chipiron.players.boardevaluators.neural_networks.models package

Submodules

chipiron.players.boardevaluators.neural_networks.models.multi_layer_perceptron module

This module defines the neural network model NetPP2D2_2_PRELU.

The NetPP2D2_2_PRELU class is a subclass of ChiNN and implements the forward pass of the neural network. It consists of two fully connected layers with PReLU activation functions.

chipiron.players.boardevaluators.neural_networks.models.multi_layer_perceptron.__init__()

Initializes the NetPP2D2_2_PRELU model.

forward(x

torch.Tensor) -> torch.Tensor: Performs the forward pass of the neural network.

init_weights(file

str) -> None: Initializes the weights of the model.

chipiron.players.boardevaluators.neural_networks.models.multi_layer_perceptron.print_param() → None

Prints the parameters of the model.

print_input(input

torch.Tensor) -> None: Prints the input tensor.

Helper Functions:

print_piece_param(i: int, vec: torch.Tensor) -> None: Prints the parameters of a specific piece.

class chipiron.players.boardevaluators.neural_networks.models.multi_layer_perceptron.MultiLayerPerceptron(args: MultiLayerPerceptronArgs)

Bases: ChiNN

Neural network model for board evaluation using 2D2_2_PRELU architecture.

fc1

The first fully connected layer.

Type:

nn.Linear

relu_1

The first PReLU activation function.

Type:

nn.PReLU

fc2

The second fully connected layer.

Type:

nn.Linear

tanh

The tanh activation function.

Type:

nn.Tanh

forward(x: Tensor) → Tensor

Forward pass of the neural network.

Parameters:

x (torch.Tensor) – The input tensor.

Returns:

The output tensor.

Return type:

torch.Tensor

init_weights() → None

Initialize the weights of the neural network.

log_readable_model_weights_to_file(file_path: str) → None

Writes the model weights and biases into a YAML file.

Parameters:

file_path (str) – The path where the YAML file will be saved.

print_input(input: Tensor) → None

Print the input tensor.

Parameters:

input (torch.Tensor) – The input tensor.

print_param() → None

Prints the model weights and biases to the console in YAML format.

Parameters:

model (nn.Sequential) – The Sequential model.

class chipiron.players.boardevaluators.neural_networks.models.multi_layer_perceptron.MultiLayerPerceptronArgs(type: Literal[<NNModelType.MultiLayerPerceptron: 'multi_layer_perceptron'>], number_neurons_per_layer: list[int], list_of_activation_functions: list[chipiron.players.boardevaluators.neural_networks.NNModelType.ActivationFunctionType])

Bases: object

filename() → str

list_of_activation_functions: list[chipiron.players.boardevaluators.neural_networks.NNModelType.ActivationFunctionType]

number_neurons_per_layer: list[int]

type: MultiLayerPerceptron: 'multi_layer_perceptron'>]

chipiron.players.boardevaluators.neural_networks.models.multi_layer_perceptron.build_sequential(layers: list[int], activations: list[chipiron.players.boardevaluators.neural_networks.NNModelType.ActivationFunctionType]) → Sequential

chipiron.players.boardevaluators.neural_networks.models.multi_layer_perceptron.extract_sequential_model_data(model: Sequential) → Dict[str, Any]

Extracts the weights and biases from a Sequential model in a layered dictionary format.

Parameters:

model (nn.Sequential) – The Sequential model to extract from.

Returns:

A dictionary with layer names as keys and parameter data as values.

Return type:

Dict[str, Any]

chipiron.players.boardevaluators.neural_networks.models.multi_layer_perceptron.print_piece_param(i: int, vec: Tensor) → None

Prints the piece parameter at index i from the given tensor.

Parameters:

• i (int) – The index of the piece parameter to print.

• vec (torch.Tensor) – The tensor containing the piece parameters.

Returns:

None

chipiron.players.boardevaluators.neural_networks.models.transformer_one module

class chipiron.players.boardevaluators.neural_networks.models.transformer_one.Block(n_embd: int, n_head: int, dropout_ratio: float)

Bases: Module

Transformer block: communication followed by computation

forward(x: Tensor) → Tensor

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class chipiron.players.boardevaluators.neural_networks.models.transformer_one.FeedFoward(n_embd: int, dropout_ratio: float)

Bases: Module

a simple linear layer followed by a non-linearity

forward(x: Tensor) → Tensor

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class chipiron.players.boardevaluators.neural_networks.models.transformer_one.Head(n_embd: int, head_size: int, dropout_ratio: float)

Bases: Module

one head of self-attention

forward(x: Tensor) → Tensor

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class chipiron.players.boardevaluators.neural_networks.models.transformer_one.MultiHeadAttention(num_heads: int, head_size: int, dropout_ratio: float, n_embd: int)

Bases: Module

multiple heads of self-attention in parallel

forward(x: Tensor) → Tensor

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class chipiron.players.boardevaluators.neural_networks.models.transformer_one.TransformerArgs(number_occupancy_types: int = 13, len_square_tensor: int = 832, number_pieces_types: int = 6, type: Literal[<NNModelType.Transformer: 'transformer'>] = <NNModelType.Transformer: 'transformer'>, n_embd: int = 27, n_head: int = 3, n_layer: int = 2, dropout_ratio: float = 0.0)

Bases: object

dropout_ratio: float = 0.0

filename() → str

Generates a filename based on the TransformerArgs instance.

len_square_tensor: int = 832

n_embd: int = 27

n_head: int = 3

n_layer: int = 2

number_occupancy_types: int = 13

number_pieces_types: int = 6

type: Transformer: 'transformer'>] = 'transformer'

class chipiron.players.boardevaluators.neural_networks.models.transformer_one.TransformerOne(args: TransformerArgs)

Bases: ChiNN

compute_representation(node: Any, parent_node: Any, board_modifications: Any) → None

Compute the input representation for the given node.

Parameters:

• node (Any) – Current node.

• parent_node (Any) – Parent node.

• board_modifications (Any) – Board modifications.

forward(indices: list[list[int]]) → Tensor

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

get_nn_input(node: Any) → None

Get the input tensor for the given node.

Parameters:

node (Any) – Current node.

Returns:

None

print_param() → None

Print the parameters of the model.

Module contents