Keras-3 Keras With Otto Group

Otto 分类问题

这里，我们将对Otto数据集进行分类。

• 本文主要参考 2.3 Introduction to Keras。个人觉得这是一个很好Keras教程，希望大家也去学习学习。
• 关于Otto，可以在 otto group 找到更多详细的材料
• 本文主要关注代码的实现，具体细节和基本概念不会详细展开

让我们开始吧

就像以前说过的那样，处理一个问题主要分为三个部分：数据准备，模型构建和模型优化

导入模块

这里遇到了新的模块

• StandardScaler 用于归一化，感觉很好使。详见StandardScaler
• LabelEncoder 配合np_utils用于One-hot编码，详见LabelEncoder。注意和OneHotEncoder的区别。
• EarlyStopping 当监测值不再改善时，该回调函数将中止训练。详见EarlyStopping
• ModelCheckpoint 保存模型。详见ModelCheckpoint

import numpy as npimport pandas as pdfrom sklearn.preprocessing import StandardScalerfrom sklearn.preprocessing import LabelEncoderfrom keras.utils import np_utilsfrom keras.models import Sequentialfrom keras.layers.core import Dense, Activation, Dropoutfrom keras.callbacks import EarlyStopping, ModelCheckpoint

## 数据准备读取数据。数据可以在 [otto group](https://www.kaggle.com/c/otto-group-product-classification-challenge/data) 找到

train_path = './data/train.csv'test_path = './data/test.csv'df = pd.read_csv(train_path)

观察数据。有93个特征，最后一列是种类，第一列的id对于训练没有任何作用。

df.head()

id feat_1 feat_2 feat_3 feat_4 feat_5 feat_6 feat_7 feat_8 feat_9 … feat_85 feat_86 feat_87 feat_88 feat_89 feat_90 feat_91 feat_92 feat_93 target 0 1 1 0 0 0 0 0 0 0 0 … 1 0 0 0 0 0 0 0 0 Class_1 1 2 0 0 0 0 0 0 0 1 0 … 0 0 0 0 0 0 0 0 0 Class_1 2 3 0 0 0 0 0 0 0 1 0 … 0 0 0 0 0 0 0 0 0 Class_1 3 4 1 0 0 1 6 1 5 0 0 … 0 1 2 0 0 0 0 0 0 Class_1 4 5 0 0 0 0 0 0 0 0 0 … 1 0 0 0 0 1 0 0 0 Class_1

5 rows × 95 columns

导入数据。

• 第一列id对训练没用，所以我们不需要它。
• train 和 test 两个文件有所区别(test中没有给出target)

def load_data(path, train=True):    df = pd.read_csv(path)    X = df.values.copy()    if train:        np.random.shuffle(X)        X, label = X[:, 1:-1].astype(np.float32), X[:, -1]        return X, label    else:        X, ids = X[:, 1:].astype(np.float32), X[:, 0].astype(str)        return X, ids

X_train, y_train = load_data(train_path)X_test, ids = load_data(test_path, train=False)

预处理，训练数据和测试数据一起归一化，以免忘记了

def preprocess_data(X, scaler=None):    if not scaler:        scaler = StandardScaler()        scaler.fit(X)    X = scaler.transform(X)    return X, scaler

X_train, scaler = preprocess_data(X_train)X_test, _ = preprocess_data(X_test, scaler)

One-hot 编码

def preprocess_label(labels, encoder=None, categorical=True):    if not encoder:        encoder = LabelEncoder()        encoder.fit(labels)    y = encoder.transform(labels).astype(np.int32)    if categorical:        y = np_utils.to_categorical(y)    return y, encoder

y_train, encoder = preprocess_label(y_train)

搭建网络模型

dim = X_train.shape[1]print(dim, 'dims')print('Building model')nb_classes = y_train.shape[1]model = Sequential()model.add(Dense(256, input_shape=(dim, )))model.add(Activation('relu'))model.add(Dropout(0.5))model.add(Dense(128))model.add(Activation('relu'))model.add(Dropout(0.5))model.add(Dense(64))model.add(Activation('relu'))model.add(Dropout(0.5))model.add(Dense(nb_classes))model.add(Activation('softmax'))

93 dimsBuilding model

model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

batch_size = 128epochs = 2

训练，同时保持最佳模型

fBestModel = 'best_model.h5'early_stop = EarlyStopping(monitor='val_acc', patience=5, verbose=1)best_model = ModelCheckpoint(fBestModel, verbose=0, save_best_only=True)model.fit(X_train, y_train, epochs=epochs, batch_size=batch_size, verbose=1, validation_split=0.1, callbacks=[best_model, early_stop])

Train on 55690 samples, validate on 6188 samplesEpoch 1/255690/55690 [==============================] - 2s 42us/step - loss: 0.5256 - acc: 0.7967 - val_loss: 0.5268 - val_acc: 0.7982Epoch 2/255690/55690 [==============================] - 2s 42us/step - loss: 0.5251 - acc: 0.7991 - val_loss: 0.5256 - val_acc: 0.8017<keras.callbacks.History at 0x13551adaef0>

预测并保存结果。将结果保存为Kaggle上要求的格式，然后提交了测试结果，得到了0.5左右的分数，据说大概前50%左右

prediction = model.predict(X_test)

num_pre = prediction.shape[0]columns = ['Class_'+str(post+1) for post in range(9)]df2 = pd.DataFrame({'id' : range(1,num_pre+1)})df3 = pd.DataFrame(prediction, columns=columns)df_pre = pd.concat([df2, df3], axis=1)

df_pre.to_csv('predition.csv', index=False)