Learn R Programming
keras3 (version 1.3.0)

loss_circle: Computes Circle Loss between integer labels and L2-normalized embeddings.

Description

It is designed to minimize within-class distances and maximize between-class distances in L2 normalized embedding space.

This is a metric learning loss designed to minimize within-class distance and maximize between-class distance in a flexible manner by dynamically adjusting the penalty strength based on optimization status of each similarity score.

To use Circle Loss effectively, the model should output embeddings without an activation function (such as a Dense layer with activation=NULL) followed by UnitNormalization layer to ensure unit-norm embeddings.

Usage

loss_circle(
  y_true,
  y_pred,
  ref_labels = NULL,
  ref_embeddings = NULL,
  remove_diagonal = TRUE,
  gamma = 80L,
  margin = 0.4,
  ...,
  reduction = "sum_over_batch_size",
  name = "circle",
  dtype = NULL
)

Value

Circle loss value.

Arguments

y_true

Tensor with ground truth labels in integer format.

y_pred

Tensor with predicted L2 normalized embeddings.

ref_labels

Optional integer tensor with labels for reference embeddings. If NULL, defaults to y_true.

ref_embeddings

Optional tensor with L2 normalized reference embeddings. If NULL, defaults to y_pred.

remove_diagonal

Boolean, whether to remove self-similarities from the positive mask. Defaults to TRUE.

gamma

Scaling factor that determines the largest scale of each similarity score. Defaults to 80.

margin

The relaxation factor, below this distance, negatives are up weighted and positives are down weighted. Similarly, above this distance negatives are down weighted and positive are up weighted. Defaults to 0.4.

...

For forward/backward compatability.

reduction

Type of reduction to apply to the loss. In almost all cases this should be "sum_over_batch_size". Supported options are "sum", "sum_over_batch_size", "mean", "mean_with_sample_weight" or NULL. "sum" sums the loss, "sum_over_batch_size" and "mean" sum the loss and divide by the sample size, and "mean_with_sample_weight" sums the loss and divides by the sum of the sample weights. "none" and NULL perform no aggregation. Defaults to "sum_over_batch_size".

name

Optional name for the loss instance.

dtype

The dtype of the loss's computations. Defaults to NULL, which means using config_floatx(). config_floatx() is a "float32" unless set to different value (via config_set_floatx()). If a keras.DTypePolicy is provided, then the compute_dtype will be utilized.

Examples

Usage with the compile() API:

model <- keras_model_sequential(input_shape = c(224, 224, 3)) |>
  layer_conv_2d(16, c(3, 3), activation = 'relu') |>
  layer_flatten() |>
  layer_dense(64, activation = NULL) |>   # No activation
  layer_unit_normalization()  # L2 normalization
model |>
  compile(optimizer = "adam", loss = loss_circle())

See Also

Other losses:
Loss()
loss_binary_crossentropy()
loss_binary_focal_crossentropy()
loss_categorical_crossentropy()
loss_categorical_focal_crossentropy()
loss_categorical_hinge()
loss_cosine_similarity()
loss_ctc()
loss_dice()
loss_hinge()
loss_huber()
loss_kl_divergence()
loss_log_cosh()
loss_mean_absolute_error()
loss_mean_absolute_percentage_error()
loss_mean_squared_error()
loss_mean_squared_logarithmic_error()
loss_poisson()
loss_sparse_categorical_crossentropy()
loss_squared_hinge()
loss_tversky()
metric_binary_crossentropy()
metric_binary_focal_crossentropy()
metric_categorical_crossentropy()
metric_categorical_focal_crossentropy()
metric_categorical_hinge()
metric_hinge()
metric_huber()
metric_kl_divergence()
metric_log_cosh()
metric_mean_absolute_error()
metric_mean_absolute_percentage_error()
metric_mean_squared_error()
metric_mean_squared_logarithmic_error()
metric_poisson()
metric_sparse_categorical_crossentropy()
metric_squared_hinge()