Saving the whole model vs only state_dict

[grolysaini]

When loading a model's state_dict which has been saved earlier, we have to first create an instance of the model. Would that mean that you have to have the model in code somewhere? i.e. The model must be defined somewhere to use it? Will TorchScript avoid this requirement of having to have the model class available before loading the model/state_dict?

[fawern]

Hi,
When you save only a state_dict, you’re serializing just the parameter tensors (weights, biases, buffers). You still need the model’s class available at load time.

model.load_state_dict(torch.load("weights.pth"))