Spaces:
Runtime error
Runtime error
add app
Browse files
app.py
ADDED
@@ -0,0 +1,163 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import transformers
|
2 |
+
import torch
|
3 |
+
|
4 |
+
import streamlit as st
|
5 |
+
|
6 |
+
from transformers import BertTokenizer
|
7 |
+
|
8 |
+
st.markdown("### Из какой серии статья")
|
9 |
+
# st.markdown("<img width=200px src='https://rozetked.me/images/uploads/dwoilp3BVjlE.jpg'>", unsafe_allow_html=True)
|
10 |
+
|
11 |
+
# from transformers import pipeline
|
12 |
+
|
13 |
+
# pipe = pipeline("ner", "Davlan/distilbert-base-multilingual-cased-ner-hrl")
|
14 |
+
|
15 |
+
|
16 |
+
num_classes = 8
|
17 |
+
class BERTClass(torch.nn.Module):
|
18 |
+
def __init__(self, n_hid1 = 1024, n_out=num_classes, bert_path='bert-base-uncased'):
|
19 |
+
super(BERTClass, self).__init__()
|
20 |
+
self.l1 = transformers.BertModel.from_pretrained(bert_path)
|
21 |
+
self.l2 = torch.nn.Dropout(0.3)
|
22 |
+
self.l3 = torch.nn.Linear(768, n_hid1)
|
23 |
+
self.l4 = torch.nn.ReLU()
|
24 |
+
self.l5 = torch.nn.Dropout(0.2)
|
25 |
+
self.l6 = torch.nn.Linear(n_hid1, n_out)
|
26 |
+
|
27 |
+
def forward(self, ids, mask, token_type_ids):
|
28 |
+
# _, output_1= self.l1(ids, attention_mask = mask, token_type_ids = token_type_ids)
|
29 |
+
out = self.l1(ids, attention_mask = mask, token_type_ids = token_type_ids)
|
30 |
+
out = self.l2(out[1])
|
31 |
+
out = self.l3(out)
|
32 |
+
out = self.l4(out)
|
33 |
+
out = self.l5(out)
|
34 |
+
out = self.l6(out)
|
35 |
+
return out
|
36 |
+
|
37 |
+
@st.cache
|
38 |
+
def load_bert():
|
39 |
+
model = BERTClass(bert_path='bert_pretrained')
|
40 |
+
model.load_state_dict(torch.load('bert_pretrained.pt'))
|
41 |
+
model.eval()
|
42 |
+
|
43 |
+
tokenizer = BertTokenizer.from_pretrained('bert_tokenizer')
|
44 |
+
|
45 |
+
return model, tokenizer
|
46 |
+
|
47 |
+
|
48 |
+
def apply_bert(text, model, tokenizer):
|
49 |
+
"""returns probabilities"""
|
50 |
+
MAX_LEN = 200
|
51 |
+
ins = tokenizer.encode_plus(text, None, add_special_tokens=True,
|
52 |
+
max_length=MAX_LEN,
|
53 |
+
pad_to_max_length=True,
|
54 |
+
return_token_type_ids=True
|
55 |
+
)
|
56 |
+
ids = torch.tensor(ins['input_ids']).unsqueeze(0)
|
57 |
+
mask = torch.tensor(ins['attention_mask']).unsqueeze(0)
|
58 |
+
token_type_ids = torch.tensor(ins["token_type_ids"])
|
59 |
+
out = model(ids, mask, token_type_ids)
|
60 |
+
return torch.sigmoid(out).flatten().detach()
|
61 |
+
|
62 |
+
|
63 |
+
class TinyBERTClass(torch.nn.Module):
|
64 |
+
def __init__(self, n_hid1 = 1024, n_out=num_classes, path='distilbert-base-uncased'):
|
65 |
+
super(TinyBERTClass, self).__init__()
|
66 |
+
self.l1 = transformers.DistilBertModel.from_pretrained(path)
|
67 |
+
self.l2 = torch.nn.Dropout(0.3)
|
68 |
+
self.l3 = torch.nn.Linear(768, n_hid1)
|
69 |
+
self.l4 = torch.nn.ReLU()
|
70 |
+
self.l5 = torch.nn.Dropout(0.2)
|
71 |
+
self.l6 = torch.nn.Linear(n_hid1, n_out)
|
72 |
+
|
73 |
+
def forward(self, ids, mask):
|
74 |
+
# _, output_1= self.l1(ids, attention_mask = mask, token_type_ids = token_type_ids)
|
75 |
+
out = self.l1(ids, attention_mask = mask)
|
76 |
+
out = self.l2(out.last_hidden_state[:,0,:])
|
77 |
+
out = self.l3(out)
|
78 |
+
out = self.l4(out)
|
79 |
+
out = self.l5(out)
|
80 |
+
out = self.l6(out)
|
81 |
+
return out
|
82 |
+
|
83 |
+
|
84 |
+
@st.cache(suppress_st_warning=True)
|
85 |
+
def load_tiny_bert():
|
86 |
+
model = TinyBERTClass(path = 'tiny_bert_pretrained')
|
87 |
+
model.load_state_dict(torch.load('tiny_bert.pt'))
|
88 |
+
model.eval()
|
89 |
+
|
90 |
+
tokenizer = transformers.DistilBertTokenizer.from_pretrained('tiny_bert_tokenizer')
|
91 |
+
|
92 |
+
return model, tokenizer
|
93 |
+
|
94 |
+
|
95 |
+
def apply_tiny_bert(text, model, tokenizer):
|
96 |
+
encoded_input = tokenizer(text, return_tensors='pt')
|
97 |
+
out = model(encoded_input['input_ids'], encoded_input['attention_mask'])
|
98 |
+
|
99 |
+
return torch.sigmoid(out).flatten().detach()
|
100 |
+
|
101 |
+
|
102 |
+
|
103 |
+
title = st.text_area("Название статьи")
|
104 |
+
if not title.endswith('.') and title:
|
105 |
+
title += '.'
|
106 |
+
|
107 |
+
summary = st.text_area("Аннотация статьи")
|
108 |
+
|
109 |
+
calc_button = st.button('Угадать тематику')
|
110 |
+
|
111 |
+
bert_model, bert_tokenizer = load_bert()
|
112 |
+
tiny_bert, tiny_bert_tokenizer = load_tiny_bert()
|
113 |
+
|
114 |
+
# calculate ================================================================
|
115 |
+
if calc_button:
|
116 |
+
print('title')
|
117 |
+
print(title)
|
118 |
+
print('=' * 80)
|
119 |
+
# print(text)
|
120 |
+
|
121 |
+
if summary:
|
122 |
+
text = title + summary
|
123 |
+
out = apply_bert(text, bert_model, bert_tokenizer)
|
124 |
+
else:
|
125 |
+
out = apply_tiny_bert(title, tiny_bert, tiny_bert_tokenizer)
|
126 |
+
|
127 |
+
|
128 |
+
RU_NAMES = ['компьютерным наукам'
|
129 |
+
,'экономике'
|
130 |
+
,'электротехнике и системотехнике'
|
131 |
+
,'математике'
|
132 |
+
,'физике'
|
133 |
+
,'количественной биологии'
|
134 |
+
,'количественным финансам'
|
135 |
+
,'статистике'
|
136 |
+
]
|
137 |
+
|
138 |
+
def get_classes(out, bandwidth = 0.5):
|
139 |
+
res = []
|
140 |
+
for i in range(out.size()[0]):
|
141 |
+
if out[i] >= bandwidth:
|
142 |
+
res.append(i)
|
143 |
+
|
144 |
+
ans = ''
|
145 |
+
total = 0
|
146 |
+
for i in res:
|
147 |
+
total += out[i].item()
|
148 |
+
if not ans:
|
149 |
+
ans += f'\nэто статья по {RU_NAMES[i]} с вероятностью {out[i].item():.2f}'
|
150 |
+
else:
|
151 |
+
ans += f',\nтакже она по {RU_NAMES[i]} с вероятностью {out[i].item():.2f}'
|
152 |
+
|
153 |
+
ans = 'Э' + ans[2:]
|
154 |
+
if total >= 1.0:
|
155 |
+
ans += '.\n(Решалась задача мультиклассификации, поэтому сумма вероятностей получилась больше 1.)'
|
156 |
+
|
157 |
+
if ans == 'Э':
|
158 |
+
return 'Не похоже на что-то научное, Вы уверены что это взято из статьи?'
|
159 |
+
return ans
|
160 |
+
|
161 |
+
res = get_classes(out)
|
162 |
+
|
163 |
+
st.markdown(f"{res}")
|