Motivation

When using timm as a backbone library, users often switch between different model architectures, such as CNNs and Vision Transformers. These models frequently have different __init__ signatures. For instance, a ViT might accept an img_size argument, while many CNN models do not.

Currently, passing an unsupported keyword argument to timm.create_model results in a TypeError, causing the program to crash. This forces users to write boilerplate conditional logic to manage keyword arguments for each model type, which undermines the simple, abstract interface that create_model is designed to provide.

This PR aims to enhance the robustness and flexibility of create_model by allowing it to handle varied model arguments gracefully, thus improving the user experience.

Changes

1. Improved _ignore_kwargs Utility: The _ignore_kwargs utility function in timm.models._builder has been updated. It now identifies and logs a warning message specifying which keyword arguments are being filtered out because they are not accepted by the target function (model_cls.__init__).

2. Proactive Kwarg Filtering: The build_model_with_cfg function was modified to call _ignore_kwargs immediately before the model class is instantiated. This preemptively removes any extraneous arguments from the kwargs dictionary that would otherwise cause a TypeError.

Outcome

With these changes, create_model now behaves as follows instead of crashing with a TypeError:

• It successfully creates the model by silently ignoring unsupported keyword arguments.
• It logs a clear warning message that informs the user which arguments were ignored.

This improves compatibility across different model types and strengthens the abstraction of the create_model factory. As a result, users can more easily experiment with various architectures without writing cumbersome boilerplate code.

Test Code

import timm

vit_model = timm.create_model('vit_base_patch16_224', pretrained=False, img_size=(256, 256))
conv_model = timm.create_model('convnext_base_in22ft1k', pretrained=False, img_size=(256, 256), good="key")

log

Ignored attempt to pass arguments (img_size, good) to function <function ConvNeXt.__init__ at 0x000001A734F26FC0>.