def test_npair_loss(self):
embedding1 = tf.reshape(self.get_embedding1(), [20, 3])
embedding2 = tf.reshape(self.get_embedding2(), [20, 3])
labels_onehot = tf.one_hot(tf.reshape(self.get_labels1(), [20]),
depth=3)
valid_labels = tf.greater_equal(tf.reshape(self.get_labels1(), [20]),
0)
embedding1 = tf.boolean_mask(embedding1, valid_labels)
embedding2 = tf.boolean_mask(embedding2, valid_labels)
labels_onehot = tf.boolean_mask(labels_onehot, valid_labels)
loss11 = mll.npair_loss(embedding1, labels_onehot, 'dotproduct',
'softmax')
loss12 = mll.npair_loss(embedding2, labels_onehot, 'dotproduct',
'softmax')
loss21 = mll.npair_loss(embedding1, labels_onehot, 'dotproduct',
'sigmoid')
loss22 = mll.npair_loss(embedding2, labels_onehot, 'dotproduct',
'sigmoid')
loss31 = mll.npair_loss(embedding1, labels_onehot, 'distance',
'softmax')
loss32 = mll.npair_loss(embedding2, labels_onehot, 'distance',
'softmax')
loss41 = mll.npair_loss(embedding1, labels_onehot, 'distance',
'sigmoid')
loss42 = mll.npair_loss(embedding2, labels_onehot, 'distance',
'sigmoid')
self.assertLess(loss11.numpy(), loss12.numpy())
self.assertLess(loss21.numpy(), loss22.numpy())
self.assertLess(loss31.numpy(), loss32.numpy())
self.assertLess(loss41.numpy(), loss42.numpy())
def npair_loss_i():
return metric_learning_losses.npair_loss(
embedding=sampled_embeddings_i,
target=target_i,
similarity_strategy=similarity_strategy,
loss_strategy=loss_strategy)