def testLoadWordVectorsWithoutFileRaisesError(self):
vocab = {"foo", "bar", "baz", "qux", "."}
with self.assertRaisesRegexp(
ValueError, "Cannot find GloVe embedding file at"):
data.load_word_vectors(self._temp_data_dir, vocab)
os.makedirs(os.path.join(self._temp_data_dir, "glove"))
with self.assertRaisesRegexp(
ValueError, "Cannot find GloVe embedding file at"):
data.load_word_vectors(self._temp_data_dir, vocab)
def testSnliData(self):
snli_1_0_dir = os.path.join(self._temp_data_dir, "snli/snli_1.0")
fake_train_file = os.path.join(snli_1_0_dir, "snli_1.0_train.txt")
os.makedirs(snli_1_0_dir)
self._createFakeSnliData(fake_train_file)
glove_dir = os.path.join(self._temp_data_dir, "glove")
os.makedirs(glove_dir)
glove_file = os.path.join(glove_dir, "glove.42B.300d.txt")
self._createFakeGloveData(glove_file)
vocab = data.load_vocabulary(self._temp_data_dir)
word2index, _ = data.load_word_vectors(self._temp_data_dir, vocab)
train_data = data.SnliData(fake_train_file, word2index)
self.assertEqual(4, train_data.num_batches(1))
self.assertEqual(2, train_data.num_batches(2))
self.assertEqual(2, train_data.num_batches(3))
self.assertEqual(1, train_data.num_batches(4))
generator = train_data.get_generator(2)()
for _ in range(2):
label, prem, prem_trans, hypo, hypo_trans = next(generator)
self.assertEqual(2, len(label))
self.assertEqual((4, 2), prem.shape)
self.assertEqual((5, 2), prem_trans.shape)
self.assertEqual((3, 2), hypo.shape)
self.assertEqual((3, 2), hypo_trans.shape)
def testLoadWordVectors(self):
glove_dir = os.path.join(self._temp_data_dir, "glove")
os.makedirs(glove_dir)
glove_file = os.path.join(glove_dir, "glove.42B.300d.txt")
words = [".", ",", "foo", "bar", "baz"]
with open(glove_file, "wt") as f:
for i, word in enumerate(words):
f.write("%s " % word)
for j in range(data.WORD_VECTOR_LEN):
f.write("%.5f" % (i * 0.1))
if j < data.WORD_VECTOR_LEN - 1:
f.write(" ")
else:
f.write("\n")
vocab = {"foo", "bar", "baz", "qux", "."}
# Notice that "qux" is not present in `words`.
word2index, embed = data.load_word_vectors(self._temp_data_dir, vocab)
self.assertEqual(6, len(word2index))
self.assertEqual(0, word2index["<unk>"])
self.assertEqual(1, word2index["<pad>"])
self.assertEqual(2, word2index["."])
self.assertEqual(3, word2index["foo"])
self.assertEqual(4, word2index["bar"])
self.assertEqual(5, word2index["baz"])
self.assertEqual((6, data.WORD_VECTOR_LEN), embed.shape)
self.assertAllClose([0.0] * data.WORD_VECTOR_LEN, embed[0, :])
self.assertAllClose([0.0] * data.WORD_VECTOR_LEN, embed[1, :])
self.assertAllClose([0.0] * data.WORD_VECTOR_LEN, embed[2, :])
self.assertAllClose([0.2] * data.WORD_VECTOR_LEN, embed[3, :])
self.assertAllClose([0.3] * data.WORD_VECTOR_LEN, embed[4, :])
self.assertAllClose([0.4] * data.WORD_VECTOR_LEN, embed[5, :])
def testInferSpinnThrowsErrorIfOnlyOneSentenceIsSpecified(self):
snli_1_0_dir = os.path.join(self._temp_data_dir, "snli/snli_1.0")
self._create_test_data(snli_1_0_dir)
vocab = data.load_vocabulary(self._temp_data_dir)
word2index, embed = data.load_word_vectors(self._temp_data_dir, vocab)
config = _test_spinn_config(
data.WORD_VECTOR_LEN, 4,
logdir=os.path.join(self._temp_data_dir, "logdir"),
inference_sentences=("( foo ( bar . ) )", None))
with self.assertRaises(ValueError):
spinn.train_or_infer_spinn(embed, word2index, None, None, None, config)
def testInferSpinnWorks(self):
"""Test inference with the spinn model."""
snli_1_0_dir = os.path.join(self._temp_data_dir, "snli/snli_1.0")
self._create_test_data(snli_1_0_dir)
vocab = data.load_vocabulary(self._temp_data_dir)
word2index, embed = data.load_word_vectors(self._temp_data_dir, vocab)
config = _test_spinn_config(
data.WORD_VECTOR_LEN, 4,
logdir=os.path.join(self._temp_data_dir, "logdir"),
inference_sentences=("( foo ( bar . ) )", "( bar ( foo . ) )"))
logits = spinn.train_or_infer_spinn(
embed, word2index, None, None, None, config)
self.assertEqual(tf.float32, logits.dtype)
self.assertEqual((3,), logits.shape)
def testTrainSpinn(self):
"""Test with fake toy SNLI data and GloVe vectors."""
# 1. Create and load a fake SNLI data file and a fake GloVe embedding file.
snli_1_0_dir = os.path.join(self._temp_data_dir, "snli/snli_1.0")
fake_train_file = self._create_test_data(snli_1_0_dir)
vocab = data.load_vocabulary(self._temp_data_dir)
word2index, embed = data.load_word_vectors(self._temp_data_dir, vocab)
train_data = data.SnliData(fake_train_file, word2index)
dev_data = data.SnliData(fake_train_file, word2index)
test_data = data.SnliData(fake_train_file, word2index)
# 2. Create a fake config.
config = _test_spinn_config(
data.WORD_VECTOR_LEN, 4,
logdir=os.path.join(self._temp_data_dir, "logdir"))
# 3. Test training of a SPINN model.
trainer = spinn.train_or_infer_spinn(
embed, word2index, train_data, dev_data, test_data, config)
# 4. Load train loss values from the summary files and verify that they
# decrease with training.
summary_file = glob.glob(os.path.join(config.logdir, "events.out.*"))[0]
events = summary_test_util.events_from_file(summary_file)
train_losses = [event.summary.value[0].simple_value for event in events
if event.summary.value
and event.summary.value[0].tag == "train/loss"]
self.assertEqual(config.epochs, len(train_losses))
# 5. Verify that checkpoints exist and contains all the expected variables.
self.assertTrue(glob.glob(os.path.join(config.logdir, "ckpt*")))
object_graph = trackable_utils.object_metadata(
checkpoint_management.latest_checkpoint(config.logdir))
ckpt_variable_names = set()
for node in object_graph.nodes:
for attribute in node.attributes:
ckpt_variable_names.add(attribute.full_name)
self.assertIn("global_step", ckpt_variable_names)
for v in trainer.variables:
variable_name = v.name[:v.name.index(":")] if ":" in v.name else v.name
self.assertIn(variable_name, ckpt_variable_names)
def testEncodeSingleSentence(self):
snli_1_0_dir = os.path.join(self._temp_data_dir, "snli/snli_1.0")
fake_train_file = os.path.join(snli_1_0_dir, "snli_1.0_train.txt")
os.makedirs(snli_1_0_dir)
self._createFakeSnliData(fake_train_file)
vocab = data.load_vocabulary(self._temp_data_dir)
glove_dir = os.path.join(self._temp_data_dir, "glove")
os.makedirs(glove_dir)
glove_file = os.path.join(glove_dir, "glove.42B.300d.txt")
self._createFakeGloveData(glove_file)
word2index, _ = data.load_word_vectors(self._temp_data_dir, vocab)
sentence_variants = [
"( Foo ( ( bar baz ) . ) )",
" ( Foo ( ( bar baz ) . ) ) ",
"( Foo ( ( bar baz ) . ) )"]
for sentence in sentence_variants:
word_indices, shift_reduce = data.encode_sentence(sentence, word2index)
self.assertEqual(np.int64, word_indices.dtype)
self.assertEqual((5, 1), word_indices.shape)
self.assertAllClose(
np.array([[3, 3, 3, 2, 3, 2, 2]], dtype=np.int64).T, shift_reduce)