我愿意为带有隐藏层的LinearRegression实现张量流模型。
我已经参考了这些教程来使用张量流进行线性回归，
这是我的代码：
class LinearRegression:
def __init__(self, x, y, data_type, learning_rate=0.01, epochs=100):
"""
Constructor to initialize the instance variables
:param x: independent variable X
:param y: dependent variable y
:param data_type: The data-type of dependent and independent variable
:param learning_rate: The learning rate of the neural network. Default: 0.01
:param epochs: Number of epochs. Default 100
"""
data_types = {
"int8": tf.int8,
"int16": tf.int16,
"int32": tf.int32,
"int64": tf.int64,
"float16": tf.float16,
"float32": tf.float32,
"float64": tf.float64
}
# Assign independent and dependent variables
self.__independent_variable = x
self.__dependent_variable = y
# Assign data-types, learning rate and epochs
self.data_type = data_types[data_type]
self.learning_rate = learning_rate
self.epochs = epochs
# Get the shape of input
# self.input_shape = None
# try:
# self.input_shape = self.__independent_variable.shape[1]
# except KeyError:
# self.input_shape = 1
# hidden layers
self.__hidden_layers = []
# placeholders
self.X = tf.placeholder(self.data_type)
self.y = tf.placeholder(self.data_type)
# cost function
self.cost_function = None
# optimizer
self.optimizer = None
def add_hidden_layer(self, weight_shape, activation_function="relu", nodes=100, bias=None):
"""
This method will add the hidden layer to our neural net. Note last hidden layer you are adding will be an
output layer.
:param weight_shape: The shape of the weight
:param activation_function: The activation function to be used. Supported: relu(Default), sigmoid, tanh
:param nodes: No of nodes for the bias, default is 100
:param bias: Default is 0's of no of nodes.
:return: True if the layer is added else will return False
"""
if bias is None:
bias = tf.Variable(tf.zeros(weight_shape[1]))
weight = tf.Variable(tf.random_uniform(weight_shape, minval=-1.0, maxval=1.0))
layer = None
# Count of hidden layer
hidden_layer_count = len(self.__hidden_layers)
# For the formula Y = mx+b
if activation_function == "relu":
if hidden_layer_count==0:
layer = tf.nn.relu(tf.matmul(self.X, weight)+bias)
else:
layer = tf.nn.relu(tf.matmul(self.__hidden_layers[hidden_layer_count-1][0], weight)+bias)
elif activation_function == "sigmoid":
if hidden_layer_count == 0:
layer = tf.sigmoid(tf.matmul(self.X, weight) + bias)
else:
layer = tf.sigmoid(tf.matmul(self.__hidden_layers[hidden_layer_count-1][0], weight) + bias)
elif activation_function == "tanh":
if hidden_layer_count == 0:
layer = tf.tanh(tf.matmul(self.X, weight) + bias)
else:
layer = tf.tanh(tf.matmul(self.__hidden_layers[hidden_layer_count-1][0], weight) + bias)
self.__hidden_layers.append([layer, weight, bias])
print("Added layer")
def train(self, optimizer="adam"):
"""
This method will train the model. Only two famous optimizers are supported
1. adam (Default)
2. gradient_descent
:param optimizer: Optizers will reduce the cost function. Supported: adam(Default), gradient_descent
:return: None
"""
hidden_layer_count = len(self.__hidden_layers)
self.cost_function = tf.reduce_sum(tf.pow(self.__hidden_layers[hidden_layer_count-1][0]-self.y, 2)
/ (2*self.__dependent_variable.shape[0]))
if optimizer == "adam":
self.optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(self.cost_function)
else:
self.optimizer = tf.train.GradientDescentOptimizer(learning_rate=self.learning_rate).minimize(self.cost_function)
W = tf.Variable(0.40, name="weight")
b = tf.Variable(0.40, name="bias")
# Initialize the global variables
init = tf.global_variables_initializer()
with tf.Session() as session:
session.run(init)
for epoch in range(self.epochs):
session.run(self.optimizer, feed_dict={self.X: self.__independent_variable, self.y: self.__dependent_variable})
c = session.run(self.cost_function, feed_dict={self.X: self.__independent_variable, self.y: self.__dependent_variable})
weight = session.run(self.__hidden_layers[hidden_layer_count-1][1])
bias = session.run(self.__hidden_layers[hidden_layer_count-1][2])
if epoch%100==0:
print("Weight: ", weight, "Cost: ", c, "Bias: ", bias)
return weight, bias
我已经用一些示例输入进行了测试：
import numpy
# Dataset
train_X = numpy.asarray([[ 3.3 ],
[ 4.4 ],
[ 5.5 ],
[ 6.71 ],
[ 6.93 ],
[ 4.168],
[ 9.779],
[ 6.182],
[ 7.59 ],
[ 2.167],
[ 7.042],
[10.791],
[ 5.313],
[ 7.997],
[ 5.654],
[ 9.27 ],
[ 3.1 ]])
train_Y = numpy.asarray([[1.7 ],
[2.76 ],
[2.09 ],
[3.19 ],
[1.694],
[1.573],
[3.366],
[2.596],
[2.53 ],
[1.221],
[2.827],
[3.465],
[1.65 ],
[2.904],
[2.42 ],
[2.94 ],
[1.3 ]])
l = LinearRegression(train_X, train_Y, "float32", epochs=1000)
l.add_hidden_layer([1, 17])
#l.add_hidden_layer([10, 10])
l.add_hidden_layer([17, 1])
weight, bias = l.train(optimizer="adam")
我得到这样的输出：
Weight: [[ 0.7321346 ]
[ 0.30761123]
[-0.49645996]
[ 0.8884759 ]
[ 0.43539715]
[ 0.6100311 ]
[-0.21072604]
[ 0.8822198 ]
[-0.8709421 ]
[ 0.1667776 ]
[ 0.8926308 ]
[-0.90756726]
[-0.41226315]
[ 0.21427989]
[ 0.17000389]
[-0.9373965 ]
[ 0.16003537]] Cost: 7.7810783 Bias: [-0.01]
Weight: [[ 0.7321346 ]
[ 0.30761123]
[-0.5783508 ]
[ 0.8262489 ]
[ 0.43539715]
[ 0.6100311 ]
[-0.28867453]
[ 0.77843475]
[-0.8709421 ]
[ 0.05801708]
[ 0.78117526]
[-0.90756726]
[-0.41226315]
[ 0.21427989]
[ 0.17000389]
[-0.983428 ]
[ 0.16003537]] Cost: 0.119424164 Bias: [-0.00681436]
Weight: [[ 0.7321346 ]
[ 0.30761123]
[-0.5923004 ]
[ 0.82624584]
[ 0.43539715]
[ 0.6100311 ]
[-0.30400994]
[ 0.76900154]
[-0.8709421 ]
[ 0.04737379]
[ 0.7730713 ]
[-0.90756726]
[-0.41226315]
[ 0.21427989]
[ 0.17000389]
[-1.0034373 ]
[ 0.16003537]] Cost: 0.08574117 Bias: [0.09325708]
我可以清楚地看到成本函数正在减少，但是我在Tensorflow中经验不足，所以我不知道我的实现是否正确（我尽了最大努力！），也不知道如何实现预测未来的阵列。您能帮我还是提供蚂蚁指导？预先谢谢你。