在tensorflow回归上接收随机成本输出-Python [英] Receiving random cost output on tensorflow regression- python

问题描述

我对tensorflow还是比较陌生，我试图改编教程中的一些代码来处理自己的数据.

I am relatively new to tensorflow and I have attempted to adapt some code from a tutorial to process my own data.

可在此处找到数据: https://github.com/z12332/tensorflow-test-1/blob/master/export.csv

请记住，要馈送的数据集仅由第9列到第27列(将nan转换为0)和第30列作为标签

Keep in mind, that the dataset being fed consists of only columns 9 though 27 (with nan's converted to 0) and column 30 as the labels

这是教程代码的链接: https://github.com/llSourcell/How_to_use_Tensorflow_for_classification-LIVE/blob/master/demo.ipynb

here is the link to the tutorial code: https://github.com/llSourcell/How_to_use_Tensorflow_for_classification-LIVE/blob/master/demo.ipynb

我能够在没有错误消息的情况下运行该程序，但是由于某种原因，它会输出200个相对随机成本值的训练步骤.作为示例，这是前几个步骤:

I am able to get the program to run without an error message, but for some reason, it outputs 200 training steps of relatively random cost values. As and example, here are the first few steps:

Training step: 0000 cost= 0.039999638
Training step: 0000 cost= 0.159996599
Training step: 0000 cost= 0.000000002
Training step: 0000 cost= 0.000000004
Training step: 0000 cost= 0.000000001
Training step: 0000 cost= 0.039994366
Training step: 0000 cost= 0.000000005
Training step: 0000 cost= 0.039997347
Training step: 0000 cost= 0.359970629
Training step: 0000 cost= 1.959837437
Training step: 0000 cost= 3.239814520
Training step: 0000 cost= 0.000000195
Training step: 0000 cost= 0.000000228
Training step: 0000 cost= 0.000000003
Training step: 0000 cost= 0.000000388
Training step: 0000 cost= 0.039958697
Training step: 0000 cost= 0.159986690
Training step: 0000 cost= 0.159973413

Training cost= 2.70406e-05 W= [  2.38201610e-05   1.96683395e-05   3.69497479e-06   2.77944509e-05
   2.02058782e-05   3.82550934e-05   3.37507554e-05   2.18498894e-06
   2.92303273e-04   7.17514267e-05   2.34498725e-06   3.40497172e-06
   6.25661269e-05   5.59996465e-07   8.81450160e-06   3.44998034e-06] b= [  2.62004360e-05] 

这是我的完整代码，因为任何人都知道为什么会这样或如何调试:

Here is my code in full, for anyone knows why this is happening or how to debug:

import tensorflow as tf
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

df = pd.read_csv('/Users/benny/desktop/export.csv')
data_ = df.iloc[1:,9:27]
data_['CRISPR'] = df.iloc[:,30]
data_ = data_.drop(['Diseases'],axis=1)

dim = 16
learning_rate = 0.0000001
display_step = 50

X = tf.placeholder(tf.float32, [None, dim])
Y = tf.placeholder(tf.float32)

train_X = data_.iloc[:200, :-2].as_matrix()

''' 
dimensions = [200,16]

array([[ 25.,   2.,   3., ...,  nan,  nan,   2.],
       [  5.,  13.,   3., ...,  nan,  19.,   2.],
       [ 25.,  13.,   3., ...,  nan,  nan,   2.],
       ..., 
       [ 25.,  13.,   3., ...,  nan,  nan,   2.],
       [ 25.,  13.,   3., ...,  nan,  nan,   2.],
       [ nan,  13.,   3., ...,  nan,  19.,   3.]])
'''
train_X = train_X.fillna(value=0)

train_Y = data_.iloc[:200, -1].as_matrix()

'''
dimensions = [200]

array([ 1,  2,  0,  0,  0,  1,  0,  1,  3,  7,  9,  0,  0,  0,  0,  1,  2,
        2,  0,  0,  0,  7,  2,  2,  2,  0,  4,  0,  0,  0,  0,  9,  5,  2,
        1,  2,  1,  0,  0,  1,  0,  2,  1,  2,  0,  1,  1,  1,  0,  0,  0,
        1,  3,  1,  2,  4,  1,  0,  1,  6,  2,  1,  0,  0,  1,  0,  1,  1,
        1,  7,  7,  4,  1,  1,  6,  4,  0,  0,  1,  1,  0,  1,  1,  1,  2,
        0,  0,  2,  0,  0,  0,  3,  2,  3,  1,  1,  9,  7,  4,  1,  1,  1,
        0,  1,  5,  4,  2,  1,  1,  1,  1,  1,  0,  4,  1,  0,  1,  0,  0,
        1,  2,  1,  4,  0, 10,  2,  0,  1,  2,  3,  0,  0,  0,  0,  0,  0,
        3,  1,  1,  2,  0,  7,  0,  2,  0,  2,  0,  0,  2,  3,  1,  0,  7,
        3,  2,  9,  1,  0,  0,  2,  1,  0,  2,  2,  1,  1,  2,  4,  0,  0,
        0,  0,  0,  0,  1,  0,  1,  4,  1,  0,  0,  1, 15,  1,  0,  1,  2,
        0,  0,  1,  0,  2,  0,  0,  0,  2,  1,  0,  1, 11])

'''


test_X = data_.iloc[200:320, :-2].as_matrix()
test_Y = data_.iloc[200:320, -1].as_matrix()

n_samples = train_Y.size

W= tf.Variable(tf.zeros([dim]), name="weight")
b = tf.Variable(tf.zeros([1]), name="bias")

activation = tf.add(tf.mul(X, W), b)


cost = tf.reduce_sum(tf.pow(activation-Y, 2))/(2*n_samples)
optimizer =  tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)

hm_epochs = 10

init = tf.initialize_all_variables()

xy = zip(train_X, train_Y) 
sess = tf.Session()
sess.run(init)

for (x, y) in xy:
    for i in range(hm_epochs):
        sess.run(optimizer, feed_dict={X: x[np.newaxis, ...],
                                       Y: y[np.newaxis, ...]})

        if (i) % display_step == 0:
            cc = sess.run(cost, feed_dict={X: x[np.newaxis, ...],
                                           Y: y[np.newaxis, ...]})
            print "Training step:", '%04d' % (i), "cost=", "{:.9f}".format(cc)

print "Optimization Finished!"
training_cost = sess.run(cost, feed_dict={X:x[np.newaxis, ...],
                                          Y:y[np.newaxis, ...]})
print "Training cost=", training_cost, "W=", sess.run(W), "b=", sess.run(b), '\n'

随机值是否可能来自缺失值(在数据集中更改为0)?另外，我现在将如何应用test_X，Y数据帧进行预测?

Could the random values be from the missing values (changed to 0 in the dataset)? Also, how would I now apply the test_X,Y dataframes to predict?

无论哪种方式，我都在这里学习，谢谢您的帮助！

Either way Im here to learn so thanks for your help!

推荐答案

这是该代码的有效版本，但是在我提供一些注释之前.

Here is a working version of the code, but before I'll offer some notes.

1)可以在 tensorflow mnist教程中进行阅读. 我特别想知道为什么您的占位符大小不正确以及为什么 我们将为此任务使用标签的单一编码版本.

1) One do some reading at tensorflow mnist tutorial. Im particular see why your placeholder sizes are not correct and why we are going to use a one-hot-encoded version of the labels for this task.

2)考虑使用交叉熵代价.这是一个更合适的成本 为此多类任务.

2) Consider using cross-entropy cost. It is a more well suited cost for this multiclass task.

3)尽量不要被基本模型的性能所影响 (效果不佳).考虑探索数据以寻找导入特征，并查看该数据集的最新性能 可能是.

3) try not to be too underwhelmed by the performance of this basic model (it does not perform well). Consider exploring the data looking for import features and also look around for what the state of the art performance on this dataset might be.

import tensorflow as tf
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

df = pd.read_csv('/Users/benny/desktop/export.csv')
data_ = df.iloc[1:,9:27]
data_['CRISPR'] = df.iloc[:,30]
data_ = data_.drop(['Diseases'],axis=1)
# we will need the number of classes to be predict
# the nunique methods gets us the number of unique labels
nclasses = data["CRISPR"].nunique()
# lets collect of labels here. We'll one-hot-encode them
# using pandas.get_dummies()
inputY = pd.get_dummies(data_.iloc[:, -1])

dim = 16
learning_rate = 0.0000001
display_step = 50

X = tf.placeholder(tf.float32, [None, dim])
# Y should define the shape of your labels. 
# As discussed we're going to need one hot encoded labels for
# this prediction task. this line does not define the shape of your input.
# we'll define later
# Y = tf.placeholder(tf.float32)


train_X = data_.iloc[:200, :-2].as_matrix()
train_X = train_X.fillna(value=0)
train_Y = inputY[:200].as_matrix()

test_X = data_.iloc[200:320, :-2].as_matrix()
test_Y = inputY[200:320].as_matrix()

n_samples = train_Y.size

# Its important we get the shape of the weight and bias matrices
# correct, the version in code is:
# W = tf.Variable(tf.zeros([dim]), name="weight")
# that wont work since we want to be able to multiply [X, W]
# to produce a evidence vector for each each example. 
# the shape of X is [200 x dim] - there should be a weight for each 
# feature and there 10 classes so  W is [dim, nclasses],
W = tf.Variable(tf.zeros([dim, nclasses]))

# for the bias, there should be one for each class.
# b = tf.Variable(tf.zeros([1]), name="bias")
b = tf.Variable(tf.zeros([nclasses]))

# the correct operation here is tf.matmal, I suspect you introduced
# this to make your early matrix definiton work in the graph
# activation = tf.add(tf.mul(X, W), b)
activation = tf.add(tf.matmul(X, W), b) 

# you forgot the actual model! Assuming you want to do
# softmax classification let do:
y = tf.nn.softmax(activation)
# Now let's define or input labels ( we could have called them Y )
# as you had them. Notice what we are saying here is expect a matrix
# of floats with any number of examples and nclasses number of columns
# which is exactly the size of train_Y.  
y_ = tf.placeholder(tf.float32, [None, nclasses])


# we define the cost reflect the fact that our model output is called
# y not activations (anymore)
# cost = tf.reduce_sum(tf.pow(activation-Y, 2))/(2*n_samples)
cost = tf.reduce_sum(tf.pow(y_ - y, 2))/(2*n_samples)
optimizer =  tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)

hm_epochs = 1000

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

# I imagine what you want to do here is stochastic gradient descent. 
# I am not sure this is the way to do it. End to check the code I 
# will train over the entire training data for 1000 repetitions, 
# similar to the tutorial code.
# .....
for i in range(hm_epochs):
    sess.run(optimizer, feed_dict={X: train_X,
                               y_: train_Y})

    if (i) % display_step == 0:
        cc = sess.run(cost, feed_dict={X: train_X,y_: train_Y})
        print "Training step:", '%04d' % (i), "cost=", "{:.9f}".format(cc)


# To check the accuracy ( this is one way of measuring the performance 
# of an algorithm on a classification task) we will do: (This is 
# adapted from the tensflow mnist example [code][1]) 

correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
print(sess.run(accuracy, feed_dict={X: test_X,
                              y_: test_Y}))

这篇关于在tensorflow回归上接收随机成本输出-Python的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！