newsreclib.models.components package

Subpackages

• newsreclib.models.components.encoders package
  • Subpackages
    • newsreclib.models.components.encoders.news package
    • newsreclib.models.components.encoders.user package
• newsreclib.models.components.layers package
  • Submodules
  • newsreclib.models.components.layers.attention module
  • newsreclib.models.components.layers.click_predictor module
  • newsreclib.models.components.layers.projection module
  • Module contents

Submodules

newsreclib.models.components.losses module

class newsreclib.models.components.losses.SupConLoss(temperature=0.1, **kwargs)

Bases: SupConLoss

compute_loss(embeddings, labels, indices_tuple, ref_emb, ref_labels)

This has to be implemented and is what actually computes the loss.

training: bool

newsreclib.models.components.utils module

newsreclib.models.components.utils.pairwise_cosine_similarity(x: Tensor, y: Tensor, zero_diagonal: bool = False) → Tensor

Implemented from https://github.com/duynguyen-0203/miner/blob/master/src/utils.py

Calculates the pairwise cosine similarity matrix.

Parameters:

• x – (batch_size, M, d)

• y – (batch_size, N, d)

• zero_diagonal – Determines if the diagonal of the distance matrix should be set to zero.

Returns:

A single-value tensor with the pairwise cosine similarity between x and y.

Module contents