-
Notifications
You must be signed in to change notification settings - Fork 0
/
study.py
205 lines (167 loc) · 7.98 KB
/
study.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
import os.path
import pandas as pd
import spacy
import torchtext
from loguru import logger
from torchtext.vocab import Vocab
from config.path import dataset_dir, download_corpus, eng_lang_file, lang_file
from train import train_autoencoder
from utils.dataset import sequence2index
from utils.processing import nlp_pipeline
from utils.utils import download_from_url, read_file, read_json
spacy.load('de_core_news_sm')
spacy.load('en_core_web_sm')
def ablation_study(corpus_4_model_training: pd.DataFrame(),
model_config_file: str,
vocab_fr: Vocab,
vocab_it: Vocab,
model_file: str,
plot_file: str,
study_result_dir: str):
"""
The ablation_study function is used to test the effect of different model configurations on the performance of
the autoencoder.
:param corpus_4_model_training: pd.DataFrame(): Pass in the training data
:param model_config_file: str: Read the model configuration file
:param vocab_fr: Vocab: Pass the vocabulary of the french language
:param vocab_it: Vocab: Pass the vocabulary of the italian language
:param model_file: str: Save the model
:param plot_file: str: Save the plot of the loss function
:param study_result_dir: str: Save the results of the ablation study
"""
config = read_json(model_config_file)
# ------- train without dropout ------- #
logger.info("[ablation_study_eval] training without dropout")
config_without_dropout = config
config_without_dropout['enc_dropout'] = 0
config_without_dropout['dec_dropout'] = 0
study_result_dir = os.path.join(study_result_dir, "dropout")
train_autoencoder(config_without_dropout,
corpus_4_model_training,
vocab_fr,
vocab_it,
model_file,
plot_file,
study_result_dir,
optimize=False,
ablation_study=True)
# ------- train with just one layer of lstm ------- #
logger.info("[ablation_study_eval] training with just one layer of lstm")
config_one_layer = config
config_one_layer['num_layers'] = 1
study_result_dir = os.path.join(study_result_dir, "num_layers")
train_autoencoder(config_one_layer,
corpus_4_model_training,
vocab_fr,
vocab_it,
model_file,
plot_file,
study_result_dir,
optimize=False,
ablation_study=True)
# ------- train without contrastive loss ------- #
logger.info("[ablation_study_eval] training without contrastive loss")
config_zero_alpha = config
config_zero_alpha['alpha'] = 0
study_result_dir = os.path.join(study_result_dir, "cl_loss")
train_autoencoder(config_zero_alpha,
corpus_4_model_training,
vocab_fr,
vocab_it,
model_file,
plot_file,
study_result_dir,
optimize=False,
ablation_study=True)
# ------- train without reconstruction loss ------- #
logger.info("[ablation_study_eval] training without reconstruction loss")
config_zero_beta = config
config_zero_beta['beta'] = 0
study_result_dir = os.path.join(study_result_dir, "reconstruction_loss")
train_autoencoder(config_zero_beta,
corpus_4_model_training,
vocab_fr,
vocab_it,
model_file,
plot_file,
study_result_dir,
optimize=False,
ablation_study=True)
# ------- test on english/german dataset ------- #
logger.info("[ablation_study_eval] training with english/german")
study_result_dir = os.path.join(study_result_dir, "english_german")
test_english_german(config, model_file, plot_file, study_result_dir)
# ------- test on english/french dataset ------- #
logger.info("[ablation_study_eval] training with english/french")
study_result_dir = os.path.join(study_result_dir, "english_french")
test_english_french(config, model_file, plot_file, study_result_dir)
def test_english_french(config, model_file, plot_file, study_result_dir):
fr_file = read_file(lang_file.format(lang="fr"))
eng_fr_file = read_file(eng_lang_file.format(lang="fr"))
corpus = pd.DataFrame(data={'french': fr_file.split("\n"), 'english': eng_fr_file.split("\n")})
corpus['french'] = nlp_pipeline(corpus['french'])
corpus['english'] = nlp_pipeline(corpus['english'])
tokenizer_fr = torchtext.data.utils.get_tokenizer('spacy', language="fr_core_news_sm")
tokenizer_en = torchtext.data.utils.get_tokenizer('spacy', language="en_core_web_sm")
tokenized_dataset_fr = corpus['french'].apply(lambda x: tokenizer_fr(x))
tokenized_dataset_en = corpus['english'].apply(lambda x: tokenizer_en(x))
vocab_fr = torchtext.vocab.build_vocab_from_iterator(tokenized_dataset_fr, max_tokens=30000)
vocab_fr.insert_token('<pad>', 0)
vocab_fr.insert_token('<eos>', 1)
vocab_fr.insert_token('<unk>', 2)
vocab_fr.set_default_index(vocab_fr['<unk>'])
vocab_en = torchtext.vocab.build_vocab_from_iterator(tokenized_dataset_en, max_tokens=30000)
vocab_en.insert_token('<pad>', 0)
vocab_en.insert_token('<eos>', 1)
vocab_en.insert_token('<unk>', 2)
vocab_en.set_default_index(vocab_en['<unk>'])
tokenized_dataset_en = tokenized_dataset_en.apply(lambda x: x + ["<eos>"])
tokenized_dataset_fr = tokenized_dataset_fr.apply(lambda x: x + ["<eos>"])
corpus_tokenized = pd.DataFrame({
'english': sequence2index(tokenized_dataset_en, vocab_en),
'french' : sequence2index(tokenized_dataset_fr, vocab_fr)})
corpus_tokenized = corpus_tokenized[
corpus_tokenized.apply(lambda row: len(row['french']) <= 100 and len(row['english']) <= 100 and len(
row['french']) > 10 and len(row['english']) > 10, axis=1)]
corpus_tokenized['french'] = corpus_tokenized['french'].apply(lambda x: x[:100] + [0] * (100 - len(x)))
corpus_tokenized['english'] = corpus_tokenized['english'].apply(lambda x: x[:100] + [0] * (100 - len(x)))
config_en_fr = config
train_autoencoder(config_en_fr, corpus_tokenized, vocab_fr, # vocab fr
vocab_en, # vocab it
model_file, plot_file, study_result_dir, optimize=False, ablation_study=True)
def test_english_german(config, model_file, plot_file, study_result_dir):
if not os.path.exists(download_corpus.format(lang="de")):
download_from_url(download_corpus.format(lang="de"), dataset_dir, "de")
de_file = read_file(lang_file.format(lang="de"))
eng_de_file = read_file(eng_lang_file.format(lang="de"))
corpus = pd.DataFrame(data={'german': de_file.split("\n"), 'english': eng_de_file.split("\n")})
corpus['german'] = nlp_pipeline(corpus['german'])
corpus['english'] = nlp_pipeline(corpus['english'])
tokenizer_de = torchtext.data.utils.get_tokenizer('spacy', language="de_core_news_sm")
tokenizer_en = torchtext.data.utils.get_tokenizer('spacy', language="en_core_web_sm")
tokenized_dataset_de = corpus['german'].apply(lambda x: tokenizer_de(x))
tokenized_dataset_en = corpus['english'].apply(lambda x: tokenizer_en(x))
vocab_de = torchtext.vocab.build_vocab_from_iterator(tokenized_dataset_de, max_tokens=30000)
vocab_de.insert_token('<pad>', 0)
vocab_de.insert_token('<eos>', 1)
vocab_de.insert_token('<unk>', 2)
vocab_de.set_default_index(vocab_de['<unk>'])
vocab_en = torchtext.vocab.build_vocab_from_iterator(tokenized_dataset_en, max_tokens=30000)
vocab_en.insert_token('<pad>', 0)
vocab_en.insert_token('<eos>', 1)
vocab_en.insert_token('<unk>', 2)
vocab_en.set_default_index(vocab_en['<unk>'])
tokenized_dataset_en = tokenized_dataset_en.apply(lambda x: x + ["<eos>"])
tokenized_dataset_de = tokenized_dataset_de.apply(lambda x: x + ["<eos>"])
corpus_tokenized = pd.DataFrame({
'english': sequence2index(tokenized_dataset_en, vocab_en),
'german' : sequence2index(tokenized_dataset_de, vocab_de)})
corpus_tokenized = corpus_tokenized[
corpus_tokenized.apply(lambda row: len(row['german']) <= 100 and len(row['english']) <= 100 and len(
row['german']) > 10 and len(row['english']) > 10, axis=1)]
corpus_tokenized['german'] = corpus_tokenized['german'].apply(lambda x: x[:100] + [0] * (100 - len(x)))
corpus_tokenized['english'] = corpus_tokenized['english'].apply(lambda x: x[:100] + [0] * (100 - len(x)))
config_en_de = config
train_autoencoder(config_en_de, corpus_tokenized, vocab_de, # vocab fr
vocab_en, # vocab it
model_file, plot_file, study_result_dir, optimize=False, ablation_study=True)