1.1 讀取資料

import pandas as pd
import numpy as np

df = pd.read_csv('C:/Users/User/OneDrive - student.nsysu.edu.tw/Educations/Contests/thu_bigdata/初賽/train model/train.csv')
df.head()

from sklearn.model_selection import train_test_split
data1 = df[df['y'].isin([5])].iloc[0:3,:]
data2 = df[df['y'].isin([8])]

train_nor = data1
test_nor = df[df['y'].isin([5])].iloc[5:14,:]
train_fra, test_fra = train_test_split(data2, test_size = 0.5)
data_train = pd.concat([train_nor,train_fra], axis=0)
data_test = pd.concat([test_nor,test_fra], axis=0)

pd.DataFrame(np.transpose(data1)).plot()
pd.DataFrame(np.transpose(data2.head())).plot()

<matplotlib.axes._subplots.AxesSubplot at 0x168ba4d4710>

1.2 Non Sampling

先嘗試使用不採樣的方式建模型，觀察直接做分類的效果。

from sklearn import ensemble
from sklearn import metrics

train_X = data_train.iloc[:,0:449]
test_X = data_test.iloc[:,0:449]
train_y = data_train["y"]
test_y = data_test["y"]

forest = ensemble.RandomForestClassifier(n_estimators = 100)
forest_fit = forest.fit(train_X, train_y)

test_y_predicted = forest.predict(test_X)
accuracy_rf = metrics.accuracy_score(test_y, test_y_predicted)
print(accuracy_rf)

test_auc = metrics.roc_auc_score(test_y, test_y_predicted)
print (test_auc)

0.993103448275862 0.9444444444444444

可以發現auc高達：0.944，因此這筆資料可能特徵太過容易做分類，因此後續的平衡資料分析可能效果不會太過顯著。

2.1 建立GAN

# import modules
%matplotlib inline
import os
import random
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from tqdm import tqdm_notebook as tqdm
from keras.models import Model
from keras.layers import Input, Reshape
from keras.layers.core import Dense, Activation, Dropout, Flatten
from keras.layers.normalization import BatchNormalization
from keras.layers.convolutional import UpSampling1D, Conv1D
from keras.layers.advanced_activations import LeakyReLU
from keras.optimizers import Adam, SGD
from keras.callbacks import TensorBoard
from sklearn.preprocessing import StandardScaler

# set parameters
dim = 450
num = 3
g_data = data1

# Standard Scaler
ss = StandardScaler()
g_data = pd.DataFrame(ss.fit_transform(g_data))


# generator
def get_generative(G_in, dense_dim=200, out_dim= dim, lr=1e-3):
    x = Dense(dense_dim)(G_in)
    x = Activation('tanh')(x)
    G_out = Dense(out_dim, activation='tanh')(x)
    G = Model(G_in, G_out)
    opt = SGD(lr=lr)
    G.compile(loss='binary_crossentropy', optimizer=opt)
    return G, G_out

G_in = Input(shape=[10])
G, G_out = get_generative(G_in)
G.summary()

# discriminator
def get_discriminative(D_in, lr=1e-3, drate=.25, n_channels= dim, conv_sz=5, leak=.2):
    x = Reshape((-1, 1))(D_in)
    x = Conv1D(n_channels, conv_sz, activation='relu')(x)
    x = Dropout(drate)(x)
    x = Flatten()(x)
    x = Dense(n_channels)(x)
    D_out = Dense(2, activation='sigmoid')(x)
    D = Model(D_in, D_out)
    dopt = Adam(lr=lr)
    D.compile(loss='binary_crossentropy', optimizer=dopt)
    return D, D_out

D_in = Input(shape=[dim])
D, D_out = get_discriminative(D_in)
D.summary()

# set up gan
def set_trainability(model, trainable=False):
    model.trainable = trainable
    for layer in model.layers:
        layer.trainable = trainable
        
def make_gan(GAN_in, G, D):
    set_trainability(D, False)
    x = G(GAN_in)
    GAN_out = D(x)
    GAN = Model(GAN_in, GAN_out)
    GAN.compile(loss='binary_crossentropy', optimizer=G.optimizer)
    return GAN, GAN_out

GAN_in = Input([10])
GAN, GAN_out = make_gan(GAN_in, G, D)
GAN.summary()

# pre train
def sample_data_and_gen(G, noise_dim=10, n_samples= num):
    XT = np.array(g_data)
    XN_noise = np.random.uniform(0, 1, size=[n_samples, noise_dim])
    XN = G.predict(XN_noise)
    X = np.concatenate((XT, XN))
    y = np.zeros((2*n_samples, 2))
    y[:n_samples, 1] = 1
    y[n_samples:, 0] = 1
    return X, y

def pretrain(G, D, noise_dim=10, n_samples = num, batch_size=32):
    X, y = sample_data_and_gen(G, n_samples=n_samples, noise_dim=noise_dim)
    set_trainability(D, True)
    D.fit(X, y, epochs=1, batch_size=batch_size)
    
pretrain(G, D)

def sample_noise(G, noise_dim=10, n_samples=num):
    X = np.random.uniform(0, 1, size=[n_samples, noise_dim])
    y = np.zeros((n_samples, 2))
    y[:, 1] = 1
    return X, y

Using TensorFlow backend.

WARNING:tensorflow:From C:3-gpu-packages_backend.py:74: The name tf.get_default_graph is deprecated. Please use tf.compat.v1.get_default_graph instead.

WARNING:tensorflow:From C:3-gpu-packages_backend.py:517: The name tf.placeholder is deprecated. Please use tf.compat.v1.placeholder instead.

WARNING:tensorflow:From C:3-gpu-packages_backend.py:4138: The name tf.random_uniform is deprecated. Please use tf.random.uniform instead.

WARNING:tensorflow:From C:3-gpu-packages.py:790: The name tf.train.Optimizer is deprecated. Please use tf.compat.v1.train.Optimizer instead.

WARNING:tensorflow:From C:3-gpu-packages_backend.py:3376: The name tf.log is deprecated. Please use tf.math.log instead.

WARNING:tensorflow:From C:3-gpu-packagesimpl.py:180: add_dispatch_support..wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version. Instructions for updating: Use tf.where in 2.0, which has the same broadcast rule as np.where ________________________________________________________________ Layer (type) Output Shape Param #
================================================================= input_1 (InputLayer) (None, 10) 0
_________________________________________________________________ dense_1 (Dense) (None, 200) 2200
_________________________________________________________________ activation_1 (Activation) (None, 200) 0
_________________________________________________________________ dense_2 (Dense) (None, 450) 90450
================================================================= Total params: 92,650 Trainable params: 92,650 Non-trainable params: 0 _________________________________________________________________ WARNING:tensorflow:From C:3-gpu-packages_backend.py:133: The name tf.placeholder_with_default is deprecated. Please use tf.compat.v1.placeholder_with_default instead.

WARNING:tensorflow:From C:3-gpu-packagesbackend.py:3445: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version. Instructions for updating: Please use rate instead of keep_prob. Rate should be set to rate = 1 - keep_prob. ________________________________________________________________ Layer (type) Output Shape Param #
================================================================= input_2 (InputLayer) (None, 450) 0
_________________________________________________________________ reshape_1 (Reshape) (None, 450, 1) 0
_________________________________________________________________ conv1d_1 (Conv1D) (None, 446, 450) 2700
_________________________________________________________________ dropout_1 (Dropout) (None, 446, 450) 0
_________________________________________________________________ flatten_1 (Flatten) (None, 200700) 0
_________________________________________________________________ dense_3 (Dense) (None, 450) 90315450
_________________________________________________________________ dense_4 (Dense) (None, 2) 902
================================================================= Total params: 90,319,052 Trainable params: 90,319,052 Non-trainable params: 0 _________________________________________________________________ _________________________________________________________________ Layer (type) Output Shape Param #
================================================================= input_3 (InputLayer) (None, 10) 0
_________________________________________________________________ model_1 (Model) (None, 450) 92650
_________________________________________________________________ model_2 (Model) (None, 2) 90319052
================================================================= Total params: 90,411,702 Trainable params: 92,650 Non-trainable params: 90,319,052 _________________________________________________________________ Epoch 1/1 6/6 [==============================] - 2s 342ms/step - loss: 0.6915

2.2 訓練GAN

# training
def train(GAN, G, D, epochs=300, n_samples= num, noise_dim=10, batch_size=32, verbose=False, v_freq=dim,):
    d_loss = []
    g_loss = []
    e_range = range(epochs)
    if verbose:
        e_range = tqdm(e_range)
    for epoch in e_range:
        X, y = sample_data_and_gen(G, n_samples=n_samples, noise_dim=noise_dim)
        set_trainability(D, True)
        d_loss.append(D.train_on_batch(X, y))
        xx,yy = X,y
        
        X, y = sample_noise(G, n_samples=n_samples, noise_dim=noise_dim)
        set_trainability(D, False)
        g_loss.append(GAN.train_on_batch(X, y))
        if verbose and (epoch + 1) % v_freq == 0:
            print("Epoch #{}: Generative Loss: {}, Discriminative Loss: {}".format(epoch + 1, g_loss[-1], d_loss[-1]))
    return d_loss, g_loss, xx, yy

d_loss, g_loss ,xx,yy= train(GAN, G, D, verbose=True)

HBox(children=(IntProgress(value=0, max=1), HTML(value=’’)))

2.3 One Class SVM

將標準化後的原始資料建立一個One Class SVM

from sklearn import svm

clf = svm.OneClassSVM(kernel='linear', gamma='auto').fit(xx[0:3])
origin = pd.DataFrame(clf.score_samples(xx[0:3]))
origin.describe()

使用生成資料計算它們的score

new = pd.DataFrame(clf.score_samples(xx[3:6]))
new.describe()

occ = pd.concat([pd.DataFrame(new[0] < origin[0].min()),pd.DataFrame(new[0] > origin[0].max())], axis=1)
occ['ava'] = pd.DataFrame(occ.iloc[:,1:2] == occ.iloc[:,0:1])
occ

2.4 計算生成異常率

err = sum(occ['ava'] == False)/len(occ['ava'])
err

0.0

畫出生成的資料圖，第一張圖片為原始3張資料，下三張資料則為迭代後所生成的資料，基本肉眼難以分辨真偽。

pd.DataFrame(np.transpose(pd.DataFrame(ss.inverse_transform(xx[0:3])))).plot()
pd.DataFrame(np.transpose(pd.DataFrame(ss.inverse_transform(xx[3:4])))).plot()
pd.DataFrame(np.transpose(pd.DataFrame(ss.inverse_transform(xx[4:5])))).plot()
pd.DataFrame(np.transpose(data_train1.iloc[6:9,:])).plot()

<matplotlib.axes._subplots.AxesSubplot at 0x169afe20748>

3.1 Balance data

生成符合標準的資料並合併舊有資料，以產生平衡的訓練資料。這裡我們生成45次*3的第5類樣本，並與舊的訓練資料做合併。

# balance train data
re = 45
new_data = pd.DataFrame(xx[3:6])
new_data.columns = g_data.columns
data = pd.concat([g_data,new_data], axis=0)
for i in range(re):
    d_loss, g_loss ,xx,yy= train(GAN, G, D, verbose=True)
    new_data = pd.DataFrame((xx[3:6]))
    data = pd.concat([data,new_data], axis=0)

# anti Scaler
data_train1  = pd.DataFrame(ss.inverse_transform(data))
data_train1 = data_train1.iloc[:,0:449]
data_train1['y'] = 5
data_train1.head()

data_train1.columns = df.columns

data_train1 = pd.concat([data_train1,train_fra], axis=0)
data_train1

3.2　Validate

如同第一部分的方式，我們一樣用隨機森林來做為分類分法，但這次訓練資料我們使用以GAN平衡後的資料當作測試集，其餘步驟街與第一部分一致。

train_X = data_train1.iloc[:,0:449]
test_X = data_test.iloc[:,0:449]
train_y = data_train1["y"]
test_y = data_test["y"]

forest = ensemble.RandomForestClassifier(n_estimators = 100)
forest_fit = forest.fit(train_X, train_y)

test_y_predicted = forest.predict(test_X)
accuracy_rf = metrics.accuracy_score(test_y, test_y_predicted)
print(accuracy_rf)

test_auc = metrics.roc_auc_score(test_y, test_y_predicted)
print (test_auc)

1.0 1.0

結果上來看，以GAN平衡的結果較好，但仍存在一些問題。
第一：隨機森林參數並無調整，因此在這個差距上來看並不能說平衡後的模型較佳；
第二：在未平衡前資料分類就可達到相當水準，因此可見此份資料的分類並不需要平衡資料來達成。

但可以確定的是，使用這個模式下，我們可以生成不錯的偽資料，至少在欺騙人眼的作用上是辦的到的。但也發現生成資料的另個問題為，幾乎無較為偏離的資料，意旨我們所產生的資料變異程度遠遠低於原先資料的變異程度，因此使用GAN的資料平衡後會降低該類別資料的變異程度，這可能會導致模型的泛化能力降低。