python - 使用 Tensorflow 的线性回归预制估计器得到错误的答案

Question

我是堆栈溢出和张量流的新手。我试图使用预制的线性回归估计器重做机器学习简介（Andrew Ng 的 Coursera 课程）中的简单线性回归。

我使用 numpy 和 scikit-learn 在 python 中编写了线性回归模型，并成功找到了模型参数 [theta0, theta1] = [-3.6303, 1.1664]。这是通过正规方程和常规梯度下降来完成的。

我无法使用 Tensorflow 的线性回归预制估计器来产生相同的结果。我正在使用 Google 机器学习速成课程中确定的基本方法——TensorFlow 的第一步（以及此处：https ://medium.com/datadriveninvestor/machine-learning-part-iv-efecd2f61f35 ）。

我把数据放在这里：https ://github.com/ChristianHaeuber/TensorFlowData

有人可以告诉我我做错了什么吗？

from __future__ import print_function

import math

from IPython import display
from matplotlib import cm
from matplotlib import gridspec
from matplotlib import pyplot as plt
import numpy as np
import pandas as pd
from sklearn import metrics
import tensorflow as tf
from tensorflow.python.data import Dataset

tf.logging.set_verbosity(tf.logging.ERROR)
pd.options.display.max_rows = 10
pd.options.display.float_format = '{:.1f}'.format

data = pd.read_csv('ex1data1.txt')

batch = data.shape[0]

feature_columns = [tf.feature_column.numeric_column('population')]

targets = data['profit']

my_optimizer=tf.train.GradientDescentOptimizer(learning_rate=0.01)

linear_regressor = tf.estimator.LinearRegressor(
        feature_columns=feature_columns,
        optimizer=my_optimizer
        )

def input_fn(ft, t, batch=1, shuffle=True, epochs=None):
    ft = {k:np.array(v) for k,v in dict(ft).items()}
    ds = Dataset.from_tensor_slices((ft, t))
    ds = ds.batch(batch).repeat(epochs)

    if shuffle:
        ds=ds.shuffle(buffer_size=10000)

    ft, lb = ds.make_one_shot_iterator().get_next()

    return ft, lb

ft = data[['population']]
input_fn_1 = lambda: input_fn(ft, targets)

linear_regressor.train(
        input_fn = input_fn_1,
        steps=1
        )

input_fn_2 = lambda: input_fn(ft, targets, shuffle=False, epochs=1)

p = linear_regressor.predict(input_fn = input_fn_2)

p = np.array([item['predictions'][0] for item in p])

mse = metrics.mean_squared_error(p, targets)

print("MSE: %0.3f" % mse)

print("Bias Weight: %0.3f" % 
      linear_regressor.get_variable_value('linear/linear_model/bias_weights').flatten())
print("Weight %0.3f" % 
      linear_regressor.get_variable_value('linear/linear_model/population/weights').flatten())

标签： pythontensorflowmachine-learninglinear-regression

Answer 1

机器学习简介课程在每次迭代中使用所有训练示例进行批量梯度下降，然后使用多次迭代收敛。上面的代码将只使用一个训练示例（batch=1）并且迭代次数（步数）是永远的（基于 tf.estimator.LinearRegressor.train 文档）。

通过一些更改，我能够复制机器学习简介课程的结果。

from __future__ import print_function

import math
from matplotlib import cm
from matplotlib import gridspec
from matplotlib import pyplot as plt
import numpy as np
import pandas as pd
from sklearn import metrics
import tensorflow as tf
from tensorflow.python.data import Dataset

tf.logging.set_verbosity(tf.logging.ERROR)
pd.options.display.max_rows = 10
pd.options.display.float_format = '{:.1f}'.format

def my_input_fn(features, labels, batch_size=1, num_epochs=None):

    features = {key:np.array(value) for key,value in         
                dict(features).items()}

    ds = Dataset.from_tensor_slices((features,labels))
    ds = ds.batch(batch_size).repeat(num_epochs)

    features, labels = ds.make_one_shot_iterator().get_next()

    return features, labels

ex1_data_df = pd.read_csv('ex1data1.txt')

features = ex1_data_df['population']
my_features = ex1_data_df[['population']]
feature_columns = [tf.feature_column.numeric_column('population')]
labels = ex1_data_df['profit']

my_optimizer=tf.train.GradientDescentOptimizer(learning_rate=0.0001)

linear_regressor = tf.estimator.LinearRegressor(
        feature_columns = feature_columns,
        optimizer=my_optimizer)

_ = linear_regressor.train(
        input_fn = lambda:my_input_fn(my_features, labels, 
                                      batch_size=ex1_data_df.shape[0]), 
        steps=2000
        )

predictions = linear_regressor.predict(
        input_fn=lambda:my_input_fn(my_features,labels,
                                    batch_size=1,num_epochs=1)
        )

predictions = np.array([item['predictions'][0] for item in     predictions])

mean_squared_error = metrics.mean_squared_error(predictions, labels)
print("Mean Squared Error (on training data):     {}".format(mean_squared_error))

weight =     linear_regressor.get_variable_value('linear/linear_model/population/weights')
bias = linear_regressor.get_variable_value('linear/linear_model/bias_weights')
print("Feature weight: {0}\t Bias weight: {1}".format(weight, bias))