Spaces:
Running
on
T4
Running
on
T4
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoTokenizer, AutoModel
|
2 |
+
import torch
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import hnswlib
|
5 |
+
import gradio as gr
|
6 |
+
import numpy as np
|
7 |
+
|
8 |
+
seperator = "-HFSEP-"
|
9 |
+
base_name="intfloat/e5-large-v2"
|
10 |
+
device="cuda"
|
11 |
+
max_length=512
|
12 |
+
tokenizer = AutoTokenizer.from_pretrained(base_name)
|
13 |
+
model = AutoModel.from_pretrained(base_name).to(device)
|
14 |
+
|
15 |
+
def get_embeddings(input_texts):
|
16 |
+
batch_dict = tokenizer(
|
17 |
+
input_texts,
|
18 |
+
max_length=max_length,
|
19 |
+
padding=True,
|
20 |
+
truncation=True,
|
21 |
+
return_tensors='pt'
|
22 |
+
).to(device)
|
23 |
+
|
24 |
+
with torch.no_grad():
|
25 |
+
outputs = model(**batch_dict)
|
26 |
+
embeddings = _average_pool(
|
27 |
+
outputs.last_hidden_state, batch_dict['attention_mask']
|
28 |
+
)
|
29 |
+
embeddings = F.normalize(embeddings, p=2, dim=1)
|
30 |
+
embeddings_np = embeddings.cpu().numpy()
|
31 |
+
|
32 |
+
if device == "cuda":
|
33 |
+
del embeddings
|
34 |
+
torch.cuda.empty_cache()
|
35 |
+
|
36 |
+
return embeddings_np
|
37 |
+
|
38 |
+
def _average_pool(
|
39 |
+
last_hidden_states,
|
40 |
+
attention_mask
|
41 |
+
):
|
42 |
+
last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
|
43 |
+
return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
|
44 |
+
|
45 |
+
def create_hnsw_index(embeddings_np, space='ip', ef_construction=100, M=16):
|
46 |
+
index = hnswlib.Index(space=space, dim=len(embeddings_np[0]))
|
47 |
+
index.init_index(max_elements=len(embeddings_np), ef_construction=ef_construction, M=M)
|
48 |
+
ids = np.arange(embeddings_np.shape[0])
|
49 |
+
index.add_items(embeddings_np, ids)
|
50 |
+
return index
|
51 |
+
|
52 |
+
def gradio_function(query, paragraph_chunks, top_k):
|
53 |
+
paragraph_chunks = paragraph_chunks.split(seperator) # Split the comma-separated values into a list
|
54 |
+
paragraph_chunks = [item.strip() for item in paragraph_chunks] # Trim whitespace from each item
|
55 |
+
|
56 |
+
print("creating embeddings")
|
57 |
+
embeddings_np = get_embeddings([query]+paragraph_chunks)
|
58 |
+
query_embedding, chunks_embeddings = embeddings_np[0], embeddings_np[1:]
|
59 |
+
|
60 |
+
print("creating index")
|
61 |
+
search_index = create_hnsw_index(chunks_embeddings)
|
62 |
+
print("searching index")
|
63 |
+
labels, _ = search_index.knn_query(query_embedding, k=min(int(top_k), len(chunks_embeddings)))
|
64 |
+
return f"The closes labels are: {labels}"
|
65 |
+
|
66 |
+
interface = gr.Interface(
|
67 |
+
fn=gradio_function,
|
68 |
+
inputs=[
|
69 |
+
gr.Textbox(placeholder="Enter a user query..."),
|
70 |
+
gr.Textbox(placeholder="Enter comma-separated strings..."),
|
71 |
+
gr.Number()
|
72 |
+
],
|
73 |
+
outputs="text"
|
74 |
+
)
|
75 |
+
|
76 |
+
interface.launch()
|