About loading model weights about ultralytics HOT 4 OPEN

@glenn-jocher I came up with a solution,
Only the pure model parameters are saved in version 8.1.106, and then the pure model parameters are obtained under 8.1.44.

from ultralytics.

Hi there! It seems like the issue might be related to incompatible versions between the one used to save the model (ultralytics==8.1.106) and the one you're trying to load it with (ultralytics==8.1.44). You might be able to resolve this error by ensuring both the saving and loading are performed with the same version of the library. If you cannot run the same older version, consider upgrading your current environment to match the version used for saving the model. Here’s how you can upgrade the library:

pip install ultralytics==8.1.106

Just make sure that the version number matches exactly to the one used for saving the model weights. Let me know if this helps or if you encounter any other issues! 🚀

from ultralytics.

I am aware that the upgrade to 8.1.106 is capable of reading this weight, but I would like to read this weight under 8.1.44 for retraining.

from ultralytics.

@tongchangD hi there! That sounds like a great approach! By saving only the model parameters in version 8.1.106 and then loading them in version 8.1.44, you should be able to bypass the module compatibility issues. Here's a quick example of how you can do this:

Saving model parameters in 8.1.106

# Assuming 'model' is your trained model
torch.save(model.state_dict(), 'model_weights.pth')

Loading model parameters in 8.1.44

# Make sure to initialize the model architecture similarly as it was in 8.1.106
model.load_state_dict(torch.load('model_weights.pth', map_location=torch.device('cpu')))
model.eval()

This method ensures that you're only dealing with the raw weights and biases, independent of any specific module structure that might have changed between versions. Let me know if this works for you! 😊

from ultralytics.