def testWordEmbeddingInitializerFailIfBothTaskContextAndVocabulary(self):
with self.test_session():
with self.assertRaises(Exception):
gen_parser_ops.word_embedding_initializer(
vectors='/dev/null',
task_context='/dev/null',
vocabulary='/dev/null').eval()
def testWordEmbeddingInitializerRepeatability(self):
records_path = os.path.join(FLAGS.test_tmpdir, 'records2')
writer = tf.python_io.TFRecordWriter(records_path)
writer.write(self._token_embedding('.', [1, 2, 3])) # 3 dims
del writer
# As long as there is one non-zero seed, the result should be repeatable.
for seed1, seed2 in [(0, 1), (1, 0), (123, 456)]:
with tf.Graph().as_default(), self.test_session():
embeddings1 = gen_parser_ops.word_embedding_initializer(
vectors=records_path,
task_context=self._task_context,
seed=seed1,
seed2=seed2)
embeddings2 = gen_parser_ops.word_embedding_initializer(
vectors=records_path,
task_context=self._task_context,
seed=seed1,
seed2=seed2)
# The number of terms is based on the word map, which may change if the
# test corpus is updated. Just assert that there are some terms.
self.assertGreater(tf.shape(embeddings1)[0].eval(), 0)
self.assertGreater(tf.shape(embeddings2)[0].eval(), 0)
self.assertEqual(tf.shape(embeddings1)[1].eval(), 3)
self.assertEqual(tf.shape(embeddings2)[1].eval(), 3)
self.assertAllEqual(embeddings1.eval(), embeddings2.eval())
def testWordEmbeddingInitializerRepeatability(self):
records_path = os.path.join(FLAGS.test_tmpdir, 'records2')
writer = tf.python_io.TFRecordWriter(records_path)
writer.write(self._token_embedding('.', [1, 2, 3])) # 3 dims
del writer
# As long as there is one non-zero seed, the result should be repeatable.
for seed1, seed2 in [(0, 1), (1, 0), (123, 456)]:
with tf.Graph().as_default(), self.test_session():
embeddings1 = gen_parser_ops.word_embedding_initializer(
vectors=records_path,
task_context=self._task_context,
seed=seed1,
seed2=seed2)
embeddings2 = gen_parser_ops.word_embedding_initializer(
vectors=records_path,
task_context=self._task_context,
seed=seed1,
seed2=seed2)
# The number of terms is based on the word map, which may change if the
# test corpus is updated. Just assert that there are some terms.
self.assertGreater(tf.shape(embeddings1)[0].eval(), 0)
self.assertGreater(tf.shape(embeddings2)[0].eval(), 0)
self.assertEqual(tf.shape(embeddings1)[1].eval(), 3)
self.assertEqual(tf.shape(embeddings2)[1].eval(), 3)
self.assertAllEqual(embeddings1.eval(), embeddings2.eval())
def testWordEmbeddingInitializer(self):
def _TokenEmbedding(token, embedding):
e = dictionary_pb2.TokenEmbedding()
e.token = token
e.vector.values.extend(embedding)
return e.SerializeToString()
# Provide embeddings for the first three words in the word map.
records_path = os.path.join(FLAGS.test_tmpdir, "sstable-00000-of-00001")
writer = tf.python_io.TFRecordWriter(records_path)
writer.write(_TokenEmbedding(".", [1, 2]))
writer.write(_TokenEmbedding(",", [3, 4]))
writer.write(_TokenEmbedding("the", [5, 6]))
del writer
with self.test_session():
embeddings = gen_parser_ops.word_embedding_initializer(
vectors=records_path, task_context=self._task_context
).eval()
self.assertAllClose(
np.array(
[
[1.0 / (1 + 4) ** 0.5, 2.0 / (1 + 4) ** 0.5],
[3.0 / (9 + 16) ** 0.5, 4.0 / (9 + 16) ** 0.5],
[5.0 / (25 + 36) ** 0.5, 6.0 / (25 + 36) ** 0.5],
]
),
embeddings[:3,],
)
def _Initializer(shape, dtype=tf.float32, partition_info=None):
unused_dtype = dtype
t = gen_parser_ops.word_embedding_initializer(
vectors=embeddings_path,
task_context=task_context,
embedding_init=self._embedding_init)
t.set_shape(shape)
return t
def testWordEmbeddingInitializerVocabularyFileWithDuplicates(self):
records_path = os.path.join(FLAGS.test_tmpdir, 'records4')
writer = tf.python_io.TFRecordWriter(records_path)
writer.write(self._token_embedding('a', [1, 2, 3]))
writer.write(self._token_embedding('b', [2, 3, 4]))
writer.write(self._token_embedding('c', [3, 4, 5]))
writer.write(self._token_embedding('d', [4, 5, 6]))
writer.write(self._token_embedding('e', [5, 6, 7]))
del writer
vocabulary_path = os.path.join(FLAGS.test_tmpdir, 'vocabulary4')
with open(vocabulary_path, 'w') as vocabulary_file:
vocabulary_file.write('a\nc\ne\nx\ny\nx') # 'x' duplicated
with self.test_session():
with self.assertRaises(Exception):
gen_parser_ops.word_embedding_initializer(
vectors=records_path, vocabulary=vocabulary_path).eval()
def _Initializer(shape, dtype=tf.float32, partition_info=None):
"""Variable initializer that loads pretrained embeddings."""
unused_dtype = dtype
seed1, seed2 = tf.get_seed(self._seed)
t = gen_parser_ops.word_embedding_initializer(
vectors=embeddings_path,
task_context=task_context,
embedding_init=self._embedding_init,
seed=seed1,
seed2=seed2)
t.set_shape(shape)
return t
def _Initializer(shape, dtype=tf.float32, partition_info=None):
"""Variable initializer that loads pretrained embeddings."""
unused_dtype = dtype
seed1, seed2 = tf.get_seed(self._seed)
t = gen_parser_ops.word_embedding_initializer(
vectors=embeddings_path,
task_context=task_context,
embedding_init=self._embedding_init,
seed=seed1,
seed2=seed2)
t.set_shape(shape)
return t
def testWordEmbeddingInitializer(self):
# Provide embeddings for the first three words in the word map.
records_path = os.path.join(test_flags.temp_dir(), 'records1')
writer = tf.python_io.TFRecordWriter(records_path)
writer.write(self._token_embedding('.', [1, 2]))
writer.write(self._token_embedding(',', [3, 4]))
writer.write(self._token_embedding('the', [5, 6]))
del writer
with self.test_session():
embeddings = gen_parser_ops.word_embedding_initializer(
vectors=records_path, task_context=self._task_context).eval()
self.assertAllClose(
np.array([[1. / (1 + 4)**.5, 2. / (1 + 4)**.5],
[3. / (9 + 16)**.5, 4. / (9 + 16)**.5],
[5. / (25 + 36)**.5, 6. / (25 + 36)**.5]]),
embeddings[:3, ])
def testWordEmbeddingInitializerPresetRowNumber(self):
records_path = os.path.join(test_flags.temp_dir(), 'records3')
writer = tf.python_io.TFRecordWriter(records_path)
writer.write(self._token_embedding('a', [1, 2, 3]))
writer.write(self._token_embedding('b', [2, 3, 4]))
writer.write(self._token_embedding('c', [3, 4, 5]))
writer.write(self._token_embedding('d', [4, 5, 6]))
writer.write(self._token_embedding('e', [5, 6, 7]))
del writer
vocabulary_path = os.path.join(test_flags.temp_dir(), 'vocabulary3')
with open(vocabulary_path, 'w') as vocabulary_file:
vocabulary_file.write(
'a\nc\ne\nx\n') # 'x' not in pretrained embeddings
# Enumerate a variety of configurations.
for cache_vectors_locally in [False, True]:
for num_special_embeddings in [None, 1, 2,
5]: # None = use default of 3
for override_num_embeddings in [-1, 8, 10]:
with self.test_session():
embeddings = gen_parser_ops.word_embedding_initializer(
vectors=records_path,
vocabulary=vocabulary_path,
override_num_embeddings=override_num_embeddings,
cache_vectors_locally=cache_vectors_locally,
num_special_embeddings=num_special_embeddings)
# Expect 4 embeddings from the vocabulary plus special embeddings.
expected_num_embeddings = 4 + (num_special_embeddings
or 3)
if override_num_embeddings > 0:
expected_num_embeddings = override_num_embeddings
self.assertAllEqual([expected_num_embeddings, 3],
tf.shape(embeddings).eval())
# The first 3 embeddings should be pretrained.
norm_a = (1.0 + 4.0 + 9.0)**0.5
norm_c = (9.0 + 16.0 + 25.0)**0.5
norm_e = (25.0 + 36.0 + 49.0)**0.5
self.assertAllClose(
[[1.0 / norm_a, 2.0 / norm_a, 3.0 / norm_a],
[3.0 / norm_c, 4.0 / norm_c, 5.0 / norm_c],
[5.0 / norm_e, 6.0 / norm_e, 7.0 / norm_e]],
embeddings[:3].eval())
def testWordEmbeddingInitializer(self):
# Provide embeddings for the first three words in the word map.
records_path = os.path.join(FLAGS.test_tmpdir, 'records1')
writer = tf.python_io.TFRecordWriter(records_path)
writer.write(self._token_embedding('.', [1, 2]))
writer.write(self._token_embedding(',', [3, 4]))
writer.write(self._token_embedding('the', [5, 6]))
del writer
with self.test_session():
embeddings = gen_parser_ops.word_embedding_initializer(
vectors=records_path,
task_context=self._task_context).eval()
self.assertAllClose(
np.array([[1. / (1 + 4) ** .5, 2. / (1 + 4) ** .5],
[3. / (9 + 16) ** .5, 4. / (9 + 16) ** .5],
[5. / (25 + 36) ** .5, 6. / (25 + 36) ** .5]]),
embeddings[:3,])
def testWordEmbeddingInitializerPresetRowNumber(self):
records_path = os.path.join(test_flags.temp_dir(), 'records3')
writer = tf.python_io.TFRecordWriter(records_path)
writer.write(self._token_embedding('a', [1, 2, 3]))
writer.write(self._token_embedding('b', [2, 3, 4]))
writer.write(self._token_embedding('c', [3, 4, 5]))
writer.write(self._token_embedding('d', [4, 5, 6]))
writer.write(self._token_embedding('e', [5, 6, 7]))
del writer
vocabulary_path = os.path.join(test_flags.temp_dir(), 'vocabulary3')
with open(vocabulary_path, 'w') as vocabulary_file:
vocabulary_file.write('a\nc\ne\nx\n') # 'x' not in pretrained embeddings
# Enumerate a variety of configurations.
for cache_vectors_locally in [False, True]:
for num_special_embeddings in [None, 1, 2, 5]: # None = use default of 3
for override_num_embeddings in [-1, 8, 10]:
with self.test_session():
embeddings = gen_parser_ops.word_embedding_initializer(
vectors=records_path,
vocabulary=vocabulary_path,
override_num_embeddings=override_num_embeddings,
cache_vectors_locally=cache_vectors_locally,
num_special_embeddings=num_special_embeddings)
# Expect 4 embeddings from the vocabulary plus special embeddings.
expected_num_embeddings = 4 + (num_special_embeddings or 3)
if override_num_embeddings > 0:
expected_num_embeddings = override_num_embeddings
self.assertAllEqual([expected_num_embeddings, 3],
tf.shape(embeddings).eval())
# The first 3 embeddings should be pretrained.
norm_a = (1.0 + 4.0 + 9.0)**0.5
norm_c = (9.0 + 16.0 + 25.0)**0.5
norm_e = (25.0 + 36.0 + 49.0)**0.5
self.assertAllClose([[1.0 / norm_a, 2.0 / norm_a, 3.0 / norm_a], [
3.0 / norm_c, 4.0 / norm_c, 5.0 / norm_c
], [5.0 / norm_e, 6.0 / norm_e, 7.0 / norm_e]],
embeddings[:3].eval())
def testWordEmbeddingInitializer(self):
def _TokenEmbedding(token, embedding):
e = dictionary_pb2.TokenEmbedding()
e.token = token
e.vector.values.extend(embedding)
return e.SerializeToString()
# Provide embeddings for the first three words in the word map.
records_path = os.path.join(FLAGS.test_tmpdir,
'sstable-00000-of-00001')
writer = tf.python_io.TFRecordWriter(records_path)
writer.write(_TokenEmbedding('.', [1, 2]))
writer.write(_TokenEmbedding(',', [3, 4]))
writer.write(_TokenEmbedding('the', [5, 6]))
del writer
with self.test_session():
embeddings = gen_parser_ops.word_embedding_initializer(
vectors=records_path, task_context=self._task_context).eval()
self.assertAllClose(
np.array([[1. / (1 + 4)**.5, 2. / (1 + 4)**.5],
[3. / (9 + 16)**.5, 4. / (9 + 16)**.5],
[5. / (25 + 36)**.5, 6. / (25 + 36)**.5]]),
embeddings[:3, ])