load OpenCLIP model and tokenizer from local path

[YushunXiang]

Load OpenCLIP model and tokenizer from local path

Usage:

model, processor = open_clip.create_model_from_pretrained(
    model_name=f'hf-hub:{model_id}',
    precision='fp16',
    cache_dir=cache_dir,
)
print(f"Loading tokenizer from {model_args.model_path}")
tokenizer = open_clip.get_tokenizer(
    model_name=f'hf-hub:{model_id}',
    cache_dir=cache_dir,
)

The cache_dir is your local path.

[rwightman]

so this will interfere with normal use of the cache by overriding part of the functionality. I've made some improvements to the cache operation. Explicitly loading from local folders is a bit of a different idea and I'd like to support that but should be done through pretrained argument (folder) and not by overriding cache behaviour