Checkpoint a model
Persist trained weights for later inference or continued training.
PyTorch · Module
model.state_dict
Returns a dictionary-like mapping of a module’s learnable parameters and registered buffers. Use it to save, inspect, modify, or restore model state.
How it works
Calling state_dict() produces a dictionary-like snapshot of a module’s learnable parameters and registered buffers, using names such as weight and bias as keys. The result can be serialized for checkpointing and later supplied to load_state_dict() on a compatible model instance.
Syntax
Module.state_dict(*args, destination=None, prefix='', keep_vars=False)Returns
A Python dictionary-like state mapping containing the module’s parameters and registered buffers.
Key points
- The mapping uses module or layer names as keys and stores the corresponding parameter tensors as values.
- Its dictionary representation makes checkpoints straightforward to serialize and manipulate.
- The documented save pattern serializes the result of model.state_dict() with torch.save().
Common mistakes
- The result includes learnable layers and registered buffers, but not arbitrary attributes that were assigned to the module.
Example: Save and restore module weights
import torch
from torch import nn
model = nn.Linear(2, 1)
state = model.state_dict()
print(list(state.keys()))
path = "model_weights.pth"
torch.save(state, path)
restored = nn.Linear(2, 1)
restored.load_state_dict(torch.load(path, weights_only=True))Creates a linear model, obtains its state dictionary, saves it, and loads it into a matching model.
When to use it
Inspect module state
Inspect parameter names and tensor shapes during debugging.
Restore model parameters
Transfer learned weights into a newly constructed compatible model.
Interview questions
QWhat does Module.state_dict() return?
state_dict() returns a Python dictionary-like mapping from parameter or buffer names to tensors. It is the standard object used to persist a module’s learned state separately from its code.
Follow-up: How would you restore it? Instantiate a compatible module, then call load_state_dict() with the deserialized mapping.
QWhich values appear in a module’s state_dict()?
It contains entries for learnable parameters and registered buffers, such as BatchNorm running statistics. Ordinary attributes that are not registered as parameters or buffers are not included.
Follow-up: Why are buffers useful here? They preserve non-learnable module state needed for correct inference, such as running statistics.
QWhy is state_dict() commonly used for saving models?
A state dictionary is a Python dictionary object, so it can be saved, changed, and restored. The usual checkpoint pattern is torch.save(model.state_dict(), path), followed by loading it into a matching model.
Follow-up: Why save the state dictionary rather than the entire model? It separates weights from the model’s class definition and is the documented modular checkpointing pattern.
Related
torch.nn.Module.load_state_dicttorch.savetorch.loadtorch.nn.Module.parameters