def arcface_layer(self, inputs, labels, output_num, weights):
''' ArcFace layer. '''
params = self.netconf['arcface_params']
s = params['scale']
m = params['margin']
limit_to_pi = params['limit_to_pi']
return arcface_loss(
inputs, labels, output_num, weights, s=s, m=m, limit_to_pi=limit_to_pi)
def test_arcface_loss(self):
''' test arcface loss '''
def gen_fake_data(batch_size, embedding_size, num_spks):
''' generate fake embeddings and labels '''
assert batch_size == embedding_size
assert num_spks == embedding_size
embeddings = np.eye(batch_size, dtype='float32')
labels = np.zeros((batch_size,), dtype='int32')
for spk in range(batch_size):
labels[spk] = spk
return embeddings, labels
with self.cached_session():
batch_size = 4
embedding_size = 4
num_spks = 4
embeddings, labels = gen_fake_data(batch_size, embedding_size, num_spks)
weights = embeddings # use whatever data is (somehow) trivial
weights_tensor = tf.constant(weights)
embeddings_tensor = tf.constant(embeddings)
labels_tensor = tf.constant(labels)
output_true = np.asarray(
[[56.165283, 0., 0., 0.], [0., 56.165283, 0., 0.],
[0., 0., 56.165283, 0.], [0., 0., 0., 56.165283]],
dtype='float32')
output = loss_utils.arcface_loss(
embeddings_tensor,
labels_tensor,
num_spks,
weights_tensor,
s=64.0,
m=0.5,
limit_to_pi=True)
self.assertAllClose(output.eval(), output_true)
