用张量流简单实现泰坦尼克号

Question

我是机器学习和张量流的新手。我试图以一种简单的方式在不使用任何架构、没有太多转换的情况下实现这一点，但我被困住了，不惜一切代价获得 NAN。下面是代码，如果我在这里做错了什么，请告诉我。

这段代码是我从其中一门学习课程中得到的，他们已经为 IRIS 数据实现了它，只是试图对泰坦尼克号数据做同样的事情。

! kaggle competitions download -c 'titanic'
import pandas as pd
train_data = pd.read_csv("train.csv")
train_data.head(), train_data.columns
n_input =7 ##7 valid columns
n_output=2 ## {0,1} survived or not survived
import tensorflow as tf
tf.reset_default_graph()
input_shape = [None,n_input]
inputplaceholder = tf.placeholder(dtype=tf.float32, shape=input_shape, name="input_placeholder")
weights = tf.Variable(tf.random_normal([n_input,n_output]), name="weights")
biases = tf.Variable(tf.zeros([n_output]), name="biases")
layer_1 = tf.matmul(inputplaceholder, weights)
layer_2 = tf.add(layer_1, biases)
outputlayer = tf.nn.sigmoid(layer_2)
learning_rate = 0.001
labelsplaceholder = tf.placeholder(dtype=tf.float32, shape=[None,n_output], name="labels_placeholder")
cost = tf.losses.mean_squared_error(labelsplaceholder, outputlayer)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cost) 
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
scaled_data=train_data[[ 'Pclass',  'Sex', 'Age', 'SibSp','Parch', 'Fare',  'Embarked']]
genderMap = {"male":1,"female":2,"":""}
embarkMap = {"C":1,"Q":2,"S":3,"":""}
scaled_data['Sex'].replace(genderMap, inplace=True)
scaled_data['Embarked'].replace(embarkMap, inplace=True)
scaled_data['Sex'] = scaled_data['Sex'].astype('float32')
scaled_data['Pclass'] = scaled_data['Pclass'].astype('float32')
scaled_data['SibSp'] = scaled_data['SibSp'].astype('float32')
scaled_data['Parch'] = scaled_data['Parch'].astype('float32')
scaled_data['Embarked'] = scaled_data['Embarked'].astype('float32')
scaled_data['Age'] = scaled_data['Age'].astype('float32')
scaled_data['Fare'] = scaled_data['Fare'].astype('float32')
import random
mydata = list(zip(scaled_data.values, train_data.Survived))
batch_size = 891
iterations = 400

history_loss = list()

for _ in range(iterations):
    inputdata = list()
    output_data = list()
    for _ in range(batch_size):
        input_output_pairs = random.choice(mydata)
        inputdata.append(input_output_pairs[0])
        output_one_hot = [0,0]
        output_one_hot[input_output_pairs[1]] = 1
        output_data.append(output_one_hot)
    res_optimizer, res_cost = sess.run([optimizer, cost], feed_dict={inputplaceholder: inputdata, labelsplaceholder: output_data})
    print(res_cost)
    history_loss.append(res_cost)    

运行这个我期待一些成本数据，但将所有数据都作为 NAN。我确实尝试了较低的学习率，即 0.00001 和 0.00005，但结果仍然相同

Answer 1

解决的问题是没有从原始数据中删除 NAN，使用“dropna”后我们能够获得一些成本值