Spaces:
Running
on
T4
Running
on
T4
Separate functions
Browse files
app.py
CHANGED
@@ -41,6 +41,34 @@ def get_docs_from_parquet(parquet_urls, column, offset, limit):
|
|
41 |
|
42 |
|
43 |
@spaces.GPU
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
def generate_topics(dataset, config, split, column, nested_column):
|
45 |
logging.info(
|
46 |
f"Generating topics for {dataset} with config {config} {split} {column} {nested_column}"
|
@@ -67,43 +95,45 @@ def generate_topics(dataset, config, split, column, nested_column):
|
|
67 |
while True:
|
68 |
docs = get_docs_from_parquet(parquet_urls, column, offset, chunk_size)
|
69 |
logging.info(f"------------> New chunk data {offset=} {chunk_size=}")
|
70 |
-
embeddings = sentence_model
|
71 |
-
logging.info(f"Embeddings shape: {embeddings.shape}")
|
72 |
offset = offset + chunk_size
|
73 |
if not docs or offset >= limit:
|
74 |
break
|
75 |
|
76 |
-
new_model = BERTopic(
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
)
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
updated_model = BERTopic.merge_models([base_model, new_model])
|
87 |
-
nr_new_topics = len(set(updated_model.topics_)) - len(
|
88 |
-
set(base_model.topics_)
|
89 |
-
)
|
90 |
-
new_topics = list(updated_model.topic_labels_.values())[-nr_new_topics:]
|
91 |
-
logging.info("The following topics are newly found:")
|
92 |
-
logging.info(f"{new_topics}\n")
|
93 |
-
base_model = updated_model
|
94 |
-
else:
|
95 |
-
base_model = new_model
|
96 |
-
logging.info(base_model.get_topic_info())
|
97 |
-
reduced_embeddings = UMAP(
|
98 |
-
n_neighbors=10, n_components=2, min_dist=0.0, metric="cosine"
|
99 |
-
).fit_transform(embeddings)
|
100 |
-
logging.info(f"Reduced embeddings shape: {reduced_embeddings.shape}")
|
101 |
yield (
|
102 |
base_model.get_topic_info(),
|
103 |
new_model.visualize_documents(
|
104 |
docs, embeddings=embeddings
|
105 |
-
),
|
106 |
-
)
|
107 |
logging.info("Finished processing all data")
|
108 |
return base_model.get_topic_info(), base_model.visualize_topics()
|
109 |
|
|
|
41 |
|
42 |
|
43 |
@spaces.GPU
|
44 |
+
def calculate_embeddings(sentence_model, docs):
|
45 |
+
embeddings = sentence_model.encode(docs, show_progress_bar=True, batch_size=100)
|
46 |
+
logging.info(f"Embeddings shape: {embeddings.shape}")
|
47 |
+
return embeddings
|
48 |
+
|
49 |
+
|
50 |
+
@spaces.GPU
|
51 |
+
def fit_model(base_model, sentence_model, representation_model, docs, embeddings):
|
52 |
+
new_model = BERTopic(
|
53 |
+
"english",
|
54 |
+
embedding_model=sentence_model,
|
55 |
+
representation_model=representation_model,
|
56 |
+
min_topic_size=15, # umap_model=umap_model, hdbscan_model=hdbscan_model
|
57 |
+
)
|
58 |
+
logging.info("Fitting new model")
|
59 |
+
new_model.fit(docs, embeddings)
|
60 |
+
logging.info("End fitting new model")
|
61 |
+
if base_model is None:
|
62 |
+
return new_model, new_model
|
63 |
+
|
64 |
+
updated_model = BERTopic.merge_models([base_model, new_model])
|
65 |
+
nr_new_topics = len(set(updated_model.topics_)) - len(set(base_model.topics_))
|
66 |
+
new_topics = list(updated_model.topic_labels_.values())[-nr_new_topics:]
|
67 |
+
logging.info("The following topics are newly found:")
|
68 |
+
logging.info(f"{new_topics}\n")
|
69 |
+
return updated_model, new_model
|
70 |
+
|
71 |
+
|
72 |
def generate_topics(dataset, config, split, column, nested_column):
|
73 |
logging.info(
|
74 |
f"Generating topics for {dataset} with config {config} {split} {column} {nested_column}"
|
|
|
95 |
while True:
|
96 |
docs = get_docs_from_parquet(parquet_urls, column, offset, chunk_size)
|
97 |
logging.info(f"------------> New chunk data {offset=} {chunk_size=}")
|
98 |
+
embeddings = calculate_embeddings(sentence_model, docs)
|
|
|
99 |
offset = offset + chunk_size
|
100 |
if not docs or offset >= limit:
|
101 |
break
|
102 |
|
103 |
+
# new_model = BERTopic(
|
104 |
+
# "english",
|
105 |
+
# embedding_model=sentence_model,
|
106 |
+
# representation_model=representation_model,
|
107 |
+
# min_topic_size=15, # umap_model=umap_model, hdbscan_model=hdbscan_model
|
108 |
+
# )
|
109 |
+
# logging.info("Fitting new model")
|
110 |
+
# new_model.fit(docs, embeddings)
|
111 |
+
# logging.info("End fitting new model")
|
112 |
+
# if base_model is not None:
|
113 |
+
# updated_model = BERTopic.merge_models([base_model, new_model])
|
114 |
+
# nr_new_topics = len(set(updated_model.topics_)) - len(
|
115 |
+
# set(base_model.topics_)
|
116 |
+
# )
|
117 |
+
# new_topics = list(updated_model.topic_labels_.values())[-nr_new_topics:]
|
118 |
+
# logging.info("The following topics are newly found:")
|
119 |
+
# logging.info(f"{new_topics}\n")
|
120 |
+
# base_model = updated_model
|
121 |
+
# else:
|
122 |
+
# base_model = new_model
|
123 |
+
# logging.info(base_model.get_topic_info())
|
124 |
+
base_model, new_model = fit_model(
|
125 |
+
base_model, sentence_model, representation_model, docs, embeddings
|
126 |
)
|
127 |
+
# reduced_embeddings = UMAP(
|
128 |
+
# n_neighbors=10, n_components=2, min_dist=0.0, metric="cosine"
|
129 |
+
# ).fit_transform(embeddings)
|
130 |
+
# logging.info(f"Reduced embeddings shape: {reduced_embeddings.shape}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
131 |
yield (
|
132 |
base_model.get_topic_info(),
|
133 |
new_model.visualize_documents(
|
134 |
docs, embeddings=embeddings
|
135 |
+
), # TODO: Visualize the merged models
|
136 |
+
)
|
137 |
logging.info("Finished processing all data")
|
138 |
return base_model.get_topic_info(), base_model.visualize_topics()
|
139 |
|