ilyshi commited on
Commit
8922b9c
1 Parent(s): 399f3ee

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +4 -2
app.py CHANGED
@@ -7,6 +7,7 @@ from transformers import TrainingArguments, \
7
  categories = ['Biology', 'Computer science', 'Economics', 'Electrics', 'Finance',
8
  'Math', 'Physics', 'Statistics']
9
  labels = [i for i in range(len(categories))]
 
10
  def print_probs(logits):
11
  probs = torch.nn.functional.softmax(logits, dim=0).numpy()*100
12
  ans = list(zip(probs,labels))
@@ -19,7 +20,8 @@ def print_probs(logits):
19
  st.write(text)
20
  sum+=prob
21
  i+=1
22
-
 
23
  def make_prediction(text):
24
  tokenized_text = tokenizer(text, return_tensors='pt')
25
  with torch.no_grad():
@@ -54,6 +56,6 @@ text = st.text_area("Введите название статьи", height=50)
54
  st.markdown("### Article Abstract")
55
  text = st.text_area("Введите описание статьи", height=400)
56
 
57
- @st.cache
58
  make_prediction(text)
59
 
 
7
  categories = ['Biology', 'Computer science', 'Economics', 'Electrics', 'Finance',
8
  'Math', 'Physics', 'Statistics']
9
  labels = [i for i in range(len(categories))]
10
+
11
  def print_probs(logits):
12
  probs = torch.nn.functional.softmax(logits, dim=0).numpy()*100
13
  ans = list(zip(probs,labels))
 
20
  st.write(text)
21
  sum+=prob
22
  i+=1
23
+
24
+ @st.cache
25
  def make_prediction(text):
26
  tokenized_text = tokenizer(text, return_tensors='pt')
27
  with torch.no_grad():
 
56
  st.markdown("### Article Abstract")
57
  text = st.text_area("Введите описание статьи", height=400)
58
 
59
+
60
  make_prediction(text)
61