fl4health.losses.fenda_loss_config module

class ConstrainedFendaLossContainer(perfcl_loss_config, cosine_similarity_loss_config, contrastive_loss_config)

Bases: object

__init__(perfcl_loss_config, cosine_similarity_loss_config, contrastive_loss_config)

Container to gather all of the possible loss functions used in constrained FENDA model optimization

Parameters:

• perfcl_loss_config (PerFclLossContainer | None) – PerFCL loss container. If none, the loss isn not used.

• cosine_similarity_loss_config (CosineSimilarityLossContainer | None) – Cosine similarity loss container. If none the loss is not used.

• contrastive_loss_config (MoonContrastiveLossContainer | None) – Contrastive loss container. If none, the loss is not used.

compute_contrastive_loss(features, positive_pairs, negative_pairs)

Compute the contrastive loss, if it exists, using the configuration

Parameters:

• features (torch.Tensor) – features from the model

• positive_pairs (torch.Tensor) – positive pair features to compare to

• negative_pairs (torch.Tensor) – negative pair features to compare to.

Returns:

loss function

Return type:

torch.Tensor

compute_cosine_similarity_loss(first_features, second_features)

Compute the cosine loss, if it exists, using the configuration

Parameters:

• first_features (torch.Tensor) – first set of features in the cosine comparison

• second_features (torch.Tensor) – second set of features in the cosine comparison

Returns:

cosine similarity loss between the provided features.

Return type:

torch.Tensor

compute_perfcl_loss(local_features, old_local_features, global_features, old_global_features, initial_global_features)

Compute the PerFCL loss, if it exists, using the configuration

Parameters:

• local_features (torch.Tensor) – See PerFCL loss documentation

• old_local_features (torch.Tensor) – See PerFCL loss documentation

• global_features (torch.Tensor) – See PerFCL loss documentation

• old_global_features (torch.Tensor) – See PerFCL loss documentation

• initial_global_features (torch.Tensor) – See PerFCL loss documentation

Returns:

PerFCL loss based on the input values

Return type:

tuple[torch.Tensor, torch.Tensor]

has_contrastive_loss()

Return type:

bool

has_cosine_similarity_loss()

Return type:

bool

has_perfcl_loss()

Return type:

bool

class CosineSimilarityLossContainer(device, cos_sim_loss_weight)

Bases: object

__init__(device, cos_sim_loss_weight)

Container to hold the different pieces associated with cosine similarity.

Parameters:

• device (torch.device) – Device to which the loss will be sent and computed on.

• cos_sim_loss_weight (float) – Weight associated with the cosine loss function in optimization.

class MoonContrastiveLossContainer(device, contrastive_loss_weight, temperature=0.5)

Bases: object

__init__(device, contrastive_loss_weight, temperature=0.5)

_summary_

Parameters:

• device (torch.device) – Device to which the loss will be sent and computed on.

• contrastive_loss_weight (float) – Weight associated with the contrastive loss function in optimization.

• temperature (float, optional) – Temperature parameter on the contrastive loss function. Defaults to 0.5.

class PerFclLossContainer(device, global_feature_contrastive_loss_weight, local_feature_contrastive_loss_weight, global_feature_loss_temperature=0.5, local_feature_loss_temperature=0.5)

Bases: object

__init__(device, global_feature_contrastive_loss_weight, local_feature_contrastive_loss_weight, global_feature_loss_temperature=0.5, local_feature_loss_temperature=0.5)

Container to hold the different pieces associated with PerFCL Loss

Parameters:

• device (torch.device) – Device to which the loss will be sent and computed on.

• global_feature_contrastive_loss_weight (float) – Weight on the global contrastive loss function.

• local_feature_contrastive_loss_weight (float) – Weight on the local model contrastive loss function.

• global_feature_loss_temperature (float, optional) – Temperature parameter on the global contrastive loss function. Defaults to 0.5.

• local_feature_loss_temperature (float, optional) – Temperature parameter on the local contrastive loss function. Defaults to 0.5.