Commit
·
31f1c86
1
Parent(s):
5d81f4d
Autoencoder not tested
Browse files- autoencoder.py +205 -1
- n_grams.py +7 -9
- script.sh +7 -0
autoencoder.py
CHANGED
@@ -1,5 +1,209 @@
|
|
|
|
1 |
import logging
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
|
4 |
def test():
|
5 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
import logging
|
3 |
+
from transformers import BertModel, BertTokenizerFast
|
4 |
+
import os
|
5 |
+
from pathlib import Path
|
6 |
+
import pandas as pd
|
7 |
+
from datasets import load_dataset
|
8 |
+
from torch.utils.data import DataLoader
|
9 |
+
from tqdm import tqdm
|
10 |
+
|
11 |
+
CURRENT_PATH = Path(__file__).parent
|
12 |
+
|
13 |
+
|
14 |
+
def tokenize(dataset):
|
15 |
+
BERT_MAX_LEN = 512
|
16 |
+
|
17 |
+
tokenizer = BertTokenizerFast.from_pretrained(
|
18 |
+
"neuralmind/bert-base-portuguese-cased", max_length=BERT_MAX_LEN)
|
19 |
+
|
20 |
+
dataset = dataset.map(lambda example: tokenizer(
|
21 |
+
example["text"], truncation=True, padding="max_length", max_length=BERT_MAX_LEN))
|
22 |
+
|
23 |
+
return dataset
|
24 |
+
|
25 |
+
|
26 |
+
def create_dataloader(dataset, shuffle=True):
|
27 |
+
return DataLoader(dataset, batch_size=8, shuffle=shuffle, num_workers=8, drop_last=True)
|
28 |
+
|
29 |
+
|
30 |
+
class AutoEncoder(torch.nn.Module):
|
31 |
+
def __init__(self):
|
32 |
+
super().__init__()
|
33 |
+
|
34 |
+
self.device = torch.device(
|
35 |
+
'cuda' if torch.cuda.is_available() else 'cpu')
|
36 |
+
|
37 |
+
self.bert = BertModel.from_pretrained(
|
38 |
+
'neuralmind/bert-base-portuguese-cased').to(self.device)
|
39 |
+
|
40 |
+
# Freeze BERT
|
41 |
+
for param in self.bert.parameters():
|
42 |
+
param.requires_grad = False
|
43 |
+
|
44 |
+
self.encoder = torch.nn.Sequential(
|
45 |
+
torch.nn.Linear(self.bert.config.hidden_size,
|
46 |
+
self.bert.config.hidden_size // 5),
|
47 |
+
torch.nn.ReLU(),
|
48 |
+
torch.nn.Linear(self.bert.config.hidden_size // 5,
|
49 |
+
self.bert.config.hidden_size // 10),
|
50 |
+
torch.nn.ReLU(),
|
51 |
+
torch.nn.Linear(self.bert.config.hidden_size // 10,
|
52 |
+
self.bert.config.hidden_size // 30),
|
53 |
+
torch.nn.ReLU(),
|
54 |
+
).to(self.device)
|
55 |
+
|
56 |
+
self.decoder = torch.nn.Sequential(
|
57 |
+
torch.nn.Linear(self.bert.config.hidden_size // 30,
|
58 |
+
self.bert.config.hidden_size // 10),
|
59 |
+
torch.nn.ReLU(),
|
60 |
+
torch.nn.Linear(self.bert.config.hidden_size // 10,
|
61 |
+
self.bert.config.hidden_size // 5),
|
62 |
+
torch.nn.ReLU(),
|
63 |
+
torch.nn.Linear(self.bert.config.hidden_size //
|
64 |
+
5, self.bert.config.hidden_size),
|
65 |
+
torch.nn.Sigmoid()
|
66 |
+
).to(self.device)
|
67 |
+
|
68 |
+
def forward(self, input_ids, attention_mask):
|
69 |
+
bert_output = self.bert(input_ids=input_ids,
|
70 |
+
attention_mask=attention_mask).last_hidden_state[:, 0, :]
|
71 |
+
|
72 |
+
encoded = self.encoder(bert_output)
|
73 |
+
|
74 |
+
decoded = self.decoder(encoded)
|
75 |
+
|
76 |
+
return bert_output, decoded
|
77 |
+
|
78 |
+
|
79 |
+
def load_models():
|
80 |
+
models = []
|
81 |
+
|
82 |
+
for domain in ['politics', 'news', 'law', 'social_media', 'literature', 'web']:
|
83 |
+
logging.info(f"Loading {domain} model...")
|
84 |
+
|
85 |
+
accumulator = []
|
86 |
+
|
87 |
+
for lang in ['brazilian', 'european']:
|
88 |
+
model = AutoEncoder()
|
89 |
+
model.load_state_dict(torch.load(os.path.join(
|
90 |
+
CURRENT_PATH, 'models', 'autoencoder', f'{domain}_{lang}_model.pt')))
|
91 |
+
accumulator.append(model)
|
92 |
+
|
93 |
+
models.append({
|
94 |
+
'models': accumulator,
|
95 |
+
'train_domain': domain,
|
96 |
+
})
|
97 |
+
|
98 |
+
return models
|
99 |
+
|
100 |
+
|
101 |
+
def benchmark(model, debug=False):
|
102 |
+
|
103 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
104 |
+
|
105 |
+
df_results = pd.DataFrame(
|
106 |
+
columns=['train_domain', 'test_domain', 'accuracy', 'f1', 'precision', 'recall'])
|
107 |
+
|
108 |
+
train_domain = model['train_domain']
|
109 |
+
|
110 |
+
brazilian_model = model['models'][0]
|
111 |
+
|
112 |
+
european_model = model['models'][1]
|
113 |
+
|
114 |
+
brazilian_model.eval()
|
115 |
+
european_model.eval()
|
116 |
+
|
117 |
+
brazilian_model.to(device)
|
118 |
+
european_model.to(device)
|
119 |
+
|
120 |
+
for test_domain in ['politics', 'news', 'law', 'social_media', 'literature', 'web']:
|
121 |
+
dataset = load_dataset(
|
122 |
+
'arubenruben/Portuguese_Language_Identification', test_domain, split='test')
|
123 |
+
|
124 |
+
if debug:
|
125 |
+
logging.info(f"Debugging {test_domain} dataset...")
|
126 |
+
dataset = dataset.select(range(100))
|
127 |
+
else:
|
128 |
+
dataset = dataset.shuffle().select(range(min(50_000, len(dataset))))
|
129 |
+
|
130 |
+
dataset = tokenize(dataset)
|
131 |
+
|
132 |
+
dataset.set_format(type='torch', columns=[
|
133 |
+
'input_ids', 'attention_mask', 'label'])
|
134 |
+
|
135 |
+
dataset = create_dataloader(dataset)
|
136 |
+
|
137 |
+
predictions = []
|
138 |
+
labels = []
|
139 |
+
|
140 |
+
reconstruction_loss = torch.nn.MSELoss(reduction='none')
|
141 |
+
|
142 |
+
with torch.no_grad():
|
143 |
+
for batch in tqdm(dataset):
|
144 |
+
input_ids = batch['input_ids'].to(device)
|
145 |
+
|
146 |
+
attention_mask = batch['attention_mask'].to(device)
|
147 |
+
|
148 |
+
label = batch['label'].to(device)
|
149 |
+
|
150 |
+
bert_european, reconstruction_european = european_model(
|
151 |
+
input_ids=input_ids, attention_mask=attention_mask)
|
152 |
+
|
153 |
+
bert_brazilian, reconstruction_brazilian = brazilian_model(
|
154 |
+
input_ids=input_ids, attention_mask=attention_mask)
|
155 |
+
|
156 |
+
test_loss_european = reconstruction_loss(
|
157 |
+
reconstruction_european, bert_european)
|
158 |
+
|
159 |
+
test_loss_brazilian = reconstruction_loss(
|
160 |
+
reconstruction_brazilian, bert_brazilian)
|
161 |
+
|
162 |
+
for loss_european, loss_brazilian in zip(test_loss_european, test_loss_brazilian):
|
163 |
+
|
164 |
+
if loss_european.mean().item() < loss_brazilian.mean().item():
|
165 |
+
predictions.append(0)
|
166 |
+
total_loss += loss_european.mean().item() / len(test_loss_european)
|
167 |
+
|
168 |
+
else:
|
169 |
+
predictions.append(1)
|
170 |
+
total_loss += loss_brazilian.mean().item() / len(test_loss_brazilian)
|
171 |
+
|
172 |
+
labels.extend(label.tolist())
|
173 |
+
|
174 |
+
accuracy = accuracy.compute(
|
175 |
+
predictions=predictions, references=labels)['accuracy']
|
176 |
+
f1 = f1.compute(predictions=predictions, references=labels)['f1']
|
177 |
+
precision = precision.compute(
|
178 |
+
predictions=predictions, references=labels)['precision']
|
179 |
+
recall = recall.compute(predictions=predictions,
|
180 |
+
references=labels)['recall']
|
181 |
+
|
182 |
+
df_results = pd.concat([df_results, pd.DataFrame(
|
183 |
+
[[train_domain, test_domain, accuracy, f1, precision, recall]], columns=df_results.columns)], ignore_index=True)
|
184 |
+
|
185 |
+
return df_results
|
186 |
|
187 |
|
188 |
def test():
|
189 |
+
DEBUG = True
|
190 |
+
|
191 |
+
models = load_models()
|
192 |
+
|
193 |
+
df_results = pd.DataFrame(
|
194 |
+
columns=['train_domain', 'test_domain', 'accuracy', 'f1', 'precision', 'recall'])
|
195 |
+
|
196 |
+
for model in models:
|
197 |
+
logging.info(f"Train Domain {model['train_domain']}...")
|
198 |
+
|
199 |
+
df_results = pd.concat([df_results, benchmark(
|
200 |
+
model, debug=DEBUG)], ignore_index=True)
|
201 |
+
|
202 |
+
logging.info(f"Saving results...")
|
203 |
+
|
204 |
+
df_results.to_json(os.path.join(CURRENT_PATH, 'results',
|
205 |
+
'autoencoder.json'), orient='records', indent=4, force_ascii=False)
|
206 |
+
|
207 |
+
|
208 |
+
if __name__ == '__main__':
|
209 |
+
test()
|
n_grams.py
CHANGED
@@ -40,11 +40,6 @@ def load_pipelines():
|
|
40 |
|
41 |
def benchmark(pipeline, debug=False):
|
42 |
|
43 |
-
accuracy_evaluator = evaluate.load('accuracy')
|
44 |
-
f1_evaluator = evaluate.load('f1')
|
45 |
-
precision_evaluator = evaluate.load('precision')
|
46 |
-
recall_evaluator = evaluate.load('recall')
|
47 |
-
|
48 |
df_results = pd.DataFrame(
|
49 |
columns=['train_domain', 'test_domain', 'accuracy', 'f1', 'precision', 'recall'])
|
50 |
|
@@ -67,13 +62,16 @@ def benchmark(pipeline, debug=False):
|
|
67 |
|
68 |
y = pipeline.predict(dataset['text'])
|
69 |
|
70 |
-
accuracy =
|
71 |
predictions=y, references=dataset['label'])['accuracy']
|
72 |
-
|
|
|
73 |
predictions=y, references=dataset['label'])['f1']
|
74 |
-
|
|
|
75 |
predictions=y, references=dataset['label'])['precision']
|
76 |
-
|
|
|
77 |
predictions=y, references=dataset['label'])['recall']
|
78 |
|
79 |
logging.info(
|
|
|
40 |
|
41 |
def benchmark(pipeline, debug=False):
|
42 |
|
|
|
|
|
|
|
|
|
|
|
43 |
df_results = pd.DataFrame(
|
44 |
columns=['train_domain', 'test_domain', 'accuracy', 'f1', 'precision', 'recall'])
|
45 |
|
|
|
62 |
|
63 |
y = pipeline.predict(dataset['text'])
|
64 |
|
65 |
+
accuracy = evaluate.load('accuracy').compute(
|
66 |
predictions=y, references=dataset['label'])['accuracy']
|
67 |
+
|
68 |
+
f1 = evaluate.load('f1').compute(
|
69 |
predictions=y, references=dataset['label'])['f1']
|
70 |
+
|
71 |
+
precision = evaluate.load('precision').compute(
|
72 |
predictions=y, references=dataset['label'])['precision']
|
73 |
+
|
74 |
+
recall = evaluate.load('recall').compute(
|
75 |
predictions=y, references=dataset['label'])['recall']
|
76 |
|
77 |
logging.info(
|
script.sh
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
pip install -U -r requirements.txt
|
2 |
+
|
3 |
+
python n_grams.py
|
4 |
+
|
5 |
+
python embeddings.py
|
6 |
+
|
7 |
+
python autoencoder.py
|