {
"model": "cc-by-nc-4.0",
"weights": "cc-by-nc-4.0",
}

Describe alternatives you've considered

Currently it is just manual work everytime you want to try a new model.

Additional context

None

[rwightman]

@alexanderdann as you noted the info is usually there and correct on the Hub model card, so someone could probably create an automated script to merge the info with the code default_cfgs... some default_cfgs, especially recent models that have an unusual or restrictive license have the 'license' field but most do not.

[alexanderdann]

Thanks for the quick follow-up. Is there a preference on your side how you would like to handle this? Since I had to do a deep dive into the source code anyways, I could also help you there.

[rwightman]

There's already a 'license' field in the pretrained cfg (default_cfg) schema, you can see it used in a few (but not many cases) by grepping for 'license':

So the idea would be to fill those out in the default cfg entries for each model weight, which is going to be a bit of work as there are thousands of them. Some can be short circuited by adding to the _cfg() helpers for various model families and then overriding for some models that differ in the individual entries.

I'm pretty sure if you worked with the claude code agent and gave it this info + the idea to cross reference with the HF hub API you could probably get it to semi-automate much of that. The license strings should ideally be lower case and match the ones used in the HF hub (where they overlap, most of them should).

You can programatically get the license from the HF hub entries like below, and the existing pretrain_cfgs will have the model (repo) ids in them for almost all of the models.

from huggingface_hub import model_info
# Specify the model ID
model_id = "timm/resnet50.a1_in1k"
# Get model information
info = model_info(repo_id=model_id,)
# Access the license
license = info.card_data.get("license") if info.card_data else None
print(f"License for {model_id}: {license}")

[alexanderdann]

Thanks! I added a pull request which should go into the direction of your suggestions: #2585