Commit
·
6162fcf
1
Parent(s):
8d6722d
Upload sentiment_analysis.py
Browse files- sentiment_analysis.py +602 -0
sentiment_analysis.py
ADDED
@@ -0,0 +1,602 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""Sentiment_analysis.ipynb
|
3 |
+
|
4 |
+
Automatically generated by Colaboratory.
|
5 |
+
|
6 |
+
Original file is located at
|
7 |
+
https://colab.research.google.com/drive/1EHgMQQJzwbNja0JVMM2DVvrVTMHIS3Vg
|
8 |
+
"""
|
9 |
+
|
10 |
+
!pip install transformers
|
11 |
+
|
12 |
+
import pandas as pd
|
13 |
+
from wordcloud import WordCloud
|
14 |
+
import seaborn as sns
|
15 |
+
import re
|
16 |
+
import string
|
17 |
+
from collections import Counter, defaultdict
|
18 |
+
|
19 |
+
from sklearn.feature_extraction.text import CountVectorizer, TfidfVectorizer
|
20 |
+
|
21 |
+
import plotly.express as px
|
22 |
+
from plotly.subplots import make_subplots
|
23 |
+
import plotly.graph_objects as go
|
24 |
+
from plotly.offline import plot
|
25 |
+
|
26 |
+
import matplotlib.gridspec as gridspec
|
27 |
+
from matplotlib.ticker import MaxNLocator
|
28 |
+
import matplotlib.patches as mpatches
|
29 |
+
import matplotlib.pyplot as plt
|
30 |
+
import warnings
|
31 |
+
warnings.filterwarnings('ignore')
|
32 |
+
import nltk
|
33 |
+
nltk.download('stopwords')
|
34 |
+
from nltk.corpus import stopwords
|
35 |
+
stopWords_nltk = set(stopwords.words('english'))
|
36 |
+
|
37 |
+
|
38 |
+
import re
|
39 |
+
from typing import Union, List
|
40 |
+
|
41 |
+
class CleanText():
|
42 |
+
""" clearing text except digits () . , word character """
|
43 |
+
|
44 |
+
def __init__(self, clean_pattern = r"[^A-ZĞÜŞİÖÇIa-zğüı'şöç0-9.\"',()]"):
|
45 |
+
self.clean_pattern =clean_pattern
|
46 |
+
|
47 |
+
def __call__(self, text: Union[str, list]) -> str:
|
48 |
+
|
49 |
+
if isinstance(text, str):
|
50 |
+
docs = [[text]]
|
51 |
+
|
52 |
+
if isinstance(text, list):
|
53 |
+
docs = text
|
54 |
+
|
55 |
+
text = [[re.sub(self.clean_pattern, " ", sent) for sent in sents] for sents in docs]
|
56 |
+
|
57 |
+
# Join the list of lists into a single string
|
58 |
+
text = ' '.join([' '.join(sents) for sents in text])
|
59 |
+
|
60 |
+
return text
|
61 |
+
|
62 |
+
def remove_emoji(data):
|
63 |
+
emoj = re.compile("["
|
64 |
+
u"\U0001F600-\U0001F64F" # emoticons
|
65 |
+
u"\U0001F300-\U0001F5FF" # symbols & pictographs
|
66 |
+
u"\U0001F680-\U0001F6FF" # transport & map symbols
|
67 |
+
u"\U0001F1E0-\U0001F1FF" # flags (iOS)
|
68 |
+
u"\U00002500-\U00002BEF"
|
69 |
+
u"\U00002702-\U000027B0"
|
70 |
+
u"\U00002702-\U000027B0"
|
71 |
+
u"\U000024C2-\U0001F251"
|
72 |
+
u"\U0001f926-\U0001f937"
|
73 |
+
u"\U00010000-\U0010ffff"
|
74 |
+
u"\u2640-\u2642"
|
75 |
+
u"\u2600-\u2B55"
|
76 |
+
u"\u200d"
|
77 |
+
u"\u23cf"
|
78 |
+
u"\u23e9"
|
79 |
+
u"\u231a"
|
80 |
+
u"\ufe0f" # dingbats
|
81 |
+
u"\u3030"
|
82 |
+
"]+", re.UNICODE)
|
83 |
+
return re.sub(emoj, '', data)
|
84 |
+
|
85 |
+
def tokenize(text):
|
86 |
+
""" basic tokenize method with word character, non word character and digits """
|
87 |
+
text = re.sub(r" +", " ", str(text))
|
88 |
+
text = re.split(r"(\d+|[a-zA-ZğüşıöçĞÜŞİÖÇ]+|\W)", text)
|
89 |
+
text = list(filter(lambda x: x != '' and x != ' ', text))
|
90 |
+
sent_tokenized = ' '.join(text)
|
91 |
+
return sent_tokenized
|
92 |
+
|
93 |
+
regex = re.compile('[%s]' % re.escape(string.punctuation))
|
94 |
+
|
95 |
+
def remove_punct(text):
|
96 |
+
text = regex.sub(" ", text)
|
97 |
+
return text
|
98 |
+
|
99 |
+
clean = CleanText()
|
100 |
+
|
101 |
+
def label_encode(x):
|
102 |
+
if x == 1 or x == 2:
|
103 |
+
return 0
|
104 |
+
if x == 3:
|
105 |
+
return 1
|
106 |
+
if x == 5 or x == 4:
|
107 |
+
return 2
|
108 |
+
|
109 |
+
def label2name(x):
|
110 |
+
if x == 0:
|
111 |
+
return "Negative"
|
112 |
+
if x == 1:
|
113 |
+
return "Neutral"
|
114 |
+
if x == 2:
|
115 |
+
return "Positive"
|
116 |
+
|
117 |
+
from google.colab import files
|
118 |
+
uploaded = files.upload()
|
119 |
+
df = pd.read_csv('tripadvisor_hotel_reviews.csv')
|
120 |
+
|
121 |
+
print("df.columns: ", df.columns)
|
122 |
+
|
123 |
+
fig = px.histogram(df,
|
124 |
+
x = 'Rating',
|
125 |
+
title = 'Histogram of Review Rating',
|
126 |
+
template = 'ggplot2',
|
127 |
+
color = 'Rating',
|
128 |
+
color_discrete_sequence= px.colors.sequential.Blues_r,
|
129 |
+
opacity = 0.8,
|
130 |
+
height = 525,
|
131 |
+
width = 835,
|
132 |
+
)
|
133 |
+
|
134 |
+
fig.update_yaxes(title='Count')
|
135 |
+
fig.show()
|
136 |
+
|
137 |
+
df.info()
|
138 |
+
|
139 |
+
df["label"] = df["Rating"].apply(lambda x: label_encode(x))
|
140 |
+
df["label_name"] = df["label"].apply(lambda x: label2name(x))
|
141 |
+
|
142 |
+
df["Review"] = df["Review"].apply(lambda x: remove_punct(clean(remove_emoji(x).lower())[0][0]))
|
143 |
+
|
144 |
+
df.head()
|
145 |
+
|
146 |
+
fig = make_subplots(rows=1, cols=2, specs=[[{"type": "pie"}, {"type": "bar"}]])
|
147 |
+
colors = ['gold', 'mediumturquoise', 'lightgreen'] # darkorange
|
148 |
+
fig.add_trace(go.Pie(labels=df.label_name.value_counts().index,
|
149 |
+
values=df.label.value_counts().values), 1, 1)
|
150 |
+
|
151 |
+
fig.update_traces(hoverinfo='label+percent', textfont_size=20,
|
152 |
+
marker=dict(colors=colors, line=dict(color='#000000', width=2)))
|
153 |
+
|
154 |
+
fig.add_trace(go.Bar(x=df.label_name.value_counts().index, y=df.label.value_counts().values, marker_color = colors), 1,2)
|
155 |
+
|
156 |
+
fig.show()
|
157 |
+
|
158 |
+
import pandas as pd
|
159 |
+
import numpy as np
|
160 |
+
import os
|
161 |
+
import random
|
162 |
+
from pathlib import Path
|
163 |
+
import json
|
164 |
+
|
165 |
+
import torch
|
166 |
+
from tqdm.notebook import tqdm
|
167 |
+
|
168 |
+
from transformers import BertTokenizer
|
169 |
+
from torch.utils.data import TensorDataset
|
170 |
+
|
171 |
+
from transformers import BertForSequenceClassification
|
172 |
+
|
173 |
+
class Config():
|
174 |
+
seed_val = 17
|
175 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
176 |
+
epochs = 5
|
177 |
+
batch_size = 6
|
178 |
+
seq_length = 512
|
179 |
+
lr = 2e-5
|
180 |
+
eps = 1e-8
|
181 |
+
pretrained_model = 'bert-base-uncased'
|
182 |
+
test_size=0.15
|
183 |
+
random_state=42
|
184 |
+
add_special_tokens=True
|
185 |
+
return_attention_mask=True
|
186 |
+
pad_to_max_length=True
|
187 |
+
do_lower_case=False
|
188 |
+
return_tensors='pt'
|
189 |
+
config = Config()
|
190 |
+
|
191 |
+
# params will be saved after training
|
192 |
+
params = {"seed_val": config.seed_val,
|
193 |
+
"device":str(config.device),
|
194 |
+
"epochs":config.epochs,
|
195 |
+
"batch_size":config.batch_size,
|
196 |
+
"seq_length":config.seq_length,
|
197 |
+
"lr":config.lr,
|
198 |
+
"eps":config.eps,
|
199 |
+
"pretrained_model": config.pretrained_model,
|
200 |
+
"test_size":config.test_size,
|
201 |
+
"random_state":config.random_state,
|
202 |
+
"add_special_tokens":config.add_special_tokens,
|
203 |
+
"return_attention_mask":config.return_attention_mask,
|
204 |
+
"pad_to_max_length":config.pad_to_max_length,
|
205 |
+
"do_lower_case":config.do_lower_case,
|
206 |
+
"return_tensors":config.return_tensors,
|
207 |
+
}
|
208 |
+
|
209 |
+
import random
|
210 |
+
|
211 |
+
device = config.device
|
212 |
+
|
213 |
+
random.seed(config.seed_val)
|
214 |
+
np.random.seed(config.seed_val)
|
215 |
+
torch.manual_seed(config.seed_val)
|
216 |
+
torch.cuda.manual_seed_all(config.seed_val)
|
217 |
+
|
218 |
+
df.head()
|
219 |
+
|
220 |
+
from sklearn.model_selection import train_test_split
|
221 |
+
|
222 |
+
train_df_, val_df = train_test_split(df,
|
223 |
+
test_size=0.10,
|
224 |
+
random_state=config.random_state,
|
225 |
+
stratify=df.label.values)
|
226 |
+
|
227 |
+
train_df_.head()
|
228 |
+
|
229 |
+
train_df, test_df = train_test_split(train_df_,
|
230 |
+
test_size=0.10,
|
231 |
+
random_state=42,
|
232 |
+
stratify=train_df_.label.values)
|
233 |
+
|
234 |
+
print(len(train_df['label'].unique()))
|
235 |
+
print(train_df.shape)
|
236 |
+
|
237 |
+
print(len(val_df['label'].unique()))
|
238 |
+
print(val_df.shape)
|
239 |
+
|
240 |
+
print(len(test_df['label'].unique()))
|
241 |
+
print(test_df.shape)
|
242 |
+
|
243 |
+
tokenizer = BertTokenizer.from_pretrained(config.pretrained_model,
|
244 |
+
do_lower_case=config.do_lower_case)
|
245 |
+
|
246 |
+
encoded_data_train = tokenizer.batch_encode_plus(
|
247 |
+
train_df.Review.values,
|
248 |
+
add_special_tokens=config.add_special_tokens,
|
249 |
+
return_attention_mask=config.return_attention_mask,
|
250 |
+
pad_to_max_length=config.pad_to_max_length,
|
251 |
+
max_length=config.seq_length,
|
252 |
+
return_tensors=config.return_tensors
|
253 |
+
)
|
254 |
+
encoded_data_val = tokenizer.batch_encode_plus(
|
255 |
+
val_df.Review.values,
|
256 |
+
add_special_tokens=config.add_special_tokens,
|
257 |
+
return_attention_mask=config.return_attention_mask,
|
258 |
+
pad_to_max_length=config.pad_to_max_length,
|
259 |
+
max_length=config.seq_length,
|
260 |
+
return_tensors=config.return_tensors
|
261 |
+
)
|
262 |
+
|
263 |
+
input_ids_train = encoded_data_train['input_ids']
|
264 |
+
attention_masks_train = encoded_data_train['attention_mask']
|
265 |
+
labels_train = torch.tensor(train_df.label.values)
|
266 |
+
|
267 |
+
input_ids_val = encoded_data_val['input_ids']
|
268 |
+
attention_masks_val = encoded_data_val['attention_mask']
|
269 |
+
labels_val = torch.tensor(val_df.label.values)
|
270 |
+
|
271 |
+
dataset_train = TensorDataset(input_ids_train, attention_masks_train, labels_train)
|
272 |
+
dataset_val = TensorDataset(input_ids_val, attention_masks_val, labels_val)
|
273 |
+
|
274 |
+
model = BertForSequenceClassification.from_pretrained(config.pretrained_model,
|
275 |
+
num_labels=3,
|
276 |
+
output_attentions=False,
|
277 |
+
output_hidden_states=False)
|
278 |
+
|
279 |
+
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
|
280 |
+
|
281 |
+
dataloader_train = DataLoader(dataset_train,
|
282 |
+
sampler=RandomSampler(dataset_train),
|
283 |
+
batch_size=config.batch_size)
|
284 |
+
|
285 |
+
dataloader_validation = DataLoader(dataset_val,
|
286 |
+
sampler=SequentialSampler(dataset_val),
|
287 |
+
batch_size=config.batch_size)
|
288 |
+
|
289 |
+
from transformers import AdamW, get_linear_schedule_with_warmup
|
290 |
+
|
291 |
+
optimizer = AdamW(model.parameters(),
|
292 |
+
lr=config.lr,
|
293 |
+
eps=config.eps)
|
294 |
+
|
295 |
+
|
296 |
+
scheduler = get_linear_schedule_with_warmup(optimizer,
|
297 |
+
num_warmup_steps=0,
|
298 |
+
num_training_steps=len(dataloader_train)*config.epochs)
|
299 |
+
|
300 |
+
from sklearn.metrics import f1_score
|
301 |
+
|
302 |
+
def f1_score_func(preds, labels):
|
303 |
+
preds_flat = np.argmax(preds, axis=1).flatten()
|
304 |
+
labels_flat = labels.flatten()
|
305 |
+
return f1_score(labels_flat, preds_flat, average='weighted')
|
306 |
+
|
307 |
+
def accuracy_per_class(preds, labels, label_dict):
|
308 |
+
label_dict_inverse = {v: k for k, v in label_dict.items()}
|
309 |
+
|
310 |
+
preds_flat = np.argmax(preds, axis=1).flatten()
|
311 |
+
labels_flat = labels.flatten()
|
312 |
+
|
313 |
+
for label in np.unique(labels_flat):
|
314 |
+
y_preds = preds_flat[labels_flat==label]
|
315 |
+
y_true = labels_flat[labels_flat==label]
|
316 |
+
print(f'Class: {label_dict_inverse[label]}')
|
317 |
+
print(f'Accuracy: {len(y_preds[y_preds==label])}/{len(y_true)}\n')
|
318 |
+
|
319 |
+
def evaluate(dataloader_val):
|
320 |
+
|
321 |
+
model.eval()
|
322 |
+
|
323 |
+
loss_val_total = 0
|
324 |
+
predictions, true_vals = [], []
|
325 |
+
|
326 |
+
for batch in dataloader_val:
|
327 |
+
|
328 |
+
batch = tuple(b.to(config.device) for b in batch)
|
329 |
+
|
330 |
+
inputs = {'input_ids': batch[0],
|
331 |
+
'attention_mask': batch[1],
|
332 |
+
'labels': batch[2],
|
333 |
+
}
|
334 |
+
|
335 |
+
with torch.no_grad():
|
336 |
+
outputs = model(**inputs)
|
337 |
+
|
338 |
+
loss = outputs[0]
|
339 |
+
logits = outputs[1]
|
340 |
+
loss_val_total += loss.item()
|
341 |
+
|
342 |
+
logits = logits.detach().cpu().numpy()
|
343 |
+
label_ids = inputs['labels'].cpu().numpy()
|
344 |
+
predictions.append(logits)
|
345 |
+
true_vals.append(label_ids)
|
346 |
+
|
347 |
+
# calculate avareage val loss
|
348 |
+
loss_val_avg = loss_val_total/len(dataloader_val)
|
349 |
+
|
350 |
+
predictions = np.concatenate(predictions, axis=0)
|
351 |
+
true_vals = np.concatenate(true_vals, axis=0)
|
352 |
+
|
353 |
+
return loss_val_avg, predictions, true_vals
|
354 |
+
|
355 |
+
config.device
|
356 |
+
|
357 |
+
model.to(config.device)
|
358 |
+
|
359 |
+
for epoch in tqdm(range(1, config.epochs+1)):
|
360 |
+
|
361 |
+
model.train()
|
362 |
+
|
363 |
+
loss_train_total = 0
|
364 |
+
# allows you to see the progress of the training
|
365 |
+
progress_bar = tqdm(dataloader_train, desc='Epoch {:1d}'.format(epoch), leave=False, disable=False)
|
366 |
+
|
367 |
+
for batch in progress_bar:
|
368 |
+
|
369 |
+
model.zero_grad()
|
370 |
+
|
371 |
+
batch = tuple(b.to(config.device) for b in batch)
|
372 |
+
|
373 |
+
inputs = {'input_ids': batch[0],
|
374 |
+
'attention_mask': batch[1],
|
375 |
+
'labels': batch[2],
|
376 |
+
}
|
377 |
+
|
378 |
+
outputs = model(**inputs)
|
379 |
+
|
380 |
+
loss = outputs[0]
|
381 |
+
loss_train_total += loss.item()
|
382 |
+
loss.backward()
|
383 |
+
|
384 |
+
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
|
385 |
+
|
386 |
+
optimizer.step()
|
387 |
+
scheduler.step()
|
388 |
+
progress_bar.set_postfix({'training_loss': '{:.3f}'.format(loss.item()/len(batch))})
|
389 |
+
|
390 |
+
|
391 |
+
torch.save(model.state_dict(), f'_BERT_epoch_{epoch}.model')
|
392 |
+
|
393 |
+
tqdm.write(f'\nEpoch {epoch}')
|
394 |
+
|
395 |
+
loss_train_avg = loss_train_total/len(dataloader_train)
|
396 |
+
tqdm.write(f'Training loss: {loss_train_avg}')
|
397 |
+
|
398 |
+
val_loss, predictions, true_vals = evaluate(dataloader_validation)
|
399 |
+
val_f1 = f1_score_func(predictions, true_vals)
|
400 |
+
tqdm.write(f'Validation loss: {val_loss}')
|
401 |
+
|
402 |
+
tqdm.write(f'F1 Score (Weighted): {val_f1}');
|
403 |
+
# save model params and other configs
|
404 |
+
with Path('params.json').open("w") as f:
|
405 |
+
json.dump(params, f, ensure_ascii=False, indent=4)
|
406 |
+
|
407 |
+
model.load_state_dict(torch.load(f'./_BERT_epoch_3.model', map_location=torch.device('cpu')))
|
408 |
+
|
409 |
+
from sklearn.metrics import classification_report
|
410 |
+
|
411 |
+
preds_flat = np.argmax(predictions, axis=1).flatten()
|
412 |
+
print(classification_report(preds_flat, true_vals))
|
413 |
+
|
414 |
+
pred_final = []
|
415 |
+
|
416 |
+
for i, row in tqdm(val_df.iterrows(), total=val_df.shape[0]):
|
417 |
+
predictions = []
|
418 |
+
|
419 |
+
review = row["Review"]
|
420 |
+
encoded_data_test_single = tokenizer.batch_encode_plus(
|
421 |
+
[review],
|
422 |
+
add_special_tokens=config.add_special_tokens,
|
423 |
+
return_attention_mask=config.return_attention_mask,
|
424 |
+
pad_to_max_length=config.pad_to_max_length,
|
425 |
+
max_length=config.seq_length,
|
426 |
+
return_tensors=config.return_tensors
|
427 |
+
)
|
428 |
+
input_ids_test = encoded_data_test_single['input_ids']
|
429 |
+
attention_masks_test = encoded_data_test_single['attention_mask']
|
430 |
+
|
431 |
+
|
432 |
+
inputs = {'input_ids': input_ids_test.to(device),
|
433 |
+
'attention_mask':attention_masks_test.to(device),
|
434 |
+
}
|
435 |
+
|
436 |
+
with torch.no_grad():
|
437 |
+
outputs = model(**inputs)
|
438 |
+
|
439 |
+
logits = outputs[0]
|
440 |
+
logits = logits.detach().cpu().numpy()
|
441 |
+
predictions.append(logits)
|
442 |
+
predictions = np.concatenate(predictions, axis=0)
|
443 |
+
pred_final.append(np.argmax(predictions, axis=1).flatten()[0])
|
444 |
+
|
445 |
+
val_df["pred"] = pred_final
|
446 |
+
# Add control column for easier wrong and right predictions
|
447 |
+
control = val_df.pred.values == val_df.label.values
|
448 |
+
val_df["control"] = control
|
449 |
+
# filtering false predictions
|
450 |
+
val_df = val_df[val_df.control == False]
|
451 |
+
|
452 |
+
|
453 |
+
|
454 |
+
name2label = {"Negative":0,
|
455 |
+
"Neutral":1,
|
456 |
+
"Positive":2
|
457 |
+
}
|
458 |
+
label2name = {v: k for k, v in name2label.items()}
|
459 |
+
|
460 |
+
val_df["pred_name"] = val_df.pred.apply(lambda x: label2name.get(x))
|
461 |
+
from sklearn.metrics import confusion_matrix
|
462 |
+
|
463 |
+
# We create a confusion matrix to better observe the classes that the model confuses.
|
464 |
+
pred_name_values = val_df.pred_name.values
|
465 |
+
label_values = val_df.label_name.values
|
466 |
+
confmat = confusion_matrix(label_values, pred_name_values, labels=list(name2label.keys()))
|
467 |
+
|
468 |
+
confmat
|
469 |
+
|
470 |
+
df_confusion_val = pd.crosstab(label_values, pred_name_values)
|
471 |
+
df_confusion_val
|
472 |
+
|
473 |
+
df_confusion_val.to_csv("val_df_confusion.csv")
|
474 |
+
|
475 |
+
test_df.head()
|
476 |
+
|
477 |
+
encoded_data_test = tokenizer.batch_encode_plus(
|
478 |
+
test_df.Review.values,
|
479 |
+
add_special_tokens=config.add_special_tokens,
|
480 |
+
return_attention_mask=config.return_attention_mask,
|
481 |
+
pad_to_max_length=config.pad_to_max_length,
|
482 |
+
max_length=config.seq_length,
|
483 |
+
return_tensors=config.return_tensors
|
484 |
+
)
|
485 |
+
input_ids_test = encoded_data_test['input_ids']
|
486 |
+
attention_masks_test = encoded_data_test['attention_mask']
|
487 |
+
labels_test = torch.tensor(test_df.label.values)
|
488 |
+
|
489 |
+
model = BertForSequenceClassification.from_pretrained(config.pretrained_model,
|
490 |
+
num_labels=3,
|
491 |
+
output_attentions=False,
|
492 |
+
output_hidden_states=False)
|
493 |
+
|
494 |
+
model.to(config.device)
|
495 |
+
|
496 |
+
model.load_state_dict(torch.load(f'./_BERT_epoch_3.model', map_location=torch.device('cpu')))
|
497 |
+
|
498 |
+
_, predictions_test, true_vals_test = evaluate(dataloader_validation)
|
499 |
+
# accuracy_per_class(predictions, true_vals, intent2label)
|
500 |
+
|
501 |
+
def predict_sentiment(text):
|
502 |
+
# Prétraitement du texte
|
503 |
+
encoded_text = tokenizer.encode_plus(
|
504 |
+
text,
|
505 |
+
add_special_tokens=config.add_special_tokens,
|
506 |
+
return_attention_mask=config.return_attention_mask,
|
507 |
+
pad_to_max_length=config.pad_to_max_length,
|
508 |
+
max_length=config.seq_length,
|
509 |
+
return_tensors=config.return_tensors
|
510 |
+
)
|
511 |
+
|
512 |
+
# Convertir les entrées en tenseurs et les déplacer vers le bon appareil
|
513 |
+
input_ids = encoded_text['input_ids'].to(config.device)
|
514 |
+
attention_mask = encoded_text['attention_mask'].to(config.device)
|
515 |
+
|
516 |
+
# Mettre le modèle en mode d'évaluation et obtenir les prédictions
|
517 |
+
model.eval()
|
518 |
+
with torch.no_grad():
|
519 |
+
outputs = model(input_ids, attention_mask)
|
520 |
+
|
521 |
+
# Obtenir la prédiction du modèle
|
522 |
+
logits = outputs[0]
|
523 |
+
logits = logits.detach().cpu().numpy()
|
524 |
+
|
525 |
+
# Extraire la classe avec la probabilité la plus élevée
|
526 |
+
pred = np.argmax(logits, axis=1).flatten()[0]
|
527 |
+
|
528 |
+
# Convertir le label numérique en son nom correspondant
|
529 |
+
pred_name = label2name.get(pred)
|
530 |
+
|
531 |
+
return pred_name
|
532 |
+
|
533 |
+
text = "Your text here"
|
534 |
+
prediction = predict_sentiment(text)
|
535 |
+
print(f"The sentiment of the text is: {prediction}")
|
536 |
+
|
537 |
+
from sklearn.metrics import classification_report
|
538 |
+
|
539 |
+
preds_flat_test = np.argmax(predictions_test, axis=1).flatten()
|
540 |
+
print(classification_report(preds_flat_test, true_vals_test))
|
541 |
+
|
542 |
+
pred_final = []
|
543 |
+
|
544 |
+
for i, row in tqdm(test_df.iterrows(), total=test_df.shape[0]):
|
545 |
+
predictions = []
|
546 |
+
|
547 |
+
review = row["Review"]
|
548 |
+
encoded_data_test_single = tokenizer.batch_encode_plus(
|
549 |
+
[review],
|
550 |
+
add_special_tokens=config.add_special_tokens,
|
551 |
+
return_attention_mask=config.return_attention_mask,
|
552 |
+
pad_to_max_length=config.pad_to_max_length,
|
553 |
+
max_length=config.seq_length,
|
554 |
+
return_tensors=config.return_tensors
|
555 |
+
)
|
556 |
+
input_ids_test = encoded_data_test_single['input_ids']
|
557 |
+
attention_masks_test = encoded_data_test_single['attention_mask']
|
558 |
+
|
559 |
+
inputs = {'input_ids': input_ids_test.to(device),
|
560 |
+
'attention_mask':attention_masks_test.to(device),
|
561 |
+
}
|
562 |
+
|
563 |
+
with torch.no_grad():
|
564 |
+
outputs = model(**inputs)
|
565 |
+
|
566 |
+
logits = outputs[0]
|
567 |
+
logits = logits.detach().cpu().numpy()
|
568 |
+
predictions.append(logits)
|
569 |
+
predictions = np.concatenate(predictions, axis=0)
|
570 |
+
pred_final.append(np.argmax(predictions, axis=1).flatten()[0])
|
571 |
+
|
572 |
+
# add pred into test
|
573 |
+
test_df["pred"] = pred_final
|
574 |
+
# Add control column for easier wrong and right predictions
|
575 |
+
control = test_df.pred.values == test_df.label.values
|
576 |
+
test_df["control"] = control
|
577 |
+
# filtering false predictions
|
578 |
+
test_df = test_df[test_df.control == False]
|
579 |
+
test_df["pred_name"] = test_df.pred.apply(lambda x: label2name.get(x))
|
580 |
+
|
581 |
+
from sklearn.metrics import confusion_matrix
|
582 |
+
|
583 |
+
# We create a confusion matrix to better observe the classes that the model confuses.
|
584 |
+
pred_name_values = test_df.pred_name.values
|
585 |
+
label_values = test_df.label_name.values
|
586 |
+
confmat = confusion_matrix(label_values, pred_name_values, labels=list(name2label.keys()))
|
587 |
+
confmat
|
588 |
+
|
589 |
+
df_confusion_test = pd.crosstab(label_values, pred_name_values)
|
590 |
+
df_confusion_test
|
591 |
+
|
592 |
+
import matplotlib.pyplot as plt
|
593 |
+
import seaborn as sns
|
594 |
+
|
595 |
+
# Supposons que 'confmat' est votre matrice de confusion
|
596 |
+
|
597 |
+
fig, ax = plt.subplots(figsize=(10,10)) # changez la taille selon vos besoins
|
598 |
+
sns.heatmap(confmat, annot=True, fmt='d',
|
599 |
+
xticklabels=name2label.keys(), yticklabels=name2label.keys())
|
600 |
+
plt.ylabel('Vraies valeurs')
|
601 |
+
plt.xlabel('Prédictions')
|
602 |
+
plt.show()
|