Option to provide path to `harnet`

[muschellij2]

Can you provide a link/path to the SSL net that has already been downloaded as opposed to using torch cache in

asleep/src/asleep/sslmodel.py

Line 75 in e544846

def get_sslnet(tag='v1.0.0', pretrained=False):

?

[angerhang]

Can you provide a link/path to the SSL net that has already been downloaded as opposed to using torch cache in

asleep/src/asleep/sslmodel.py

Line 75 in e544846

def get_sslnet(tag='v1.0.0', pretrained=False):
?

Can you elaborate a bit on your use case?

Are you looking for the cached model checkpoint path?

[muschellij2]

If I already have harnet locally downloaded, there is no way to pass that in to get_sleep as I see it other than through a cached version in torch cache. The local_repo_path seems only for this repo, not ssl-wearables.

[angerhang]

Ah thank you. We will take a look at this some time this week.

[angerhang]

@muschellij2 after a quick look, indeed the local path option is not supported yet. Do you this feature because you want to run the model in an environment without the internet or you want to test different model weights?

[muschellij2]

1. I was getting errors when downloading the file in certain runs
2. I was using a temporary environment with uv when processing but didn't want to download the model each time
3. For reproducibility (in case harnet version changes in future)

[angerhang]

1. I was getting errors when downloading the file in certain runs
2. I was using a temporary environment with uv when processing but didn't want to download the model each time
3. For reproducibility (in case harnet version changes in future)

gotchu, yeah I can imagine it will be some on-demand systems like the AWS. Don't think harnet will be changed within the foreseeable future. I can see that you can trigger rerun if certain run fails due to this.

Essentially this feature requires changing the ssl load function and potentially adding some of the model definition to this to support local loading. This use case wasn't considered in conception.

We are a bit limited on the bandwidth to prioritise this feature given that most users won't have on-demand systems. But very happy to give you some code pointers and review your pull requests if you can make one?