I am trying to use multiprocessing to get performance boosts and increase speeds in AWS Lambda for ML inference with PyTorch, I am going of a blog post on how to effectively use multiprocessing in AWS lambda, since AWS Lambda is pretty bare-bones and has no shared memory (no /dev/shm), i have to use multiprocessing.Pipe and multiprocessing.Process - BLOG POST: Parallel processing in python with aws lambda. I have followed what the post said, and the comments of that post, but when i push the code below to AWS lambda, it deadlocks and hangs, The deadlocking happens when the program stalls and is still running but no error is returned (Lambda times out once this happens of course), and no error message is left. The strange thing is that my code runs perfectly fine on my Macbook and google colab (link below).
If anybody has any ideas why this doesn't work in AWS Lambda, please let me know. Any help would be greatly appreciated.
GOOGLE COLAB
https://colab.research.google.com/drive/1UUEC9UDeNwHmkS2MwMAm8lGa2OqnZqYl?usp=sharing
CODE
# -*- coding: utf-8 -*-
from sentence_splitter import SentenceSplitter
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM
from re import sub
from multiprocessing import Process, Pipe, set_start_method
set_start_method('fork', force=True)
model_path = "sshleifer/distilbart-cnn-12-6"
print('getting tokenizer and model for summarization...')
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForSeq2SeqLM.from_pretrained(model_path)
print('got tokenizer and model for summarization')
def serverless_pipeline():
def predict(fulltext):
# precossesing text (not important)
def preprocess_plain_text(x):
x = sub(r"https*\S+", " ", x) # url
x = sub(r"#\S+", " ", x) # mentions
x = sub(r"\s{2,}", " ", x) # over spaces
x = x.replace('–', '-')
x = x.replace('–', '-')
x = x.replace('’', "'")
x = sub("[^.,[.*?\]!$#(%?'-)}&+{=;:A-Za-z0-9]+", " ", x) # special charachters except .,!$(%?'-)&+=;:
return x
# chunking text (not important)
def chunk_clean_text(text):
newtext = preprocess_plain_text(text)
splitter = SentenceSplitter(language='en')
sentences = (splitter.split(text=newtext))
current_chunk = 0
chunks = []
for sentence in sentences:
if len(chunks) == current_chunk + 1:
if len(chunks[current_chunk]) + len(sentence.split(" ")) <= 450:
chunks[current_chunk].extend(sentence.split(" "))
else:
current_chunk += 1
chunks.append(sentence.split(" "))
else:
chunks.append(sentence.split(" "))
for chunk_id in range(len(chunks)):
chunks[chunk_id] = " ".join(chunks[chunk_id])
return chunks
clean_text = chunk_clean_text(fulltext)
def sumamrize_text(chunked_text, minimum_length, maximum_length):
print("starting actully summarizing the chunk...")
print('getting inputs in tokenizer()...')
inputs = tokenizer(str(chunked_text), return_tensors="pt")
print('got inputs in tokenizer()')
print('getting outputs in model.generate()...')
# DEADLOCKS HERE IN AWS LAMBDA
modelgenerate = model.generate(inputs["input_ids"], min_length=minimum_length, max_length=maximum_length, num_beams=4)[0]
print('got outputs in model.generate()')
chunk_summary = tokenizer.decode(modelgenerate)
print("finished actully summarizing chunk")
summarychunk = str(chunk_summary)
summarychunkend = chunk_summary[:-4]
summarychunk = summarychunkend[7:]
return str(summarychunk)
def summarize_norm(num_words, chunked_text, id, conn):
if num_words > 100:
print('starting summarize_norm')
minimum_length = num_words / 4.5
minimum_length = int(minimum_length)
maximum_length = int(num_words*100)
summarytext = sumamrize_text(chunked_text, minimum_length, maximum_length)
result = id, summarytext
print('finished making summary from summarize_norm')
print('sending result')
conn.send([result])
print('result sent')
print('clossing connection')
conn.close()
print('connection closed')
def summarize_short(summary_list, num_words, chunked_text):
if num_words <= 100 and num_words > 40:
minimum_length = num_words / 2
minimum_length = int(minimum_length)
maximum_length = int(num_words*100)
summarytext = sumamrize_text(chunked_text, minimum_length, maximum_length)
print(summarytext)
summary_list.append(str(summarytext))
def appendsummary(summary_list, num_words, chunked_text):
if num_words <= 40:
print(chunked_text)
summary_list.append(chunked_text)
def summarize():
summary_list = []
processes = []
parent_connections = []
result_list = []
for chunked_text in clean_text:
parent_conn, child_conn = Pipe(duplex=True)
parent_connections.append(parent_conn)
num_words = chunked_text.split()
num_words = len(num_words)
if num_words > 975:
return 'Make sure your prompt is properly punctuated.. And try again.'
if sub(r'[^\w\s]', '', chunked_text) == chunked_text:
return 'Make sure your prompt is properly punctuated.. And try again.'
id = clean_text.index(chunked_text)
process = Process(target=summarize_norm, args=(num_words, chunked_text, id, child_conn,))
processes.append(process)
process.start()
print('process started')
child_conn.close()
for parent_connection in parent_connections:
print('recving parent_connection results and appeding them...')
result_list.append(parent_connection.recv()[0])
print('recved parent_connection results and appeded them')
for process in processes:
print('joining proceses')
process.join()
print('joined proceses')
appendsummary(summary_list, num_words, clean_text[-1])
summarize_short(summary_list, num_words, clean_text[-1])
for summaries in sorted(result_list):
summaries = list(summaries)
summaries.remove(summaries[0])
summaries = str(summaries)
summaries = summaries[2:]
summaries = summaries[:-2]
summary_list.append(summaries)
finalsummary = ' '.join(summary_list)
return finalsummary
summary = summarize()
return summary
return predict
summarizing_pipeline = serverless_pipeline()
def handler(event, context):
# context = event['text']
print(context)
context = str(context)
summary = summarizing_pipeline(context)
return {'summary' : summary}
# Test it on random article
handler(1, '''
An article to summarize and test here
''')
Working on a project to scrape billboard top 100 over multiple weeks, look up song audio features using Spotify's API, and save the info in a new pandas df.
I got this to work for up to 100 searches at a time (the spotify api only allows 100 ids), but I am having trouble writing code for iterating through the song ids 100 at a time, running the api, and saving into a new df.
Below is the working code for 100 id searches at a time:
df_import = pd.read_csv(r'xxx/Billboard_Top_100.csv')
track_id_list = []
artist_name_list = []
track_name_list = []
for item, row in df_import.head(100).iterrows():
artist = row['Artist']
track = row['Song']
try:
spotify_response = sp.search(q='artist:' + artist + ' track:' + track, type='track')
#artist name
artist_name = spotify_response['tracks']['items'][0]['artists'][0]['name']
#song name
track_name = spotify_response['tracks']['items'][0]['name']
#unique sportify track id used for audio feautre search
track_id = spotify_response['tracks']['items'][0]['uri']
#splits string to search for features
track_id_split = str.split(track_id, 'spotify:track:')
track_id_list.append(track_id_split[1])
artist_name_list.append(row['Artist'])
track_name_list.append(row['Song'])
except:
DNF_song_search = sp.search(q=track)
artist_name = DNF_song_search['tracks']['items'][0]['artists'][0]['name']
if search(artist_name, artist):
#song name
track_name = DNF_song_search['tracks']['items'][0]['name']
#unique sportify track id used for audio feautre search
track_id = DNF_song_search['tracks']['items'][0]['uri']
#splits string to search for features
track_id_split = str.split(track_id, 'spotify:track:')
track_id_list.append(track_id_split[1])
artist_name_list.append(row['Artist'])
track_name_list.append(row['Song'])
else:
print('Inconsistent artist match on: ' + artist + ' ' + artist_name + ' for song ' + track)
#spotify api to save song features based on track ids
features = sp.audio_features(track_id_list)
#save features list into pandas df
features_df = pd.DataFrame(data = features)
#add artist and song columns from imported billboard df
features_df['Artist'] = artist_name_list
features_df['Song'] = track_name_list
#combine the two dataframes
df_merged = pd.merge(df_import, features_df, on = 'Song', how = 'left')
df_merged.to_csv('merged.csv')
I have tried saving all of the songs ids into a list, and then executing the api 100 ids at a time, but I get various errors when I try to save into a new dataframe.
solved myself
track_id_list = []
artist_name_list = []
track_name_list = []
for n in range(len(df_import) // 100):
for r in range(99):
artist = df_import.iloc[r+(n*100),3]
track = df_import.iloc[r+(n*100),4]
try:
spotify_response = sp.search(q='artist:' + artist + ' track:' + track, type='track')
artist_name = spotify_response['tracks']['items'][0]['artists'][0]['name']
track_name = spotify_response['tracks']['items'][0]['name']
#unique spotify track id used for audio feature search
track_id = spotify_response['tracks']['items'][0]['uri']
#splits string to search for features
track_id_split = str.split(track_id, 'spotify:track:')
track_id_list.append(track_id_split[1])
artist_name_list.append(artist)
track_name_list.append(track)
except:
DNF_song_search = sp.search(q=track)
artist_name = DNF_song_search['tracks']['items'][0]['artists'][0]['name']
if search(artist_name, artist):
track_name = DNF_song_search['tracks']['items'][0]['name']
track_id = DNF_song_search['tracks']['items'][0]['uri']
track_id_split = str.split(track_id, 'spotify:track:')
track_id_list.append(track_id_split[1])
artist_name_list.append(artist)
track_name_list.append(track)
else:
print('Inconsistent artist match on: ' + artist + ' ' + artist_name + ' for song ' + track)
features_df = pd.DataFrame()
for num in range(len(track_id_list) // 100 + 1):
features = sp.audio_features(track_id_list[(num*100):(num+1)*100])
features_df = features_df.append(pd.DataFrame(features))
#add artist and song columns from imported billboard df
features_df['Artist'] = artist_name_list
features_df['Song'] = track_name_list
#combine the two dataframes
df_merged = pd.merge(df_import, features_df.drop_duplicates(), on = 'Song', how = 'left')
df_merged.to_csv('mergedv2.csv')
I'm working on a Deep Learning project where I use a bidirectional attention flow model (allennlp pretrained model)to make a question answering system.It uses squad dataset.The bidaf model extracts the answer span from paragraph.Is there any way to determine the confidence score(accuracy)or any other metrics of the answer extracted by the model?
I have used the subcommand evaluate from the allennlp package but it determines only score of the model after testing.I was hoping there is a much easier way to solve the issue using other such command.
Attaching the code and the terminal output below.
from rake_nltk import Rake
from string import punctuation
from nltk.corpus import stopwords
from allennlp.predictors.predictor import Predictor
import spacy
import wikipedia
import re
import requests
from requests_html import HTMLSession
from bs4 import BeautifulSoup
import traceback
from nltk.stem import SnowballStemmer
from nltk.util import ngrams
from math import log10
from flask import Flask, request, jsonify, render_template
from gevent.pywsgi import WSGIServer
import time
import multiprocessing as mp
from gtts import gTTS
import os
NLP = spacy.load('en_core_web_md')
stop = stopwords.words('english')
symbol = r"""!#$%^&*();:\n\t\\\"!\{\}\[\]<>-\?"""
stemmer = SnowballStemmer('english')
wikipedia.set_rate_limiting(True)
session = HTMLSession()
results = 5
try:
predictor = Predictor.from_path("bidaf-model-2017.09.15-charpad.tar.gz")
except:
predictor = Predictor.from_path("https://storage.googleapis.com/allennlp-public-models/bidaf-elmo-model-2018.11.30-charpad.tar.gz")
try:
srl = Predictor.from_path('srl-model-2018.05.25.tar.gz')
except:
srl = Predictor.from_path('https://s3-us-west-2.amazonaws.com/allennlp/models/bert-base-srl-2019.06.17.tar.gz')
key = Rake(min_length=1, stopwords=stop, punctuations=punctuation, max_length=6)
wh_words = "who|what|how|where|when|why|which|whom|whose|explain".split('|')
stop.extend(wh_words)
session = HTMLSession()
output = mp.Queue()
def termFrequency(term, doc):
normalizeTermFreq = re.sub('[\[\]\{\}\(\)]', '', doc.lower()).split()
normalizeTermFreq = [stemmer.stem(i) for i in normalizeTermFreq]
dl = len(normalizeTermFreq)
normalizeTermFreq = ' '.join(normalizeTermFreq)
term_in_document = normalizeTermFreq.count(term)
#len_of_document = len(normalizeTermFreq )
#normalized_tf = term_in_document / len_of_document
normalized_tf = term_in_document
return normalized_tf, normalizeTermFreq, dl#, n_unique_term
def inverseDocumentFrequency(term, allDocs):
num_docs_with_given_term = 0
for doc in allDocs:
if term in doc:
num_docs_with_given_term += 1
if num_docs_with_given_term > 0:
total_num_docs = len(allDocs)
idf_val = log10(((total_num_docs+1) / num_docs_with_given_term))
term_split = term.split()
if len(term_split) == 3:
if len([term_split[i] for i in [0, 2] if term_split[i] not in stop]) == 2:
return idf_val*1.5
return idf_val
return idf_val
else:
return 0
def sent_formation(question, answer):
tags_doc = NLP(question)
tags_doc_cased = NLP(question.title())
tags_dict_cased = {i.lower_:i.pos_ for i in tags_doc_cased}
tags_dict = {i.lower_:i.pos_ for i in tags_doc}
question_cased = []
for i in question[:-1].split():
if tags_dict[i] == 'PROPN' or tags_dict[i] == 'NOUN':
question_cased.append(i.title())
else:
question_cased.append(i.lower())
question_cased.append('?')
question_cased = ' '.join(question_cased)
#del tags_dict,tags_doc, tags_doc_cased
pre = srl.predict(question_cased)
verbs = []
arg1 = []
for i in pre['verbs']:
verbs.append(i['verb'])
if 'B-ARG1' in i['tags']:
arg1.append((i['tags'].index('B-ARG1'), i['tags'].count('I-ARG1'))\
if not pre['words'][i['tags'].index('B-ARG1')].lower() in wh_words else \
(i['tags'].index('B-ARG2'), i['tags'].count('I-ARG2')))
arg1 = arg1[0] if arg1 else []
if not arg1:
verb_idx = pre['verbs'][0]['tags'].index('B-V')
verb = pre['words'][verb_idx] if pre['words'][verb_idx] != answer.split()[0].lower() else ''
subj_uncased = pre['words'][verb_idx+1:] if pre['words'][-1] not in symbol else \
pre['words'][verb_idx+1:-1]
else:
verb = ' '.join(verbs)
subj_uncased = pre['words'][arg1[0]:arg1[0]+arg1[1]+1]
conj = ''
if question.split()[0].lower() == 'when':
conj = ' on' if len(answer.split()) > 1 else ' in'
subj = []
for n, i in enumerate(subj_uncased):
if tags_dict_cased[i.lower()] == 'PROPN' and tags_dict[i.lower()] != 'VERB' or n == 0:
subj.append(i.title())
else:
subj.append(i.lower())
subj[0] = subj[0].title()
print(subj)
print(pre)
subj = ' '.join(subj)
sent = "{} {}{} {}.".format(subj, verb, conj, answer if answer[-1] != '.' else answer[:-1])
return sent
class extractAnswer:
def __init__(self):
self.wiki_error = (wikipedia.exceptions.DisambiguationError,
wikipedia.exceptions.HTTPTimeoutError,
wikipedia.exceptions.WikipediaException)
self.article_title = None
# symbol = """!#$%^&*();:\n\t\\\"!\{\}\[\]<>-\?"""
def extractAnswer_model(self, passage, question, s=0.4, e=0.3, wiki=False):
if type(passage) == list:
passage = " ".join(passage)
if not question[-1] == '?':
question = question+'?'
pre = predictor.predict(passage=passage, question=question)
if wiki:
if max(pre['span_end_probs']) > 0.5:
s = 0.12
elif max(pre['span_end_probs']) > 0.4:
s = 0.13
elif max(pre['span_end_probs']) > 0.35:
s = 0.14
if max(pre['span_start_probs']) > 0.5:
e = 0.12
elif max(pre['span_start_probs']) > 0.4:
e = 0.14
elif max(pre['span_start_probs']) > 0.3:
e = 0.15
if max(pre['span_start_probs']) > s and max(pre['span_end_probs']) > e:
key.extract_keywords_from_text(question)
ques_key = [stemmer.stem(i) for i in ' '.join(key.get_ranked_phrases())]
key.extract_keywords_from_text(passage)
pass_key = [stemmer.stem(i) for i in ' '.join(key.get_ranked_phrases())]
l = len(ques_key)
c = 0
for i in ques_key:
if i in pass_key:
c += 1
if c >= l/2:
print(max(pre['span_start_probs']),
max(pre['span_end_probs']))
if wiki:
return pre['best_span_str'], max(pre['span_start_probs']) + max(pre['span_end_probs'])
try:
ans = sent_formation(question, pre['best_span_str'])
except:
ans = pre['best_span_str']
print(traceback.format_exc())
return ans
print(ques_key, c, l)
print(max(pre['span_start_probs']), max(pre['span_end_probs']))
return 0, 0
else:
print(max(pre['span_start_probs']), max(pre['span_end_probs']), pre['best_span_str'])
return 0, 0
def wiki_search_api(self, query):
article_list = []
try:
article_list.extend(wikipedia.search(query, results=results))
print(article_list)
return article_list
except self.wiki_error:
params = {'search': query, 'profile': 'engine_autoselect',
'format': 'json', 'limit': results}
article_list.extend(requests.get('https://en.wikipedia.org/w/api.php?action=opensearch',
params=params).json()[1])
return article_list
except:
print('Wikipedia search error!')
print(traceback.format_exc())
return 0
def wiki_passage_api(self, article_title, article_list, output):
# Disambiguation_title = {}
try:
passage = wikipedia.summary(article_title)
output.put((article_title, self.passage_pre(passage)))
except wikipedia.exceptions.DisambiguationError as e:
print(e.options[0], e.options)
Disambiguation_pass = {}
for p in range(2 if len(e.options) > 1 else len(e.options)):
params = {'search':e.options[p], 'profile':'engine_autoselect', 'format':'json'}
article_url = requests.get('https://en.wikipedia.org/w/api.php?action=opensearch',
params=params).json()
if not article_url[3]:
continue
article_url = article_url[3][0]
r = session.get(article_url)
soup = BeautifulSoup(r.html.raw_html)
print(soup.title.string)
article_title_dis = soup.title.string.rsplit('-')[0].strip()
if article_title_dis in article_list:
print('continue')
continue
try:
url = "https://en.wikipedia.org/w/api.php?format=json&action=query&prop=extracts&exintro&explaintext&redirects=1&titles={}".format(article_title_dis)
passage = requests.get(url).json()['query']['pages']
for i in passage.keys():
if 'extract' in passage[i]:
Disambiguation_pass[article_title_dis] = self.passage_pre(passage[i]['extract'])
except wikipedia.exceptions.HTTPTimeoutError:
passage = wikipedia.summary(article_title_dis)
Disambiguation_pass[article_title_dis] = self.passage_pre(passage)
except:
Disambiguation_pass[article_title_dis] = ''
continue
output.put((article_title, Disambiguation_pass))
except:
output.put((article_title, ''))
print(traceback.format_exc())
def sorting(self, article, question, topic):
processes = [mp.Process(target=self.wiki_passage_api, args=(article[x], article, output))\
for x in range(len(article))]
for p in processes:
p.start()
for p in processes:
p.join(timeout=3)
results_p = [output.get() for p in processes]
article_list = []
passage_list = []
for i, j in results_p:
if type(j) != dict and j:
article_list.append(i)
passage_list.append(j)
elif type(j) == dict and j:
for k, l in j.items():
if l:
article_list.append(k)
passage_list.append(l)
normalize_passage_list = []
start = time.time()
keywords = " ".join(self.noun+self.ques_key+[topic.lower()])
keywords = re.sub('[{0}]'.format(symbol), ' ', keywords).split()
question = question+' '+topic
ques_tokens = [stemmer.stem(i.lower()) for i in question.split() \
if i.lower() not in wh_words]
print(ques_tokens)
keywords_bigram = [' '.join(i) for i in list(ngrams(ques_tokens, 2)) \
if i[0] not in stop and i[1] not in stop]
if len(ques_tokens) > 3:
keywords_trigram = [' '.join(i) for i in list(ngrams(ques_tokens, 3)) \
if (i[0] in stop) + (i[2] in stop) + (i[1] in stop) < 3]
else:
keywords_trigram = []
if len(ques_tokens) > 5:
keywords_4gram = [' '.join(i) for i in list(ngrams(ques_tokens, 4)) \
if (i[0] in stop) + (i[2] in stop) +(i[1] in stop)+(i[3] in stop) < 4]
else:
keywords_4gram = []
keywords_unigram = list(set([stemmer.stem(i.lower()) for i in keywords \
if i.lower() not in stop]))
keywords = keywords_unigram+list(set(keywords_bigram))+keywords_trigram+keywords_4gram
tf = []
if not passage_list:
return 0
pass_len = []
#n_u_t=[]
#key_dict = {i: keywords.count(i) for i in keywords}
print('Extraction complete')
#remove_pass={}
#for n,i in enumerate(passage_list):
#if len(i)<200 or not i:
#remove_pass[article_list[n]]=i
#print(n, article_list[n])
#passage_list=[i for i in passage_list if i not in remove_pass.values()]
#article_list=[i for i in article_list if i not in remove_pass.keys()]
passage_list_copy = passage_list.copy()
article_list_copy = article_list.copy()
for i in range(len(passage_list_copy)):
if passage_list.count(passage_list_copy[i]) > 1:
passage_list.remove(passage_list_copy[i])
article_list.remove(article_list_copy[i])
print('Copy:', article_list_copy[i])
del passage_list_copy
del article_list_copy
for n, i in enumerate(passage_list):
temp_tf = {}
c = 0
for j in keywords:
temp_tf[j], temp_pass, temp_len = termFrequency(j, i + ' ' + article_list[n])
if temp_tf[j]:
c += 1
normalize_passage_list.append(temp_pass)
pass_len.append(temp_len)
temp_tf['key_match'] = c
tf.append(temp_tf)
print(pass_len)
print(keywords)
idf = {}
for i in keywords:
idf[i] = inverseDocumentFrequency(i, normalize_passage_list)
#print(tf, idf)
tfidf = []
#b=0.333 #for PLN
b, k = 0.75, 1.2 #for BM25
avg_pass_len = sum(pass_len)/len(pass_len)
#pivot=sum(n_u_t)/len(n_u_t)
for n, i in enumerate(tf):
tf_idf = 0
#avg_tf=sum(i.values())/len(i)
key_match_ratio = i['key_match']/len(keywords)
for j in keywords:
#tf_idf+=idf[j]*((log(1+log(1+i[j])))/(1-b+(b*pass_len[n]/avg_pass_len))) #PLN
tf_idf += idf[j]*(((k+1)*i[j])/(i[j]+k*(1-b+(b*pass_len[n]/avg_pass_len)))) #BM25
tfidf.append(tf_idf*key_match_ratio)
tfidf = [i/sum(tfidf)*100 for i in tfidf if any(tfidf)]
if not tfidf:
return 0, 0, 0, 0, 0
print(tfidf)
print(article_list, len(passage_list))
if len(passage_list) > 1:
sorted_tfidf = sorted(tfidf, reverse=1)
idx1 = tfidf.index(sorted_tfidf[0])
passage1 = passage_list[idx1]
#article_title=
tfidf1 = sorted_tfidf[0]
idx2 = tfidf.index(sorted_tfidf[1])
passage2 = passage_list[idx2]
article_title = (article_list[idx1], article_list[idx2])
tfidf2 = sorted_tfidf[1]
else:
article_title = 0
tfidf2 = 0
if passage_list:
passage1 = passage_list[0]
tfidf1 = tfidf[0]
passage2 = 0
else:
passage1 = 0
passage2 = 0
tfidf1, tfidf2 = 0, 0
end = time.time()
print('TFIDF time:', end-start)
return passage1, passage2, article_title, tfidf1, tfidf2
def passage_pre(self, passage):
#passage=re.findall("[\da-zA-z\.\,\'\-\/\–\(\)]*", passage)
passage = re.sub('\n', ' ', passage)
passage = re.sub('\[[^\]]+\]', '', passage)
passage = re.sub('pronunciation', '', passage)
passage = re.sub('\\\\.+\\\\', '', passage)
passage = re.sub('{.+}', '', passage)
passage = re.sub(' +', ' ', passage)
return passage
def wiki(self, question, topic=''):
if not question:
return 0
question = re.sub(' +', ' ', question)
question = question.title()
key.extract_keywords_from_text(question)
self.ques_key = key.get_ranked_phrases()
doc = NLP(question)
self.noun = [str(i).lower() for i in doc.noun_chunks if str(i).lower() not in wh_words]
print(self.ques_key, self.noun)
question = re.sub('[{0}]'.format(symbol), ' ', question)
if not self.noun + self.ques_key:
return 0
article_list = None
question = question.lower()
if self.noun:
if len(self.noun) == 2 and len(" ".join(self.noun).split()) < 6:
#question1=question
self.noun = " ".join(self.noun).split()
if self.noun[0] in stop:
self.noun.pop(0)
self.noun = question[question.index(self.noun[0]):question.index(self.noun[-1]) \
+len(self.noun[-1])+1].split()
#del question1
print(self.noun)
article_list = self.wiki_search_api(' '.join(self.noun))
if self.ques_key and not article_list:
article_list = self.wiki_search_api(self.ques_key[0])
if not article_list:
article_list = self.wiki_search_api(' '.join(self.ques_key))
if not article_list:
print('Article not found on wikipedia.')
return 0, 0
article_list = list(set(article_list))
passage1, passage2, article_title, tfidf1, tfidf2 = self.sorting(article_list,
question, topic)
if passage1:
ans1, conf1 = self.extractAnswer_model(passage1, question, s=0.20, e=0.20, wiki=True)
else:
ans1, conf1 = 0, 0
if ans1:
conf2 = 0
if len(ans1) > 600:
print(ans1)
print('Repeat')
ans1, conf1 = self.extractAnswer_model(ans1, question, s=0.20, e=0.20, wiki=True)
threshhold = 0.3 if not ((tfidf1- tfidf2) <= 10) else 0.2
if round(tfidf1- tfidf2) < 5:
threshhold = 0
if (tfidf1- tfidf2) > 20:
threshhold = 0.35
if (tfidf1- tfidf2) > 50:
threshhold = 1
if (passage2 and conf1 < 1.5) or (tfidf1 - tfidf2) < 10:
ans2, conf2 = self.extractAnswer_model(passage2, question, s=0.20, e=0.20,
wiki=True) if passage2 else (0, 0)
title = 0
if round(conf1, 2) > round(conf2, 2) - threshhold:
print('ans1')
ans = ans1
title = article_title[0] if article_title else 0
else:
print('ans2')
title = article_title[1] if article_title else 0
ans = ans2
if not question[-1] == '?':
question = question+'?'
try:
ans = sent_formation(question, ans)
except:
print(traceback.format_exc())
print(ans, '\n', '\n', article_title)
return ans, title
extractor = extractAnswer()
app = Flask(__name__)
#app.route("/", methods=["POST", "get"])
#app.route("/ans")
def ans():
start = time.time()
question = request.args.get('question')
topic = request.args.get('topic')
passage = request.args.get('passage')
if not question:
return render_template('p.html')
if not topic:
topic = ''
if passage:
answer = extractor.extractAnswer_model(passage, question)
else:
answer, title = extractor.wiki(question, topic)
end = time.time()
if answer:
mytext = str(answer)
language = 'en'
myobj = gTTS(text=mytext, lang=language, slow=False)
myobj.save("welcome.mp3")
# prevName = 'welcome.mp3'
#newName = 'static/welcome.mp3'
#os.rename(prevName,newName)
return render_template('pro.html', answer=answer)
else:
return jsonify(Status='E', Answer=answer, Time=end-start)
#app.route("/audio_del/", methods=["POST", "get"])
def audio_del():
return render_template('p.html');
#app.route("/audio_play/", methods=["POST", "get"])
def audio_play():
os.system("mpg321 welcome.mp3")
return render_template('white.html')
if __name__ == "__main__":
PORT = 7091
HTTP_SERVER = WSGIServer(('0.0.0.0', PORT), app)
print('Running on',PORT, '...')
HTTP_SERVER.serve_forever()
I came across a possible solution to this after deeply looking into the output returned by the model. Although this, is probably not something you can accurately rely on, it seemed to have done the task in my case:
Note that the text answer which is "best_span_str" is always a subarray of the passage. It spans the range which is stored in "best_span".
i.e., "best_span" contains the start and end index of the answer.
Now, the output data contains a property named "span_end_probs".
"span_end_probs" contains a list of values that correspond to all the words present in the text input.
If you look closely for various inputs, the value is always maximum at one of the indexes within the starting and ending range that "best_span" contains. This value seemed to be very similar to the confidence levels that we need. Let's call this value score. All you need to do now is to try some inputs and find a suitable method to use this score as a metric.
e.g.: if you need a threshold value for some application, you can try a number of test inputs and find a value that is most accurate. In my case, this was around 0.35.
i.e. if score is lesser than 0.35, it prints answer not found and if greater than or equal 0.35, prints string in "best_span_str".
Here's my code snippet:
from allennlp.predictors.predictor import Predictor
passage = '--INPUT PASSAGE--'
question = '--INPUT QUESTION--'
predictor = Predictor.from_path("https://storage.googleapis.com/allennlp-public-models/bidaf-elmo.2021-02-11.tar.gz")
output = predictor.predict(
passage = passage,
question = question
)
score = max(output["span_end_probs"])
if score < 0.35:
print('Answer not found')
else:
print(output["best_span_str"])
You can readily see the example input and output here.
I'm working on a Sentiment Analysis project using Twitter Data, and I've encountered a small problem regarding Dates. The code itself runs fine, but I don't know how to build custom time blocks for grouping my final data. Right now, it is defaulting to grouping them by the second, which is not very useful. I want to be able to group them in half-hour, hour, and day segments...
Feel free to skip to the bottom of the code to see where the issue lies!
Here is the code:
import tweepy
API_KEY = "XXXXX"
API_SECRET = XXXXXX"
auth = tweepy.AppAuthHandler(API_KEY, API_SECRET)
api = tweepy.API(auth, wait_on_rate_limit = True, wait_on_rate_limit_notify = True)
import sklearn as sk
import pandas as pd
import got3
#"Get Old Tweets" to find older data
tweetCriteria = got3.manager.TweetCriteria()
tweetCriteria.setQuerySearch("Kentucky Derby")
tweetCriteria.setSince("2016-05-07")
tweetCriteria.setUntil("2016-05-08")
tweetCriteria.setMaxTweets(1000)
TweetCriteria = got3.manager.TweetCriteria()
KYDerby_tweets = got3.manager.TweetManager.getTweets(tweetCriteria)
from afinn import Afinn
afinn = Afinn()
#getting afinn library to use for sentiment polarity analysis
for x in KYDerby_tweets:
Text = x.text
Retweets = x.retweets
Favorites = x.favorites
Date = x.date
Id = x.id
print(Text)
AllText = []
AllRetweets = []
AllFavorites = []
AllDates = []
AllIDs = []
for x in KYDerby_tweets:
Text = x.text
Retweets = x.retweets
Favorites = x.favorites
Date = x.date
AllText.append(Text)
AllRetweets.append(Retweets)
AllFavorites.append(Favorites)
AllDates.append(Date)
AllIDs.append(Id)
data_set = [[x.id, x.date, x.text, x.retweets, x.favorites]
for x in KYDerby_tweets]
df = pd.DataFrame(data=data_set, columns=["Id", "Date", "Text", "Favorites", "Retweets"])
#I now have a DataFrame with my basic info in it
pscore = []
for x in KYDerby_tweets:
afinn.score(x.text)
pscore.append(afinn.score(x.text))
df['P Score'] = pscore
#I now have the pscores for each Tweet in the DataFrame
nrc = pd.read_csv('C:\\users\\andrew.smith\\downloads\\NRC-emotion-lexicon-wordlevel-alphabetized-v0.92.txt', sep="\t", names=["word", "emotion", "association"], skiprows=45)
#import NRC emotion lexicon
nrc = nrc[nrc["association"]==1]
nrc = nrc[nrc["emotion"].isin(["positive", "negative"]) == False]
#cleaned it up a bit
from nltk import TweetTokenizer
tt = TweetTokenizer()
tokenized = [x.lower() for x in tokenized]
#built my Tweet-specific, NRC-ready tokenizer
emotions = list(set(nrc["emotion"]))
index2emotion = {}
emotion2index = {}
for i in range(len(emotions)):
index2emotion[i] = emotions[i]
emotion2index[emotions[i]] = i
cv = [0] * len(emotions)
#built indices showing locations of emotions
for token in tokenized:
sub = nrc[nrc['word'] == token]
token_emotions = sub['emotion']
for e in token_emotions:
position_index = emotion2index[e]
cv[position_index]+=1
emotions = list(set(nrc['emotion']))
index2emotion = {}
emotion2index = {}
for i in range(len(emotions)):
index2emotion[i] = emotions[i]
emotion2index[emotions[i]] = i
def makeEmoVector(tweettext):
cv = [0] * len(emotions)
tokenized = tt.tokenize(tweettext)
tokenized = [x.lower() for x in tokenized]
for token in tokenized:
sub = nrc[nrc['word'] == token]
token_emotions = sub['emotion']
for e in token_emotions:
position_index = emotion2index[e]
cv[position_index] += 1
return cv
tweettext = df.iloc[14,:]['Text']
emotion_vectors = []
for text in df['Text']:
emotion_vector = makeEmoVector(text)
emotion_vectors.append(emotion_vector)
ev = pd.DataFrame(emotion_vectors, index=df.index, columns=emotions)
#Now I have a DataFrame with all of the emotion counts for each tweet
Date_Group = df.groupby("Date")
Date_Group[emotions].agg("sum")
#Finally, we arrive at the problem! When I run this, I end up with tweets that are grouped *by the second. What I want is to be able to group them: a) by the half-hour, b) by the hour, and c) by the day
Since, the default date format for tweets with the Tweepy API is "2017-04-14 18:41:56". To get tweets grouped by hour, you can do something as simple as this:
# This will get the time parameter
time = [item.split(" ")[1] for item in df['date'].values]
# This will get the hour parameter
hour = [item.split(":")[0] for item in time]
df['time'] = hour
grouped_tweets = df[['time', 'number_tweets']].groupby('time')
tweet_growth_hour = grouped_tweets.sum()
tweet_growth_hour['time']= tweet_growth_hour.index
print tweet_growth_hour
To group by date, you can do something similiar like:
days = [item.split(" ")[0] for item in df['date'].values]
df['days'] = days
grouped_tweets = df[['days', 'number_tweets']].groupby('days')
tweet_growth_days = grouped_tweets.sum()
tweet_growth_days['days']= tweet_growth_days.index
print tweet_growth_days
Sorry for the unsophisticated question title but I need help desperately:
My objective at work is to create a script that pulls all the records from exacttarget salesforce marketing cloud API. I have successfully setup the API calls, and successfully imported the data into DataFrames.
The problem I am running into is two-fold that I need to keep pulling records till "Results_Message" in my code stops reading "MoreDataAvailable" and I need to setup logic which allows me to control the date from either within the API call or from parsing the DataFrame.
My code is getting stuck at line 44 where "print Results_Message" is looping around the string "MoreDataAvailable"
Here is my code so far, on lines 94 and 95 you will see my attempt at parsing the date directly from the dataframe but no luck and no luck on line 32 where I have specified the date:
import ET_Client
import pandas as pd
AggreateDF = pd.DataFrame()
Data_Aggregator = pd.DataFrame()
#Start_Date = "2016-02-20"
#End_Date = "2016-02-25"
#retrieveDate = '2016-07-25T13:00:00.000'
Export_Dir = 'C:/temp/'
try:
debug = False
stubObj = ET_Client.ET_Client(False, debug)
print '>>>BounceEvents'
getBounceEvent = ET_Client.ET_BounceEvent()
getBounceEvent.auth_stub = stubObj
getBounceEvent.search_filter = {'Property' : 'EventDate','SimpleOperator' : 'greaterThan','Value' : '2016-02-22T13:00:00.000'}
getResponse1 = getBounceEvent.get()
ResponseResultsBounces = getResponse1.results
Results_Message = getResponse1.message
print(Results_Message)
#EventDate = "2016-05-09"
print "This is orginial " + str(Results_Message)
#print ResponseResultsBounces
i = 1
while (Results_Message == 'MoreDataAvailable'):
#if i > 5: break
print Results_Message
results1 = getResponse1.results
#print(results1)
i = i + 1
ClientIDBounces = []
partner_keys1 = []
created_dates1 = []
modified_date1 = []
ID1 = []
ObjectID1 = []
SendID1 = []
SubscriberKey1 = []
EventDate1 = []
EventType1 = []
TriggeredSendDefinitionObjectID1 = []
BatchID1 = []
SMTPCode = []
BounceCategory = []
SMTPReason = []
BounceType = []
for BounceEvent in ResponseResultsBounces:
ClientIDBounces.append(str(BounceEvent['Client']['ID']))
partner_keys1.append(BounceEvent['PartnerKey'])
created_dates1.append(BounceEvent['CreatedDate'])
modified_date1.append(BounceEvent['ModifiedDate'])
ID1.append(BounceEvent['ID'])
ObjectID1.append(BounceEvent['ObjectID'])
SendID1.append(BounceEvent['SendID'])
SubscriberKey1.append(BounceEvent['SubscriberKey'])
EventDate1.append(BounceEvent['EventDate'])
EventType1.append(BounceEvent['EventType'])
TriggeredSendDefinitionObjectID1.append(BounceEvent['TriggeredSendDefinitionObjectID'])
BatchID1.append(BounceEvent['BatchID'])
SMTPCode.append(BounceEvent['SMTPCode'])
BounceCategory.append(BounceEvent['BounceCategory'])
SMTPReason.append(BounceEvent['SMTPReason'])
BounceType.append(BounceEvent['BounceType'])
df1 = pd.DataFrame({'ClientID': ClientIDBounces, 'PartnerKey': partner_keys1,
'CreatedDate' : created_dates1, 'ModifiedDate': modified_date1,
'ID':ID1, 'ObjectID': ObjectID1,'SendID':SendID1,'SubscriberKey':SubscriberKey1,
'EventDate':EventDate1,'EventType':EventType1,'TriggeredSendDefinitionObjectID':TriggeredSendDefinitionObjectID1,
'BatchID':BatchID1,'SMTPCode':SMTPCode,'BounceCategory':BounceCategory,'SMTPReason':SMTPReason,'BounceType':BounceType})
#print df1
#df1 = df1[(df1.EventDate > "2016-02-20") & (df1.EventDate < "2016-02-25")]
#AggreateDF = AggreateDF[(AggreateDF.EventDate > Start_Date) and (AggreateDF.EventDate < End_Date)]
print(df1['ID'].max())
AggreateDF = AggreateDF.append(df1)
print(AggreateDF.shape)
#df1 = df1[(df1.EventDate > "2016-02-20") and (df1.EventDate < "2016-03-25")]
#AggreateDF = AggreateDF[(AggreateDF.EventDate > Start_Date) and (AggreateDF.EventDate < End_Date)]
print("Final Aggregate DF is: " + str(AggreateDF.shape))
#EXPORT TO CSV
AggreateDF.to_csv(Export_Dir +'DataTest1.csv')
#with pd.option_context('display.max_rows',10000):
#print (df_masked1.shape)
#print df_masked1
except Exception as e:
print 'Caught exception: ' + str(e.message)
print e
Before my code parses the data, the orginal format I get of the data is a SOAP response, this is what it look like(below). Is it possible to directly parse records based on EventDate from the SOAP response?
}, (BounceEvent){
Client =
(ClientID){
ID = 1111111
}
PartnerKey = None
CreatedDate = 2016-05-12 07:32:20.000937
ModifiedDate = 2016-05-12 07:32:20.000937
ID = 1111111
ObjectID = "1111111"
SendID = 1111111
SubscriberKey = "aaa#aaaa.com"
EventDate = 2016-05-12 07:32:20.000937
EventType = "HardBounce"
TriggeredSendDefinitionObjectID = "aa111aaa"
BatchID = 1111111
SMTPCode = "1111111"
BounceCategory = "Hard bounce - User Unknown"
SMTPReason = "aaaa"
BounceType = "immediate"
Hope this makes sense, this is my desperately plea for help.
Thank you in advance!
You don't seem to be updating Results_Message in your loop, so it's always going to have the value it gets in line 29: Results_Message = getResponse1.message. Unless there's code involved that you didn't share, that is.