关于

情绪检测，是脑科学研究中的一个常见和热门的方向。在进行情绪检测的分类中，真实数据不足，经常导致情绪检测模型的性能不佳。因此，对数据进行增强，成为了一个提升下游任务的重要的手段。本项目通过DCGAN模型实现脑电信号的扩充。

 图片来源：https://www.medicalnewstoday.com/articles/seizure-eeg

工具

数据

方法实现

DCGAN速递：https://arxiv.org/abs/1511.06434

数据加载和预处理

import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
from sklearn.preprocessing import StandardScaler
from tensorflow.keras.utils import to_categorical
from sklearn.model_selection import train_test_split
import tensorflow as tf
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Dropout
from tensorflow.keras.layers import Embedding
from tensorflow.keras.layers import LSTM
from tensorflow.keras.optimizers import SGD
from sklearn.metrics import accuracy_score
from model_DCGAN import DCGAN
from tensorflow.keras.optimizers import Adam, SGD, RMSprop
from sklearn.utils import shuffle
from sklearn.ensemble import GradientBoostingClassifieruse_feature_reduction = Truetf.keras.backend.clear_session()df=pd.read_csv('dataset/emotions.csv')encode = ({'NEUTRAL': 0, 'POSITIVE': 1, 'NEGATIVE': 2} )
#new dataset with replaced values
df_encoded = df.replace(encode)print(df_encoded.head())
print(df_encoded['label'].value_counts()),x=df_encoded.drop(["label"]  ,axis=1)
y = df_encoded.loc[:,'label'].valuesscaler = StandardScaler()
scaler.fit(x)
x = scaler.transform(x)
y = to_categorical(y)x_train, x_test, y_train, y_test = train_test_split(x, y, test_size = 0.2, random_state = 4)if use_feature_reduction:# Feature reduction partest = GradientBoostingClassifier(n_estimators=10, learning_rate=0.1, random_state=0).fit(x_train,y_train.argmax(-1))# Obtain feature importance results from Gradient Boosting Regressorfeature_importance = est.feature_importances_epsilon_feature = 1e-2x_train = x_train[:, feature_importance > epsilon_feature]x_test = x_test[:, feature_importance > epsilon_feature]

设置DCGAN优化器

# setup optimzers
gen_optim = Adam(1e-4, beta_1=0.5)
disc_optim = RMSprop(5e-4)

 训练GAN生成类别0脑电数据

# generate samples for class 0
generator_class = 0
dcgan = DCGAN(gen_optim, disc_optim, noise_dim=100, dropout=0.3, input_dim=x_train.shape[2])
x_train_class_0 = x_train[y_train[:,generator_class]==1,:]
loss_history_class_0, acc_history_class_0, grads_history_class_0 = dcgan.train(x_train_class_0, epochs=100)
print("Class 0 fake samples are generating")
generator_class_0 = dcgan.generator
generated_samples_class_0, _ = dcgan.generate_fake_data(N=len(x_train_class_0))

  训练GAN生成类别1脑电数据

# generate samples for class 1
generator_class = 1
dcgan = DCGAN(gen_optim, disc_optim, noise_dim=100, dropout=0.3, input_dim=x_train.shape[2])
x_train_class_1 = x_train[y_train[:,generator_class]==1,:]
loss_history_class_1, acc_history_class_1, grads_history_class_1 = dcgan.train(x_train_class_1, epochs=100)
print("Class 1 fake samples are generating")
generator_class_1 = dcgan.generator
generated_samples_class_1, _ = dcgan.generate_fake_data(N=len(x_train_class_1))

 训练GAN生成类别2脑电数据

# generate samples for class 2
generator_class = 2
dcgan = DCGAN(gen_optim, disc_optim, noise_dim=100, dropout=0.3, input_dim=x_train.shape[2])
x_train_class_2 = x_train[y_train[:,generator_class]==1,:]
loss_history_class_2, acc_history_class_2, grads_history_class_2 = dcgan.train(x_train_class_2,epochs=100)
print("Class 2 fake samples are generating")
generator_class_2 = dcgan.generator
generated_samples_class_2, _ = dcgan.generate_fake_data(N=len(x_train_class_2))

合成数据融入真实训练数据集

generated_samples = np.concatenate((generated_samples_class_0,generated_samples_class_1,generated_samples_class_2),axis=0)
generated_y =np.concatenate((np.zeros((len(x_train_class_0),),dtype=np.int32),np.ones((len(x_train_class_1),),dtype=np.int32),2 * np.ones((len(x_train_class_2),),dtype=np.int32)),axis=0)generated_y = to_categorical(generated_y)x_train_all = np.concatenate((x_train,generated_samples),axis=0)
y_train_all = np.concatenate((y_train,generated_y), axis=0)#shuffle training data
x_train_all, y_train_all = shuffle(x_train_all,y_train_all)

 基于数据增强的LSTM模型情绪检测

model = Sequential()
model.add(LSTM(64, input_shape=(1,x_train_all.shape[2]),activation="relu",return_sequences=True))
model.add(Dropout(0.2))
model.add(LSTM(32,activation="sigmoid"))
model.add(Dropout(0.2))
model.add(Dense(3, activation='sigmoid'))
model.compile(loss = 'categorical_crossentropy', optimizer = "adam", metrics = ['accuracy'])
model.summary()history = model.fit(x_train_all, y_train_all, epochs = 250, validation_data= (x_test, y_test))
score, acc = model.evaluate(x_test, y_test)pred = model.predict(x_test)
predict_classes = np.argmax(pred,axis=1)
expected_classes = np.argmax(y_test,axis=1)
print(expected_classes.shape)
print(predict_classes.shape)
correct = accuracy_score(expected_classes,predict_classes)
print(f"Test Accuracy: {correct}")

已附DCGAN模型

相关项目和代码问题，欢迎沟通交流。