How to loads weights (as pretrained) from local files (supporting head adaptation, features_only, etc)

[RealAntonVoronov]

Describe the bug
The bug happens when trying to specify local path to pretrained weights (in order to run model on devices with no access to the Internet). When passing a custom config as an argument to a function code falls when initializing an EfficientNet instance with the keyword argument 'cfg', which is not present in EfficientNet class init.

To Reproduce

from timm.models.efficientnet import _cfg, tf_efficientnet_b7_ns

config = _cfg(url='', file='../tf_efficientnet_b7_ns-1dbc32de.pth',
              input_size=(3, 600, 600), pool_size=(19, 19), crop_pct=0.949)
model = tf_efficientnet_b7_ns(pretrained=True, model_cfg=config, drop_path_rate=0.2)

---

TypeError Traceback (most recent call last)
/tmp/ipykernel_4061624/3684105051.py in
3 config = _cfg(url='', file='../tf_efficientnet_b7_ns-1dbc32de.pth',
4 input_size=(3, 600, 600), pool_size=(19, 19), crop_pct=0.949)
----> 5 model = tf_efficientnet_b7_ns(pretrained=True, model_cfg=config, drop_path_rate=0.2)

~/.imgenv-fakedetection-voronov-0/lib/python3.7/site-packages/timm/models/efficientnet.py in tf_efficientnet_b7_ns(pretrained, **kwargs)
1916 kwargs['pad_type'] = 'same'
1917 model = _gen_efficientnet(
-> 1918 'tf_efficientnet_b7_ns', channel_multiplier=2.0, depth_multiplier=3.1, pretrained=pretrained, **kwargs)
1919 return model
1920

~/.imgenv-fakedetection-voronov-0/lib/python3.7/site-packages/timm/models/efficientnet.py in _gen_efficientnet(variant, channel_multiplier, depth_multiplier, pretrained, **kwargs)
842 **kwargs,
843 )
--> 844 model = _create_effnet(variant, pretrained, **model_kwargs)
845 return model
846

~/.imgenv-fakedetection-voronov-0/lib/python3.7/site-packages/timm/models/efficientnet.py in _create_effnet(variant, pretrained, **kwargs)
597 pretrained_strict=not features_only,
598 kwargs_filter=kwargs_filter,
--> 599 **kwargs)
600 if features_only:
601 model.default_cfg = default_cfg_for_features(model.default_cfg)

~/.imgenv-fakedetection-voronov-0/lib/python3.7/site-packages/timm/models/helpers.py in build_model_with_cfg(model_cls, variant, pretrained, default_cfg, model_cfg, feature_cfg, pretrained_strict, pretrained_filter_fn, pretrained_custom_load, kwargs_filter, **kwargs)
451
452 # Build the model
--> 453 model = model_cls(**kwargs) if model_cfg is None else model_cls(cfg=model_cfg, **kwargs)
454 model.default_cfg = default_cfg
455

TypeError: init() got an unexpected keyword argument 'cfg'

Expected behavior
I expect the model to be created with weights loaded from local file.

Desktop (please complete the following information):

• OS: Ubuntu 18.04
• timm==0.5.4
• PyTorch version: 1.9.0+cu111

[rwightman]

@RealAntonVoronov that's now how the cfg args are used (model_cfg is for architectures that support specifcying the full arch via a config without any intemediary), the support for 'file' in pretrained_cfg for local pretrained weights is a WIP (ie not fully complte & tested), that said, that config is to be passed via pretrained_cfg argument. For now it's best to continue as before and make sure the downloaded weight is in the cache dir

So there is a chance this will work...

pretrained_config = _cfg(url='', file='../tf_efficientnet_b7_ns-1dbc32de.pth',
              input_size=(3, 600, 600), pool_size=(19, 19), crop_pct=0.949)
model = tf_efficientnet_b7_ns(pretrained=True, pretrained_cfg=pretrained_config, drop_path_rate=0.2)

[RealAntonVoronov]

No, this does not work as well. Falls with the same error message. As far as I understand the problem is that model_cls is an instance of EfficientNet class and it tries to initialize it with passing the argument pretrained_cfg or in my original case model_cfg, but this class does not have any of this arguments. Thanks for the reply anyway, I got that this feature is not implemented yet, I'll workaround it some other way.

[rwightman]

@RealAntonVoronov I took a closer look at this since it's not supposed to be completely broken... it seems to work for me

>>> ptcfg = timm.models.registry.get_pretrained_cfg('tf_efficientnet_b7_ns')
>>> ptcfg.pop('url')
'https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_efficientnet_b7_ns-1dbc32de.pth'
>>> ptcfg['file']='./tf_efficientnet_b7_ns-1dbc32de.pth'
>>> ptcfg
{'num_classes': 1000, 'input_size': (3, 600, 600), 'pool_size': (19, 19), 'crop_pct': 0.949, 'interpolation': 'bicubic', 'mean': (0.485, 0.456, 0.406), 'std': (0.229, 0.224, 0.225), 'first_conv': 'conv_stem', 'classifier': 'classifier', 'file': './tf_efficientnet_b7_ns-1dbc32de.pth'}
>>> m = timm.models.tf_efficientnet_b7_ns(pretrained=True, pretrained_cfg=ptcfg, drop_path_rate=0.2)
>>> m.classifier.weight
Parameter containing:
tensor([[ 0.0050, -0.0276,  0.0390,  ..., -0.0076, -0.0338,  0.0537],
        [-0.0399,  0.0463,  0.0057,  ..., -0.0437, -0.0358,  0.0621],
        [-0.0185, -0.0606,  0.0494,  ..., -0.0119, -0.0221,  0.0491],
        ...,
        [-0.0137,  0.0394, -0.0450,  ...,  0.0036, -0.0685,  0.0440],
        [-0.0168,  0.0478, -0.0789,  ..., -0.0590, -0.0488, -0.0713],
        [-0.0700,  0.0158,  0.0114,  ..., -0.0023,  0.0537,  0.0528]],
       requires_grad=True)

Pretrained via defaults (url):

>>> m = timm.models.tf_efficientnet_b7_ns(pretrained=True, drop_path_rate=0.2)
>>> m.classifier.weight
Parameter containing:
tensor([[ 0.0050, -0.0276,  0.0390,  ..., -0.0076, -0.0338,  0.0537],
        [-0.0399,  0.0463,  0.0057,  ..., -0.0437, -0.0358,  0.0621],
        [-0.0185, -0.0606,  0.0494,  ..., -0.0119, -0.0221,  0.0491],
        ...,
        [-0.0137,  0.0394, -0.0450,  ...,  0.0036, -0.0685,  0.0440],
        [-0.0168,  0.0478, -0.0789,  ..., -0.0590, -0.0488, -0.0713],
        [-0.0700,  0.0158,  0.0114,  ..., -0.0023,  0.0537,  0.0528]],
       requires_grad=True)

Random init:

>>> m = timm.models.tf_efficientnet_b7_ns(pretrained=False, drop_path_rate=0.2)
>>> m.classifier.weight
Parameter containing:
tensor([[ 2.3909e-02, -2.5098e-02,  3.1557e-02,  ..., -4.3423e-04,
          2.6137e-03,  6.8434e-03],
        [-3.7444e-03, -1.7700e-03, -1.9717e-02,  ..., -8.9100e-03,
          9.8918e-03, -1.9784e-02],
        [-7.5210e-05, -2.5257e-02,  2.5059e-02,  ...,  1.6234e-02,
          5.3525e-03, -2.8231e-02],
        ...,
        [ 1.2436e-02, -7.4817e-03, -1.1816e-02,  ...,  2.0292e-02,
          1.9143e-02,  2.9750e-02],
        [ 2.5358e-02,  2.3163e-02,  1.0472e-03,  ..., -2.0183e-02,
         -1.3123e-02,  1.7490e-02],
        [ 1.1625e-02,  1.0379e-02,  3.1121e-02,  ...,  2.0877e-03,
          2.7863e-02, -2.6154e-02]], requires_grad=True)

[rwightman]

Note has been recent work as mentioned, so it's possible it was broken if you aren't up to date with the very latest. But let me know if there are issues as I want to have this working soon.

[rwightman]

I'm going to move this to discussion as I feel some others might be interested... also related to #488 #967