def test_unweighted():
nl_obj = npairs.NpairsLoss()
# batch size = 4, hidden size = 2
y_true = tf.constant([0, 1, 2, 3], dtype=tf.int64)
# features of anchors
f = tf.constant(
[[1.0, 1.0], [1.0, -1.0], [-1.0, 1.0], [-1.0, -1.0]], dtype=tf.float32
)
# features of positive samples
fp = tf.constant(
[[1.0, 1.0], [1.0, -1.0], [-1.0, 1.0], [-1.0, -1.0]], dtype=tf.float32
)
# similarity matrix
y_pred = tf.matmul(f, fp, transpose_a=False, transpose_b=True)
loss = nl_obj(y_true, y_pred)
# Loss = 1/4 * \sum_i log(1 + \sum_{j != i} exp(f_i*fp_j^T-f_i*f_i^T))
# Compute loss for i = 0, 1, 2, 3 without multiplier 1/4
# i = 0 => log(1 + sum([exp(-2), exp(-2), exp(-4)])) = 0.253846
# i = 1 => log(1 + sum([exp(-2), exp(-4), exp(-2)])) = 0.253846
# i = 2 => log(1 + sum([exp(-2), exp(-4), exp(-2)])) = 0.253846
# i = 3 => log(1 + sum([exp(-4), exp(-2), exp(-2)])) = 0.253846
# Loss = (0.253856 + 0.253856 + 0.253856 + 0.253856) / 4 = 0.253856
np.testing.assert_allclose(loss, 0.253856, rtol=1e-06, atol=1e-06)
def test_config(self):
nl_obj = npairs.NpairsLoss(name="nl")
self.assertEqual(nl_obj.name, "nl")
self.assertEqual(nl_obj.reduction, tf.keras.losses.Reduction.NONE)
def test_config():
nl_obj = npairs.NpairsLoss(name="nl")
assert nl_obj.name == "nl"
assert nl_obj.reduction == tf.keras.losses.Reduction.NONE