def test_histogram_loss_incompatible_shapes(self, descs_n, labels_n):
assume(descs_n != labels_n)
descs_pl = tf.placeholder(tf.float32, [descs_n, None])
labels_pl = tf.placeholder(tf.int32, [labels_n])
with pytest.raises(ValueError):
histogram_loss(descs_pl, labels_pl, 128)
def test_histogram_loss_wrong_descriptors_dims(self, descs_shape):
assume(len(descs_shape) != 2)
descs_pl = tf.placeholder(tf.float32, descs_shape)
labels_pl = tf.placeholder(tf.int32, [None])
with pytest.raises(ValueError):
histogram_loss(descs_pl, labels_pl, 128)
def test_eager_histogram_loss_adding_constants_to_labels_invariance(
self, descs, labels, n_bins, offset):
# discard uniform labels
assume(len(np.unique(labels)) != 1)
# l2 normalize
norm = np.sqrt(np.sum(descs**2, axis=1) + 1e-12)
descs /= np.expand_dims(norm, -1)
with eager_mode():
desired = histogram_loss(descs, labels, n_bins)
actual = histogram_loss(descs, labels + offset, n_bins)
np.testing.assert_almost_equal(actual, desired)
def test_histogram_loss_eager_graph_consistency(self, graph, descs,
labels):
# discard uniform labels
assume(len(np.unique(labels)) != 1)
# l2 normalize
norm = np.sqrt(np.sum(descs**2, axis=1) + 1e-12)
descs /= np.expand_dims(norm, -1)
# compute eager output
with eager_mode():
eager_output = histogram_loss(descs, labels, 128)
# compute graph output
descs_pl, labels_pl, loss = graph
with tf.Session() as sess:
graph_output = sess.run(loss,
feed_dict={
descs_pl: descs,
labels_pl: labels
})
np.testing.assert_almost_equal(np.array(eager_output), graph_output)
def graph():
descs_pl = tf.placeholder(tf.float32, [None, None])
labels_pl = tf.placeholder(tf.int32, [None])
loss = histogram_loss(descs_pl, labels_pl, 128)
yield descs_pl, labels_pl, loss
def test_histogram_loss_wrong_labels_dims(self, labels_shape):
descs_pl = tf.placeholder(tf.float32, [None, None])
labels_pl = tf.placeholder(tf.int32, labels_shape)
with pytest.raises(ValueError):
histogram_loss(descs_pl, labels_pl, 128)