-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
368 lines (337 loc) · 14.3 KB
/
utils.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
import pandas as pd
import numpy as np
import os
import io
from tqdm import tqdm
import re
import string
import numpy as np
from time import time
import pickle as pkl
from copy import deepcopy
from collections import Counter
import operator
import matplotlib.pyplot as plt
from keras.preprocessing.text import Tokenizer
from keras.utils import to_categorical
from keras.preprocessing.sequence import pad_sequences
from keras.models import load_model, model_from_yaml
from keras import backend as K
############################################## N-grams based modelling ##############################################
def read_data(name):
"""
Read the data
:param name: str, exact name of data set to read
:return: pd.DataFrame with 1 column containing the sentences
"""
d = pd.read_csv(open(os.path.join('./Data/', name), 'r'), sep=" .\n", header=None,
engine='python')
d.columns = ['text']
if 'char' in name:
print('Cleaning to redefine (character-level data set)')
return d
def read_split(name):
"""
Read train, test, validation sets
:param name: str, choose between 'wiki', 'char', 'tatoeba'
"""
if name not in ['wiki', 'char', 'tatoeba']:
raise AssertionError("Please choose between: 'wiki', 'char' or 'tatoeba'")
train, test, valid = [read_data(name + '.' + d + '.txt') for d in ['train', 'test', 'valid']]
print('Loading ' + name + ' data set...')
print('Number of training sentences: {:10.0f}'.format(train.shape[0]))
print('Number of validation sentences: {:10.0f}'.format(valid.shape[0]))
print('Number of testing sentences: {:10.0f}'.format(test.shape[0]))
print('Done.')
return train, test, valid
words_replace = dict()
for char, sub in zip([
"wasn 't", "isn 't",
"you' re", "we' re", "they' re",
"can 't",
"don 't'", "doesn 't", "didn 't",
"haven' t", "hasn 't'",
"shouldn 't", "wouldn 't", "won 't",
"' ll",
"I' m"
],
['was not', 'is not',
'you are', 'we are', 'they are',
'can not',
'do not', 'does not', 'did not',
'have not', 'has not',
'should not', 'would not', 'will not',
' will',
'I am'
]):
words_replace[char] = sub
flatten = lambda l: [item for sublist in l for item in sublist]
def ngram_nltk(s, n):
"""
For a given sentence s, return list of n-grams
"""
# Break sentence in the token, remove empty tokens
tokens = [token for token in s.split(" ") if token != ""]
# Use the zip function to help us generate n-grams
# Concatentate the tokens into ngrams and return
ngrams = zip(*[tokens[i:] for i in range(n)])
return [" ".join(ngram) for ngram in ngrams]
punct_replace = dict({char: ' ' for char in list(string.punctuation)})
dict_replace = {**words_replace, **punct_replace}
def clean_text(df):
"""
Clean sentences by:
- lowering characters
- removing non alphabetic characters except brackets ('<unk>')
- adding start and end-symbol (resp. <s> and </s>)
"""
data = deepcopy(df)
data.text = data.text.apply(lambda s: s[:-2].lower())
data.text = data.text.apply(lambda s: s.replace('<unk>', '<UNK>'))
for item in dict_replace.items():
char, sub = item
data.text = data.text.apply(lambda s: s.replace(char, sub))
data.text = data.text.apply(lambda s: re.sub(r'[^a-zA-Z0-9\s]', ' ', s))
data.text = data.text.apply(lambda s: re.sub(' +',' ', s))
data.text = data.text.apply(lambda s: s.replace('UNK', '<unk>'))
data.text = data.text.apply(lambda s: '<s> '+s+' </s>')
return data
def words_count(data):
"""
Print some statistics about the data:
- Total number of words
- Number of unique words
- Percentage of OOV
"""
split = data.text.apply(lambda s: s.split())
words = dict(Counter(flatten(list(split))))
words_count = sum(words.values())
unique_count = len(set(words.keys()))
unk_count = words['<unk>']
p_unk = unk_count/words_count * 100
print('Total number of words: {:5.0f}\nNumber of unique words: {:3.0f}\nPercentage of OOV: {:2.3f}%'.format(words_count,
unique_count,
p_unk))
return words
def split_and_join(s, n, n1=0):
"""
For a given sentence s = 'I am not a genius', n=2 returns 'I am', n=3 returns 'I am not', etc...
"""
s = s.split()[n1:n]
return " ".join(s)
def sort_dict(dict_):
"""
Sort a dictionary by key
"""
return sorted(dict_.items(), key=operator.itemgetter(1))[::-1]
############################################## Deep Learning based modelling ##############################################
def load_vectors(fname, vocab, nmax=2*10**6):
"""
Load pre-trained words embeddings from FastText:
A. Joulin, E. Grave, P. Bojanowski, M. Douze, H. Jégou, T. Mikolov, 'FastText.zip: Compressing text classification models'
"""
word2vec = {}
with io.open(fname, encoding='utf-8') as f:
next(f)
for i, line in tqdm(enumerate(f), total=nmax):
word, vec = line.split(' ', 1)
if word in vocab:
word2vec[word] = np.fromstring(vec, sep=' ')
if i == (nmax - 1):
break
print('Loaded %s pretrained word vectors' % (len(word2vec)))
return word2vec
class keras_preprocess():
"""
Preprocess the training, validation and testing data sets.
Returns: np.arrays, X_train, X_valid, X_test, y_train, y_valid and y_test
"""
def __init__(self, train, valid, test, ngram, verbose=True):
"""
If ngram = 3, then X.shape = (number of trigrams, 2), y.shape = (number of trigrams, 1)
X contains the 2 first words, y the last one.
"""
self.t0 = time()
self.train = train
self.valid = valid
self.test = test
self.ngram = ngram
self.verboseprint = print if verbose else lambda *a, **k: None
self.verboseprint('Tokenizing...')
self.tokenize()
self.verboseprint('Creating n-grams...')
self.sequences_train = self.create_ngrams(self.train)
self.sequences_valid = self.create_ngrams(self.valid)
self.sequences_test = self.create_ngrams(self.test)
def tokenize(self):
"""
Fit tokenizer
"""
# Lower, remove non alphabetic characters except brackets and antislash
self.tokenizer = Tokenizer(num_words=None,
filters='!"#$%&()*+,-.:;=?@^_`{|}~\t\n',
lower=True,
split=' ',
char_level=False,
oov_token='<unk>',
document_count=0)
# Fit on train
self.tokenizer.fit_on_texts(self.train.text)
self.vocab = self.tokenizer.word_index
self.vocab_size = len(self.vocab)+2
def create_ngrams(self, data):
"""
Return sequences (ngrams). If self.ngram=3, then output.shape = (number of sequences, 3)
"""
sequences = list()
for line in data.text:
encoded = self.tokenizer.texts_to_sequences([line])[0]
for i in range(self.ngram, len(encoded)):
sequence = encoded[i-self.ngram:i]
sequences.append(sequence)
return np.stack(sequences, axis=0)
def split_X_y(self, sequences):
"""
Split sequences into X and y sets (X: first self.ngram-1 words, y: last words)
"""
X, y = sequences[:, :-1], sequences[:, -1]
#y = to_categorical(y, num_classes=self.vocab_size)
return X, y
def output(self):
"""
Store the quantities of interest
"""
self.verboseprint('Split into X and y...')
self.X_train, self.y_train = self.split_X_y(self.sequences_train)
self.X_valid, self.y_valid = self.split_X_y(self.sequences_valid)
self.X_test, self.y_test = self.split_X_y(self.sequences_test)
self.verboseprint('Done. Time elapsed: {:3.1f} sec.'.format(time()-self.t0))
return self.X_train, self.X_valid, self.X_test, self.y_train, self.y_valid, self.y_test
def keras_print_score(X, y, model, dataset):
"""
Print metrics for a given tuple (X, y)
dataset: string, 'Train', 'Test' or 'Valid'
"""
l, p, a = model.evaluate(X, y, verbose=0)
print(dataset + ' set - Loss: {:.3f} | Perplexity: {:.2f} | Accuracy: {:3f}'.format(l, p, a))
return l, p, a
def keras_perplexity(y_true, y_pred):
"""
Compute perplexity using keras.backend
Source: https://nbviewer.jupyter.org/github/chakki-works/chariot/blob/master/notebooks/language%20modeling.ipynb
"""
cross_entropy = K.mean(K.sparse_categorical_crossentropy(y_true, y_pred))
perplexity = K.exp(cross_entropy)
return perplexity
def keras_generate_seq(model, tokenizer, max_length, seed_text, n_words):
"""
Generate sequence based on output probabilities
:param model: keras.engine.sequential.Sequential instance
:param tokenizer: keras_preprocessing.text.Tokenizer instance
:param max_length: int, size of ngrams
:param seed_text: string, first words of the sentence
:param n_words: int, number of words to predict
"""
in_text = seed_text
# generate a fixed number of words
for _ in range(n_words):
# encode the text as integer
encoded = tokenizer.texts_to_sequences([in_text])[0]
# pre-pad sequences to a fixed length
encoded = pad_sequences([encoded], maxlen=max_length, padding='pre')
# predict probabilities for each word
yhat = model.predict_classes(encoded, verbose=0)
# map predicted word index to word
out_word = ''
for word, index in tokenizer.word_index.items():
if index == yhat:
out_word = word
break
# append to input
in_text += ' ' + out_word
return in_text
def keras_load_model(modelname):
'''
Load model from modelname
'''
# load YAML
print('Loading model...')
yaml_file = open(os.path.join('./Models', modelname+'.yaml'), 'r')
loaded_model_yaml = yaml_file.read()
yaml_file.close()
loaded_model = model_from_yaml(loaded_model_yaml)
# load weights into new model
loaded_model.load_weights(os.path.join('./Models/', modelname+".h5"))
print("Loaded model from disk.")
return loaded_model
def load_results(layer, dataset='tatoeba', best=False):
"""
For a given recurrent layer, store and return all results obtained from the experiments
:param layer: string
:param dataset: string, choose between tatoeba or wiki
:param best: Boolean, if comparison of best models
"""
layer_ = '_'+layer+'_'
path = './Results/'
path_ = path if not best else './Results/best'
entries = []
for entry in os.scandir(path):
condition = ('best' in entry.path) if best else ('best' not in entry.path)
if (layer_ in entry.path) and condition and (dataset in entry.path):
entries.append(entry.path)
n = len(entries)
zo = np.zeros(n+1)
print('Loading %d results for %s...' %(n, layer))
# Then create dataframe
results = pd.DataFrame({'dataset': zo,'layer': zo,'ngram': zo, 'embedding': zo, 'mcells': zo, 'batchsize': zo, 'dropout2': zo,
'loss_train': zo, 'per_train': zo, 'acc_train': zo,
'loss_val': zo, 'per_val': zo, 'acc_val': zo,
'loss_test': zo, 'per_test': zo, 'acc_test': zo})
# Finally, fill it
j = 0
for entry in os.scandir(path):
condition = ('best' in entry.path) if best else ('best' not in entry.path)
if (layer_ in entry.path) and condition and (dataset in entry.path):
j += 1
params = entry.path.split('.pk')[0].split(path_)[1].split('_')[1*best:]
dataset = params[0]
layer = params[1]
ngram = int(params[2][5:])
emb = int(params[3][3:])
mcells = int(params[4][6:])
batchsize = int(params[5][5:])
dropout2 = int(params[7][-1]) * 0.1
r = list(pkl.load(open(entry.path, 'rb')).values())
results.iloc[j, :7] = [dataset, layer, ngram, emb, mcells, batchsize, dropout2]
results.iloc[j, 7:] = r
results = results[results['dataset'] != 0]
return results
def plot_comp(results, param, layer, metric):
"""
Plot the mean metric values as a function of a given parameter, for both train, test and val sets
along with error bars.
:param results: pd.Dataframe
:param param: string, choose between 'ngram', 'embedding', 'mcells' or 'dropout2'
:param layer: string, 'lstm' or 'bilstm', etc... (layer only serves for the title)
:param metric: string, choose between 'per' (perplexity), 'acc' (accuracy) or 'loss' (cross-entropy)
"""
# First we apply groupby + mean or + std
if 'per' in metric:
rm = results.groupby(param).mean()[['per_train', 'per_val', 'per_test']]
rstd = results.groupby(param).std()[['per_train', 'per_val', 'per_test']]
elif 'acc' in metric:
rm = results.groupby(param).mean()[['acc_train', 'acc_val', 'acc_test']]
rstd = results.groupby(param).std()[['acc_train', 'acc_val', 'acc_test']]
else:
rm = results.groupby(param).mean()[['loss_train', 'loss_val', 'loss_test']]
rstd = results.groupby(param).std()[['loss_train', 'loss_val', 'loss_test']]
# Then we draw the errorbars
plt.errorbar(x=rm.index, y=rm[metric+'_'+'train'], yerr=rstd[metric+'_'+'train'], marker='o', label='Train')
plt.errorbar(x=rm.index, y=rm[metric+'_'+'val'], yerr=rstd[metric+'_'+'val'], marker='o', label='Validation')
plt.errorbar(x=rm.index, y=rm[metric+'_'+'test'], yerr=rstd[metric+'_'+'test'], marker='o', label='Test')
plt.legend(loc='best')
# Change metric name for the title
metric_ = 'Accuracy' if metric=='acc' else 'Perplexity' if metric=='per' else 'Entropy'
plt.title('%s as a function of %s (%s)' % (metric_, param, layer))
# Be sure to convert into
plt.xticks(rm.index, np.round(rm.index, 1))