-
Notifications
You must be signed in to change notification settings - Fork 0
/
seq2seq_model.py
421 lines (368 loc) · 18 KB
/
seq2seq_model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from pygments.lexer import words
from setuptools.command.dist_info import dist_info
from spacy.syntax.nn_parser import precompute_hiddens
from torch.autograd import Variable
import pandas as pd
import random
import itertools
import nltk
from nltk.corpus import stopwords
import random
from collections import Counter
import string
import re
from Early_Stopping import EarlyStopping
import itertools
from operator import itemgetter
import spacy
import pandas as pd
from gensim.corpora import Dictionary
import textcleaner as tc
from docutils.nodes import section
from gensim.models import Word2Vec
from nltk.tokenize import sent_tokenize,word_tokenize
import string
import torch.nn.functional as F
import pickle
from spacy.language import Language
import EncoderRNN as en
import DecoderRNN as dc
from torch import *
from spacy.vocab import Vocab
from spacy.lang.en import English
import spacy
import os
import pandas as pd
import gensim
from nltk.corpus import subjectivity
#pd.options.display.max_colwidth = 100
#os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
device = torch.device("cuda" if torch.cuda.is_available() else"cpu")
nlp = spacy.load('en_core_web_sm')
# # Load pre-trained Word2Vec model.
# Word2vec_model = gensim.models.Word2Vec.load("models/Word2vec/Word2Vec.model")
# word2vec_dict = Dictionary.load('models/Word2vec/Word2Vec_dict.pkl')
# word2vec_dict.add_documents([['EOS','SOS']])
stop_words = set(stopwords.words('english'))
custom_stop_words = {'please'}
final_stop_words = stop_words.union(custom_stop_words)
add_tok = {'UNK':1, 'PAD':0, 'SOS':2,'EOS':3}
pretrained=False
load_representation = False
def create_dictionary(sentences):
translator = str.maketrans('', '', string.punctuation)
full_sentense=" ".join(sentences).translate(translator).lower().split()
word_list = [word for word in set(full_sentense) if word not in final_stop_words]
word_dict = {w : i for i, w in enumerate(word_list)}
final_dict = {**word_dict,**add_tok}
#Save Dict
pickle_out = open("models/Seq2Seq_LSTM/dict.pickle","wb")
pickle.dump(final_dict, pickle_out)
pickle_out.close()
def text_preprocessing(sentences:[str]):
lema_text = new_lemmatization(sentences)
input_text = list(tc.document(lema_text).remove_numbers().remove_stpwrds().remove_symbols().lower_all())
return input_text
def lemmatization(sentences:[str], allowed_postags=['NOUN', 'ADJ', 'VERB', 'ADV']):
"""https://spacy.io/api/annotation"""
texts_out = []
for sent in sentences:
doc = nlp(sent.lower())
texts_out.append(" ".join([token.lemma_ if token.lemma_ not in ['-PRON-'] else '' for token in doc if token.pos_ in allowed_postags]))
return texts_out
def new_lemmatization(sentences:[str],allowed_postags=['NN','NNS', 'NNP','NNPS','RB','RBR','RBS','VB','VBD','VBG','VBN','VBP','VBZ','JJ','JJR','JJS']):
texts_out = []
for sent in sentences:
tokens = word_tokenize(sent)
tagged = nltk.pos_tag(tokens)
pos_cleaned_sent = " ".join([token for (token, pos) in tagged if pos in allowed_postags])
doc = nlp(pos_cleaned_sent)
# Extract the lemma for each token and join
texts_out.append(" ".join([token.lemma_ for token in doc]))
return texts_out
def remove_null_sentence(sentences:[str]):
return [x for x in sentences if x is not '']
def word_tokenizer(sentences:[str]):
sentence=[]
for raw_sentence in sentences:
# If a sentence is empty, skip it
if len(raw_sentence) > 0:
# Otherwise, get a list of words
sentence.append(word_tokenize(raw_sentence))
return sentence
# Return the list of sentences
# so this returns a list of lists
def load_dictionary(path):
final_dict= pickle.load(open(path, "rb"))
number_dict = {i : w for w, i in final_dict.items()}
return final_dict,number_dict
def getCleanText(clean_text):
clean_text_result = re.sub(r"[{}–!='™‘’‘|?,-:@#%&$/1234567890()]+/*", " ", clean_text)
return clean_text_result
def make_batch(sentences,dictionary):
input_batch = []
target_batch = []
pairs=[]
for index,sen in enumerate(sentences):
s=[]
#clean_sen=lemmatization(sen).strip()
#sen= 'SOS '+clean_sen+' EOS'
word = sen.split()
#process input sentence
# for n in word:
# if n not in list(dictionary.keys()):
# continue
# else:
# s.append(dictionary[n])
s=dictionary.doc2idx(word,unknown_word_index=random.randint(1,len(dictionary.token2id)))
src = torch.tensor(s[1:-1],dtype=torch.long).to(device)
trg= torch.tensor(s[2:],dtype=torch.long).to(device)
# if src.shape[0] == 0:
# print(index)
input_batch.append(src)
target_batch.append(trg)
pairs.append((src,trg))
return input_batch, target_batch,pairs
# torch.save(input_batch,'input.pth')
# torch.save(target_batch,'target.pth')
# input_tensor = torch.load('input.pth')
# target_tensor = torch.loa bd('target.pth')
teacher_forcing_ratio = 0.5
def train(input_tensor, target_tensor, encoder, decoder, encoder_optimizer, decoder_optimizer, criterion, max_length=30):
encoder.train()
decoder.train()
#encoder_hidden = encoder.initHidden()
encoder_optimizer.zero_grad()
decoder_optimizer.zero_grad()
input_length = input_tensor.size(0)
target_length = target_tensor.size(0)
#encoder_outputs = torch.zeros(max_length, encoder.hidden_size, device=device)
loss = 0
seq_length = input_length if input_length<max_length else max_length
for ei in range(seq_length):
input_embedding = Word2vec_model.wv.word_vec(number_dict[input_tensor[ei].item()])
encoder_output, encoder_hidden = encoder(torch.from_numpy(input_embedding).float().to(device)) #encoder_hidden)
#encoder_outputs[ei] = encoder_output[0, 0]
# if target_length < 2:
# decoder_input = torch.tensor([[add_tok['SOS']]], device=device)
# else:
decoder_hidden = encoder_hidden
use_teacher_forcing = True if random.random() < teacher_forcing_ratio else False
if use_teacher_forcing:
# Teacher forcing: Feed the target as the next input
for di in range(target_length):
decoder_input = Word2vec_model.wv.word_vec(number_dict[input_tensor[di].item()])
decoder_output, decoder_hidden = decoder(torch.from_numpy(decoder_input).float().to(device), decoder_hidden)
loss += criterion(decoder_output, target_tensor[di].unsqueeze(0))
# Teacher forcing
else:
# Without teacher forcing: use its own predictions as the next input
decoder_input = Word2vec_model.wv.word_vec(number_dict[input_tensor[0].item()])
for di in range(target_length):
decoder_output, decoder_hidden= decoder(torch.from_numpy(decoder_input).float().to(device), decoder_hidden)
topv, topi = decoder_output.topk(1)
decoder_input_token = topi.squeeze().detach() # detach from history as input
decoder_input = Word2vec_model.wv.word_vec(number_dict[decoder_input_token.item()])
loss += criterion(decoder_output, target_tensor[di].unsqueeze(0))
loss.backward()
encoder_optimizer.step()
decoder_optimizer.step()
return loss.item() / target_length
def trainIters(encoder, decoder,pairs,encoder_optimizer,decoder_optimizer,criterion, print_every=1000,epoch_id=1):
print_loss_total = 0 # Reset every print_every
training_pairs = pairs
for iter in range(len(training_pairs)):
pair = training_pairs[iter]
input_tensor = pair[0]
target_tensor = pair[1]
if len(input_tensor)>0:
loss = train(input_tensor, target_tensor, encoder,decoder, encoder_optimizer, decoder_optimizer, criterion)
print_loss_total += loss
if iter % print_every == 0:
print_loss_avg = print_loss_total / print_every
print_loss_total = 0
print('Epoch : %d || Iteration : %d || completeness : %d%% || Training_Loss : %.4f' % (epoch_id,iter, iter / len(training_pairs)* 100, print_loss_avg))
print("Loss during {} epoch : {}".format(epoch_id,print_loss_total/len(training_pairs)))
return print_loss_total/len(training_pairs)
def evaluate(encoder, sentence):
encoder.eval()
with torch.no_grad():
input_tensor=sentence
input_length = input_tensor.shape[0]
#encoder_hidden = encoder.initHidden()
#encoder_outputs = torch.zeros(max_length, encoder.hidden_size, device=device)
encoder_outputs=[]
for ei in range(input_length):
input_embedding = Word2vec_model.wv.word_vec(number_dict[input_tensor[ei].item()])
encoder_output, encoder_hidden = encoder(torch.from_numpy(input_embedding).float().to(device))#,encoder_hidden)
#encoder_output, encoder_hidden = encoder(input_tensor[ei],encoder_hidden)
encoder_outputs.append(encoder_output[0, 0])
last_hidden = encoder_hidden
#decoder_input = torch.tensor([[add_tok['SOS']]], device=device) # SOS
# decoded_words = []
#
# for di in range(max_length):
# decoder_output, decoder_hidden= decoder(decoder_input, decoder_hidden)
# topv, topi = decoder_output.data.topk(1)
# if topi.item() == add_tok['EOS']:
# decoded_words.append('EOS')
# break
# else:
# decoded_words.append(number_dict[topi.item()])
# decoder_input = topi.squeeze().detach()
return last_hidden,torch.stack(encoder_outputs,dim=0)
def evaluateRandomly(encoder,sentences,dictionary,method=None):
input_batch,_,_= make_batch(sentences,dictionary)
temp=[]
if method=='middle_last':
for i in input_batch:
if i.size(0) == 0:
print("=============*****************================********************=================")
representation = torch.zeros(400).to(device)
print("=============*****************================********************=================")
elif i.size(0) < 5 and i.size(0)>0:
print('>', i)
last_hidden, embeddings = evaluate(encoder, i)
zero_vec = torch.zeros(200).to(device)
print(embeddings.shape)
representation = torch.cat((zero_vec,last_hidden.view(-1))).flatten().to(device)
print('======================================================================================================')
else:
print('>',i)
last_hidden,embeddings= evaluate(encoder, i)
print(embeddings.shape)
print('')
indices = torch.tensor([int(embeddings.size(0) / 2)-1, int(embeddings.size(0)) - 1]).to(device)
representation=torch.index_select(embeddings, 0, indices).flatten().to(device)
print('======================================================================================================')
temp.append(representation)
else:
for i in input_batch:
if i.size(0) == 0:
representation = torch.zeros(200).to(device)
else:
last_hidden, embeddings = evaluate(encoder, i)
representation = last_hidden.view(-1)
temp.append(representation)
return temp
def save_checkpoint(Path,encoder,encoder_optimizer,epoch_id,train_loss):
'''Saves model when validation loss decrease.'''
SAVE_DIR = Path
# #MODEL_SAVE_PATH = os.path.join(SAVE_DIR, 'model.pt')
torch.save({
'epoch': epoch_id,
'state_dict': encoder.state_dict(),
'optimizer': encoder_optimizer.state_dict(),
'train_loss': train_loss,
}, os.path.join(SAVE_DIR ,'train_loss-{:.4f}_epoch-{}_checkpoint.pth.tar'.format(train_loss,epoch_id)))
print("model saved at : {}".format(SAVE_DIR,':train_loss-{:.4f}_epoch-{}_checkpoint.pth.tar'.format(train_loss,epoch_id)))
def load_checkpoint(checkpoint_path, model, optimizer):
""" loads state into model and optimizer and returns:
epoch, model, optimizer
"""
#model_path = 'models/seq2seq/without_batchnorm'
if os.path.isfile(checkpoint_path):
print("=> loading checkpoint '{}'".format(checkpoint_path))
checkpoint = torch.load(checkpoint_path)
epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(epoch, checkpoint['epoch']))
return epoch,model, optimizer
else:
print("=> no checkpoint found at '{}'".format(checkpoint_path))
def testing(sentense,representation,dictionary,encoder,df,processed_sentences,k=15):
sentences=[]
sentences.append(sentense)
preprocessed_sentences = remove_null_sentence(text_preprocessing(sentences=sentences))
print(preprocessed_sentences)
input_representation = evaluateRandomly(encoder=encoder,sentences=preprocessed_sentences,dictionary=dictionary)
cosine_distance = nn.CosineSimilarity(dim=0, eps=1e-6)
distance=[]
for i in representation:
output = cosine_distance(input_representation[0], i)
distance.append(output)
distance = np.array(distance)
index = distance.argsort()[-k:][::-1]
for i in index:
print(i,"------->",df.iloc[i,:].values[0],"------->",processed_sentences[i],"------->",distance[i].item())
#return df.iloc[index, :]
#
# def mixup_bert_word2vec_embedding(bert_file_path,word2vec_file_path):
#
# with open(file=bert_file_path, mode='rb') as handle:
# bert_embeddings = pickle.load(handle)
# with open(file=word2vec_file_path, mode='rb') as handle:
# word2vec_embeddings = pickle.load(handle)
# print(type(bert_embeddings),type(word2vec_embeddings))
# return torch.from_numpy(np.concatenate((bert_embeddings,word2vec_embedding),axis=1)).to(device)
df = pd.read_excel('Tickets.xlsx').values.tolist()
sentences = list(itertools.chain.from_iterable(df))
preprocessed_sentences = remove_null_sentence(text_preprocessing(sentences=sentences))
input_sentence = word_tokenizer(preprocessed_sentences)
# dictionary = Dictionary(input_sentence)
# dictionary.save("models/Word2vec/word2vec_dict_with_new_lemma.pkl")
dictionary = Dictionary.load("models/Word2vec/word2vec_dict_with_new_lemma.pkl")
#create_dictionary(sentences)
dictionary[0]
final_dict, number_dict = dictionary.token2id,dictionary.id2token #load_dictionary(path='models/Seq2Seq_LSTM/dict.pickle')
Word2vec_model = Word2Vec(input_sentence, size=100, window=5, min_count=1, workers=8,sg=1) #replace with bigram_sent,sentence for bigram model, unigram evaluation
# print(Word2vec_model)
Word2vec_model.train(input_sentence,total_examples=len(preprocessed_sentences),epochs=10,compute_loss=True)
result = Word2vec_model.wv.most_similar(positive='bvoip',topn=20)
Word2vec_model.init_sims(replace=True)
# Word2vec_model.save("models/Word2vec/Word2Vec_with_new_lemma.model")
Word2vec_model = gensim.models.Word2Vec.load("models/Word2vec/Word2Vec_with_new_lemma.model")
## Hyper-parameter
INPUT_DIM = len(final_dict)
HID_DIM = 200
EPOCH = 100
def main():
input_batch, target_batch, pairs = make_batch(sentences=preprocessed_sentences,dictionary=dictionary)
encoder = en.EncoderRNN(hidden_size=HID_DIM).to(device)
n_enc_parms = sum([p.numel() for _, p in encoder.named_parameters() if p.requires_grad == True])
print(encoder,n_enc_parms)
decoder = dc.DecoderRNN(hidden_size=HID_DIM, output_size=INPUT_DIM).to(device)
n_dec_parms = sum([p.numel() for _, p in decoder.named_parameters() if p.requires_grad == True])
print(decoder,n_dec_parms)
encoder_optimizer = optim.Adam(encoder.parameters())
decoder_optimizer = optim.Adam(decoder.parameters())
criterion = nn.CrossEntropyLoss()
early_stopping = EarlyStopping(PATH='models/Seq2Seq_LSTM', patience=10,verbose=True)
if pretrained:
epoch,encoder,encoder_optimizer =load_checkpoint('models/Seq2Seq_LSTM/train_loss-0.04279935139991392_epoch-4_checkpoint.pth.tar',encoder,encoder_optimizer)
else:
for i in range(EPOCH):
train_loss = trainIters(encoder,decoder,pairs, print_every=5,epoch_id=(i+1),encoder_optimizer=encoder_optimizer,decoder_optimizer=decoder_optimizer,criterion=criterion)
early_stopping(encoder,train_loss,encoder_optimizer,i)
if early_stopping.early_stop:
print("Early stopping")
break
#save_checkpoint(Path='models/Seq2Seq_LSTM',encoder=encoder, encoder_optimizer=encoder_optimizer,epoch_id=i,train_loss=loss)
if not load_representation:
representations = evaluateRandomly(encoder=encoder,sentences=preprocessed_sentences,dictionary=dictionary)
# with open("models/Seq2Seq_LSTM/representations.pkl", "wb") as handle:
# pickle.dump(representation, handle)
torch.save(torch.stack(representations).to(device),'models/Seq2Seq_LSTM/representations.pth')
print("total {} representation saved".format(len(representations)))
else:
# with open("models/Seq2Seq_LSTM/representations.pkl", "rb") as handle:
# representation = pickle.load(handle)
representations = torch.load('models/Seq2Seq_LSTM/representations.pth')
print("=============== Representation Loaded =============")
#mixed_representation = mixup_bert_word2vec_embedding(bert_file_path='models/BERT/Bert_representation.pickle',word2vec_file_path='models/Seq2Seq_LSTM/representations.pkl')
#print(mixed_representation.shape)
print("================Testing========================")
df=pd.read_excel('Tickets.xlsx')
pd.set_option('display.max_columns', 500)
while True:
input_ticket = input('Type Ticket text : ')
#result=
testing(sentense=input_ticket,representation=representations,dictionary=dictionary,encoder=encoder,df=df,processed_sentences=preprocessed_sentences,k=20)
#print(result.values)
if __name__ == "__main__":
main()