Spaces:
Runtime error
Runtime error
AlejandraVento2
commited on
Commit
•
70ead2f
1
Parent(s):
8103ee5
sentiment analist
Browse files
README.md
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
---
|
2 |
title: Analisis De Sentimientos
|
3 |
-
emoji:
|
4 |
colorFrom: yellow
|
5 |
colorTo: indigo
|
6 |
sdk: streamlit
|
@@ -9,4 +9,57 @@ app_file: app.py
|
|
9 |
pinned: false
|
10 |
---
|
11 |
|
12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
title: Analisis De Sentimientos
|
3 |
+
emoji: 🥺😡
|
4 |
colorFrom: yellow
|
5 |
colorTo: indigo
|
6 |
sdk: streamlit
|
|
|
9 |
pinned: false
|
10 |
---
|
11 |
|
12 |
+
|
13 |
+
# Modelo Clasificatorio de sentiminentos
|
14 |
+
|
15 |
+
Mi proyecto es un modelo multi-label que clasifica sentimientos en tristeza (sadness), enojo(anger), alegria(joy), miedo(fear) y sorpresa(surprise) fue entrenado y validado a partir de una tabla de texto y fue modelado con BERT y PyTorch Lightning.
|
16 |
+
|
17 |
+
## Decisions
|
18 |
+
|
19 |
+
`1` Clasificar sentimientos (anger, sadness, joy, fear, surprise, neutral).
|
20 |
+
|
21 |
+
`2` Usar de referencia estos modelos:
|
22 |
+
|
23 |
+
- https://github.com/curiousily/Getting-Things-Done-with-Pytorch/blob/master/11.multi-label-text-classification-with-bert.ipynb
|
24 |
+
- https://www.youtube.com/watch?v=UJGxCsZgalA&ab_channel=VenelinValkov
|
25 |
+
- https://github.com/theartificialguy/NLP-with-Deep-Learning/blob/master/BERT/Multi%20Label%20Text%20Classification%20using%20BERT%20PyTorch/bert_multilabel_pytorch_standard.ipynb
|
26 |
+
|
27 |
+
`3` Añadir un dropout en las capas para que el modelo no se sobreajuste.
|
28 |
+
|
29 |
+
`4` Anadir checkpoints en el entrenamiento para evitar que la ram se ocupe totalmente.
|
30 |
+
|
31 |
+
`5` Utilizar la funcion de activacion sigmoid en al ultima capa para un mejor resultado final.
|
32 |
+
|
33 |
+
`6` Utilizar optimizador AdamW y funcion de perdida binary_crossentropy con BCELoss como criterio, que son de los mas recomendados y usados para estos modelos.
|
34 |
+
|
35 |
+
`7` Utilizar 1625 ensayos de entrenamiento y 86 ensayos de validacion.
|
36 |
+
|
37 |
+
## Data Sources
|
38 |
+
|
39 |
+
Use la tabla de entrenamiento y de validacion sobre sentimientos de esta fuente [messages_train_ready_for_WS.tsv](https://github.com/caisa-lab/wassa-empathy-adapters/tree/main/data).
|
40 |
+
|
41 |
+
## Features
|
42 |
+
|
43 |
+
La entrada es texto a trave de un campo de texto, para predecir y/o clasificar los sentimientos que mas se asemejan al texto.
|
44 |
+
|
45 |
+
## Data Collection
|
46 |
+
|
47 |
+
Decidi utilizar esta coleccion de datos porque tiene una gran variedad de texto en diferentes contextos y con una gran variedad de sentimientos que se midieron a aprtir de varias columnas de metricas.
|
48 |
+
|
49 |
+
## Value Proposition
|
50 |
+
|
51 |
+
Puede ser utilizado en aplicaciones de reconocimiento de sentimientos, como filtro de textos empresariales por ejemplo, en programas psicologicos que permitan ayudar a personas que tengan algun problema en entender los sentimientos de otros o personas con discapacidad en general, por ejemplo con discapacidad en el habla que se comunica a traves de texto como entrada a un sistema que genera sonidos, podria modificar el tono de voz a partir de esta prediccion y reflejar los sentimientos de la persona.
|
52 |
+
|
53 |
+
# Environment requirements to run the model
|
54 |
+
|
55 |
+
Transformers 4.5.1
|
56 |
+
|
57 |
+
Pytorch lightning 1.2.8
|
58 |
+
|
59 |
+
Numpy
|
60 |
+
|
61 |
+
Pandas
|
62 |
+
|
63 |
+
Torch
|
64 |
+
|
65 |
+
Sklearn
|
app.py
ADDED
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
from transformers import pipeline
|
3 |
+
import run_sentiment_analysis from model.py
|
4 |
+
|
5 |
+
st.title("Analisis de Sentimientos")
|
6 |
+
|
7 |
+
txt = st.text_area(label="Please write what you want to analyze...")
|
8 |
+
predictions = run_sentiment_analysis(txt)
|
9 |
+
|
10 |
+
for prediction in predictions:
|
11 |
+
st.write(prediction)
|
12 |
+
col1, col2 = st.columns(2)
|
13 |
+
|
14 |
+
image = Image.open(file_name)
|
15 |
+
col1.image(image, use_column_width=True)
|
16 |
+
predictions = pipeline(image)
|
17 |
+
|
18 |
+
col2.header("Probabilities")
|
19 |
+
for p in predictions:
|
20 |
+
col2.subheader(f"{ p['label'] }: { round(p['score'] * 100, 1)}%")
|
model.py
ADDED
@@ -0,0 +1,361 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""
|
3 |
+
Automatically generated by Colaboratory.
|
4 |
+
|
5 |
+
Original file is located at
|
6 |
+
https://colab.research.google.com/drive/193Qwk9yyPHgI0H84JJOchTovg_CELJuw
|
7 |
+
"""
|
8 |
+
|
9 |
+
import pandas as pd
|
10 |
+
import numpy as np
|
11 |
+
import torch
|
12 |
+
import torch.nn as nn
|
13 |
+
from torch.utils.data import Dataset, DataLoader
|
14 |
+
from transformers import BertTokenizer, BertModel, AdamW, get_linear_schedule_with_warmup
|
15 |
+
import pytorch_lightning as pl
|
16 |
+
from pytorch_lightning.callbacks import ModelCheckpoint, EarlyStopping
|
17 |
+
from sklearn.model_selection import train_test_split
|
18 |
+
|
19 |
+
RANDOM_SEED = 42
|
20 |
+
np.random.seed(RANDOM_SEED)
|
21 |
+
torch.manual_seed(RANDOM_SEED)
|
22 |
+
|
23 |
+
# Preparing training data
|
24 |
+
train_file_path = '/content/sample_data/train_data.csv'
|
25 |
+
train_data = pd.read_csv(train_file_path)
|
26 |
+
|
27 |
+
filtro = (train_data['emotion'] == 'anger') | (train_data['emotion'] == 'fear') | (train_data['emotion'] == 'joy') | (train_data['emotion'] == 'sadness') | (train_data['emotion'] == 'neutral') | (train_data['emotion'] == 'surprise')
|
28 |
+
df = train_data[filtro]
|
29 |
+
|
30 |
+
angerColumn = []
|
31 |
+
fearColumn = []
|
32 |
+
surpriseColumn = []
|
33 |
+
sadnessColumn = []
|
34 |
+
joyColumn = []
|
35 |
+
neutralColumn = []
|
36 |
+
|
37 |
+
for e in df['emotion']:
|
38 |
+
if e == 'anger':
|
39 |
+
angerColumn.append(1)
|
40 |
+
joyColumn.append(0)
|
41 |
+
sadnessColumn.append(0)
|
42 |
+
fearColumn.append(0)
|
43 |
+
surpriseColumn.append(0)
|
44 |
+
neutralColumn.append(0)
|
45 |
+
elif e == 'joy':
|
46 |
+
joyColumn.append(1)
|
47 |
+
angerColumn.append(0)
|
48 |
+
sadnessColumn.append(0)
|
49 |
+
fearColumn.append(0)
|
50 |
+
surpriseColumn.append(0)
|
51 |
+
neutralColumn.append(0)
|
52 |
+
elif e == 'sadness':
|
53 |
+
sadnessColumn.append(1)
|
54 |
+
angerColumn.append(0)
|
55 |
+
joyColumn.append(0)
|
56 |
+
fearColumn.append(0)
|
57 |
+
surpriseColumn.append(0)
|
58 |
+
neutralColumn.append(0)
|
59 |
+
elif e == 'fear':
|
60 |
+
fearColumn.append(1)
|
61 |
+
angerColumn.append(0)
|
62 |
+
joyColumn.append(0)
|
63 |
+
sadnessColumn.append(0)
|
64 |
+
surpriseColumn.append(0)
|
65 |
+
neutralColumn.append(0)
|
66 |
+
elif e == 'surprise':
|
67 |
+
surpriseColumn.append(1)
|
68 |
+
angerColumn.append(0)
|
69 |
+
joyColumn.append(0)
|
70 |
+
sadnessColumn.append(0)
|
71 |
+
fearColumn.append(0)
|
72 |
+
neutralColumn.append(0)
|
73 |
+
elif e == 'neutral':
|
74 |
+
neutralColumn.append(1)
|
75 |
+
surpriseColumn.append(0)
|
76 |
+
angerColumn.append(0)
|
77 |
+
joyColumn.append(0)
|
78 |
+
sadnessColumn.append(0)
|
79 |
+
fearColumn.append(0)
|
80 |
+
|
81 |
+
df['anger'] = angerColumn
|
82 |
+
df['fear'] = fearColumn
|
83 |
+
df['surprise'] = surpriseColumn
|
84 |
+
df['joy'] = joyColumn
|
85 |
+
df['sadness'] = sadnessColumn
|
86 |
+
df['neutral'] = neutralColumn
|
87 |
+
|
88 |
+
df.drop(['emotion', 'message_id', 'response_id', 'article_id', 'empathy', 'distress',
|
89 |
+
'empathy_bin', 'distress_bin', 'gender', 'education','race', 'age','income','personality_conscientiousness',
|
90 |
+
'personality_openess','personality_extraversion','personality_agreeableness','personality_stability',
|
91 |
+
'iri_perspective_taking','iri_personal_distress', 'iri_fantasy', 'iri_empathatic_concern','raw_input_emotions'],
|
92 |
+
axis=1, inplace=True)
|
93 |
+
|
94 |
+
print(df.head())
|
95 |
+
|
96 |
+
train_df, val_df = train_test_split(df, test_size=0.05)
|
97 |
+
train_df.shape, val_df.shape
|
98 |
+
|
99 |
+
LABEL_COLUMNS = ['anger','joy','fear','surprise','sadness', 'neutral']
|
100 |
+
|
101 |
+
sample_row = train_df.iloc[16]
|
102 |
+
sample_comment = sample_row.essay
|
103 |
+
sample_labels = sample_row[LABEL_COLUMNS]
|
104 |
+
print(sample_comment)
|
105 |
+
print(sample_labels.to_dict())
|
106 |
+
|
107 |
+
BERT_MODEL_NAME = 'bert-base-cased'
|
108 |
+
tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_NAME)
|
109 |
+
|
110 |
+
encoding = tokenizer.encode_plus(
|
111 |
+
sample_comment,
|
112 |
+
add_special_tokens=True,
|
113 |
+
max_length=512,
|
114 |
+
return_token_type_ids=False,
|
115 |
+
padding="max_length",
|
116 |
+
return_attention_mask=True,
|
117 |
+
return_tensors='pt',
|
118 |
+
)
|
119 |
+
|
120 |
+
encoding.keys()
|
121 |
+
|
122 |
+
encoding["input_ids"].shape, encoding["attention_mask"].shape
|
123 |
+
|
124 |
+
encoding["input_ids"].squeeze()[:20]
|
125 |
+
|
126 |
+
encoding["attention_mask"].squeeze()[:20]
|
127 |
+
|
128 |
+
print(tokenizer.convert_ids_to_tokens(encoding["input_ids"].squeeze())[:20])
|
129 |
+
|
130 |
+
class EmotionDataset(Dataset):
|
131 |
+
def __init__(
|
132 |
+
self,
|
133 |
+
data: pd.DataFrame,
|
134 |
+
tokenizer: BertTokenizer,
|
135 |
+
max_token_len: int = 128
|
136 |
+
):
|
137 |
+
self.tokenizer = tokenizer
|
138 |
+
self.data = data
|
139 |
+
self.max_token_len = max_token_len
|
140 |
+
|
141 |
+
def __len__(self):
|
142 |
+
return len(self.data)
|
143 |
+
|
144 |
+
def __getitem__(self, index: int):
|
145 |
+
data_row = self.data.iloc[index]
|
146 |
+
|
147 |
+
comment_text = data_row.essay
|
148 |
+
labels = data_row[LABEL_COLUMNS]
|
149 |
+
|
150 |
+
encoding = self.tokenizer.encode_plus(
|
151 |
+
comment_text,
|
152 |
+
add_special_tokens=True,
|
153 |
+
max_length=self.max_token_len,
|
154 |
+
return_token_type_ids=False,
|
155 |
+
padding="max_length",
|
156 |
+
truncation=True,
|
157 |
+
return_attention_mask=True,
|
158 |
+
return_tensors='pt',
|
159 |
+
)
|
160 |
+
|
161 |
+
return dict(
|
162 |
+
comment_text=comment_text,
|
163 |
+
input_ids=encoding["input_ids"].flatten(),
|
164 |
+
attention_mask=encoding["attention_mask"].flatten(),
|
165 |
+
labels=torch.FloatTensor(labels)
|
166 |
+
)
|
167 |
+
|
168 |
+
bert_model = BertModel.from_pretrained(BERT_MODEL_NAME, return_dict=True)
|
169 |
+
|
170 |
+
train_dataset = EmotionDataset(train_df,tokenizer)
|
171 |
+
sample_item = train_dataset[0]
|
172 |
+
sample_item.keys()
|
173 |
+
|
174 |
+
sample_batch = next(iter(DataLoader(train_dataset, batch_size=8, num_workers=2)))
|
175 |
+
sample_batch["input_ids"].shape, sample_batch["attention_mask"].shape
|
176 |
+
|
177 |
+
output = bert_model(sample_batch["input_ids"], sample_batch["attention_mask"])
|
178 |
+
|
179 |
+
output.last_hidden_state.shape, output.pooler_output.shape
|
180 |
+
|
181 |
+
class EmotionDataModule(pl.LightningDataModule):
|
182 |
+
|
183 |
+
def __init__(self, train_df, test_df, tokenizer, batch_size=8, max_token_len=128):
|
184 |
+
super().__init__()
|
185 |
+
self.batch_size = batch_size
|
186 |
+
self.train_df = train_df
|
187 |
+
self.test_df = test_df
|
188 |
+
self.tokenizer = tokenizer
|
189 |
+
self.max_token_len = max_token_len
|
190 |
+
|
191 |
+
def setup(self, stage=None):
|
192 |
+
self.train_dataset = EmotionDataset(
|
193 |
+
self.train_df,
|
194 |
+
self.tokenizer,
|
195 |
+
self.max_token_len
|
196 |
+
)
|
197 |
+
|
198 |
+
self.test_dataset = EmotionDataset(
|
199 |
+
self.test_df,
|
200 |
+
self.tokenizer,
|
201 |
+
self.max_token_len
|
202 |
+
)
|
203 |
+
|
204 |
+
def train_dataloader(self):
|
205 |
+
return DataLoader(
|
206 |
+
self.train_dataset,
|
207 |
+
batch_size=self.batch_size,
|
208 |
+
shuffle=True,
|
209 |
+
num_workers=2
|
210 |
+
)
|
211 |
+
|
212 |
+
def val_dataloader(self):
|
213 |
+
return DataLoader(
|
214 |
+
self.test_dataset,
|
215 |
+
batch_size=self.batch_size,
|
216 |
+
num_workers=2
|
217 |
+
)
|
218 |
+
|
219 |
+
def test_dataloader(self):
|
220 |
+
return DataLoader(
|
221 |
+
self.test_dataset,
|
222 |
+
batch_size=self.batch_size,
|
223 |
+
num_workers=2
|
224 |
+
)
|
225 |
+
|
226 |
+
N_EPOCHS = 10
|
227 |
+
BATCH_SIZE = 12
|
228 |
+
MAX_TOKEN_COUNT = 512
|
229 |
+
|
230 |
+
data_module = EmotionDataModule(
|
231 |
+
train_df,
|
232 |
+
val_df,
|
233 |
+
tokenizer,
|
234 |
+
batch_size=BATCH_SIZE,
|
235 |
+
max_token_len=MAX_TOKEN_COUNT
|
236 |
+
)
|
237 |
+
|
238 |
+
class EmotionTagger(pl.LightningModule):
|
239 |
+
def __init__(self, n_classes: int, n_training_steps=None, n_warmup_steps=None):
|
240 |
+
super().__init__()
|
241 |
+
self.bert = BertModel.from_pretrained(BERT_MODEL_NAME, return_dict=True)
|
242 |
+
self.classifier = nn.Linear(self.bert.config.hidden_size, n_classes)
|
243 |
+
self.n_training_steps = n_training_steps
|
244 |
+
self.n_warmup_steps = n_warmup_steps
|
245 |
+
self.criterion = nn.BCELoss()
|
246 |
+
|
247 |
+
def forward(self, input_ids, attention_mask, labels=None):
|
248 |
+
output = self.bert(input_ids, attention_mask=attention_mask)
|
249 |
+
output = self.classifier(output.pooler_output)
|
250 |
+
output = torch.sigmoid(output)
|
251 |
+
loss = 0
|
252 |
+
if labels is not None:
|
253 |
+
loss = self.criterion(output, labels)
|
254 |
+
return loss, output
|
255 |
+
|
256 |
+
def training_step(self, batch, batch_idx):
|
257 |
+
input_ids = batch["input_ids"]
|
258 |
+
attention_mask = batch["attention_mask"]
|
259 |
+
labels = batch["labels"]
|
260 |
+
loss, outputs = self(input_ids, attention_mask, labels)
|
261 |
+
self.log("train_loss", loss, prog_bar=True, logger=True)
|
262 |
+
return {"loss": loss, "predictions": outputs, "labels": labels}
|
263 |
+
|
264 |
+
def validation_step(self, batch, batch_idx):
|
265 |
+
input_ids = batch["input_ids"]
|
266 |
+
attention_mask = batch["attention_mask"]
|
267 |
+
labels = batch["labels"]
|
268 |
+
loss, outputs = self(input_ids, attention_mask, labels)
|
269 |
+
self.log("val_loss", loss, prog_bar=True, logger=True)
|
270 |
+
return loss
|
271 |
+
|
272 |
+
def test_step(self, batch, batch_idx):
|
273 |
+
input_ids = batch["input_ids"]
|
274 |
+
attention_mask = batch["attention_mask"]
|
275 |
+
labels = batch["labels"]
|
276 |
+
loss, outputs = self(input_ids, attention_mask, labels)
|
277 |
+
self.log("test_loss", loss, prog_bar=True, logger=True)
|
278 |
+
return loss
|
279 |
+
|
280 |
+
for i, name in enumerate(LABEL_COLUMNS):
|
281 |
+
class_roc_auc = pytorch_lightning.metrics.functional.auroc(predictions[:, i], labels[:, i])
|
282 |
+
self.logger.experiment.add_scalar(f"{name}_roc_auc/Train", class_roc_auc, self.current_epoch)
|
283 |
+
|
284 |
+
def configure_optimizers(self):
|
285 |
+
optimizer = AdamW(self.parameters(), lr=2e-5)
|
286 |
+
|
287 |
+
scheduler = get_linear_schedule_with_warmup(
|
288 |
+
optimizer,
|
289 |
+
num_warmup_steps=self.n_warmup_steps,
|
290 |
+
num_training_steps=self.n_training_steps
|
291 |
+
)
|
292 |
+
|
293 |
+
return dict(
|
294 |
+
optimizer=optimizer,
|
295 |
+
lr_scheduler=dict(
|
296 |
+
scheduler=scheduler,
|
297 |
+
interval='step'
|
298 |
+
)
|
299 |
+
)
|
300 |
+
|
301 |
+
steps_per_epoch=len(train_df) // BATCH_SIZE
|
302 |
+
total_training_steps = steps_per_epoch * N_EPOCHS
|
303 |
+
warmup_steps = total_training_steps // 5
|
304 |
+
|
305 |
+
model = EmotionTagger(
|
306 |
+
n_classes=len(LABEL_COLUMNS),
|
307 |
+
n_warmup_steps=warmup_steps,
|
308 |
+
n_training_steps=total_training_steps
|
309 |
+
)
|
310 |
+
|
311 |
+
!rm -rf lightning_logs/
|
312 |
+
!rm -rf checkpoints/
|
313 |
+
|
314 |
+
checkpoint_callback = ModelCheckpoint(
|
315 |
+
dirpath="checkpoints",
|
316 |
+
filename="best-checkpoint",
|
317 |
+
save_top_k=1,
|
318 |
+
verbose=True,
|
319 |
+
monitor="val_loss",
|
320 |
+
mode="min"
|
321 |
+
)
|
322 |
+
|
323 |
+
early_stopping_callback = EarlyStopping(monitor='val_loss', patience=2)
|
324 |
+
|
325 |
+
trainer = pl.Trainer(
|
326 |
+
max_epochs=N_EPOCHS,
|
327 |
+
callbacks=[early_stopping_callback,checkpoint_callback],)
|
328 |
+
|
329 |
+
trainer.fit(model, data_module)
|
330 |
+
|
331 |
+
trained_model = EmotionTagger.load_from_checkpoint(
|
332 |
+
trainer.checkpoint_callback.best_model_path,
|
333 |
+
n_classes=len(LABEL_COLUMNS)
|
334 |
+
)
|
335 |
+
trained_model.eval()
|
336 |
+
trained_model.freeze()
|
337 |
+
|
338 |
+
|
339 |
+
def run_sentiment_analysis (txt) :
|
340 |
+
THRESHOLD = 0.5
|
341 |
+
|
342 |
+
encoding = tokenizer.encode_plus(
|
343 |
+
txt,
|
344 |
+
add_special_tokens=True,
|
345 |
+
max_length=512,
|
346 |
+
return_token_type_ids=False,
|
347 |
+
padding="max_length",
|
348 |
+
return_attention_mask=True,
|
349 |
+
return_tensors='pt',
|
350 |
+
)
|
351 |
+
|
352 |
+
_, test_prediction = trained_model(encoding["input_ids"], encoding["attention_mask"])
|
353 |
+
test_prediction = test_prediction.flatten().numpy()
|
354 |
+
|
355 |
+
predictions = []
|
356 |
+
|
357 |
+
for label, prediction in zip(LABEL_COLUMNS, test_prediction):
|
358 |
+
if prediction < THRESHOLD:
|
359 |
+
continue
|
360 |
+
predictions.append("{label}: {prediction}")
|
361 |
+
return predictions
|
requirements.txt
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
transformers==4.5.1
|
2 |
+
pytorch-lightning==1.2.8
|
3 |
+
torch
|
4 |
+
pandas
|
5 |
+
numpy
|
6 |
+
sklearn
|