tensorflow_addons/losses/triplet.py [75:143]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    distance_metric: Union[str, Callable] = "L2",
) -> tf.Tensor:
    r"""Computes the triplet loss with semi-hard negative mining.

    Usage:

    >>> y_true = tf.convert_to_tensor([0, 0])
    >>> y_pred = tf.convert_to_tensor([[0.0, 1.0], [1.0, 0.0]])
    >>> tfa.losses.triplet_semihard_loss(y_true, y_pred, distance_metric="L2")
    <tf.Tensor: shape=(), dtype=float32, numpy=2.4142137>

    >>> # Calling with callable `distance_metric`
    >>> distance_metric = lambda x: tf.linalg.matmul(x, x, transpose_b=True)
    >>> tfa.losses.triplet_semihard_loss(y_true, y_pred, distance_metric=distance_metric)
    <tf.Tensor: shape=(), dtype=float32, numpy=1.0>

    Args:
      y_true: 1-D integer `Tensor` with shape `[batch_size]` of
        multiclass integer labels.
      y_pred: 2-D float `Tensor` of embedding vectors. Embeddings should
        be l2 normalized.
      margin: Float, margin term in the loss definition.
      distance_metric: `str` or a `Callable` that determines distance metric.
        Valid strings are "L2" for l2-norm distance,
        "squared-L2" for squared l2-norm distance,
        and "angular" for cosine similarity.

        A `Callable` should take a batch of embeddings as input and
        return the pairwise distance matrix.

    Returns:
      triplet_loss: float scalar with dtype of `y_pred`.
    """
    labels = tf.convert_to_tensor(y_true, name="labels")
    embeddings = tf.convert_to_tensor(y_pred, name="embeddings")

    convert_to_float32 = (
        embeddings.dtype == tf.dtypes.float16 or embeddings.dtype == tf.dtypes.bfloat16
    )
    precise_embeddings = (
        tf.cast(embeddings, tf.dtypes.float32) if convert_to_float32 else embeddings
    )

    # Reshape label tensor to [batch_size, 1].
    lshape = tf.shape(labels)
    labels = tf.reshape(labels, [lshape[0], 1])

    # Build pairwise squared distance matrix

    if distance_metric == "L2":
        pdist_matrix = metric_learning.pairwise_distance(
            precise_embeddings, squared=False
        )

    elif distance_metric == "squared-L2":
        pdist_matrix = metric_learning.pairwise_distance(
            precise_embeddings, squared=True
        )

    elif distance_metric == "angular":
        pdist_matrix = metric_learning.angular_distance(precise_embeddings)

    else:
        pdist_matrix = distance_metric(precise_embeddings)

    # Build pairwise binary adjacency matrix.
    adjacency = tf.math.equal(labels, tf.transpose(labels))
    # Invert so we can select negatives only.
    adjacency_not = tf.math.logical_not(adjacency)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



tensorflow_addons/losses/triplet.py [211:279]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    distance_metric: Union[str, Callable] = "L2",
) -> tf.Tensor:
    r"""Computes the triplet loss with hard negative and hard positive mining.

    Usage:

    >>> y_true = tf.convert_to_tensor([0, 0])
    >>> y_pred = tf.convert_to_tensor([[0.0, 1.0], [1.0, 0.0]])
    >>> tfa.losses.triplet_hard_loss(y_true, y_pred, distance_metric="L2")
    <tf.Tensor: shape=(), dtype=float32, numpy=1.0>

    >>> # Calling with callable `distance_metric`
    >>> distance_metric = lambda x: tf.linalg.matmul(x, x, transpose_b=True)
    >>> tfa.losses.triplet_hard_loss(y_true, y_pred, distance_metric=distance_metric)
    <tf.Tensor: shape=(), dtype=float32, numpy=0.0>

    Args:
      y_true: 1-D integer `Tensor` with shape `[batch_size]` of
        multiclass integer labels.
      y_pred: 2-D float `Tensor` of embedding vectors. Embeddings should
        be l2 normalized.
      margin: Float, margin term in the loss definition.
      soft: Boolean, if set, use the soft margin version.
      distance_metric: `str` or a `Callable` that determines distance metric.
        Valid strings are "L2" for l2-norm distance,
        "squared-L2" for squared l2-norm distance,
        and "angular" for cosine similarity.

        A `Callable` should take a batch of embeddings as input and
        return the pairwise distance matrix.

    Returns:
      triplet_loss: float scalar with dtype of `y_pred`.
    """
    labels = tf.convert_to_tensor(y_true, name="labels")
    embeddings = tf.convert_to_tensor(y_pred, name="embeddings")

    convert_to_float32 = (
        embeddings.dtype == tf.dtypes.float16 or embeddings.dtype == tf.dtypes.bfloat16
    )
    precise_embeddings = (
        tf.cast(embeddings, tf.dtypes.float32) if convert_to_float32 else embeddings
    )

    # Reshape label tensor to [batch_size, 1].
    lshape = tf.shape(labels)
    labels = tf.reshape(labels, [lshape[0], 1])

    # Build pairwise squared distance matrix.
    if distance_metric == "L2":
        pdist_matrix = metric_learning.pairwise_distance(
            precise_embeddings, squared=False
        )

    elif distance_metric == "squared-L2":
        pdist_matrix = metric_learning.pairwise_distance(
            precise_embeddings, squared=True
        )

    elif distance_metric == "angular":
        pdist_matrix = metric_learning.angular_distance(precise_embeddings)

    else:
        pdist_matrix = distance_metric(precise_embeddings)

    # Build pairwise binary adjacency matrix.
    adjacency = tf.math.equal(labels, tf.transpose(labels))
    # Invert so we can select negatives only.
    adjacency_not = tf.math.logical_not(adjacency)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -