Rename build_umls_faiss_index.py to umls_metadata.json
Browse files- build_umls_faiss_index.py +0 -103
- umls_metadata.json +15 -0
build_umls_faiss_index.py
DELETED
@@ -1,103 +0,0 @@
|
|
1 |
-
"""
|
2 |
-
Script to build a FAISS index from UMLS concept metadata.
|
3 |
-
Produces:
|
4 |
-
- `umls_embeddings.npy`: normalized vectors for each concept
|
5 |
-
- `umls_index.faiss`: FAISS index for fast similarity search
|
6 |
-
- `umls_metadata.json`: mapping from index position to concept metadata
|
7 |
-
|
8 |
-
Usage:
|
9 |
-
python build_umls_faiss_index.py \
|
10 |
-
--input concepts.csv \
|
11 |
-
--output_meta backend/umls_metadata.json \
|
12 |
-
--output_emb backend/umls_embeddings.npy \
|
13 |
-
--output_idx backend/umls_index.faiss
|
14 |
-
"""
|
15 |
-
import argparse
|
16 |
-
import csv
|
17 |
-
import json
|
18 |
-
import numpy as np
|
19 |
-
import faiss
|
20 |
-
from transformers import AutoTokenizer, AutoModel
|
21 |
-
import torch
|
22 |
-
|
23 |
-
def encode_concepts(model, tokenizer, texts, batch_size=32, device='cpu'):
|
24 |
-
embeddings = []
|
25 |
-
model.to(device)
|
26 |
-
for i in range(0, len(texts), batch_size):
|
27 |
-
batch_texts = texts[i:i+batch_size]
|
28 |
-
inputs = tokenizer(
|
29 |
-
batch_texts,
|
30 |
-
padding=True,
|
31 |
-
truncation=True,
|
32 |
-
return_tensors='pt'
|
33 |
-
).to(device)
|
34 |
-
with torch.no_grad():
|
35 |
-
outputs = model(**inputs)
|
36 |
-
cls_emb = outputs.last_hidden_state[:, 0, :].cpu().numpy()
|
37 |
-
# normalize
|
38 |
-
norms = np.linalg.norm(cls_emb, axis=1, keepdims=True)
|
39 |
-
embeddings.append(cls_emb / norms)
|
40 |
-
return np.vstack(embeddings)
|
41 |
-
|
42 |
-
def main():
|
43 |
-
parser = argparse.ArgumentParser(description="Build FAISS index for UMLS concepts.")
|
44 |
-
parser.add_argument('--input', required=True,
|
45 |
-
help='CSV file with columns: cui,name,definition,source')
|
46 |
-
parser.add_argument('--output_meta', required=True,
|
47 |
-
help='JSON metadata output path')
|
48 |
-
parser.add_argument('--output_emb', required=True,
|
49 |
-
help='NumPy embeddings output path')
|
50 |
-
parser.add_argument('--output_idx', required=True,
|
51 |
-
help='FAISS index output path')
|
52 |
-
parser.add_argument('--model',
|
53 |
-
default='microsoft/BiomedNLP-KRISSBERT-PubMed-UMLS-EL',
|
54 |
-
help='Hugging Face model name')
|
55 |
-
args = parser.parse_args()
|
56 |
-
|
57 |
-
# Load model & tokenizer
|
58 |
-
tokenizer = AutoTokenizer.from_pretrained(args.model)
|
59 |
-
model = AutoModel.from_pretrained(args.model)
|
60 |
-
model.eval()
|
61 |
-
|
62 |
-
# Read concepts CSV
|
63 |
-
cuis, names, defs, sources = [], [], [], []
|
64 |
-
with open(args.input, newline='', encoding='utf-8') as f:
|
65 |
-
reader = csv.DictReader(f)
|
66 |
-
for row in reader:
|
67 |
-
cuis.append(row['cui'])
|
68 |
-
text = row['name']
|
69 |
-
if row.get('definition'):
|
70 |
-
text += ' - ' + row['definition']
|
71 |
-
names.append(text)
|
72 |
-
defs.append(row.get('definition', ''))
|
73 |
-
sources.append(row.get('source', 'UMLS'))
|
74 |
-
|
75 |
-
# Encode all concept texts
|
76 |
-
print(f"Encoding {len(names)} concepts...")
|
77 |
-
embeddings = encode_concepts(model, tokenizer, names)
|
78 |
-
|
79 |
-
# Build FAISS index (inner-product search)
|
80 |
-
dim = embeddings.shape[1]
|
81 |
-
index = faiss.IndexFlatIP(dim)
|
82 |
-
index.add(embeddings)
|
83 |
-
|
84 |
-
# Save outputs
|
85 |
-
np.save(args.output_emb, embeddings)
|
86 |
-
faiss.write_index(index, args.output_idx)
|
87 |
-
|
88 |
-
# Build metadata mapping
|
89 |
-
metadata = {}
|
90 |
-
for idx, cui in enumerate(cuis):
|
91 |
-
metadata[str(idx)] = {
|
92 |
-
'cui': cui,
|
93 |
-
'name': names[idx],
|
94 |
-
'definition': defs[idx],
|
95 |
-
'source': sources[idx]
|
96 |
-
}
|
97 |
-
with open(args.output_meta, 'w', encoding='utf-8') as f:
|
98 |
-
json.dump(metadata, f, indent=2)
|
99 |
-
|
100 |
-
print("FAISS index, embeddings, and metadata saved.")
|
101 |
-
|
102 |
-
if __name__ == '__main__':
|
103 |
-
main()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
umls_metadata.json
ADDED
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"0": {
|
3 |
+
"cui": "C0000005",
|
4 |
+
"name": "Metformin",
|
5 |
+
"definition": "A biguanide antihyperglycemic agent.",
|
6 |
+
"source": "UMLS"
|
7 |
+
},
|
8 |
+
"1": {
|
9 |
+
"cui": "C0011849",
|
10 |
+
"name": "Diabetes Mellitus, Type 2",
|
11 |
+
"definition": "A metabolic disorder characterized by high blood sugar.",
|
12 |
+
"source": "UMLS"
|
13 |
+
}
|
14 |
+
// ... add entries for each row in umls_embeddings.npy
|
15 |
+
}
|