Give a pre-trained model that can be loaded directly using the ‘model_weights = torch.load(file_path, map_location='cpu')’

[SHIBOYA]

While trying to load the file using Python’s pickle module, I encountered an _pickle.UnpicklingError, stating that persistent IDs in protocol 0 must be ASCII strings. Here is the exact error message:

_pickle.UnpicklingError: persistent IDs in protocol 0 must be ASCII strings

I attempted to resolve the issue by employing various methods, including utilizing the persistent_load parameter with pickle.Unpickler and trying to load the file in different environments, but unfortunately, all efforts have been in vain.

Request for Assistance:
Given the circumstances, I was hoping you could provide some insights or guidance on the following points:

Creation Environment: Could you share details about the environment in which the file was created, including the Python and PyTorch versions used?
Persistent IDs: Any information or context regarding the persistent IDs encountered in the file would be immensely helpful.
Loading Method: If there is a specific method or procedure to correctly load the file, could you please share it with me?
Additional Details: Any other details or specifications about the file that you think might assist in resolving the issue would be greatly appreciated.

[gabrielilharco]

Hi @SHIBOYA. We have our full env at environment.yml. You can install with the command below. Let me know if that solves your issues

conda env create
conda activate task-vectors

[SHIBOYA]

Thank you for your prompt response. I appreciate the guidance provided.

I'd like to inform you that I am indeed running the code within the environment specified in environment.yml. However, I've encountered an issue while trying to install open-clip-torch==2.0.2 using pip. The package seems to be unavailable or there might be some compatibility issues.

Given the challenges I've encountered, I was wondering if it might be possible for you to provide an alternative version of the finetuned model. Specifically, I am looking for a format that might be more universally compatible, such as an .npy file or any other format that can be easily loaded without the specific environment dependencies.

I understand that these issues can sometimes be intricate, and I truly appreciate your patience and assistance in this matter. I'm keen on resolving this to continue with my work.

Thank you once again for your support.

[gabrielilharco]

That version of open clip exists (see https://github.com/mlfoundations/open_clip/releases/tag/v2.0.2), I can install it without problems on my side. If you have dependencies outside of this repository that are causing conflicts, one solution would be to do a clean installation of our requirements, load the model, and then save it in another format so you can load the way you want.