Question

为了通过TensorFlow后端了解Keras以进行进一步的研究项目，我尝试为简单的分类问题实现NN。我只想区分randomly distributed points in 2D into two categories depending on their coordinates（颜色表示类别）。生成数据的相关代码是：

import numpy as np    
np.random.seed(40)

def createData(N=120, M=75):
    train_x1 = np.random.random(size=N)
    train_x2 = np.random.random(size=N)

    test_x1 = np.random.random(size=M)
    test_x2 = np.random.random(size=M)

    train_x = np.zeros((N, 2))
    train_y = np.zeros((N, 1))

    test_x = np.zeros((M, 2))
    test_y = np.zeros((M, 1))

    for i in range(N):
        train_x[i][0] = train_x1[i]
        train_x[i][1] = train_x2[i]
        if train_x1[i] < 0.5:
            if train_x2[i] < 0.5:
                train_y[i][0] = 1
            else:
                train_y[i][0] = 2
        else:
            if train_x2[i] < 0.5:
                train_y[i][0] = 2
            else:
                train_y[i][0] = 1

    for j in range(M):
        test_x[j][0] = test_x1[j]
        test_x[j][1] = test_x2[j]
        if test_x1[j] < 0.5:
            if test_x2[j] < 0.5:
                test_y[j][0] = 1
            else:
                test_y[j][0] = 2
        else:
            if test_x2[j] < 0.5:
                test_y[j][0] = 2
            else:
                test_y[j][0] = 1

    return train_x, train_y, test_x, test_y

我的神经网络代码如下：

from keras.models import Sequential
from keras.layers import Dense
import numpy as np
import matplotlib.pyplot as plt

X, Y, x, y = createData()

# Model
model = Sequential()
model.add(Dense(4, input_dim=2, activation='relu'))
model.add(Dense(10, activation='sigmoid'))
model.add(Dense(1, activation='sigmoid'))

# Compile
model.compile(loss='binary_crossentropy', optimizer='sgd', metrics=['accuracy'])

# Fit
model.fit(X, Y, epochs=500, batch_size=25)

# Evaluation
scores = model.evaluate(X, Y)
print("\n%s: %.2f%%" % (model.metrics_names[1], scores[1]*100))

# Predictions
predictions = model.predict(x)
rounded = [round(z[0]) for z in predictions]
print(rounded)

我的问题是，使用上述种子和配置，神经网络达到47.5％的准确度。它指定all test coordinates to only one category。但是，其他配置（更多层，更多/更少的神经元，其他激活函数，其他损失函数等）会产生类似的结果。

＆＃34;纪元监视器的最后一行＆＃34;在shell上看起来像这样：

Epoch 490/500
120/120 [==============================] - 0s - loss: -8.3351 - acc: 0.4750     
Epoch 491/500
120/120 [==============================] - 0s - loss: -8.1866 - acc: 0.4750     
Epoch 492/500
120/120 [==============================] - 0s - loss: -8.3524 - acc: 0.4750     
Epoch 493/500
120/120 [==============================] - 0s - loss: -8.2608 - acc: 0.4750     
Epoch 494/500
120/120 [==============================] - 0s - loss: -8.3269 - acc: 0.4750     
Epoch 495/500
120/120 [==============================] - 0s - loss: -8.2039 - acc: 0.4750     
Epoch 496/500
120/120 [==============================] - 0s - loss: -8.1786 - acc: 0.4750     
Epoch 497/500
120/120 [==============================] - 0s - loss: -8.2488 - acc: 0.4750     
Epoch 498/500
120/120 [==============================] - 0s - loss: -8.3090 - acc: 0.4750     
Epoch 499/500
120/120 [==============================] - 0s - loss: -8.3457 - acc: 0.4750     
Epoch 500/500
120/120 [==============================] - 0s - loss: -8.1235 - acc: 0.4750

如何改善神经网络以摆脱这种愚蠢的行为？非常感谢您的任何意见和建议！

Answer 1

首先，我会检查您的输入数据是否已正确创建/标记。尝试打印出几个值以确保它有意义。

其次，当学习率很高时，我通常会看到这个问题。您可以尝试删除中心有10个神经元的隐藏层或降低学习率。我认为很多层的网络对于这个问题来说太过分了。尝试将此用于编译行。

model.compile(loss='binary_crossentropy', optimizer=keras.optimizers.Adam(lr=0.0001), metrics=['accuracy'])

学习率为0.001是Adam的默认值，但我发现如果0.001不会使用0.0001收敛，那么

Keras中的简单NN - 分类错误

1 个答案: