Update app.py
Browse files
app.py
CHANGED
@@ -1,18 +1,126 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
with gr.Blocks() as demo:
|
4 |
gr.Markdown("# E5 Large V2 Demo")
|
5 |
|
6 |
-
q_txt = gr.Textbox(placeholder="Enter your query")
|
7 |
|
8 |
-
p_txt1 = gr.Textbox(placeholder="Enter passage 1")
|
9 |
-
p_txt2 = gr.Textbox(placeholder="Enter passage 2")
|
10 |
-
p_txt3 = gr.Textbox(placeholder="Enter passage 3")
|
11 |
-
p_txt4 = gr.Textbox(placeholder="Enter passage 4")
|
12 |
-
p_txt5 = gr.Textbox(placeholder="Enter passage 5")
|
13 |
-
p_txt6 = gr.Textbox(placeholder="Enter passage 6")
|
14 |
|
15 |
submit = gr.Button("Submit")
|
|
|
|
|
|
|
|
|
|
|
16 |
|
17 |
o_txt = gr.Textbox(placeholder="Output", lines=10, interactive=False)
|
18 |
|
|
|
1 |
+
import json
|
2 |
+
import torch
|
3 |
+
import torch.nn.functional as F
|
4 |
+
|
5 |
+
from torch import Tensor
|
6 |
+
from transformers import AutoTokenizer, AutoModel
|
7 |
+
|
8 |
import gradio as gr
|
9 |
|
10 |
+
# instantiate tokenizer and model
|
11 |
+
def get_model(base_name='intfloat/e5-large-v2'):
|
12 |
+
tokenizer = AutoTokenizer.from_pretrained(base_name)
|
13 |
+
model = AutoModel.from_pretrained(base_name)
|
14 |
+
|
15 |
+
return tokenizer, model
|
16 |
+
|
17 |
+
# get normalized scores on input_texts, the final scores are
|
18 |
+
# reported without queries, and the number of queries should
|
19 |
+
# be denoted as in how_many_q
|
20 |
+
def get_scores(model, tokenizer, input_texts, max_length=512, how_many_q=1):
|
21 |
+
# Tokenize the input texts
|
22 |
+
batch_dict = tokenizer(
|
23 |
+
input_texts,
|
24 |
+
max_length=max_length,
|
25 |
+
padding=True,
|
26 |
+
truncation=True,
|
27 |
+
return_tensors='pt'
|
28 |
+
)
|
29 |
+
|
30 |
+
outputs = model(**batch_dict)
|
31 |
+
embeddings = average_pool(
|
32 |
+
outputs.last_hidden_state, batch_dict['attention_mask']
|
33 |
+
)
|
34 |
+
|
35 |
+
# (Optionally) normalize embeddings
|
36 |
+
embeddings = F.normalize(embeddings, p=2, dim=1)
|
37 |
+
scores = (embeddings[:how_many_q] @ embeddings[how_many_q:].T) * 100
|
38 |
+
|
39 |
+
return scores
|
40 |
+
|
41 |
+
# get top n results out of the scores. This
|
42 |
+
# function only returns the scores and indices
|
43 |
+
def get_top(scores, top_k=None):
|
44 |
+
result = torch.sort(scores, descending=True, dim=1)
|
45 |
+
top_indices = result.indices
|
46 |
+
top_values = result.values
|
47 |
+
|
48 |
+
if top_k:
|
49 |
+
top_indices = top_indices[:, :top_k]
|
50 |
+
top_values = top_values[:, :top_k]
|
51 |
+
|
52 |
+
return top_indices, top_values
|
53 |
+
|
54 |
+
# get top n results out of the scores. This function
|
55 |
+
# returns scores and indices along with the associated text
|
56 |
+
def get_human_readable_top(scores, input_texts, top_k=None):
|
57 |
+
input_texts = list(filter(lambda text: "query:" not in text, input_texts))
|
58 |
+
top_indices, top_values = get_top(scores, top_k)
|
59 |
+
|
60 |
+
result = {}
|
61 |
+
for input_idx, (indices, values) in enumerate(zip(top_indices, top_values)):
|
62 |
+
q = input_texts[input_idx]
|
63 |
+
a = []
|
64 |
+
|
65 |
+
for idx, val in zip(indices.tolist(), values.tolist()):
|
66 |
+
a.append({
|
67 |
+
"idx": idx,
|
68 |
+
"val": round(val, 3),
|
69 |
+
"text": input_texts[idx]
|
70 |
+
})
|
71 |
+
|
72 |
+
result[q] = a
|
73 |
+
|
74 |
+
return result
|
75 |
+
|
76 |
+
def average_pool(last_hidden_states: Tensor,
|
77 |
+
attention_mask: Tensor) -> Tensor:
|
78 |
+
last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
|
79 |
+
return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
|
80 |
+
|
81 |
+
def get_result(q_txt, p_txt1, p_txt2, p_txt3, p_txt4, p_txt5):
|
82 |
+
input_texts = [
|
83 |
+
f"query: {q_txt}"
|
84 |
+
]
|
85 |
+
|
86 |
+
if p_txt1 != '':
|
87 |
+
input_txt.append(f"passage: {p_txt1}")
|
88 |
+
|
89 |
+
if p_txt2 != '':
|
90 |
+
input_txt.append(f"passage: {p_txt2}")
|
91 |
+
|
92 |
+
if p_txt3 != '':
|
93 |
+
input_txt.append(f"passage: {p_txt3}")
|
94 |
+
|
95 |
+
if p_txt4 != '':
|
96 |
+
input_txt.append(f"passage: {p_txt4}")
|
97 |
+
|
98 |
+
if p_txt5 != '':
|
99 |
+
input_txt.append(f"passage: {p_txt5}")
|
100 |
+
|
101 |
+
scores = get_scores(model, tokenizer, input_texts)
|
102 |
+
result = get_human_readable_top(scores, input_texts)
|
103 |
+
return json.dumps(result, indent=4)
|
104 |
+
|
105 |
+
tokenizer, model = get_model('intfloat/e5-large-v2')
|
106 |
+
|
107 |
with gr.Blocks() as demo:
|
108 |
gr.Markdown("# E5 Large V2 Demo")
|
109 |
|
110 |
+
q_txt = gr.Textbox(placeholder="Enter your query", info="Query")
|
111 |
|
112 |
+
p_txt1 = gr.Textbox(placeholder="Enter passage 1", info="Passage 1")
|
113 |
+
p_txt2 = gr.Textbox(placeholder="Enter passage 2", info="Passage 2")
|
114 |
+
p_txt3 = gr.Textbox(placeholder="Enter passage 3", info="Passage 3")
|
115 |
+
p_txt4 = gr.Textbox(placeholder="Enter passage 4", info="Passage 4")
|
116 |
+
p_txt5 = gr.Textbox(placeholder="Enter passage 5", info="Passage 5")
|
|
|
117 |
|
118 |
submit = gr.Button("Submit")
|
119 |
+
submit.click(
|
120 |
+
get_result,
|
121 |
+
[q_txt, p_txt1, p_txt2, p_txt3, p_txt4, p_txt5],
|
122 |
+
o_txt
|
123 |
+
)
|
124 |
|
125 |
o_txt = gr.Textbox(placeholder="Output", lines=10, interactive=False)
|
126 |
|