Are all models expected to have the attribute default_cfg?

[jalabort]

Hi all, thanks for putting together this great project.

I noticed that when using some models as feature extractors some of them seem to have the attribute default_cfg defined:

model = timm.create_model("mobilenetv2_100", features_only=True, pretrained=True)
model.default_cfg

'input_size': (3, 224, 224),
 'pool_size': (7, 7),
 'crop_pct': 0.875,
 'interpolation': 'bicubic',
 'mean': (0.485, 0.456, 0.406),
 'std': (0.229, 0.224, 0.225),
 'first_conv': 'conv_stem',
 'classifier': 'classifier'}

while others do not:

model = timm.create_model("resnet18", features_only=True, pretrained=True)
model.default_cfg

---------------------------------------------------------------------------
ModuleAttributeError                      Traceback (most recent call last)
<ipython-input-52-b9386d3287d2> in <module>
----> 1 backbone.default_cfg

/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in __getattr__(self, name)
    777                 return modules[name]
    778         raise ModuleAttributeError("'{}' object has no attribute '{}'".format(
--> 779             type(self).__name__, name))
    780 
    781     def __setattr__(self, name: str, value: Union[Tensor, 'Module']) -> None:

ModuleAttributeError: 'FeatureListNet' object has no attribute 'default_cfg'

Is this the expected behaviour? I would have expected all of them to have this attribute defined but perhaps I am missing something here...

Any information you might have regarding the above will be greatly appreciated.

[rwightman]

@jalabort all models should have .default_cfg attribute if they are being used as classifiers or as backbone with the classifier removed. When they are wrapped in feature extractors I did not make an effort to maintain the default_cfg as it wasn't clear that information was generally useful in most situations where feature map extraction is used.

The difference is due to the way the features are extracted, there is variety between the models due to the way they are constructed. The EfficientNet/MobileNet family has its own mechanism that predates the extraction wrappers used for ResNets and many of the other models.

I can look into copying that attr into the feature extraction wrappers for the models that use them. All I ask is your use case so that I understand the motivation.

[jalabort]

Hi @rwightman, thanks for your reply.

I am using some of the timm models as backbones in my object detection, re-identification, etc... models. Looking at some of the training scripts (here and here) my understanding was that default_cfg contained the ImageNet mean and std that were used to normalized the training images used to train the models. There are other easy ways for me to get ImageNet mean and std; it is just that the default_cfg looked like a pretty neat way of doing it since I wasn't sure all models were actually trained by normalizing the images using those two values.

I guess there's also a chance that I have misunderstood how these values are used in your library as well 😅...

[rwightman]

@jalabort those values are used by me and others for that purpose. It would make sense to have access to them via the feature extraction wrappers as well. I'll fix it. Maybe I'll remove the ones that don't make sense like the classifier layer name, crop, num_classes, etc.

[jalabort]

Excellent! That'd make sense to me. Congrats again on the excellent project you've put together!

[rwightman]

Merged to master. Will likely push a pypi release of it in a few days time.