NN
This module contains pytorch compatible layers of the SPDNet architecture.
- class anotherspdnet.nn.BiMap(*args: Any, **kwargs: Any)[source]
- __init__(n_in: int, n_out: int, n_batches: tuple | None = None, manifold: str = 'stiefel', seed: int | None = None, dtype: torch.dtype = torch.float64, device: torch.device = torch.device, mm_mode: str = 'einsum', use_autograd: bool = False) None[source]
- BiMap layer in a SPDnet layer according to the paper:
A Riemannian Network for SPD Matrix Learning, Huang et al AAAI Conference on Artificial Intelligence, 2017
- Parameters:
n_in (int) – Number of input features.
n_out (int) – Number of output features.
n_batches (tuple) – Number of Batches of SPD matrices. It must be a tuple containing at least one batch dimension. Default is None.
manifold (str, optional) – Manifold on which the layer is initialized. Default is ‘stiefel’. choice between ‘stiefel’ and ‘sphere’.
seed (int, optional) – Seed for the initialization of the weight matrix. Default is None.
dtype (torch.dtype, optional) – Data type of the layer. Default is torch.float64.
device (torch.device, optional) – Device on which the layer is initialized. Default is ‘cpu’.
mm_mode (str, optional) – Mode for the matrix multiplication. Default is ‘einsum’. Choice between ‘einsum’ and ‘bmm’.
use_autograd (bool, optional) – Use torch autograd for the computation of the gradient rather than the analytical formula. Default is False.
- forward(X: torch.Tensor) torch.Tensor[source]
Forward pass of the BiMap layer
- Parameters:
X (torch.Tensor of shape self.n_batches + (n_matrices, n_in, n_in)) – Batches of input SPD matrices.
- Returns:
Y – The output matrices is close to SPD. They need regularization with the ReEig layer especially if n_out > n_in.
- Return type:
torch.Tensor of shape self.n_batches + (n_matrices, n_out, n_out)
- class anotherspdnet.nn.ReEigBias(*args: Any, **kwargs: Any)[source]
- __init__(dim: int, eps: float = 0.0001, use_autograd: bool = False, dtype: torch.dtype = torch.float64, device: torch.device = torch.device, seed: int | None = None) None[source]
ReEig layer with a bias term.
- Parameters:
dim (int) – Dimension of the SPD matrices.
eps (float, optional) – Value of rectification of the eigenvalues. Default is 1e-4.
use_autograd (bool, optional) – Use torch autograd for the computation of the gradient rather than the analytical formula. Default is True. FOR NOW: without autograd, the layer is not implemented.
dtype (torch.dtype, optional) – Data type of the layer. Default is torch.float64.
device (torch.device, optional) – Device on which the layer is initialized. Default is ‘cpu’.
seed (int, optional) – Seed for the initialization of the bias term. Default is None.
- forward(X: torch.Tensor) torch.Tensor[source]
Forward pass of the ReEig layer with bias
- Parameters:
X (torch.Tensor of shape (..., n_features, n_features)) – Batches of input almost-SPD matrices.
- Returns:
Y – The regularized SPD matrices.
- Return type:
torch.Tensor of shape (…, n_features, n_features)
- class anotherspdnet.nn.ReEig(*args: Any, **kwargs: Any)[source]
- __init__(eps: float = 0.0001, use_autograd: bool = False, mm_mode: str = 'einsum', eig_function: str = 'eigh', formula: str = 'brooks', dim: int | None = None) None[source]
- ReEig layer in a SPDnet layer according to the paper:
A Riemannian Network for SPD Matrix Learning, Huang et al AAAI Conference on Artificial Intelligence, 2017
- Parameters:
eps (float, optional) – Value of rectification of the eigenvalues. Default is 1e-4.
use_autograd (bool, optional) – Use torch autograd for the computation of the gradient rather than the analytical formula. Default is False.
mm_mode (str, optional) – Mode for the matrix multiplication. Default is ‘einsum’. Choice between ‘einsum’ and ‘bmm’.
eig_function (str, optional) – Function used for the computation of the eigendecomposition. Default is ‘eigh’. Choice between ‘eigh’ and ‘eig’.
formula (str, optional) –
- Formula used for the computation of the gradient. Default is
’brooks’. Choice between ‘brooks’ and ‘ionescu”.
dim (int, optional) – Dimension of the SPD matrices. Default is None. Used for logging purposes.
- class anotherspdnet.nn.LogEig(*args: Any, **kwargs: Any)[source]
- __init__(use_autograd: bool = False, mm_mode: str = 'einsum', eig_function: str = 'eigh', formula: str = 'brooks') None[source]
- LogEig layer in a SPDnet layer according to the paper:
A Riemannian Network for SPD Matrix Learning, Huang et al AAAI Conference on Artificial Intelligence, 2017
- Parameters:
use_autograd (bool, optional) – Use torch autograd for the computation of the gradient rather than the analytical formula. Default is False.
mm_mode (str, optional) – Mode for the matrix multiplication. Default is ‘einsum’. Choice between ‘einsum’ and ‘bmm’.
eig_function (str, optional) – Function used for the computation of the eigendecomposition. Default is ‘eigh’. Choice between ‘eigh’ and ‘eig’.
formula (str, optional) – Formula used for the computation of the gradient. Default is ‘brooks’. Choice between ‘brooks’ and ‘ionescu”.
- class anotherspdnet.nn.Vectorization(*args: Any, **kwargs: Any)[source]
-
- forward(X: torch.Tensor) torch.Tensor[source]
Forward pass of the Vectorization layer
- Parameters:
X (torch.Tensor of shape (..., n, k)) – Batch of matrices.
- Returns:
X_vec – Batch of vectorized matrices.
- Return type:
torch.Tensor of shape (…, n*k)
- inverse_transform(X: torch.Tensor, n_rows: int) torch.Tensor[source]
Inverse transform of the Vectorization layer
- Parameters:
X (torch.Tensor of shape (..., n_rows*k)) – Batch of vectorized matrices.
n_rows (int) – Number of rows of the original matrices.
- Returns:
X_vec – Batch of matrices.
- Return type:
torch.Tensor of shape (…, n_rows, k)
- class anotherspdnet.nn.Vech(*args: Any, **kwargs: Any)[source]