How to load a HF model from a locally downloaded version and not a hub copy?

[CoffeeVampir3]

pytorch-image-models/timm/models/_factory.py

Line 38 in 59b3d86

def create_model(

TIMM seems to have a loader for HF based models, but I can't see how I can simply load an existing downloaded model using the same config info, without uploading it to the hub, and then letting timm redownload it again. Surely I'm missing something here, cheers.

[CoffeeVampir3]

It looks as if there's no built in way, but I did hack something together:

def load_cfg_from_json(json_file):
    with open(json_file, "r", encoding="utf-8") as reader:
        text = reader.read()
    return json.loads(text)

def load_cfg(cfg_path):
    hf_config = load_cfg_from_json(cfg_path)
    if 'pretrained_cfg' not in hf_config:
        # old form, pull pretrain_cfg out of the base dict
        pretrained_cfg = hf_config
        hf_config = {}
        hf_config['architecture'] = pretrained_cfg.pop('architecture')
        hf_config['num_features'] = pretrained_cfg.pop('num_features', None)
        if 'labels' in pretrained_cfg:  # deprecated name for 'label_names'
            pretrained_cfg['label_names'] = pretrained_cfg.pop('labels')
        hf_config['pretrained_cfg'] = pretrained_cfg
    
    pretrained_cfg = hf_config['pretrained_cfg']
    pretrained_cfg['hf_hub_id'] = hf_config["architecture"]  # insert hf_hub id for pretrained weight load during model creation
    pretrained_cfg['source'] = 'hf-hub'
    
    # model should be created with base config num_classes if its exist
    if 'num_classes' in hf_config:
        pretrained_cfg['num_classes'] = hf_config['num_classes']
    
    # label meta-data in base config overrides saved pretrained_cfg on load
    if 'label_names' in hf_config:
        pretrained_cfg['label_names'] = hf_config.pop('label_names')
    if 'label_descriptions' in hf_config:
        pretrained_cfg['label_descriptions'] = hf_config.pop('label_descriptions')
    
    model_args = hf_config.get('model_args', {})
    model_name = hf_config['architecture']

    return pretrained_cfg, model_name, model_args

#"/home/blackroot/Desktop/ViTagger/model/config.json"
def load_model_but_good_this_time_srsly(model_path, model_config):
    pretrained_cfg, model_name, model_args = load_cfg(model_config)
    model = timm.create_model(
        model_name, 
        checkpoint_path=model_path, 
        pretrained_cfg=pretrained_cfg,
        **model_args)

    return model

not sure if this is worth a PR to add support for, but with the rise of HF models and their configs this might be a worthwhile addition.