使用Keras实现回归神经网络时预测值趋近单一值的问题排查

我尝试将CPFNN实现转为Keras版本，但预测效果极差——预测值几乎始终趋近单一数值。无论增加网络层数、神经元数量，还是改用卷积神经网络，情况都未改善。但使用对应模型在同一数据集上训练可得到良好结果。

实现代码 #

data1_file_path = 'GSE106648_data1.csv'
data2_file_path = 'GSE106648_data2.csv'


#read in training data
train = np.loadtxt(data1_file_path, skiprows=1, delimiter=',') 
print("Finish read training set")
#read in test data
test = np.loadtxt(data2_file_path, skiprows=1, delimiter=',')
print("Finish read test set")

#separate training/testing input features and labels
x_train = train[:,1:]
y_train = train[:,0].reshape(-1,1)
x_test = test[:,1:]
y_test = test[:,0].reshape(-1,1)

# define base model
def baseline_model():
    # create model
    model = Sequential()
    model.add(Dense(200, input_shape=(x_train.shape[1],),activation='LeakyReLU'))
    model.add(Dense(1))
    # Compile model
    model.compile(loss='mse', optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001))
    return model

baseline_model().fit(x_train,y_train,batch_size=20,epochs=500)
y_pred = baseline_model().predict(x_test)

结果展示 #

• 预测结果：
• 损失函数随epoch变化曲线：

内容的提问来源于stack exchange，提问作者Caterina