WeightedBPCallback#

class sionna.phy.fec.ldpc.WeightedBPCallback(num_edges: int, pcm: numpy.ndarray | scipy.sparse._matrix.spmatrix | None = None, precision: str | None = None, device: str | None = None, **kwargs)[source]#

Bases: sionna.phy.object.Object

Callback for the LDPCBPDecoder to enable weighted BP [Nachmani].

The BP decoder is fully differentiable and can be made trainable by following the concept of weighted BP [Nachmani] leading to

\[y_{j \to i} = 2 \operatorname{tanh}^{-1} \left( \prod_{i' \in \mathcal{N}(j) \setminus i} \operatorname{tanh} \left( \frac{\textcolor{red}{w_{i' \to j}} \cdot x_{i' \to j}}{2} \right) \right)\]

where \(w_{i \to j}\) denotes the trainable weight of message \(x_{i \to j}\). Please note that the training of some check node types may be not supported.

Can be registered as c2v_callbacks and v2c_callbacks in the LDPCBPDecoder and the LDPC5GDecoder.

Parameters:

  • num_edges (int) – Number of edges in the decoding graph.

  • pcm (numpy.ndarray | scipy.sparse._matrix.spmatrix | None) – Optional parity-check matrix. If provided, enables weighted BP in padded message format used by the decoder.

  • precision (str | None) – Precision used for internal calculations and outputs. If set to None, precision is used.

  • device (str | None) – Device for computation (e.g., ‘cpu’, ‘cuda:0’).

Inputs:

msg – [batch_size, num_vns, max_degree], torch.float. v2c messages.

Outputs:

msg – torch.float. Same as msg.

Examples

from sionna.phy.fec.ldpc import LDPCBPDecoder
from sionna.phy.fec.ldpc.utils import WeightedBPCallback
import numpy as np

# Create a simple parity-check matrix
pcm = np.array([[1, 1, 0, 1], [0, 1, 1, 1]])

# Create callback with trainable weights
weighted_cb = WeightedBPCallback(num_edges=np.sum(pcm), pcm=pcm)

# Create decoder with callback
decoder = LDPCBPDecoder(pcm, v2c_callbacks=[weighted_cb])

# Access trainable weights
print(weighted_cb.weights)

Attributes

property weights: torch.Tensor#

Trainable edge weights

Methods

show_weights(size: float = 7) → None[source]#

Show histogram of trainable weights.

Parameters:

size (float) – Figure size of the matplotlib figure.