Spaces:
Runtime error
Runtime error
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
|
3 |
+
import nltk
|
4 |
+
import math
|
5 |
+
import torch
|
6 |
+
|
7 |
+
model_name = "afnanmmir/t5-base-abstract-to-plain-language-1"
|
8 |
+
# model_name = "afnanmmir/t5-base-axriv-to-abstract-3"
|
9 |
+
max_input_length = 1024
|
10 |
+
max_output_length = 256
|
11 |
+
|
12 |
+
st.header("Generate summaries")
|
13 |
+
|
14 |
+
st_model_load = st.text('Loading summary generator model...')
|
15 |
+
|
16 |
+
# @st.cache(allow_output_mutation=True)
|
17 |
+
@st.cache_data
|
18 |
+
def load_model():
|
19 |
+
print("Loading model...")
|
20 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
21 |
+
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
|
22 |
+
nltk.download('punkt')
|
23 |
+
print("Model loaded!")
|
24 |
+
return tokenizer, model
|
25 |
+
|
26 |
+
tokenizer, model = load_model()
|
27 |
+
st.success('Model loaded!')
|
28 |
+
st_model_load.text("")
|
29 |
+
|
30 |
+
# with st.sidebar:
|
31 |
+
# st.header("Model parameters")
|
32 |
+
# if 'num_titles' not in st.session_state:
|
33 |
+
# st.session_state.num_titles = 5
|
34 |
+
# def on_change_num_titles():
|
35 |
+
# st.session_state.num_titles = num_titles
|
36 |
+
# num_titles = st.slider("Number of titles to generate", min_value=1, max_value=10, value=1, step=1, on_change=on_change_num_titles)
|
37 |
+
# if 'temperature' not in st.session_state:
|
38 |
+
# st.session_state.temperature = 0.7
|
39 |
+
# def on_change_temperatures():
|
40 |
+
# st.session_state.temperature = temperature
|
41 |
+
# temperature = st.slider("Temperature", min_value=0.1, max_value=1.5, value=0.6, step=0.05, on_change=on_change_temperatures)
|
42 |
+
# st.markdown("_High temperature means that results are more random_")
|
43 |
+
|
44 |
+
if 'text' not in st.session_state:
|
45 |
+
st.session_state.text = ""
|
46 |
+
st_text_area = st.text_area('Text to generate the summary for', value=st.session_state.text, height=500)
|
47 |
+
|
48 |
+
def generate_summary():
|
49 |
+
st.session_state.text = st_text_area
|
50 |
+
|
51 |
+
# tokenize text
|
52 |
+
inputs = ["summarize: " + st_text_area]
|
53 |
+
# print(inputs)
|
54 |
+
inputs = tokenizer(inputs, return_tensors="pt", max_length=max_input_length, truncation=True)
|
55 |
+
print("Tokenized inputs: ")
|
56 |
+
# print(inputs)
|
57 |
+
# inputs = tokenizer(inputs, return_tensors="pt")
|
58 |
+
|
59 |
+
# # compute span boundaries
|
60 |
+
# num_tokens = len(inputs["input_ids"][0])
|
61 |
+
# print(f"Input has {num_tokens} tokens")
|
62 |
+
# max_input_length = 500
|
63 |
+
# num_spans = math.ceil(num_tokens / max_input_length)
|
64 |
+
# print(f"Input has {num_spans} spans")
|
65 |
+
# overlap = math.ceil((num_spans * max_input_length - num_tokens) / max(num_spans - 1, 1))
|
66 |
+
# spans_boundaries = []
|
67 |
+
# start = 0
|
68 |
+
# for i in range(num_spans):
|
69 |
+
# spans_boundaries.append([start + max_input_length * i, start + max_input_length * (i + 1)])
|
70 |
+
# start -= overlap
|
71 |
+
# print(f"Span boundaries are {spans_boundaries}")
|
72 |
+
# spans_boundaries_selected = []
|
73 |
+
# j = 0
|
74 |
+
# for _ in range(num_titles):
|
75 |
+
# spans_boundaries_selected.append(spans_boundaries[j])
|
76 |
+
# j += 1
|
77 |
+
# if j == len(spans_boundaries):
|
78 |
+
# j = 0
|
79 |
+
# print(f"Selected span boundaries are {spans_boundaries_selected}")
|
80 |
+
|
81 |
+
# # transform input with spans
|
82 |
+
# tensor_ids = [inputs["input_ids"][0][boundary[0]:boundary[1]] for boundary in spans_boundaries_selected]
|
83 |
+
# tensor_masks = [inputs["attention_mask"][0][boundary[0]:boundary[1]] for boundary in spans_boundaries_selected]
|
84 |
+
|
85 |
+
# inputs = {
|
86 |
+
# "input_ids": torch.stack(tensor_ids),
|
87 |
+
# "attention_mask": torch.stack(tensor_masks)
|
88 |
+
# }
|
89 |
+
|
90 |
+
# compute predictions
|
91 |
+
# outputs = model.generate(**inputs, do_sample=True, temperature=temperature, max_length=max_output_length)
|
92 |
+
outputs = model.generate(**inputs, do_sample=True, max_length=max_output_length, early_stopping=True, num_beams=8, length_penalty=2.0, no_repeat_ngram_size=2, min_length=64)
|
93 |
+
# print("outputs", outputs)
|
94 |
+
decoded_outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
|
95 |
+
# print("Decoded_outputs", decoded_outputs)
|
96 |
+
predicted_summaries = nltk.sent_tokenize(decoded_outputs.strip())
|
97 |
+
# print("Predicted summaries", predicted_summaries)
|
98 |
+
# decoded_outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
99 |
+
# predicted_summaries = [nltk.sent_tokenize(decoded_output.strip())[0] for decoded_output in decoded_outputs]
|
100 |
+
|
101 |
+
st.session_state.summaries = predicted_summaries
|
102 |
+
|
103 |
+
# generate title button
|
104 |
+
st_generate_button = st.button('Generate summary', on_click=generate_summary)
|
105 |
+
|
106 |
+
# title generation labels
|
107 |
+
if 'summaries' not in st.session_state:
|
108 |
+
st.session_state.summaries = []
|
109 |
+
|
110 |
+
if len(st.session_state.summaries) > 0:
|
111 |
+
# print("In summaries if")
|
112 |
+
with st.container():
|
113 |
+
st.subheader("Generated summaries")
|
114 |
+
st.markdown(f"{' '.join(st.session_state.summaries)}")
|