之前我們其實已經寫過了，這次既然原文單獨列了一章，我們也來再寫一次好了。原文希望我們把RMSE壓到110以下，我這裡驗證集的RMSE 115左右，因隨機數種子最低降到過112左右，在設定好種子的情況下seed(1) 驗證集的rmse在116左右。

上次我們用xs1,xs2,xs3...來得到特徵集的tensor，如若columns太多，操作起來很繁瑣，這次我把tensor壓到了一個列表裡，然後通過concat來組合。

import tensorflow as tf
import numpy as np
import pandas as pd
from tensorflow.data import Dataset

df = pd.read_csv('california_housing_train.csv')
df['per_rooms'] = df['total_rooms'] / df['population']
df['median_house_value'] /= 1000
df = df.reindex(np.random.permutation(df.index))
df = df.sort_index()
tf.set_random_seed(1)

def train_validation(df):
    features = df[['longitude', 'latitude', 'housing_median_age',
                   'total_rooms', 'total_bedrooms', 'population', 'households',
         'median_income', 'per_rooms']]
    targets = df['median_house_value']
    train_features = (features.head(12000).astype('float32')).copy()
    validation_features = (features.tail(5000).astype('float32')).copy()
    train_targets = (targets.head(12000).astype('float32')).copy()
    validation_targets = (targets.tail(5000).astype('float32')).copy()
    return train_features, train_targets, validation_features, validation_targets

def get_features(xs_dict):
    feautres = list(xs_dict.values())
    outputs = []
    for i in feautres:
        outputs.append(tf.expand_dims(i, -1))
    outputs = tf.concat([i for i in outputs], -1)
    return outputs


def my_input_fn(features, labels, batch_size=1, num_epochs=1, shuffle=False):
    features = {key: value for key, value in features.items()}
    ds = Dataset.from_tensor_slices((features, labels))
    ds = ds.batch(batch_size).repeat(num_epochs)
    if shuffle:
        ds.shuffle(10000)
    features, labels = ds.make_one_shot_iterator().get_next()
    return features, labels


def add_layer(inputs, input_size, output_size, activation_function=None):
    weights = tf.Variable(tf.random_normal([input_size, output_size], stddev=0.1))
    tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(0.001)(weights))
    biases = tf.Variable(tf.zeros(output_size) + 0.1)
    wx_b = tf.matmul(inputs, weights) + biases
    if activation_function is None:
        outputs = wx_b
    else:
        outputs = activation_function(wx_b)
    return weights, biases, outputs


def loss(pred, ys, regularizer=True):
    rmse = tf.sqrt(tf.reduce_mean(tf.square(ys - pred)))
    if regularizer is True:
        loss = rmse + tf.add_n(tf.get_collection('losses'))
    else:
        loss = rmse
    return loss


def train_step(learning_rate, loss):
    train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
    return train_step


xs, ys, dx, dy =train_validation(df)
xs, ys = my_input_fn(xs, ys, batch_size=1000, num_epochs=300)
dx, dy = my_input_fn(dx, dy, batch_size=5000, num_epochs=40)
x_input = get_features(xs)
x_validation = get_features(dx)
w1, b1, l1 = add_layer(x_input, 9, 30, activation_function=tf.nn.tanh)
w2, b2, l2 = add_layer(l1, 30, 6)
w3, b3, l3 = add_layer(l2, 6, 1)
train_loss = loss(l3, ys, regularizer=True)
train = train_step(0.005, train_loss)

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

for i in range(2000):
    sess.run(train)
    if i % 50 == 0:
        pred_l1 = tf.nn.tanh(tf.matmul(x_validation, w1) + b1)
        pred_l2 = tf.matmul(pred_l1, w2) + b2
        pred = tf.matmul(pred_l2, w3) + b3
        pred_loss = loss(pred, dy, regularizer=False)
        total_loss = sess.run([train_loss, pred_loss])
        print('train loss is :', '%0.2f' % total_loss[0], 'validation loss is:', '%0.2f' % total_loss[1])

 

這裡我只調整了第一層和第二層的weights，learning_rate, 優化器,batch_size還有正則化率。當然還有很多引數可以調整，我們不妨自己動手試試，這裡我之所以把seed固定，是因為我希望結果的變化是因為我更改了某一引數而變化，而不是因為初始weights的影響。在調整的時候請不要對特徵進行操作，本次的目的不是特徵工程。而是在輸入特徵固定的情況下，調整引數來看神經網路的變化。這裡我沒有新增太多的隱藏層，你們也可以試試新增隱藏層來看看是否可以得到更好的模型。