按十位分类整数数据

import numpy as np import tensorflow as tf def set_x_data(): x_data = np.array([[50] , [60] , [70] , [80] , [90] , [110] , [120] , [130] , [140] , [150] , [160] , [170] , [180] , [190] , [200] , [210] , [220] , [230] , [240] , [250] , [260] , [270] , [280] , [290] , [300] , [310] , [320] , [330] , [340] , [350] , [360] , [370] , [380] , [390]]) return x_data def set_y_data(x): y_data = np.array([[0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [0, 1] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0] , [1, 0]]) return y_data def set_bias(efficiency): arr = np.array([efficiency]) return arr W1 = tf.Variable(tf.random_normal([1, 5]), name='weight1') W2 = tf.Variable(tf.random_normal([5, 5]), name='weight2') W3 = tf.Variable(tf.random_normal([5, 5]), name='weight3') W4 = tf.Variable(tf.random_normal([5, 5]), name='weight4') W5 = tf.Variable(tf.random_normal([5, 2]), name='weight5') def inference(input, b): hidden_layer1 = tf.sigmoid(tf.matmul(input, W1) + b) hidden_layer2 = tf.sigmoid(tf.matmul(hidden_layer1, W2) + b) hidden_layer3 = tf.sigmoid(tf.matmul(hidden_layer2, W3) + b) hidden_layer4 = tf.sigmoid(tf.matmul(hidden_layer3, W4) + b) out_layer = tf.nn.softmax(tf.matmul(hidden_layer4, W5) + b) return out_layer def loss(hypothesis, y): cross_entropy = tf.reduce_mean(-tf.reduce_sum(y * tf.log(hypothesis), reduction_indices=[1])) return cross_entropy def train(loss): optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.1) train = optimizer.minimize(loss) return train x_data = set_x_data(1) y_data = set_y_data(0) b_data = set_bias(0.8) x= tf.placeholder(tf.float32, shape=[None, 1]) y= tf.placeholder(tf.float32, shape=[None, 2]) b = tf.placeholder(tf.float32, shape=[None]) hypothesis = inference(x, b) loss = loss(hypothesis, y) train = train(loss) sess = tf.Session() init = tf.global_variables_initializer() sess.run(init) print(sess.run(W1)) for step in range(2000): sess.run(train, feed_dict={x:x_data, y:y_data, b:b_data}) print(sess.run(W1)) print(sess.run(hypothesis, feed_dict={x:np.array([[1000]]), b:b_data}))

1条回答

网友

1楼 · 发布于 2024-04-26 10:19:46

上述网络的训练中存在一些问题，但是只要做一些更改，就可以实现一个获得this decision function的网络

（The plot in the link显示2级的分数，即如果x>；200）

此网络中有待改进的问题列表：

培训数据非常缺乏（只有34分！）这通常太小，尤其是对于您的案例中的5层网络。在网络中，通常需要比参数多得多的输入样本。尝试添加更多的输入值并减少层的数量（如下面的代码所示-我使用了浮点而不是整数来获得更多的点，但我认为它仍然是兼容的）。
输入范围通常需要缩放（下面我尝试了一个除以常数的超简单缩放）。这是因为您通常希望避免高范围的变量（特别是当您通过具有软最大非线性的许多层时，这将破坏包含在非常高或非常低的值中的信息）。在更高级的情况下，您可能需要执行最小-最大缩放或z-分数。
尝试更多的时代（并尝试绘制损失函数值的演变）。在给定的历元数下，损失函数的优化没有收敛。下面我做了10倍多的纪元。看看下面的代码现在是如何在this plot中聚合的（看看2000个时代是如何不够的）：
有帮助的是对（x，y）数据进行洗牌。虽然这在这种情况下并不重要，但它收敛得更快（见Le Cun的论文“Efficient Backprop”）。在更严重的例子中，它通常是需要的。
重要的是，我想你希望b是一个参数，而不是一个常数，不是吗？网络的偏差通常也与乘性权重一起优化。（另外，对所有隐藏层使用单一的共享偏移也是不常见的。）

下面是代码。注意，可能会有进一步的改进，但这些技巧最终会得到所需的决策函数。你知道吗

我添加了一些内联注释，以指示相对于原始的更改。我希望你觉得这些建议很有见地！你知道吗

代码：

import numpy as np
import tensorflow as tf

# I've modified the functions set_x_data and set_y_data
# so as to generate a larger set of numbers. 

# Generate a range of numbers from 50 to 390
def set_x_data():
    x_data = np.arange(50, 390, 0.1)
    return x_data[:,None]

# Assign labels depending on x_data
def set_y_data(x_data):
    ydata1 = x_data >= 200
    ydata2 = x_data < 200
    return np.hstack((ydata1, ydata2))

def set_bias(efficiency):
    arr = np.array([efficiency])

    return arr

# Let's keep W1 and W5 (one hidden layer only)
# BTW, in this problem you could do with 0 hidden layers. But keeping
# 1 to show it works
W1 = tf.Variable(tf.random_normal([1, 5]), name='weight1')
W5 = tf.Variable(tf.random_normal([5, 2]), name='weight5')

# BTW, b should be a parameter, too. 
b = tf.Variable(tf.constant(0.0))

# Just keeping 1 hidden layer
def inference(input):
    hidden_layer1 = tf.sigmoid(tf.matmul(input, W1) + b)
    out_layer = tf.nn.softmax(tf.matmul(hidden_layer1, W5) + b)

    return out_layer

# This is unchanged
def loss(hypothesis, y):
    cross_entropy = tf.reduce_mean(-tf.reduce_sum(y * tf.log(hypothesis), reduction_indices=[1]))

    return cross_entropy

# This is unchanged
def train(loss):
    optimizer = 
tf.train.GradientDescentOptimizer(learning_rate=0.1)
    train = optimizer.minimize(loss)

    return train

# Using SCALE to normalize the input variables (range of inputs too big)
# This is a simple normalization in this case. Other examples are 
# Min-Max normalization or z-scores. 

SCALE = 1000
x_data = set_x_data()
y_data = set_y_data(x_data)
x_data /= SCALE

# Now only placeholders are x and y (b is a parameter)
x= tf.placeholder(tf.float32, shape=[None, 1])
y= tf.placeholder(tf.float32, shape=[None, 2])

hypothesis = inference(x)
loss = loss(hypothesis, y)
train = train(loss)

sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)

print(sess.run(W1))

# Epochs x 10, it did not converge with fewer epochs
epochs = 20000
losses = np.zeros(epochs)
for step in range(epochs):
# Shuffle data
    r = np.random.permutation(x_data.shape[0])
    x_data = x_data[r]
    y_data = y_data[r,:]
    # Small modification here to capture the loss. 
    _, l = sess.run([train, loss], feed_dict={x:x_data, y:y_data})
    losses[step] = l

print(sess.run(W1))
print(sess.run(b))

显示上述决策函数的代码：

%matplotlib inline
import matplotlib.pyplot as plt
ystar = np.arange(50, 400, 10)[:,None]
plt.plot(ystar, sess.run(hypothesis, feed_dict={x:ystar/SCALE})[:,0])

相关问题更多 >

编程相关推荐

热门问题

热门文章