finetuner.tuner.paddle.losses module#

finetuner.tuner.paddle.losses.get_distance(embeddings, distance)[source]#

Get a matrix of pairwise distances between the embedings

Return type

Tensor

class finetuner.tuner.paddle.losses.PaddleTupleLoss(*args, **kwargs)[source]#

Bases: paddle.nn.Layer, finetuner.tuner.base.BaseLoss[paddle.Tensor]

Base class for all paddle losses.

forward(embeddings, labels)[source]#
Return type

Tensor

class finetuner.tuner.paddle.losses.SiameseLoss(distance='cosine', margin=1.0, miner=None)[source]#

Bases: finetuner.tuner.paddle.losses.PaddleTupleLoss

Computes the loss for a siamese network.

The loss for a pair of objects equals

is_sim * dist + (1 - is_sim) * max(0, margin - dist)

where is_sim equals 1 if the two objects are similar, and 0 if they are not similar. The dist refers to the distance between the two objects, and margin is a number to help bound the loss for dissimilar objects.

The final loss is the average over losses for all pairs given by the indices.

Initialize the loss instance

Parameters

  • distance (str) – The type of distance to use, avalilable options are "cosine", "euclidean" and "sqeuclidean"

  • margin (float) – The margin to use in loss calculation

compute(embeddings, indices)[source]#

Compute the loss

Parameters

  • embeddings (Tensor) – An [N, d] tensor of embeddings

  • indices (Tuple[Tensor, Tensor, Tensor]) – A list of tuple indices and target, where each element in the list contains three elements: the indices of the two objects in the pair, and their similarity (which equals 1 if they are similar, and 0 if they are dissimilar)

Return type

Tensor

get_default_miner(is_session_dataset)[source]#

Get the default miner for this loss, given the dataset type

Return type

Union[SiameseMiner, SiameseSessionMiner]

class finetuner.tuner.paddle.losses.TripletLoss(distance='cosine', margin=1.0, miner=None)[source]#

Bases: finetuner.tuner.paddle.losses.PaddleTupleLoss

Compute the loss for a triplet network.

The loss for a single triplet equals:

max(dist_pos - dist_neg + margin, 0)

where dist_pos is the distance between the anchor embedding and positive embedding, dist_neg is the distance between the anchor and negative embedding, and margin represents a wedge between the desired anchor-negative and anchor-positive distances.

The final loss is the average over losses for all triplets given by the indices.

Initialize the loss instance

Parameters

  • distance (str) – The type of distance to use, avalilable options are "cosine", "euclidean" and "sqeuclidean"

  • margin (float) – The margin to use in loss calculation

compute(embeddings, indices)[source]#

Compute the loss

Parameters

  • embeddings (Tensor) – An [N, d] tensor of embeddings

  • indices (Tuple[Tensor, Tensor, Tensor]) – A list of tuple indices, where each element in the list contains three elements: the index of anchor, positive match and negative match in the embeddings tensor

Return type

Tensor

get_default_miner(is_session_dataset)[source]#

Get the default miner for this loss, given the dataset type

Return type

Union[SiameseMiner, SiameseSessionMiner]

class finetuner.tuner.paddle.losses.NTXentLoss(temperature=0.1)[source]#

Bases: paddle.nn.Layer, finetuner.tuner.base.BaseLoss[paddle.Tensor]

Compute the NTXent (Normalized Temeprature Cross-Entropy) loss.

This loss function is a temperature-adjusted cross-entropy loss, as defined in the SimCLR paper <https://arxiv.org/abs/2002.05709>. It operates on batches where there are two views of each instance

Initialize the loss instance.

Parameters

temerature – The temperature parameter

forward(embeddings, labels)[source]#

Compute the loss.

Parameters

  • embeddings (Tensor) – An [N, d] tensor of embeddings.

  • labels (Tensor) – An [N,] tensor of item labels. It is expected that each label appears two times.

Return type

Tensor