Saving Trained Model

[michael-camilleri]

Hi Gorjan

Once trained, how can I save a CACNF model? does setting the save_model_path store also the backbones (including the fine-tuned Resnet) or do I need to store these (and load them) separately using the resnet_model_path and save_backbone_path?

[gorjanradevski]

Hi Michael, using the save_model_path saves the whole model, including the classifier, useful if you want to use the same model for inference. The save_backbone_path, saves the whole model, excluding the classifier. Useful if you want to fine-tune the backbone on a different dataset, with different number of classes.

When performing inference, you should create the model model = CrossAttentionCentralNetFusion, and then load the full model that you saved with save_model_path (see inference.py). On the other hand, if you just want to load the backbone, and have a randomly initialized classifier with different number of classes, you can provide the load_backbone_path and set freeze_backbone to True or False, in case you want to fine-tune the backbone or not, respectively.

The fact that you need to always provide the resnet_model_path when creating a model is poor API design, but didn't have time to think about a workaround. Ideally, if you want to load either a backbone, or a full model, you should not be obliged to provide the resnet_model_path.

[michael-camilleri]

As long as the resnet is still loaded correctly, I don't mind passing a path.

[gorjanradevski]

I'm closing this issue, feel free to reopen if the problem persists.