m. polinsky
commited on
Create digestor.py
Browse files- digestor.py +252 -0
digestor.py
ADDED
@@ -0,0 +1,252 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# digestor.py is an implementation of a digestor that creates news digests.
|
2 |
+
# the digestor manages the creation of summaries and assembles them into one digest...
|
3 |
+
|
4 |
+
import requests, json
|
5 |
+
from collections import namedtuple
|
6 |
+
from functools import lru_cache
|
7 |
+
from typing import List
|
8 |
+
from dataclasses import dataclass, field
|
9 |
+
from datetime import datetime as dt
|
10 |
+
import streamlit as st
|
11 |
+
|
12 |
+
from codetiming import Timer
|
13 |
+
from transformers import AutoTokenizer
|
14 |
+
|
15 |
+
from source import Source, Summary
|
16 |
+
from scrape_sources import stub as stb
|
17 |
+
|
18 |
+
|
19 |
+
|
20 |
+
@dataclass
|
21 |
+
class Digestor:
|
22 |
+
timer: Timer
|
23 |
+
cache: bool = True
|
24 |
+
text: str = field(default="no_digest")
|
25 |
+
stubs: List = field(default_factory=list)
|
26 |
+
# For clarity.
|
27 |
+
# Each stub/summary has its entities.
|
28 |
+
user_choices: List =field(default_factory=list)
|
29 |
+
# The digest text
|
30 |
+
summaries: List = field(default_factory=list)
|
31 |
+
#sources:List = field(default_factory=list) # I'm thinking create a string list for easy ref
|
32 |
+
# text:str = None
|
33 |
+
|
34 |
+
digest_meta:namedtuple(
|
35 |
+
"digestMeta",
|
36 |
+
[
|
37 |
+
'digest_time',
|
38 |
+
'number_articles',
|
39 |
+
'digest_length',
|
40 |
+
'articles_per_cluster'
|
41 |
+
]) = None
|
42 |
+
|
43 |
+
# Summarization params:
|
44 |
+
token_limit: int = 512
|
45 |
+
word_limit: int = 400
|
46 |
+
SUMMARIZATION_PARAMETERS = {
|
47 |
+
"do_sample": False,
|
48 |
+
"use_cache": cache
|
49 |
+
}
|
50 |
+
|
51 |
+
# Inference parameters
|
52 |
+
API_URL = "https://api-inference.huggingface.co/models/sshleifer/distilbart-cnn-12-6"
|
53 |
+
headers = {"Authorization": f"""Bearer {st.secrets['ato']}"""}
|
54 |
+
|
55 |
+
# I would like to keep the whole scraped text separate if I can,
|
56 |
+
# which I'm not doing here
|
57 |
+
# After this runs, the digestor is populated with s
|
58 |
+
|
59 |
+
# relevance is a matter of how many chosen clusters this article belongs to.
|
60 |
+
# max relevance is the number of unique chosen entities. min is 1.
|
61 |
+
# Allows placing articles that hit more chosen topics to go higher up,
|
62 |
+
# mirroring "upside down pyramid" journalism convention, i.e. ordering facts by decreasing information content.
|
63 |
+
def relevance(self, summary):
|
64 |
+
return len(set(self.user_choices) & set(summary.cluster_list))
|
65 |
+
|
66 |
+
def digest(self):
|
67 |
+
"""Retrieves all data for user-chosen articles, builds summary object list"""
|
68 |
+
# Clear timer from previous digestion
|
69 |
+
self.timer.timers.clear()
|
70 |
+
# Start digest timer
|
71 |
+
with Timer(name=f"digest_time", text="Total digest time: {seconds:.4f} seconds"):
|
72 |
+
# Loop through stubs, collecting data and instantiating
|
73 |
+
# and collecting Summary objects.
|
74 |
+
for stub in self.stubs:
|
75 |
+
# Check to see if we already have access to this summary:
|
76 |
+
if not isinstance(stub, stb):
|
77 |
+
self.summaries.append(stub)
|
78 |
+
else:
|
79 |
+
# if not:
|
80 |
+
summary_data: List
|
81 |
+
# Get full article data
|
82 |
+
text, summary_data = stub.source.retrieve_article(stub)
|
83 |
+
# Drop problem scrapes
|
84 |
+
# Log here
|
85 |
+
if text != None and summary_data != None:
|
86 |
+
# Start chunk timer
|
87 |
+
with Timer(name=f"{stub.hed}_chunk_time", logger=None):
|
88 |
+
chunk_list = self.chunk_piece(text, self.word_limit, stub.source.source_summarization_checkpoint)
|
89 |
+
# start totoal summarization timer. Summarization queries are timed in 'perform_summarzation()'
|
90 |
+
with Timer(name=f"{stub.hed}_summary_time", text="Whole article summarization time: {:.4f} seconds"):
|
91 |
+
summary = self.perform_summarization(
|
92 |
+
stub.hed,
|
93 |
+
chunk_list,
|
94 |
+
self.API_URL,
|
95 |
+
self.headers,
|
96 |
+
cache = self.cache,
|
97 |
+
)
|
98 |
+
# return these things and instantiate a Summary object with them,
|
99 |
+
# add that summary object to a list or somesuch collection.
|
100 |
+
# There is also timer data and data on articles
|
101 |
+
|
102 |
+
self.summaries.append(
|
103 |
+
Summary(
|
104 |
+
source=summary_data[0],
|
105 |
+
cluster_list=summary_data[1],
|
106 |
+
link_ext=summary_data[2],
|
107 |
+
hed=summary_data[3],
|
108 |
+
dek=summary_data[4],
|
109 |
+
date=summary_data[5],
|
110 |
+
authors=summary_data[6],
|
111 |
+
original_length = summary_data[7],
|
112 |
+
summary_text=summary,
|
113 |
+
summary_length=len(' '.join(summary).split(' ')),
|
114 |
+
chunk_time=self.timer.timers[f'{stub.hed}_chunk_time'],
|
115 |
+
query_time=self.timer.timers[f"{stub.hed}_query_time"],
|
116 |
+
mean_query_time=self.timer.timers.mean(f'{stub.hed}_query_time'),
|
117 |
+
summary_time=self.timer.timers[f'{stub.hed}_summary_time'],
|
118 |
+
|
119 |
+
)
|
120 |
+
)
|
121 |
+
else:
|
122 |
+
print("Null article") # looog this.
|
123 |
+
|
124 |
+
|
125 |
+
# When finished, order the summaries based on the number of user-selected clusters each article appears in.
|
126 |
+
self.summaries.sort(key=self.relevance, reverse=True)
|
127 |
+
|
128 |
+
# Query the HuggingFace Inference engine.
|
129 |
+
def query(self, payload, API_URL, headers):
|
130 |
+
"""Performs summarization inference API call."""
|
131 |
+
data = json.dumps(payload)
|
132 |
+
response = requests.request("POST", API_URL, headers=headers, data=data)
|
133 |
+
return json.loads(response.content.decode("utf-8"))
|
134 |
+
|
135 |
+
|
136 |
+
def chunk_piece(self, piece, limit, tokenizer_checkpoint, include_tail=False):
|
137 |
+
"""Breaks articles into chunks that will fit the desired token length limit"""
|
138 |
+
# Get approximate word count
|
139 |
+
words = len(piece.split(' ')) # rough estimate of words. # words <= number tokens generally.
|
140 |
+
# get number of chunks by idividing number of words by chunk size (word limit)
|
141 |
+
# Create list of ints to create rangelist from
|
142 |
+
base_range = [i*limit for i in range(words//limit+1)]
|
143 |
+
# For articles less than limit in length base_range will only contain zero.
|
144 |
+
# For most articles there is a small final chunk less than the limit.
|
145 |
+
# It may make summaries less coherent.
|
146 |
+
if include_tail or base_range == [0]:
|
147 |
+
base_range.append(base_range[-1]+words%limit) # add odd part at end of text...maybe remove.
|
148 |
+
# list of int ranges
|
149 |
+
range_list = [i for i in zip(base_range,base_range[1:])]
|
150 |
+
|
151 |
+
|
152 |
+
# Setup for chunking/checking tokenized chunk length
|
153 |
+
fractured = piece.split(' ')
|
154 |
+
tokenizer = AutoTokenizer.from_pretrained(tokenizer_checkpoint)
|
155 |
+
chunk_list = []
|
156 |
+
|
157 |
+
# Finally, chunk the piece, adjusting the chunks if too long.
|
158 |
+
for i, j in range_list:
|
159 |
+
if (tokenized_len := len(tokenizer(chunk := ' '.join(fractured[i:j]).replace('\n',' ')))) <= self.token_limit:
|
160 |
+
chunk_list.append(chunk)
|
161 |
+
else: # if chunks of <limit> words are too long, back them off.
|
162 |
+
chunk_list.append(' '.join(chunk.split(' ')[: self.token_limit - tokenized_len ]).replace('\n',' '))
|
163 |
+
|
164 |
+
return chunk_list
|
165 |
+
|
166 |
+
|
167 |
+
|
168 |
+
# Returns list of summarized chunks instead of concatenating them which loses info about the process.
|
169 |
+
def perform_summarization(self, stubhead, chunklist : List[str], API_URL: str, headers: None, cache=True) -> List[str]:
|
170 |
+
"""For each in chunk_list, appends result of query(chunk) to list collection_bin."""
|
171 |
+
collection_bin = []
|
172 |
+
repeat = 0
|
173 |
+
# loop list and pass each chunk to the summarization API, storing results.
|
174 |
+
# API CALLS: consider placing the code from query() into here. * * * *
|
175 |
+
for chunk in chunklist:
|
176 |
+
safe = False
|
177 |
+
with Timer(name=f"{stubhead}_query_time", logger=None):
|
178 |
+
while not safe and repeat < 4:
|
179 |
+
try: # make these digest params.
|
180 |
+
summarized_chunk = self.query(
|
181 |
+
{
|
182 |
+
"inputs": str(chunk),
|
183 |
+
"parameters": self.SUMMARIZATION_PARAMETERS
|
184 |
+
},
|
185 |
+
API_URL,
|
186 |
+
headers,
|
187 |
+
)[0]['summary_text']
|
188 |
+
safe = True
|
189 |
+
except Exception as e:
|
190 |
+
print("Summarization error, repeating...")
|
191 |
+
print(e)
|
192 |
+
repeat+=1
|
193 |
+
collection_bin.append(summarized_chunk)
|
194 |
+
return collection_bin
|
195 |
+
|
196 |
+
|
197 |
+
|
198 |
+
# Order for display, arrange links?
|
199 |
+
def build_digest(self) -> str:
|
200 |
+
"""Called to show the digest. Also creates data dict for digest and summaries."""
|
201 |
+
# builds summaries from pieces in each object
|
202 |
+
# orders summaries according to cluster count
|
203 |
+
# above done below not
|
204 |
+
# Manages data to be presented along with digest.
|
205 |
+
# returns all as data to display method either here or in main.
|
206 |
+
digest = []
|
207 |
+
for each in self.summaries:
|
208 |
+
digest.append(' '.join(each.summary_text))
|
209 |
+
|
210 |
+
# Create dict to write out digest data for analysis
|
211 |
+
out_data = {}
|
212 |
+
datetime_str = f"""{dt.now()}"""
|
213 |
+
choices_str = ', '.join(self.user_choices)
|
214 |
+
digest_str = '\n\n'.join(digest)
|
215 |
+
|
216 |
+
|
217 |
+
# This is a long comprehension to store all the fields and values in each summary.
|
218 |
+
# integer: {
|
219 |
+
# name_of_field:value except for source,
|
220 |
+
# which is unhashable so needs explicit handling.
|
221 |
+
# }
|
222 |
+
summaries = { # k is a summary tuple, i,p = enumerate(k)
|
223 |
+
# Here we take the first dozen words of the first summary chunk as key
|
224 |
+
c: {
|
225 |
+
# field name : value unless its the source
|
226 |
+
k._fields[i]:p if k._fields[i]!='source'
|
227 |
+
else
|
228 |
+
{
|
229 |
+
'name': k.source.source_name,
|
230 |
+
'source_url': k.source.source_url,
|
231 |
+
'Summarization" Checkpoint': k.source.source_summarization_checkpoint,
|
232 |
+
'NER Checkpoint': k.source.source_ner_checkpoint,
|
233 |
+
} for i,p in enumerate(k)
|
234 |
+
} for c,k in enumerate(self.summaries)}
|
235 |
+
|
236 |
+
out_data['timestamp'] = datetime_str
|
237 |
+
out_data['choices'] = choices_str
|
238 |
+
out_data['digest_text'] = digest_str
|
239 |
+
out_data['article_count'] = len(self.summaries)
|
240 |
+
out_data['digest_length'] = len(digest_str.split(" "))
|
241 |
+
out_data['digest_time'] = self.timer.timers['digest_time']
|
242 |
+
out_data['sum_params'] = {
|
243 |
+
'token_limit':self.token_limit,
|
244 |
+
'word_limit':self.word_limit,
|
245 |
+
'params':self.SUMMARIZATION_PARAMETERS,
|
246 |
+
}
|
247 |
+
out_data['summaries'] = summaries
|
248 |
+
|
249 |
+
|
250 |
+
self.text = digest_str
|
251 |
+
|
252 |
+
return out_data
|