kaggle之路 暴力keras（回归问题）

这个东西呢是这个样子的，首先建立一个神经网络，参数从（三层两层，激活函数，优化器）里选择，
输入直接在train和test里读入修改就可以了，然后跑一边看那种参数最好，最后再预测。完全没用特征选择，但是有nlp的做不了，在kaggle房价预测里mse直接跑了0.14.太累了今天先写这么多，以后填坑。

# 加载数据分析常用库
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns

from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation
from sklearn.metrics import mean_squared_error
from sklearn.cross_validation import train_test_split
import seaborn as sns
import matplotlib.pyplot as plt
from datetime import datetime
from sklearn import preprocessing
sns.set_style('darkgrid')


def create_nn_model(input_dim, activation, layers, optimizer):
    model = Sequential()
    l_num = 0
    for l in layers:
        if l_num == 0:
            model.add(Dense(l, input_dim=input_dim, activation=activation, init='he_normal'))
        else:
            model.add(Dense(l, activation=activation, init='he_normal'))
        l_num = l_num + 1
    model.compile(optimizer=optimizer, loss='mse')
    return model


train=pd.read_csv('C:\\Users\\Administrator\\Downloads\\train (5).csv')
test=pd.read_csv('C:\\Users\\Administrator\\Downloads\\test.csv')


ID = 'Id'
TARGET = 'SalePrice'

FEATURES = train.columns.drop([ID, TARGET])
y = np.log(train[TARGET].values)
train=train.drop('SalePrice',axis=1)
all_date=pd.concat([train,test],axis=0)



f_cat = all_date[FEATURES].select_dtypes(include=['object']).columns
f_num = all_date[FEATURES].select_dtypes(exclude=['object']).columns
# Replace NAs
all_date[f_num] = all_date[f_num].fillna(all_date[f_num].mean())
all_date[f_cat] = all_date[f_cat].fillna('?')


dummy_cat = pd.get_dummies(all_date[f_cat])
all_date = pd.DataFrame(preprocessing.scale(all_date[f_num]), columns=f_num)
dummy_cat .index=all_date.index
#all_date = all_date.join(dummy_cat)
all_date = pd.concat([all_date, dummy_cat], axis=1, ignore_index=True)
X = all_date.values


activations = ['softplus', 'relu', 'linear']
# It seems 'softsign', 'tanh','sigmoid', 'hard_sigmoid' do not perform well on the House Prices data
# activations = ['softplus', 'relu', 'linear','softsign', 'tanh','sigmoid', 'hard_sigmoid']
# optimizers = ['sgd', 'rmsprop', 'adagrad', 'adadelta', 'adam', 'adamax', 'nadam']
optimizers = ['adagrad', 'rmsprop', 'adam']
layers = [[150, 1], [150, 50, 1], [200, 100, 50, 1]]
X_train, X_test, y_train, y_test = train_test_split(X[0:train.shape[0]], y, test_size=0.1, random_state=0)
nnet_params = []
nnet_score = []


for a in activations:
    for o in optimizers:
        for l in layers:
            model = create_nn_model(input_dim=X.shape[1], activation=a, layers=l, optimizer=o)
            fit = model.fit(X_train, y_train, batch_size=100, nb_epoch=100, validation_split=0.1, verbose=0)
            score = np.sqrt(model.evaluate(X_test, y_test))
            #print("\nActivation: {} Optimizer: {} Layers: {} Score: {}\n".format(a, o, l, score))
            nnet_params.append(str(a) + '-' + str(o) + '-' + str(l))
            nnet_score.append(score)
res = pd.DataFrame({'params': nnet_params, 'score': nnet_score})
res.sort_values(['score'], ascending=True, inplace=True)
# print(res)
# res.to_csv('./keras_params_.csv')
# f, ax = plt.subplots(figsize=(20, 15))
# sns.set(style="whitegrid")
# ax = sns.barplot(x='score', y='params', data=res, label="Keras parameters - RMSE")
# ax.set(xlabel='RMSE', ylabel='Parameters')



model = create_nn_model(input_dim=X.shape[1], activation='linear', layers=[200, 100, 50, 1], optimizer='adam')
fit = model.fit(X_train, y_train, batch_size=100, nb_epoch=100, validation_split=0.1, verbose=0)

result=model.predict(X[train.shape[0]:])
result=result.flatten()
zz=[pow(np.e,x) for x in result]

df=pd.DataFrame({ 'Id': [x for x in range(1461,2920)],
                   'SalePrice': zz })
df.to_csv('over.csv',index=False)