def test_nearest_neighbors(metric):
# create a features vector where 1st and 3rd item are in same direction
# and are near to each other so that both cosine and euclidean dist work
# similarly 2nd and 4th vectors are opposite in direction and far from
# the remaining two so that they are similar to each other based on both
# cosine and euclidean distance
features = np.array(
[[1.0, 1, 1], [-0.1, -0.1, -0.1], [1, 0.9, 0.9], [-0.1, -0.1, -0.2]]
)
ds = stream.DataStream(features, context=["a", "b", "c", "d"])
op = ops.neighbors.NearestNeighborsOperation(n_neighbors=2, metric=metric)
nbors_ds = ds.apply(op)
nbors = list(nbors_ds)
# distance does not matter as long as the items we expect to be same are
# returned as neighbors
assert nbors[0] == [
{"context": "a", "distance": ANY, "item_idx": 0},
{"context": "c", "distance": ANY, "item_idx": 2},
]
assert nbors[1] == [
{"context": "b", "distance": ANY, "item_idx": 1},
{"context": "d", "distance": ANY, "item_idx": 3},
]
assert nbors[2] == [
{"context": "c", "distance": ANY, "item_idx": 2},
{"context": "a", "distance": ANY, "item_idx": 0},
]
assert nbors[3] == [
{"context": "d", "distance": ANY, "item_idx": 3},
{"context": "b", "distance": ANY, "item_idx": 1},
]
def topic_model(
texts: Union[stream.DataStream, Iterable],
num_topics: int = 30,
max_words_per_topic: int = 10,
vectorize_op: ops.base.ScikitBasedOperation = None,
cleaning_ops: Optional[List[base.Operation]] = None,
topic_modeling_op: Optional[ops.topic.TopicModelingOperation] = None,
):
texts = texts if isinstance(
texts, stream.DataStream) else stream.DataStream(texts)
if cleaning_ops is None:
cleaning_ops = []
if vectorize_op and not isinstance(vectorize_op,
ops.base.ScikitBasedOperation):
raise ValueError(
"vectorize_op should be of type ops.base.ScikitBasedOperation"
f" but got {type(vectorize_op)}")
elif vectorize_op is None:
vectorize_op = ops.text.encode.tfidf(max_features=15000,
max_df=0.98,
min_df=2)
if not topic_modeling_op:
topic_modeling_op = ops.topic.lda(n_topics=num_topics, )
topics_ds = texts.apply(*cleaning_ops, vectorize_op, topic_modeling_op)
feature_names = vectorize_op.model.get_feature_names()
return topic_modeling_op.map_topics(
topics_ds,
feature_names=feature_names,
max_words_per_topic=max_words_per_topic)
def run(self, ds: stream.DataStream) -> stream.DataStream:
G = nx.Graph()
for pair in ds:
idx1, idx2 = pair[0], pair[1]
G.add_edge(idx1, idx2)
groups = list(nx.connected_components(G))
return stream.DataStream(groups, applied_ops=ds.applied_ops + [self])
def run(self, ds: stream.DataStream) -> stream.DataStream:
raw_topics_scores_ds = super().run(ds)
topics_with_ctx = self._get_topic_per_item(raw_topics_scores_ds)
topics, ctxs = more_itertools.unzip(topics_with_ctx)
return stream.DataStream(items=topics,
applied_ops=ds.applied_ops + [self],
context=ctxs)
def test_returns_a_stream_with_doc_vectors():
ds = stream.DataStream(["this is", "another", "sentence"])
vector_ds = ds.apply(ops.text.embedding.doc_embedding())
vectors = list(vector_ds)
context = list(vector_ds.context)
assert len(vectors) == ds.total_items
assert len(context) == len(vectors)
assert all(isinstance(v, np.ndarray) for v in vectors)
def test_flatten():
input_ds = stream.DataStream(items=[[1, 2], [3, 4, 5]], context=["a", "b"])
ds = input_ds.apply(ops.stream.flatten(distribute_context=True))
assert list(ds) == [1, 2, 3, 4, 5]
assert list(ds.context) == ["a", "a", "b", "b", "b"]
ds = input_ds.apply(ops.stream.flatten(distribute_context=False))
assert list(ds) == [1, 2, 3, 4, 5]
assert list(ds.context) == ["a_0", "a_1", "b_0", "b_1", "b_2"]
def run(self, ds: stream.DataStream):
if self.mode == "upper":
fn = str.upper
elif self.mode == "capitalize":
fn = str.capitalize
else:
fn = str.lower
items = map(fn, ds)
return stream.DataStream(applied_ops=ds.applied_ops + [self],
items=items,
context=ds.context)
def test_similar_pairs(metric):
features = np.array(
[[1.0, 1, 1], [-0.1, -0.1, -0.1], [1, 0.9, 0.9], [-0.1, -0.1, -0.2]]
)
# ds = stream.DataStream(features, context=[{"A": 1}, {"B": 2}, {"c": 3}, {"d": 4}])
ds = stream.DataStream(features, context=["a", "b", "c", "d"])
op = ops.neighbors.SimilarPairOperation(n_neighbors=2, metric=metric)
pairs_ds = ds.apply(op)
pairs = list(pairs_ds)
assert sorted(pairs) == sorted([("a", "c", ANY), ("b", "d", ANY)])
def test_sentence_extraction():
texts = ["this is a text. with two sentences.", "this is with single sentence"]
out_ds = stream.DataStream(items=texts).apply(ops.text.extract.sentences())
expected_items = [
["this is a text.", "with two sentences."],
["this is with single sentence"],
]
expected_context = [0, 1]
actual_items, actual_context = list(out_ds.items), list(out_ds.context)
assert actual_items == expected_items
assert actual_context == expected_context
def run(self,
ds: stream.DataStream,
fit_params: dict = {}) -> stream.DataStream:
if self.should_train:
if ds.is_countable:
train_ds = ds
pred_ds = ds
else:
train_items, pred_items = itertools.tee(ds, 2)
train_context, pred_context = itertools.tee(ds.context, 2)
train_ds = stream.DataStream(train_items,
context=train_context)
pred_ds = stream.DataStream(pred_items, context=pred_context)
self._fit(train_ds, fit_params)
else:
pred_ds = ds
predictions = self._predict(pred_ds)
return stream.DataStream(items=predictions,
applied_ops=ds.applied_ops + [self],
context=ds.context)
def run(self, ds: stream.DataStream) -> stream.DataStream:
docs_ds = self.get_docs_stream(ds)
docs = zip(docs_ds, docs_ds.context)
# match results is a tuple ((doc, matches), context)
match_results = self.matcher.pipe(docs,
return_matches=True,
as_tuples=True)
new_docs_with_context = more_itertools.map_except(
self._filter_tokens, match_results, EmptyTextError)
new_docs, context = more_itertools.unzip(new_docs_with_context)
return stream.DataStream(new_docs,
applied_ops=ds.applied_ops + [self],
context=context)
def test_topic_modeling(topic_op):
ds = stream.DataStream(["this is about playstation", "government governs"])
topics_ds = ds.apply(ops.text.encode.count(name="vec"), topic_op)
vec_op = topics_ds.applied_ops.find_by_name("vec")
feature_names = vec_op.model.get_feature_names()
topic_word_stream = topic_op.map_topics(
topics_ds=topics_ds,
feature_names=feature_names,
max_words_per_topic=2,
)
assert list(topic_word_stream) == [[ANY, ANY], [ANY, ANY]]
def test_sklearn_classifiers_work_with_multiclass_label(
classifier, feature_vectors, multiclass_labels
):
vec_ds = stream.DataStream(feature_vectors, context=multiclass_labels)
fit_params = {"y": multiclass_labels}
if classifier.supports_batch_training:
fit_params["classes"] = list(set(multiclass_labels))
pred_ds = vec_ds.apply(classifier, op_kwargs={"cls": {"fit_params": fit_params}},)
preds = list(pred_ds)
assert len(preds) == len(feature_vectors)
assert all(isinstance(p, ops.classify.ClassificationResult) for p in preds)
assert all(isinstance(p.label, str) for p in preds)
def test_sklearn_classifiers_work_with_multilabel_label(
classifier, feature_vectors, multilabel_labels
):
vec_ds = stream.DataStream(feature_vectors, context=multilabel_labels)
fit_params = {"y": multilabel_labels}
if classifier.supports_batch_training or not classifier.exclusive_classes:
fit_params["classes"] = list(
set(itertools.chain.from_iterable(multilabel_labels))
)
pred_ds = vec_ds.apply(classifier, op_kwargs={"cls": {"fit_params": fit_params}},)
preds = list(pred_ds)
assert len(preds) == len(feature_vectors)
assert all(isinstance(p, ops.classify.ClassificationResult) for p in preds)
assert all(isinstance(p.label, (list, tuple)) for p in preds)
def map_topics(
self,
topics_ds: stream.DataStream,
feature_names: Iterable[str],
max_words_per_topic: int = 5,
) -> stream.DataStream:
words_of_topics = []
for topic_vec in self.model.components_:
words_of_topic = []
for feature_id in topic_vec.argsort()[-1:-max_words_per_topic -
1:-1]:
words_of_topic.append(feature_names[feature_id])
words_of_topics.append(words_of_topic)
mapped_topics = (words_of_topics[topic_id] for topic_id in topics_ds)
return stream.DataStream(mapped_topics,
applied_ops=topics_ds.applied_ops,
context=topics_ds.context)
def test_calls_appropriate_underlying_methods_for_training_and_prediction(
model_class, should_train):
model = MagicMock(spec=model_class)
ds = stream.DataStream([[0, 1, 2], [3, 4, 5]])
op = cluster.ClusterOperation(model=model)
op.should_train = should_train
assert op.should_train == should_train
test_return_value = [[0, 1], [3, 4]]
if model_class in cluster.ALGORITHMS_NOT_SUPPORTING_NEW_INFERENCE:
model.labels_ = test_return_value
output = op.run(ds)
model.fit.assert_called_once()
assert list(output.items) == test_return_value
elif model_class in cluster.ALGORITHMS_SUPPORTING_NEW_INFERENCE:
model.predict.return_value = test_return_value
output = op.run(ds)
if should_train:
if op.supports_batch_training:
model.partial_fit.assert_called_once()
else:
model.fit.assert_called_once()
model.predict.assert_called_once()
else:
model.fit.assert_not_called()
model.predict.assert_called_once()
assert list(output.items) == test_return_value
else:
raise Exception("model class not expected")
def test_token_extraction():
ds = stream.DataStream(items=["this is a sentence", "woo hoo", ""])
out_ds = ds.apply(ops.text.extract.tokens())
expected = [["this", "is", "a", "sentence"], ["woo", "hoo"], []]
assert list(out_ds.items) == expected
def run(self, ds: stream.DataStream) -> stream.DataStream:
dist_indices_ds = super().run(ds)
pairs = self._get_pairs(dist_indices_ds)
# contexts do not make sense anymore
return stream.DataStream(pairs, applied_ops=ds.applied_ops + [self])
