Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
import torch
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import pandas as pd
|
5 |
+
from transformers import AutoModelForSequenceClassification, AutoTokenizer
|
6 |
+
from datasets import load_dataset
|
7 |
+
|
8 |
+
|
9 |
+
device = 'cpu'
|
10 |
+
|
11 |
+
@st.cache_resource
|
12 |
+
def get_model_and_tokenizer():
|
13 |
+
model_name = "FacebookAI/roberta-base"
|
14 |
+
num_labels = 157
|
15 |
+
|
16 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
17 |
+
model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=num_labels)
|
18 |
+
|
19 |
+
chkp = torch.load("arxiv_roberta_9unfrozen_5scaleloss.pt", map_location=device)
|
20 |
+
model.load_state_dict(chkp['model'])
|
21 |
+
|
22 |
+
return model, tokenizer
|
23 |
+
|
24 |
+
@st.cache_data
|
25 |
+
def get_categories():
|
26 |
+
categories = load_dataset("TimSchopf/arxiv_categories", "arxiv_category_descriptions")
|
27 |
+
|
28 |
+
cat2id = dict((cat, id) for id, cat in enumerate(categories['arxiv_category_descriptions']['tag']))
|
29 |
+
id2cat = categories['arxiv_category_descriptions']['tag']
|
30 |
+
names = categories['arxiv_category_descriptions']['name']
|
31 |
+
|
32 |
+
return cat2id, id2cat, names
|
33 |
+
|
34 |
+
model, tokenizer = get_model_and_tokenizer()
|
35 |
+
cat2id, id2cat, cat_names = get_categories()
|
36 |
+
|
37 |
+
@torch.no_grad
|
38 |
+
def predict_and_decode(model, title='', abstract=''):
|
39 |
+
model.eval()
|
40 |
+
|
41 |
+
inputs = tokenizer(title, abstract, return_tensors='pt', truncation=True, max_length=512).to(device)
|
42 |
+
logits = model(**inputs)['logits'][0].cpu()
|
43 |
+
|
44 |
+
df = pd.DataFrame([
|
45 |
+
(id2cat[cat_id], cat_names[cat_id], prob.item())
|
46 |
+
for cat_id, prob in enumerate(F.sigmoid(logits))
|
47 |
+
], columns=("tag", "name", "probability"))
|
48 |
+
df.sort_values(by="probability", ascending=False, inplace=True)
|
49 |
+
|
50 |
+
return df.reset_index(drop=True)
|
51 |
+
|
52 |
+
st.header("Paper Category Classifier")
|
53 |
+
st.text("Input title and/or abstract of a scientific paper, and get classification according to arxiv.org categories")
|
54 |
+
|
55 |
+
title_default = "Attention Is All You Need"
|
56 |
+
abstract_default = (
|
57 |
+
"The dominant sequence transduction models are based on complex recurrent or convolutional neural networks "
|
58 |
+
"in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through "
|
59 |
+
"an attention mechanism. We propose a new simple network architecture, the Transformer..."
|
60 |
+
)
|
61 |
+
|
62 |
+
line_height = 34
|
63 |
+
n_lines = 10
|
64 |
+
title = st.text_input("Paper title", value=title_default, help="Type in paper's title")
|
65 |
+
abstract = st.text_area("Paper abstract", value=abstract_default, height=line_height*n_lines, help="Type in paper's abstract")
|
66 |
+
|
67 |
+
result = predict_and_decode(model, title=title, abstract=abstract)
|
68 |
+
|
69 |
+
cnt = st.container(border=True)
|
70 |
+
with cnt:
|
71 |
+
st.markdown("#### Top category")
|
72 |
+
st.markdown(f"**{result.tag[0]}** -- {result.name[0]}")
|
73 |
+
st.markdown(f"Probability: {result.probability[0]*100:.2f}%")
|
74 |
+
|
75 |
+
threshold = 0.55
|
76 |
+
st.text("Other top categories:")
|
77 |
+
max_len = min(max(1, sum(result.iloc[1:].probability > threshold)), 5)
|
78 |
+
|
79 |
+
def format_p(example):
|
80 |
+
example.probability = f"{example.probability * 100 :.2f}%"
|
81 |
+
return example
|
82 |
+
st.table(result.iloc[1:1 + max_len].apply(format_p, axis=1))
|