Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
|
3 |
+
from transformers import MarkupLMProcessor, MarkupLMForQuestionAnswering
|
4 |
+
|
5 |
+
import requests
|
6 |
+
from bs4 import BeautifulSoup
|
7 |
+
|
8 |
+
import numpy as np
|
9 |
+
|
10 |
+
import torch
|
11 |
+
import torch.nn.functional as F
|
12 |
+
|
13 |
+
# Prediction Parameters
|
14 |
+
MAX_LEN = 512
|
15 |
+
STRIDE = 100
|
16 |
+
|
17 |
+
# Answer filtering parameters
|
18 |
+
MAX_ANSWER_LEN = 30
|
19 |
+
MIN_CONFIDENCE = 0.9
|
20 |
+
|
21 |
+
# Model name
|
22 |
+
MODEL_STR = "microsoft/markuplm-base-finetuned-websrc"
|
23 |
+
|
24 |
+
# Load markuplm model
|
25 |
+
processor = MarkupLMProcessor.from_pretrained(MODEL_STR)
|
26 |
+
model = MarkupLMForQuestionAnswering.from_pretrained(MODEL_STR)
|
27 |
+
|
28 |
+
headers = {
|
29 |
+
'user-agent': 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/86.0.4240.111 Safari/537.36',
|
30 |
+
}
|
31 |
+
|
32 |
+
# User Input
|
33 |
+
input_url = st.text_input(
|
34 |
+
label="Enter url of page to scrape",
|
35 |
+
value="https://www.opentable.com/carlo-and-johnny",
|
36 |
+
key="url",
|
37 |
+
)
|
38 |
+
|
39 |
+
input_question = st.text_input(
|
40 |
+
label="Enter Question",
|
41 |
+
value="What is the food on the menu?",
|
42 |
+
key="question",
|
43 |
+
)
|
44 |
+
|
45 |
+
st.write("Getting html page ...")
|
46 |
+
|
47 |
+
# Request page
|
48 |
+
page = requests.get(input_url, headers=headers)
|
49 |
+
|
50 |
+
# Parse page with beautifulsoup
|
51 |
+
soup = BeautifulSoup(page.content, "html.parser")
|
52 |
+
|
53 |
+
# Extract page body
|
54 |
+
body = soup.find('body')
|
55 |
+
|
56 |
+
html_string = str(body)
|
57 |
+
len(html_string)
|
58 |
+
|
59 |
+
# Process input string
|
60 |
+
encoding = processor(html_string, questions=input_question, return_tensors="pt", truncation="only_second",
|
61 |
+
stride=STRIDE, max_length=MAX_LEN, return_overflowing_tokens=True, padding=True)
|
62 |
+
|
63 |
+
# Postprocess encoding
|
64 |
+
del encoding['overflow_to_sample_mapping']
|
65 |
+
encoding['token_type_ids'] = encoding['token_type_ids'].fill_(0)
|
66 |
+
|
67 |
+
# Keep index of question for future use
|
68 |
+
n_segments = encoding['input_ids'].shape[0]
|
69 |
+
question_index = encoding[0].tokens.index('</s>')
|
70 |
+
|
71 |
+
# Run model
|
72 |
+
with torch.no_grad():
|
73 |
+
outputs = model(**encoding)
|
74 |
+
|
75 |
+
# Get start and end probabilities
|
76 |
+
start_probs = F.softmax(outputs.start_logits, dim=1).numpy()
|
77 |
+
end_probs = F.softmax(outputs.end_logits, dim=1).numpy()
|
78 |
+
|
79 |
+
# Extract and filter answers for each window
|
80 |
+
answers = []
|
81 |
+
|
82 |
+
for i in range(n_segments):
|
83 |
+
|
84 |
+
start_index = np.argmax(start_probs[i])
|
85 |
+
end_index = np.argmax(end_probs[i])
|
86 |
+
confidence = max(start_probs[i]) * max(end_probs[i])
|
87 |
+
|
88 |
+
if end_index > start_index and end_index - start_index <= MAX_ANSWER_LEN and start_index > question_index and end_index > question_index and confidence > MIN_CONFIDENCE:
|
89 |
+
|
90 |
+
predict_answer_tokens = encoding.input_ids[0, start_index : end_index + 1]
|
91 |
+
answer = processor.decode(predict_answer_tokens, skip_special_tokens=True)
|
92 |
+
|
93 |
+
answers.append({"answer": answer, "confidence": confidence})
|
94 |
+
|
95 |
+
# Print answers
|
96 |
+
for answer in answers:
|
97 |
+
st.write(answer)
|
98 |
+
|
99 |
+
st.write("Done!")
|