# checkpoint

Python API: `luxonis_train.utils.checkpoint`

The filter for the checkpoint keys of the attached modules.

A loss, a metric, or a visualizer stores its node in `_node`. A state dict that includes such a module holds the parameters and
the buffers of the node again, under the key of the module.
[CHECKPOINT_FILTERED_STATE_DICT_PATTERN](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/utils/checkpoint.md)
matches these keys, and
[filter_checkpoint_state_dict](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/utils/checkpoint.md)
drops them.

A key matches the pattern when it starts with `nodes.<node>.losses.`, `nodes.<node>.metrics.`, or `nodes.<node>.visualizers.`, and
holds `_node.` after that prefix. `<node>` is a name without a dot. For example, `nodes.head.losses.loss._node.weight` matches.

[LuxonisLightningModule](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/lightning/luxonis_lightning.md)
drops these keys when it saves a checkpoint. It also ignores them when a run resumes with strict weight loading.
[EMACallback.state_dict](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/callbacks/ema.md)
drops them from the average that it returns.
[EMACallback.on_fit_start](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/callbacks/ema.md)
ignores them when it restores an average from a checkpoint.

## Functions

### filter_checkpoint_state_dict

```python
def filter_checkpoint_state_dict(state_dict: Mapping[str, Tensor]) -> dict[str, Tensor]:
```

Drop the keys that the attached modules hold for their node.

The function returns a new dictionary without the keys that match
[CHECKPOINT_FILTERED_STATE_DICT_PATTERN](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/utils/checkpoint.md).
It does not change `state_dict`.

> **Example**
> ```pycon
>>> import torch
>>> state_dict = {
...     "nodes.head.module.weight": torch.ones(1),
...     "nodes.head.losses.loss._node.weight": torch.ones(1),
... }
>>> list(filter_checkpoint_state_dict(state_dict))
['nodes.head.module.weight']
```

Parameters

 * `state_dict` (`Mapping[str, Tensor]`): The state dict of the model, keyed by the names of the parameters and the buffers.

Returns

 * `dict[str, Tensor]`: The entries of `state_dict` with a key that does not match the pattern, in the same order.

## Attributes

### CHECKPOINT_FILTERED_STATE_DICT_PATTERN