cayleypy.models.ModelConfig

class cayleypy.models.ModelConfig(model_type: str, input_size: int, num_classes_for_one_hot: int, layers_sizes: list[int], weights_kaggle_id: str | None = None, weights_path: str | None = None, n_outputs: int = 1, tokenizer_groups: list[list[int]] | None = None, graph_hash: str | None = None)

Configuration used to describe ML model.

Fields n_outputs, tokenizer_groups and graph_hash describe capabilities added after the first version of this class. Their defaults describe a single-output model without tokenization, which is not tied to a particular graph, so configs written before these fields existed keep working.

Parameters:

• model_type – Type of the model, one of “MLP” (see MlpModel) or “RESMLP” (see ResMlpModel).

• input_size – Number of elements in one state.

• num_classes_for_one_hot – Number of distinct values one element of a state can take.

• layers_sizes – Sizes of hidden layers (for “RESMLP” these are sizes of residual blocks).

• weights_kaggle_id – Id of the Kaggle model with weights (optional).

• weights_path – Path to the file with weights (optional).

• n_outputs – Number of outputs of the model. 1 means the model predicts distance for the state it is applied to. n_generators means the model predicts distance for every child of that state (Q-model).

• tokenizer_groups – Specification of how elements of a state are grouped into tokens, as a list of [group_size, num_groups] pairs. For example, [[3, 20], [2, 30]] means 20 tokens of 3 elements followed by 30 tokens of 2 elements. None means the state is not tokenized.

• graph_hash – Hash of the graph this model was trained for, see cayleypy.models.graph_hash().

__init__(model_type: str, input_size: int, num_classes_for_one_hot: int, layers_sizes: list[int], weights_kaggle_id: str | None = None, weights_path: str | None = None, n_outputs: int = 1, tokenizer_groups: list[list[int]] | None = None, graph_hash: str | None = None) → None

Methods

__init__(model_type, input_size, ...[, ...])
build_model()	Creates model described by this config, with randomly initialized weights.
from_dict(cfg)	Creates config from Python dict.
load([device])	Creates model described by this config and loads weights.
to_dict()	Converts this config to a Python dict containing only primitive values.

Attributes

graph_hash
n_outputs
tokenizer_groups
weights_kaggle_id
weights_path
model_type
input_size
num_classes_for_one_hot
layers_sizes

build_model() → Module

Creates model described by this config, with randomly initialized weights.

static from_dict(cfg: dict[str, Any])

Creates config from Python dict.

load(device='cpu') → Module

Creates model described by this config and loads weights.

Weights are loaded from weights_path. A config with neither weights_path nor weights_kaggle_id describes an untrained model, and the returned model has randomly initialized weights.

Parameters:

device – PyTorch device to load the model to.

Returns:

The model.

to_dict() → dict[str, Any]

Converts this config to a Python dict containing only primitive values.