def __init__(self,
embedding_key,
data_keys,
score_transform=None,
top_k=None,
reducer=tf.math.reduce_mean):
"""Initializes the CacheClassificationLoss object.
Args:
embedding_key: The key containing the embedding in the cache.
data_keys: The keys containing the document data in the cache.
score_transform: Scores are transformed by this function before use.
Specifically we have scores(i, j) = score_transform(dot(query_embed_i,
doc_embed_j))
top_k: If set, the top k scoring negative elements will be mined and the
rest of the elements masked before calculating the loss.
reducer: Function that reduces the losses to a single scaler. If None,
then the elementwise losses are returned.
"""
self.embedding_key = embedding_key
self.data_keys = data_keys
self.score_transform = score_transform
self.top_k = top_k
self.reducer = reducer
self._retrieval_fn = retrieval_fns.GumbelMaxRetrievalFn()
def test_gumbel_max_retrieval_fn_with_temperature_and_mock_randomness(self):
with mock.patch.object(retrieval_fns,
'_sample_gumbel') as mock_sample_gumbel:
scores = tf.convert_to_tensor([[1.0, 0.0, 0.0]])
mock_sample_gumbel.return_value = tf.convert_to_tensor([[0.0, 0.9, 0.0]])
gumbel_max_retrieval_fn = retrieval_fns.GumbelMaxRetrievalFn(inv_temp=0.5)
expected = tf.convert_to_tensor([[1]])
actual = gumbel_max_retrieval_fn(scores)
self.assertAllEqual(expected, actual)
def __init__(self,
embedding_key,
data_keys,
score_transform=None,
top_k=None,
reducer=tf.math.reduce_mean):
self.embedding_key = embedding_key
self.data_keys = data_keys
self.score_transform = score_transform
self.top_k = top_k
self.reducer = reducer
self._retrieval_fn = retrieval_fns.GumbelMaxRetrievalFn()
def test_gumbel_max_retrieval_fn_with_mock_randomness(self):
with mock.patch.object(retrieval_fns,
'_sample_gumbel') as mock_sample_gumbel:
scores = tf.convert_to_tensor([[1.0, 0.0, 0.0], [100.0, 0.0, 0.0],
[0.0, 0.0, -1.0]])
mock_sample_gumbel.return_value = tf.convert_to_tensor(
[[0.0, 0.1, -0.1], [1.0, -2.0, 3.3], [0.1, -0.1, 2.0]])
gumbel_max_retrieval_fn = retrieval_fns.GumbelMaxRetrievalFn()
expected = tf.convert_to_tensor([[0], [0], [2]])
actual = gumbel_max_retrieval_fn(scores)
self.assertAllEqual(expected, actual)
expected_shape = tf.convert_to_tensor([3, 3])
mock_sample_gumbel.assert_called_once()
mock_args, mock_kwargs = mock_sample_gumbel.call_args
self.assertAllEqual(mock_args, (expected_shape, ))
self.assertEqual(mock_kwargs, {})
def test_gumbel_max_retrieval_fn_has_correct_output_shape(self):
scores = tf.convert_to_tensor([[1.0, 0.0, 0.0], [100.0, 0.0, 0.0],
[0.0, 0.0, -1.0]])
gumbel_max_retrieval_fn = retrieval_fns.GumbelMaxRetrievalFn()
output = gumbel_max_retrieval_fn(scores)
self.assertAllEqual([3, 1], tf.shape(output))