sos，CNN训练二分类问题，训练过程准确度反复变化应该怎么解决？

Question

如题，我所用的CNN代码和训练结果见下：我用了Adam和RMSprop，感觉可能用adam激活器要更好一点？但是感觉训练的也不是很令人满意，有大佬可以帮忙指点一下要怎么改进吗？

import tensorflow as tf
from keras.models import Model
from keras.layers.core import Dense, Dropout
from keras.layers.convolutional import Convolution1D, MaxPooling1D
from keras.layers import Embedding,BatchNormalization,LSTM, Bidirectional,Masking, Input, Multiply, Activation, Lambda, Add,Reshape
from keras import models, Sequential
from keras import backend as K
from sklearn import metrics

def cnn_bilstm_am_model(filters=32, kernel_size=3, pooling_size=2, lstm_units=32, embedding_vec=encoding_vectors, attention_size=50):
    model = Sequential()
    model.add(Embedding(len(embedding_vec), len(embedding_vec[0]), weights=[embedding_vec],input_length=2000,trainable=False) )

    model.add(Convolution1D(32, kernel_size, padding='same', activation='relu', use_bias=False) )
    model.add(BatchNormalization())
    model.add(MaxPooling1D(pool_size=pooling_size, strides=1))
    
    model.add(Convolution1D(64, kernel_size, padding='same', activation='relu', use_bias=False) )
    model.add(BatchNormalization())
    model.add(MaxPooling1D(pool_size=pooling_size, strides=1))   
    
    #model.add(Bidirectional(LSTM(lstm_units, dropout=0.1, return_sequences=True) ) )
    
    model.add(tf.keras.layers.Flatten())
    model.add(Dense(256,activation='sigmoid'))
    model.add(Dropout(0.5))
    model.add(Dense(2))
    
    from keras import optimizers
    optim = optimizers.RMSprop(lr=0.001)
    model.compile(loss='binary_crossentropy', optimizer=optim , metrics=['accuracy'],)
        
    return model  


model = cnn_bilstm_am_model()
model.summary()

from keras.callbacks import ModelCheckpoint,EarlyStopping,ReduceLROnPlateau

for fold_num, (train_indices, test_indices) in enumerate(folds):
    if fold_num == 0:
        print('Evaluating fold{}'.format(fold_num+1))
        x_train = X[train_indices]
        y_train = y[train_indices]
        x_test = X[test_indices]
        y_test = y[test_indices]
        
 filepath='./best_model/rmsprop_2cnn_2dense_weights_of_fold{}.best.hdf5'.format(fold_num)
 checkpoint = ModelCheckpoint(filepath, monitor='val_accuracy',  verbose=1,save_best_only=True,mode='max',period=1)
 reducelr = ReduceLROnPlateau(monitor='val_accuracy', factor=0.8,patience=10,verbose=1,mode="max",min_lr=0.0001)
 callbacks_list = [checkpoint,reducelr]   
 train_history = model.fit(x_train,y_train,validation_split=0.1,epochs=100,batch_size=100,verbose=2,callbacks = callbacks_list)

激活器为RMSprop

激活器为Adam

Answer 1

1.尝试调整lr学习率，减小点，建议尝试adam。
2.尝试调大epoch，改变batch_size大小，一般为8的倍数比较好。