#serialization

Question 1: Single choice

Why does PyTorch recommend saving model.state_dict() rather than the whole model with torch.save(model, PATH)?

Pickle stores only a reference to the model's class, so loading breaks if that code moves1
The state_dict also stores the architecture's code, so no class definition is needed to load it2
torch.save(model) keeps only the trainable weights and drops buffers such as BatchNorm statistics3
A state_dict file is stored in ONNX format, so any framework can load it without PyTorch4