Callback, extracts embeddings, add them to AnnoyIndex and evaluate them as recall.

Callback, extracts embeddings and add them to TensorBoard Projector.

Categorization loss function. (Softmax)

The original implementation was taken from: https://github.com/geonm/tf_ms_loss Use tf.keras.Input(shape=(1), name="input_labels") for input.

NPairLoss layer using tf.addons. Each pair in the batch must have unique label, that is why the labels are created as tf.range(batch_size

Concat pairs to list and return this list with labels.

MultiLabel/Binary loss function. (Sigmoid)

Concat triplets to list and return this list with labels.