States

States are managed with two concepts:

• StateType: Represents a family of states that are compatible with each other. In general, a model can be loaded with any pre-trained state if it matches its state_type attribute.
• StateKey: The identifier of a state instance, it should uniquely identify the result of a training activity for a model.

mozuma.states.StateType dataclass

Definition for a type of state

A state type is used to identify states that can be re-used accross models.

For instance, the weights from ResNet18 with 1000 classes pretrained on ImageNet can be reused to initialised the weights of a ResNet18 for binary classification. In this scenario, the state type of the ResNet trained on ImageNet can be re-used to partially initialize the binary classification ResNet18.

This is also used for models like key-frames extraction. Key-frames extraction does not define a new weights architecture, it is a wrapper around an image encoder. Therefore, any state that can be loaded into the image encoder, can also be loaded into the key-frame extractor. They share the same state type.

As a convention, two state types are compatible when backend and architecture attributes are the same. This is implemented in the is_compatible_with method.

Warning

All string attributes must match the VALID_NAMES pattern.

Attributes:

Name	Type	Description
backend	str	The model backend. For instance pytorch.
architecture	str	Identifier for the architecture (e.g. torchresnet18...).
extra	tuple[str, ...]	Additional information to identify architecture variants (number of output classes...).

is_compatible_with(self, other: StateType) -> bool

Tells whether two architecture are compatible with each other.

Parameters:

Name	Type	Description	Default
other	StateType	The other architecture to compare	required

Returns:

Type	Description
bool	true if backend and architecture attributes match.

mozuma.states.StateKey dataclass

Identifier for a state of a trained model

Warning

All string attributes must match the VALID_NAMES pattern.

Attributes:

Name	Type	Description
state_type	StateType	Identifies the type of state
training_id	str	Identifies the training activity that was used to get to this state.

mozuma.states.VALID_NAMES

The pattern for valid architecture name in StateType.architecture. It must be alphanumeric characters separated with dashes (i.e. [\w\-\d])