def testDeepMatcDocUsrProjection(self):
"""Tests DeepMatch with doc/usr fields projection."""
hparams = copy.copy(self.hparams)
# ftr_ext = cnn
with tf.Graph().as_default():
query = tf.constant(self.query, dtype=tf.int32)
doc_fields = self._get_constant_doc_fields()
usr_fields = self._get_constant_usr_fields()
wide_ftrs = tf.constant(self.wide_ftrs, dtype=tf.float32)
setattr(hparams, 'use_doc_projection', True)
setattr(hparams, 'use_usr_projection', True)
setattr(hparams, 'num_usr_fields', len(usr_fields))
setattr(hparams, 'explicit_empty', False)
# num_sim_ftrs should be doc projection size* (query + user projection size)
expected_num_sim_ftrs = 1 * (1 + 1)
# Test no query field when ftr_ext = bert
dm = deep_match.DeepMatch(query,
wide_ftrs,
doc_fields,
hparams,
tf.estimator.ModeKeys.EVAL,
usr_fields=usr_fields)
self.assertAllEqual(dm.deep_ftr_model.num_sim_ftrs,
expected_num_sim_ftrs)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
self.assertAllEqual(dm.deep_ftr_model.sim_ftrs.eval().shape,
self.group_size + [expected_num_sim_ftrs])
def testDeepMatchWithUsrField(self):
"""Tests DeepMatch with user fields."""
hparams = copy.copy(self.hparams)
# ftr_ext = bert
with tf.Graph().as_default():
query = tf.constant(self.query, dtype=tf.int32)
wide_ftrs = tf.constant(self.wide_ftrs, dtype=tf.float32)
doc_fields = self._get_constant_doc_fields()
usr_fields = self._get_constant_usr_fields()
setattr(hparams, 'emb_sim_func', ['concat'])
setattr(hparams, 'num_usr_fields', len(usr_fields))
# Test no query field when ftr_ext = bert
dm = deep_match.DeepMatch(query,
wide_ftrs,
doc_fields,
hparams,
tf.estimator.ModeKeys.EVAL,
usr_fields=usr_fields)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
wide_ftrs = wide_ftrs.eval()
# all_ftrs should contain information about wide_ftrs user_fields and doc_fields
self.assertAllEqual(
dm.all_ftrs.eval().shape[-1],
(len(doc_fields) + len(usr_fields) + 1) *
(dm.deep_ftr_model.ftr_size + int(hparams.explicit_empty))
+ wide_ftrs.shape[-1])
def testDeepMatchNoQuery(self):
"""Tests DeepMatch with query as None.
Two conditions must be met when query is None:
1. query_ftrs of model.text_encoding_model must be None
2. all_ftrs should only contain information about doc_ftrs and wide_ftrs
"""
hparams = copy.copy(self.hparams)
# ftr_ext = cnn
with tf.Graph().as_default():
query = None
doc_fields = self._get_constant_doc_fields()
wide_ftrs = tf.constant(self.wide_ftrs, dtype=tf.float32)
setattr(hparams, 'ftr_ext', 'cnn')
# Test no query field when ftr_ext = bert
dm = deep_match.DeepMatch(query, wide_ftrs, doc_fields, hparams,
tf.estimator.ModeKeys.EVAL)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
wide_ftrs = wide_ftrs.eval()
# all_ftrs should only contain wide_ftrs and doc_fields
doc_ftrs_size_after_cnn = len(doc_fields) * len(
hparams.filter_window_sizes) * hparams.num_filters
self.assertAllEqual(
dm.all_ftrs.eval().shape[-1], doc_ftrs_size_after_cnn +
wide_ftrs.shape[-1] + len(doc_fields))
# query_ftrs of model.text_encoding_model must be None
self.assertAllEqual(
dm.deep_ftr_model.text_encoding_model.query_ftrs, None)
def testDeepMatchNaNFtrs(self):
"""Tests DeepMatch outputs"""
query = tf.constant(self.query, dtype=tf.int32)
doc_fields = self._get_constant_doc_fields()
nan_wide_ftrs = np.copy(self.wide_ftrs)
nan_removed_wide_ftrs = np.copy(self.wide_ftrs)
nan_wide_ftrs[0] = np.nan
nan_removed_wide_ftrs[0] = 0
wide_ftrs = tf.constant(nan_wide_ftrs, dtype=tf.float32)
hparams = self.hparams
dm = deep_match.DeepMatch(query,
wide_ftrs,
doc_fields,
hparams,
tf.estimator.ModeKeys.EVAL)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
wide_ftrs_dm, scores = sess.run(
[dm._wide_ftrs,
dm.scores]
)
# check sizes and shapes
self.assertAllClose(wide_ftrs_dm, nan_removed_wide_ftrs)
self.assertAllEqual(scores.shape, [2, 3])
def testDeepMatchClassificationNoQueryNoWide(self):
"""Tests DeepMatch for classification outputs"""
doc_field1 = tf.constant(np.random.rand(2, 1, 4), dtype=tf.int32)
doc_field2 = tf.constant(np.random.rand(2, 1, 4), dtype=tf.int32)
doc_fields = [doc_field1, doc_field2]
hparamscp = copy.copy(self.hparams)
hparamscp.num_classes = 7
dm = deep_match.DeepMatch(None, None, doc_fields, hparamscp,
tf.estimator.ModeKeys.EVAL)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
scores = sess.run(dm.scores)
# Check sizes and shapes
self.assertAllEqual(scores.shape, [2, hparamscp.num_classes])
def testDeepMatch(self):
"""Tests DeepMatch outputs"""
hparams = self.hparams
query = tf.constant(self.query, dtype=tf.int32)
doc_fields = self._get_constant_doc_fields()
wide_ftrs = tf.constant(self.wide_ftrs, dtype=tf.float32)
dm = deep_match.DeepMatch(query, wide_ftrs, doc_fields, hparams,
tf.estimator.ModeKeys.EVAL)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
scores = sess.run(dm.scores)
# Check sizes and shapes
self.assertAllEqual(scores.shape, [2, 3])
def testDeepMatchMultitaskRanking(self):
"""Tests DeepMatch with multitask ranking"""
query = tf.constant(self.query, dtype=tf.int32)
doc_fields = self._get_constant_doc_fields()
wide_ftrs = tf.constant(self.wide_ftrs, dtype=tf.float32)
task_id_field = tf.constant([1, 0])
hparamscp = copy.copy(self.hparams)
hparamscp.task_ids = [0, 1]
hparamscp.task_weights = [0.2, 0.8]
dm = deep_match.DeepMatch(query=query,
wide_ftrs=wide_ftrs,
doc_fields=doc_fields,
hparams=hparamscp,
mode=tf.estimator.ModeKeys.EVAL,
task_id_field=task_id_field)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
scores = sess.run(dm.scores)
# Check sizes and shapes
self.assertAllEqual(scores.shape, [2, 3])