Spaces:
Build error
Build error
Upload 2 files
Browse files- Summarization_25Nov2022.py +357 -0
- requirements.txt +4 -0
Summarization_25Nov2022.py
ADDED
@@ -0,0 +1,357 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import streamlit as st
|
3 |
+
from streamlit import components
|
4 |
+
import pandas as pd
|
5 |
+
from transformers import BartTokenizer, BartForConditionalGeneration
|
6 |
+
from transformers import T5Tokenizer, T5ForConditionalGeneration
|
7 |
+
import evaluate
|
8 |
+
from datasets import load_dataset
|
9 |
+
from transformers import AutoTokenizer, LongT5ForConditionalGeneration
|
10 |
+
import numpy as np
|
11 |
+
from math import ceil
|
12 |
+
import en_core_web_lg
|
13 |
+
from collections import Counter
|
14 |
+
from string import punctuation
|
15 |
+
# Gensim
|
16 |
+
import gensim
|
17 |
+
from gensim.summarization import summarize
|
18 |
+
import spacy
|
19 |
+
|
20 |
+
nlp = en_core_web_lg.load()
|
21 |
+
|
22 |
+
st.set_page_config(page_title ='Clinical Note Summarization',
|
23 |
+
#page_icon= "Notes",
|
24 |
+
layout='wide')
|
25 |
+
st.title('Clinical Note Summarization')
|
26 |
+
st.sidebar.markdown('Using transformer model')
|
27 |
+
|
28 |
+
## Loading in dataset
|
29 |
+
#df = pd.read_csv('mtsamples_small.csv',index_col=0)
|
30 |
+
df = pd.read_csv("shpi_w_rouge21Nov.csv")
|
31 |
+
#df.shape
|
32 |
+
df['HADM_ID'] = df['HADM_ID'].astype(str).apply(lambda x: x.replace('.0',''))
|
33 |
+
|
34 |
+
##Renaming column
|
35 |
+
#df.rename(columns={'patient id':'Patient_ID',
|
36 |
+
# 'hospital admission id':'Admission_ID',
|
37 |
+
# 'transcription':'Original_Text'}, inplace = True)
|
38 |
+
|
39 |
+
#Renaming column
|
40 |
+
df.rename(columns={'SUBJECT_ID':'Patient_ID',
|
41 |
+
'HADM_ID':'Admission_ID',
|
42 |
+
'hpi_input_text':'Original_Text',
|
43 |
+
'hpi_reference_summary':'Reference_text'}, inplace = True)
|
44 |
+
|
45 |
+
#data.rename(columns={'gdp':'log(gdp)'}, inplace=True)
|
46 |
+
|
47 |
+
#Filter selection
|
48 |
+
st.sidebar.header("Search for Patient:")
|
49 |
+
|
50 |
+
patientid = df['Patient_ID']
|
51 |
+
patient = st.sidebar.selectbox('Select Patient ID:', patientid)
|
52 |
+
admissionid = df['Admission_ID'].loc[df['Patient_ID'] == patient]
|
53 |
+
HospitalAdmission = st.sidebar.selectbox('', admissionid)
|
54 |
+
|
55 |
+
#Another way to for filter selection
|
56 |
+
#patient = st.sidebar.multiselect(
|
57 |
+
# "Select Patient ID:",
|
58 |
+
# options=df['Patient_ID'].unique(),
|
59 |
+
# default= None
|
60 |
+
#)
|
61 |
+
|
62 |
+
|
63 |
+
#HospitalAdmission = st.sidebar.multiselect(
|
64 |
+
# "Select Hospital Admission ID:",
|
65 |
+
# options=df['Admission_ID'].unique(),
|
66 |
+
# #default=df['Admission_ID'].unique()
|
67 |
+
# default = None
|
68 |
+
#)
|
69 |
+
|
70 |
+
|
71 |
+
# List of Model available
|
72 |
+
model = st.sidebar.selectbox('Select Model', ('BART','BERT','BertGPT2','Gensim','LexRank','Long T5','Luhn','Pysummarization','SBERT Summary Tokenizer','T5','T5 Seq2Seq','T5-Base','TextRank'))
|
73 |
+
|
74 |
+
|
75 |
+
if model == 'BART':
|
76 |
+
_num_beams = 4
|
77 |
+
_no_repeat_ngram_size = 3
|
78 |
+
_length_penalty = 1
|
79 |
+
_min_length = 12
|
80 |
+
_max_length = 128
|
81 |
+
_early_stopping = True
|
82 |
+
else:
|
83 |
+
_num_beams = 4
|
84 |
+
_no_repeat_ngram_size = 3
|
85 |
+
_length_penalty = 2
|
86 |
+
_min_length = 30
|
87 |
+
_max_length = 200
|
88 |
+
_early_stopping = True
|
89 |
+
|
90 |
+
|
91 |
+
|
92 |
+
|
93 |
+
col3,col4 = st.columns(2)
|
94 |
+
patientid = col3.write(f"Patient ID: {patient} ")
|
95 |
+
admissionid =col4.write(f"Admission ID: {HospitalAdmission} ")
|
96 |
+
|
97 |
+
col1, col2 = st.columns(2)
|
98 |
+
_min_length = col1.number_input("Minimum Length", value=_min_length)
|
99 |
+
_max_length = col2.number_input("Maximun Length", value=_max_length)
|
100 |
+
##_early_stopping = col3.number_input("early_stopping", value=_early_stopping)
|
101 |
+
|
102 |
+
#text = st.text_area('Input Clinical Note here')
|
103 |
+
|
104 |
+
# Query out relevant Clinical notes
|
105 |
+
original_text = df.query(
|
106 |
+
"Patient_ID == @patient & Admission_ID == @HospitalAdmission"
|
107 |
+
)
|
108 |
+
|
109 |
+
original_text2 = original_text['Original_Text'].values
|
110 |
+
|
111 |
+
runtext =st.text_area('Input Clinical Note here:', str(original_text2), height=300)
|
112 |
+
|
113 |
+
reference_text = original_text['Reference_text'].values
|
114 |
+
|
115 |
+
|
116 |
+
|
117 |
+
## ===== to highlight text =====
|
118 |
+
from IPython.core.display import HTML, display
|
119 |
+
def visualize(title, sentence_list, best_sentences):
|
120 |
+
text = ''
|
121 |
+
|
122 |
+
#display(HTML(f'<h1>Summary - {title}</h1>'))
|
123 |
+
for sentence in sentence_list:
|
124 |
+
if sentence in best_sentences:
|
125 |
+
#text += ' ' + str(sentence).replace(sentence, f"<mark>{sentence}</mark>")
|
126 |
+
text += ' ' + str(sentence).replace(sentence, f"<span class='highlight yellow'>{sentence}</span>")
|
127 |
+
else:
|
128 |
+
text += ' ' + sentence
|
129 |
+
display(HTML(f""" {text} """))
|
130 |
+
|
131 |
+
output = ''
|
132 |
+
best_sentences = []
|
133 |
+
for sentence in output:
|
134 |
+
#print(sentence)
|
135 |
+
best_sentences.append(str(sentence))
|
136 |
+
return text
|
137 |
+
# try this web solution https://discuss.streamlit.io/t/colored-boxes-around-sections-of-a-sentence/3201/2
|
138 |
+
|
139 |
+
#===== Pysummarization =====
|
140 |
+
from pysummarization.nlpbase.auto_abstractor import AutoAbstractor
|
141 |
+
from pysummarization.tokenizabledoc.simple_tokenizer import SimpleTokenizer
|
142 |
+
from pysummarization.abstractabledoc.top_n_rank_abstractor import TopNRankAbstractor
|
143 |
+
import regex as re
|
144 |
+
|
145 |
+
auto_abstractor = AutoAbstractor()
|
146 |
+
auto_abstractor.tokenizable_doc = SimpleTokenizer()
|
147 |
+
auto_abstractor.delimiter_list = [".", "\n"]
|
148 |
+
abstractable_doc = TopNRankAbstractor()
|
149 |
+
|
150 |
+
def pysummarizer(input_text):
|
151 |
+
# print(type(text))
|
152 |
+
summary = auto_abstractor.summarize(input_text, abstractable_doc)
|
153 |
+
best_sentences=[]
|
154 |
+
#summary_clean = ''.join([str(sentence).capitalize() for sentence in summary['summarize_result'] for summary['summarize_result'] in auto_abstractor.summarize(text, abstractable_doc)])
|
155 |
+
for sentence in summary['summarize_result']:
|
156 |
+
best_sentences.append(re.sub(r'\s+', ' ', sentence).strip())
|
157 |
+
clean_summary=''.join(sentence for sentence in best_sentences)
|
158 |
+
return clean_summary
|
159 |
+
|
160 |
+
|
161 |
+
|
162 |
+
##===== BERT Summary tokenizer =====
|
163 |
+
|
164 |
+
def BertSummarizer(input_text):
|
165 |
+
from transformers import BigBirdTokenizer
|
166 |
+
from summarizer import Summarizer
|
167 |
+
|
168 |
+
bertsummarizer = Summarizer()
|
169 |
+
|
170 |
+
model = Summarizer()
|
171 |
+
result = model(input_text,ratio=0.4)
|
172 |
+
|
173 |
+
return result
|
174 |
+
|
175 |
+
|
176 |
+
##===== SBERT =====
|
177 |
+
from summarizer.sbert import SBertSummarizer
|
178 |
+
|
179 |
+
|
180 |
+
Sbertmodel = SBertSummarizer('paraphrase-MiniLM-L6-v2')
|
181 |
+
|
182 |
+
def Sbert(input_text):
|
183 |
+
|
184 |
+
# Sbertresult = Sbertmodel(text, num_sentences=3)
|
185 |
+
Sbertresult = Sbertmodel(input_text, ratio=0.4)
|
186 |
+
return Sbertresult
|
187 |
+
|
188 |
+
|
189 |
+
|
190 |
+
##===== T5 Seq2Seq =====
|
191 |
+
def t5seq2seq(input_text):
|
192 |
+
import torch
|
193 |
+
import torch.nn.functional as F
|
194 |
+
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
|
195 |
+
|
196 |
+
model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")
|
197 |
+
tokenizer = AutoTokenizer.from_pretrained("t5-base")
|
198 |
+
|
199 |
+
inputs = tokenizer("summarize: " + input_text, return_tensors="pt", max_length=512, truncation=True)
|
200 |
+
outputs = model.generate(inputs["input_ids"], max_length=150, min_length=40, length_penalty=2.0, num_beams=4, early_stopping=True)
|
201 |
+
|
202 |
+
summary= tokenizer.decode(outputs[0], skip_special_tokens=True)
|
203 |
+
|
204 |
+
return summary
|
205 |
+
|
206 |
+
def BertGPT2(input_text):
|
207 |
+
#import nlp
|
208 |
+
|
209 |
+
# BioClinicalBert with BERT2GPT2 model with GPT2 decoder
|
210 |
+
from transformers import BertTokenizer, GPT2Tokenizer, EncoderDecoderModel
|
211 |
+
from transformers import AutoTokenizer, AutoModel
|
212 |
+
|
213 |
+
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
214 |
+
model = EncoderDecoderModel.from_pretrained("patrickvonplaten/bert2gpt2-cnn_dailymail-fp16")
|
215 |
+
model.to(device)
|
216 |
+
|
217 |
+
#bert_tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
|
218 |
+
bert_tokenizer= AutoTokenizer.from_pretrained("emilyalsentzer/Bio_ClinicalBERT")
|
219 |
+
|
220 |
+
# CLS token will work as BOS token
|
221 |
+
bert_tokenizer.bos_token = bert_tokenizer.cls_token
|
222 |
+
|
223 |
+
# SEP token will work as EOS token
|
224 |
+
bert_tokenizer.eos_token = bert_tokenizer.sep_token
|
225 |
+
|
226 |
+
|
227 |
+
# make sure GPT2 appends EOS in begin and end
|
228 |
+
def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
|
229 |
+
outputs = [self.bos_token_id] + token_ids_0 + [self.eos_token_id]
|
230 |
+
return outputs
|
231 |
+
|
232 |
+
|
233 |
+
GPT2Tokenizer.build_inputs_with_special_tokens = build_inputs_with_special_tokens
|
234 |
+
gpt2_tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
|
235 |
+
# set pad_token_id to unk_token_id -> be careful here as unk_token_id == eos_token_id == bos_token_id
|
236 |
+
gpt2_tokenizer.pad_token = gpt2_tokenizer.unk_token
|
237 |
+
|
238 |
+
|
239 |
+
# set decoding params
|
240 |
+
model.config.decoder_start_token_id = gpt2_tokenizer.bos_token_id
|
241 |
+
model.config.eos_token_id = gpt2_tokenizer.eos_token_id
|
242 |
+
model.config.max_length = 142
|
243 |
+
model.config.min_length = 56
|
244 |
+
model.config.no_repeat_ngram_size = 3
|
245 |
+
model.early_stopping = True
|
246 |
+
model.length_penalty = 2.0
|
247 |
+
model.num_beams = 4
|
248 |
+
|
249 |
+
#test_dataset = nlp.load_dataset("cnn_dailymail", "3.0.0", split="test")
|
250 |
+
|
251 |
+
batch_size = 64
|
252 |
+
|
253 |
+
def Sbertmodel(batch):
|
254 |
+
# Tokenizer will automatically set [BOS] <text> [EOS]
|
255 |
+
# cut off at BERT max length 512
|
256 |
+
inputs = bert_tokenizer(batch, padding="max_length", truncation=True, max_length=512, return_tensors="pt")
|
257 |
+
input_ids = inputs.input_ids.to("cuda")
|
258 |
+
attention_mask = inputs.attention_mask.to("cuda")
|
259 |
+
|
260 |
+
outputs = model.generate(input_ids, attention_mask=attention_mask)
|
261 |
+
|
262 |
+
# all special tokens including will be removed
|
263 |
+
output_str = gpt2_tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
264 |
+
|
265 |
+
#batch["pred"] = output_str
|
266 |
+
|
267 |
+
return output_str
|
268 |
+
|
269 |
+
Sbert(input_text)
|
270 |
+
|
271 |
+
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
272 |
+
|
273 |
+
|
274 |
+
def run_model(input_text):
|
275 |
+
|
276 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
277 |
+
|
278 |
+
if model == "BART":
|
279 |
+
bart_model = BartForConditionalGeneration.from_pretrained("facebook/bart-base")
|
280 |
+
bart_tokenizer = BartTokenizer.from_pretrained("facebook/bart-base")
|
281 |
+
input_text = str(input_text)
|
282 |
+
input_text = ' '.join(input_text.split())
|
283 |
+
input_tokenized = bart_tokenizer.encode(input_text, return_tensors='pt').to(device)
|
284 |
+
summary_ids = bart_model.generate(input_tokenized,
|
285 |
+
num_beams=_num_beams,
|
286 |
+
no_repeat_ngram_size=_no_repeat_ngram_size,
|
287 |
+
length_penalty=_length_penalty,
|
288 |
+
min_length=_min_length,
|
289 |
+
max_length=_max_length,
|
290 |
+
early_stopping=_early_stopping)
|
291 |
+
|
292 |
+
output = [bart_tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summary_ids]
|
293 |
+
st.write('Summary')
|
294 |
+
st.success(output[0])
|
295 |
+
|
296 |
+
elif model == "T5":
|
297 |
+
t5_model = T5ForConditionalGeneration.from_pretrained("t5-base")
|
298 |
+
t5_tokenizer = T5Tokenizer.from_pretrained("t5-base")
|
299 |
+
input_text = str(input_text).replace('\n', '')
|
300 |
+
input_text = ' '.join(input_text.split())
|
301 |
+
input_tokenized = t5_tokenizer.encode(input_text, return_tensors="pt").to(device)
|
302 |
+
summary_task = torch.tensor([[21603, 10]]).to(device)
|
303 |
+
input_tokenized = torch.cat([summary_task, input_tokenized], dim=-1).to(device)
|
304 |
+
summary_ids = t5_model.generate(input_tokenized,
|
305 |
+
num_beams=_num_beams,
|
306 |
+
no_repeat_ngram_size=_no_repeat_ngram_size,
|
307 |
+
length_penalty=_length_penalty,
|
308 |
+
min_length=_min_length,
|
309 |
+
max_length=_max_length,
|
310 |
+
early_stopping=_early_stopping)
|
311 |
+
output = [t5_tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summary_ids]
|
312 |
+
st.write('Summary')
|
313 |
+
st.success(output[0])
|
314 |
+
|
315 |
+
|
316 |
+
elif model == "Gensim":
|
317 |
+
output=summarize(str(input_text))
|
318 |
+
#visualize('of text', input_text, output)
|
319 |
+
st.write('Summary')
|
320 |
+
st.success(output)
|
321 |
+
|
322 |
+
elif model == "Pysummarization":
|
323 |
+
output = pysummarizer(input_text)
|
324 |
+
st.write('Summary')
|
325 |
+
st.success(output)
|
326 |
+
|
327 |
+
elif model == "BERT":
|
328 |
+
output = BertSummarizer(input_text)
|
329 |
+
st.write('Summary')
|
330 |
+
st.success(output)
|
331 |
+
|
332 |
+
elif model == "SBERT Summary Tokenizer":
|
333 |
+
output = Sbert(input_text)
|
334 |
+
st.write('Summary')
|
335 |
+
st.success(output)
|
336 |
+
|
337 |
+
elif model == "T5 Seq2Seq":
|
338 |
+
output = t5seq2seq(input_text)
|
339 |
+
st.write('Summary')
|
340 |
+
st.success(output)
|
341 |
+
|
342 |
+
elif model == "BertGPT2": #Not working correctly. to work on it later on
|
343 |
+
output = BertGPT2(input_text)
|
344 |
+
st.write('Summary')
|
345 |
+
st.success(output)
|
346 |
+
|
347 |
+
|
348 |
+
if st.button('Submit'):
|
349 |
+
run_model(runtext)
|
350 |
+
|
351 |
+
# runtext2=runtext.split('.')
|
352 |
+
# reference_text2=reference_text.split('.')
|
353 |
+
|
354 |
+
st.write(visualize('of text', runtext ,reference_text))
|
355 |
+
|
356 |
+
st.text_area('Reference text', str(reference_text))
|
357 |
+
|
requirements.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
streamlit==1.14.0
|
2 |
+
pandas==1.3.5
|
3 |
+
numpy==1.20.0
|
4 |
+
regex==2022.9.13
|