Running example error

[IceIce1ce]

Hi author, thanks for your great code. When I run your example code in your README, an error happened. Can you check your code again?

Loading checkpoint shards: 100%|██████████████████| 2/2 [00:01<00:00, 1.63it/s]
Traceback (most recent call last):
File "/home/vsw/Desktop/RADIO/main.py", line 63, in
model = torch.hub.load('NVlabs/RADIO', 'radio_model', version=model_version, progress=True, skip_validation=True, adaptor_names=['siglip2-g'])
File "/home/vsw/miniconda3/envs/radio/lib/python3.10/site-packages/torch/hub.py", line 566, in load
model = _load_local(repo_or_dir, model, *args, **kwargs)
File "/home/vsw/miniconda3/envs/radio/lib/python3.10/site-packages/torch/hub.py", line 595, in _load_local
model = entry(*args, **kwargs)
File "/home/vsw/.cache/torch/hub/NVlabs_RADIO_main/hubconf.py", line 143, in radio_model
adaptor = adaptor_registry.create_adaptor(ttype, chk["args"], tconf, adaptor_state)
File "/home/vsw/.cache/torch/hub/NVlabs_RADIO_main/radio/adaptor_registry.py", line 34, in create_adaptor
return self._registry[name](main_config, adaptor_config, state)
File "/home/vsw/.cache/torch/hub/NVlabs_RADIO_main/radio/siglip2_adaptor.py", line 96, in create_siglip2_adaptor
return SigLIP2Adaptor(main_config, adaptor_config, state)
File "/home/vsw/.cache/torch/hub/NVlabs_RADIO_main/radio/siglip2_adaptor.py", line 37, in init
model = AutoModel.from_pretrained(version, trust_remote_code=True)
File "/home/vsw/miniconda3/envs/radio/lib/python3.10/site-packages/transformers/models/auto/auto_factory.py", line 563, in from_pretrained
return model_class.from_pretrained(
File "/home/vsw/miniconda3/envs/radio/lib/python3.10/site-packages/transformers/modeling_utils.py", line 3677, in from_pretrained
) = cls._load_pretrained_model(
File "/home/vsw/miniconda3/envs/radio/lib/python3.10/site-packages/transformers/modeling_utils.py", line 4155, in _load_pretrained_model
raise RuntimeError(f"Error(s) in loading state_dict for {model.class.name}:\n\t{error_msg}")
RuntimeError: Error(s) in loading state_dict for SiglipModel:
size mismatch for text_model.head.weight: copying a param with shape torch.Size([1536, 1152]) from checkpoint, the shape in current model is torch.Size([1152, 1152]).
size mismatch for text_model.head.bias: copying a param with shape torch.Size([1536]) from checkpoint, the shape in current model is torch.Size([1152]).
You may consider adding ignore_mismatched_sizes=True in the model from_pretrained method.

[mranzinger]

Which model version are you trying to run?

[IceIce1ce]

Hi author, thanks for your reply. The model version which I used is model_version="c-radio_v4-h" and the example I tried to run is the first example in your README.

[gheinrich]

Hello, I am not getting this error when I run the example in the README. This seems to relate to an inconsistency between the model definition and the shape of the weights in the checkpoint. Did you run the example command in a clean environment?

The error occurs during the loading of SigLIP2-g adapters, which may not need depend on your use case (they are normally only useful if you want to do e.g. zero-shot classification).

You may also try to force reload in this statement, i.e. replace:

model = torch.hub.load('NVlabs/RADIO', 'radio_model', version=model_version, progress=True, skip_validation=True, adaptor_names=['siglip2-g'])

With:

model = torch.hub.load('NVlabs/RADIO', 'radio_model', version=model_version, progress=True, skip_validation=True, force_reload=True, adaptor_names=['siglip2-g'])

[IceIce1ce]

Hi, I will try again. Thanks for your help.